feat(ai): add media foundation with ModelRef, Media assets, and Image rewrite

Co-authored-by: Hona <10430890+Hona@users.noreply.github.com>
Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>

.changeset/directory-page-heading.md

@@ -0,0 +1,5 @@
+---
+"@opencode/core": patch
+---
+
+Correct directory page headings when the read offset is zero.

.gitattributes

@@ -1,2 +1,5 @@
 packages/core/migration/**/snapshot.json linguist-generated
 packages/core/src/database/migration.gen.ts linguist-generated
+packages/core/src/models-dev/snapshot.txt linguist-generated
+packages/core/src/**/*.txt text eol=lf
+packages/httpapi-codegen/test/generated/*.ts text eol=lf

.github/TEAM_MEMBERS

@@ -10,6 +10,7 @@ kitlangton
 kommander
 ludvigrask
 MrMushrooooom
+neriousy
 nexxeln
 R44VC0RP
 rekram1-node
@@ -18,4 +19,4 @@ simonklee
 Slickstef11
 usrnk1
 vimtor
-starptech
+StarpTech

.github/actions/setup-bun/action.yml

@@ -1,6 +1,10 @@
 name: "Setup Bun"
 description: "Setup Bun with caching and install dependencies"
 inputs:
+  bun-version:
+    description: "Bun version to install instead of the root packageManager version"
+    required: false
+    default: ""
   install-flags:
     description: "Additional flags to pass to 'bun install'"
     required: false
@@ -20,19 +24,22 @@ runs:
       shell: bash
       run: |
         if [ "$RUNNER_ARCH" = "X64" ]; then
-          V=$(node -p "require('./package.json').packageManager.split('@')[1]")
+          V="${{ inputs.bun-version }}"
+          if [ -z "$V" ]; then V=$(node -p "require('./package.json').packageManager.split('@')[1]"); fi
+          TAG=$([ "$V" = "canary" ] && echo "canary" || echo "bun-v${V}")
           case "$RUNNER_OS" in
             macOS)   OS=darwin ;;
             Linux)   OS=linux ;;
             Windows) OS=windows ;;
           esac
-          echo "url=https://github.com/oven-sh/bun/releases/download/bun-v${V}/bun-${OS}-x64-baseline.zip" >> "$GITHUB_OUTPUT"
+          echo "url=https://github.com/oven-sh/bun/releases/download/${TAG}/bun-${OS}-x64-baseline.zip" >> "$GITHUB_OUTPUT"
         fi
 
     - name: Setup Bun
       uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
       with:
-        bun-version-file: ${{ !steps.bun-url.outputs.url && 'package.json' || '' }}
+        bun-version: ${{ !steps.bun-url.outputs.url && inputs.bun-version || '' }}
+        bun-version-file: ${{ !steps.bun-url.outputs.url && !inputs.bun-version && 'package.json' || '' }}
         bun-download-url: ${{ steps.bun-url.outputs.url }}
 
     - name: Get cache directory

.github/workflows/beta.yml

@@ -1,37 +0,0 @@
-name: beta
-
-on:
-  workflow_dispatch:
-  schedule:
-    - cron: "0 * * * *"
-
-jobs:
-  sync:
-    runs-on: blacksmith-4vcpu-ubuntu-2404
-    permissions:
-      contents: write
-      pull-requests: write
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
-        with:
-          fetch-depth: 0
-
-      - name: Setup Bun
-        uses: ./.github/actions/setup-bun
-
-      - name: Setup Git Committer
-        id: setup-git-committer
-        uses: ./.github/actions/setup-git-committer
-        with:
-          opencode-app-id: ${{ vars.OPENCODE_APP_ID }}
-          opencode-app-secret: ${{ secrets.OPENCODE_APP_SECRET }}
-
-      - name: Install OpenCode
-        run: bun i -g opencode-ai
-
-      - name: Sync beta branch
-        env:
-          GH_TOKEN: ${{ steps.setup-git-committer.outputs.token }}
-          OPENCODE_API_KEY: ${{ secrets.OPENCODE_API_KEY }}
-        run: bun script/beta.ts

.github/workflows/check.yml

@@ -0,0 +1,22 @@
+name: check
+
+on:
+  push:
+    branches: [dev, v2]
+  pull_request:
+    branches: [dev, v2]
+  workflow_dispatch:
+
+jobs:
+  check:
+    name: typecheck
+    runs-on: blacksmith-4vcpu-ubuntu-2404
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
+
+      - name: Setup Bun
+        uses: ./.github/actions/setup-bun
+
+      - name: Run checks
+        run: bun run check

.github/workflows/deploy-files.yml

@@ -0,0 +1,34 @@
+name: deploy-files
+
+on:
+  push:
+    branches:
+      - dev
+      - v2
+  workflow_dispatch:
+
+concurrency:
+  group: deploy-files-${{ github.ref_name }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+
+jobs:
+  deploy:
+    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'dev' || github.ref_name == 'v2')
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+
+      - name: Typecheck
+        working-directory: services/files
+        run: bun typecheck
+
+      - name: Deploy
+        working-directory: services/files
+        run: bun run deploy --env ${{ github.ref_name == 'v2' && 'production' || 'dev' }}
+        env:
+          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}

.github/workflows/deploy-posts.yml

@@ -0,0 +1,37 @@
+name: deploy-posts
+
+on:
+  push:
+    branches:
+      - v2
+    paths:
+      - packages/posts/**
+      - bun.lock
+      - .github/workflows/deploy-posts.yml
+  workflow_dispatch:
+
+concurrency:
+  group: deploy-posts-${{ github.ref_name }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+
+jobs:
+  deploy:
+    if: github.repository == 'anomalyco/opencode' && github.ref_name == 'v2'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+
+      - name: Build
+        working-directory: packages/posts
+        run: bun run build
+
+      - name: Deploy
+        working-directory: packages/posts
+        run: bun run deploy
+        env:
+          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}

.github/workflows/deploy-www.yml

@@ -0,0 +1,36 @@
+name: deploy-www
+
+on:
+  push:
+    branches:
+      - dev
+      - v2
+  workflow_dispatch:
+
+concurrency:
+  group: deploy-www-${{ github.ref_name }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+
+jobs:
+  deploy:
+    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'dev' || github.ref_name == 'v2')
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+
+      - name: Build
+        working-directory: services/www
+        run: bun run build
+        env:
+          CLOUDFLARE_ENV: ${{ github.ref_name == 'v2' && 'production' || 'dev' }}
+
+      - name: Deploy
+        working-directory: services/www
+        run: bun run deploy
+        env:
+          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}

.github/workflows/deploy.yml

@@ -5,6 +5,7 @@ on:
     branches:
       - dev
       - production
+      - beta
   workflow_dispatch:
 
 concurrency: ${{ github.workflow }}-${{ github.ref }}
@@ -15,7 +16,7 @@ permissions:
 
 jobs:
   deploy:
-    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'dev' || github.ref_name == 'production')
+    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'dev' || github.ref_name == 'production' || github.ref_name == 'beta')
     runs-on: ubuntu-latest
     environment: ${{ github.ref_name }}
     steps:
@@ -28,6 +29,7 @@ jobs:
           node-version: "24"
 
       - uses: aws-actions/configure-aws-credentials@7474bc4690e29a8392af63c5b98e7449536d5c3a # v4.3.1
+        if: github.ref_name != 'beta'
         with:
           role-to-assume: ${{ vars.AWS_DEPLOY_ROLE_ARN }}
           role-session-name: opencode-${{ github.run_id }}

.github/workflows/generate.yml

@@ -1,51 +0,0 @@
-name: generate
-
-on:
-  push:
-    branches:
-      - dev
-
-jobs:
-  generate:
-    runs-on: blacksmith-4vcpu-ubuntu-2404
-    permissions:
-      contents: write
-      pull-requests: write
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
-
-      - name: Setup Bun
-        uses: ./.github/actions/setup-bun
-
-      - name: Setup git committer
-        id: committer
-        uses: ./.github/actions/setup-git-committer
-        with:
-          opencode-app-id: ${{ vars.OPENCODE_APP_ID }}
-          opencode-app-secret: ${{ secrets.OPENCODE_APP_SECRET }}
-
-      - name: Generate
-        run: ./script/generate.ts
-
-      - name: Commit and push
-        run: |
-          if [ -z "$(git status --porcelain)" ]; then
-            echo "No changes to commit"
-            exit 0
-          fi
-          git add -A
-          git commit -m "chore: generate" --allow-empty
-          git push origin HEAD:${{ github.ref_name }} --no-verify
-          # if ! git push origin HEAD:${{ github.event.pull_request.head.ref || github.ref_name }} --no-verify; then
-          #   echo ""
-          #   echo "============================================"
-          #   echo "Failed to push generated code."
-          #   echo "Please run locally and push:"
-          #   echo ""
-          #   echo "  ./script/generate.ts"
-          #   echo "  git add -A && git commit -m \"chore: generate\" && git push"
-          #   echo ""
-          #   echo "============================================"
-          #   exit 1
-          # fi

.github/workflows/nix-hashes.yml

@@ -6,11 +6,12 @@ permissions:
 on:
   workflow_dispatch:
   push:
-    branches: [dev, beta]
+    branches: [dev, beta, v2]
     paths:
       - "bun.lock"
       - "package.json"
       - "packages/*/package.json"
+      - "services/*/package.json"
       - "flake.lock"
       - "nix/node_modules.nix"
       - "nix/scripts/**"

.github/workflows/pr-standards.yml

@@ -135,7 +135,16 @@ jobs:
 
             const linkedIssues = result.repository.pullRequest.closingIssuesReferences.totalCount;
 
-            if (linkedIssues === 0) {
+            // GitHub only populates closingIssuesReferences when a PR targets the repository's
+            // default branch (dev). PRs targeting other branches like v2 always return totalCount 0.
+            // Fall back to checking the PR description for closing keywords (e.g. Closes #123).
+            const body = pr.body || '';
+            const issueMatch = body.match(/### Issue for this PR\s*\n([\s\S]*?)(?=###|$)/);
+            const issueContent = issueMatch ? issueMatch[1].trim() : body;
+            const hasBodyIssueRef = /(closes|fixes|resolves)\s+#\d+/i.test(issueContent) || /#\d+/.test(issueContent);
+            const hasLinkedIssue = linkedIssues > 0 || hasBodyIssueRef;
+
+            if (!hasLinkedIssue) {
               await addLabel('needs:issue');
               await comment('issue', `Thanks for your contribution!
 

.github/workflows/publish.yml

@@ -6,8 +6,8 @@ on:
     branches:
       - ci
       - dev
-      - v2
       - beta
+      - v2
       - fix/npm-native-binary-install
       - snapshot-*
   workflow_dispatch:
@@ -25,7 +25,7 @@ on:
         required: false
         type: string
 
-concurrency: ${{ github.workflow }}-${{ github.ref }}-${{ inputs.version || inputs.bump }}
+concurrency: ${{ github.workflow }}-${{ github.ref }}-${{ (github.ref_name == 'v2' && (inputs.version || inputs.bump) && 'release') || inputs.version || inputs.bump }}
 
 permissions:
   id-token: write
@@ -33,7 +33,7 @@ permissions:
   packages: write
 
 env:
-  OPENCODE_CHANNEL: ${{ (github.ref_name == 'v2' && 'next') || '' }}
+  OPENCODE_CHANNEL: ${{ (github.ref_name == 'v2' && !inputs.bump && !inputs.version && 'dev') || '' }}
 
 jobs:
   version:
@@ -46,6 +46,13 @@ jobs:
 
       - uses: ./.github/actions/setup-bun
 
+      - name: Deploy update service
+        if: github.ref_name == 'v2'
+        working-directory: services/update
+        run: bun run deploy
+        env:
+          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}
+
       - name: Setup git committer
         id: committer
         uses: ./.github/actions/setup-git-committer
@@ -75,6 +82,7 @@ jobs:
   build-cli:
     needs: version
     runs-on: blacksmith-4vcpu-ubuntu-2404
+    timeout-minutes: 30
     if: github.repository == 'anomalyco/opencode'
     steps:
       - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
@@ -82,6 +90,8 @@ jobs:
           fetch-tags: true
 
       - uses: ./.github/actions/setup-bun
+        with:
+          bun-version: 1.4.2
 
       - name: Setup git committer
         id: committer
@@ -90,18 +100,27 @@ jobs:
           opencode-app-id: ${{ vars.OPENCODE_APP_ID }}
           opencode-app-secret: ${{ secrets.OPENCODE_APP_SECRET }}
 
-      - name: Build
-        id: build
-        run: |
-          ./packages/opencode/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
-          ./packages/cli/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
+      - name: Build legacy CLI
+        if: github.ref_name != 'v2' && github.ref_name != 'beta'
+        run: ./packages/opencode/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
         env:
           OPENCODE_VERSION: ${{ needs.version.outputs.version }}
           OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
           GH_REPO: ${{ needs.version.outputs.repo }}
           GH_TOKEN: ${{ steps.committer.outputs.token }}
 
+      - name: Build preview CLI
+        id: build
+        run: ./packages/cli/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
+        env:
+          BUN_COMPILE_RELEASE: bun-v1.4.2
+          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
+          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
+          GH_REPO: ${{ needs.version.outputs.repo }}
+          GH_TOKEN: ${{ steps.committer.outputs.token }}
+
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        if: github.ref_name != 'v2' && github.ref_name != 'beta'
         with:
           name: opencode-cli
           path: |
@@ -109,24 +128,157 @@ jobs:
             packages/opencode/dist/opencode-linux*
 
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        if: github.ref_name != 'v2' && github.ref_name != 'beta'
         with:
           name: opencode-cli-windows
           path: packages/opencode/dist/opencode-windows*
 
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
-          name: opencode-preview-cli
+          name: opencode-preview-cli-unsigned
           path: packages/cli/dist/cli-*
 
     outputs:
       version: ${{ needs.version.outputs.version }}
 
+  sign-cli-macos:
+    needs: build-cli
+    runs-on: macos-26
+    if: github.repository == 'anomalyco/opencode'
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: apple-actions/import-codesign-certs@8f3fb608891dd2244cdab3d69cd68c0d37a7fe93 # v2.0.0
+        with:
+          keychain: build
+          p12-file-base64: ${{ secrets.APPLE_CERTIFICATE }}
+          p12-password: ${{ secrets.APPLE_CERTIFICATE_PASSWORD }}
+
+      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        with:
+          name: opencode-preview-cli-unsigned
+          path: packages/cli/dist
+
+      - name: Sign macOS CLI binaries
+        run: |
+          identity=$(security find-identity -v -p codesigning build.keychain | sed -n 's/.*"\(Developer ID Application:.*\)"/\1/p' | head -n 1)
+          if [ -z "$identity" ]; then
+            echo "Developer ID Application identity not found"
+            exit 1
+          fi
+
+          found=0
+          for file in packages/cli/dist/cli-darwin-*/bin/opencode; do
+            if [ ! -f "$file" ]; then
+              continue
+            fi
+            found=1
+            codesign \
+              --force \
+              --timestamp \
+              --options runtime \
+              --entitlements packages/cli/script/entitlements.plist \
+              --sign "$identity" \
+              "$file"
+            codesign --verify --deep --strict --verbose=4 "$file"
+            codesign --display --requirements - "$file"
+          done
+
+          if [ "$found" -eq 0 ]; then
+            echo "No macOS CLI binaries found"
+            exit 1
+          fi
+
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: opencode-preview-cli-macos
+          path: packages/cli/dist/cli-*
+          if-no-files-found: error
+
+  build-node-app-archive:
+    needs: version
+    runs-on: blacksmith-4vcpu-ubuntu-2404
+    timeout-minutes: 30
+    if: github.repository == 'anomalyco/opencode' && !(github.ref_name == 'v2' && (inputs.bump || inputs.version))
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+
+      - name: Build app archive
+        run: bun packages/cli/script/build-node.ts --app-archive-only --app-archive=.cache/app-archive.bin --skip-install
+        env:
+          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
+          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
+
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: opencode-node-app-archive
+          path: packages/cli/.cache/app-archive.bin
+          if-no-files-found: error
+
+  build-node-cli:
+    needs:
+      - version
+      - build-node-app-archive
+    if: github.repository == 'anomalyco/opencode' && !(github.ref_name == 'v2' && (inputs.bump || inputs.version))
+    strategy:
+      fail-fast: false
+      matrix:
+        settings:
+          - target: linux-arm64
+            host: blacksmith-4vcpu-ubuntu-2404-arm
+          - target: linux-x64
+            host: blacksmith-4vcpu-ubuntu-2404
+          - target: darwin-arm64
+            host: macos-26
+          - target: windows-arm64
+            host: blacksmith-4vcpu-windows-2025
+            bun_install_flags: --cpu=*
+          - target: windows-x64
+            host: blacksmith-4vcpu-windows-2025
+    runs-on: ${{ matrix.settings.host }}
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+        with:
+          install-flags: ${{ matrix.settings.bun_install_flags }}
+
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
+        with:
+          node-version: "26.4.0"
+
+      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        with:
+          name: opencode-node-app-archive
+          path: packages/cli/.cache
+
+      - name: Build
+        run: bun packages/cli/script/build-node.ts --target=${{ matrix.settings.target }} --skip-install --outdir=dist/node --app-archive=.cache/app-archive.bin
+        env:
+          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
+          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
+
+      - name: Verify service lifecycle
+        if: matrix.settings.target != 'windows-arm64'
+        working-directory: packages/cli
+        run: bun run script/service-smoke.ts --node
+
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: opencode-node-cli-${{ matrix.settings.target }}
+          path: packages/cli/dist/node/cli-node-*
+          if-no-files-found: error
+
   sign-cli-windows:
     needs:
-      - build-cli
-      - version
+      - sign-cli-macos
     runs-on: blacksmith-4vcpu-windows-2025
-    if: github.repository == 'anomalyco/opencode' && github.ref_name != 'v2'
+    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'v2' || github.ref_name == 'beta')
     env:
       AZURE_CLIENT_ID: ${{ secrets.AZURE_CLIENT_ID }}
       AZURE_TENANT_ID: ${{ secrets.AZURE_TENANT_ID }}
@@ -139,15 +291,8 @@ jobs:
 
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
         with:
-          name: opencode-cli-windows
-          path: packages/opencode/dist
-
-      - name: Setup git committer
-        id: committer
-        uses: ./.github/actions/setup-git-committer
-        with:
-          opencode-app-id: ${{ vars.OPENCODE_APP_ID }}
-          opencode-app-secret: ${{ secrets.OPENCODE_APP_SECRET }}
+          name: opencode-preview-cli-macos
+          path: packages/cli/dist
 
       - name: Azure login
         uses: azure/login@a457da9ea143d694b1b9c7c869ebb04ebe844ef5 # v2.3.0
@@ -162,9 +307,9 @@ jobs:
           signing-account-name: ${{ env.AZURE_TRUSTED_SIGNING_ACCOUNT_NAME }}
           certificate-profile-name: ${{ env.AZURE_TRUSTED_SIGNING_CERTIFICATE_PROFILE }}
           files: |
-            ${{ github.workspace }}\packages\opencode\dist\opencode-windows-arm64\bin\opencode.exe
-            ${{ github.workspace }}\packages\opencode\dist\opencode-windows-x64\bin\opencode.exe
-            ${{ github.workspace }}\packages\opencode\dist\opencode-windows-x64-baseline\bin\opencode.exe
+            ${{ github.workspace }}\packages\cli\dist\cli-windows-arm64\bin\opencode.exe
+            ${{ github.workspace }}\packages\cli\dist\cli-windows-x64\bin\opencode.exe
+            ${{ github.workspace }}\packages\cli\dist\cli-windows-x64-baseline\bin\opencode.exe
           exclude-environment-credential: true
           exclude-workload-identity-credential: true
           exclude-managed-identity-credential: true
@@ -180,9 +325,9 @@ jobs:
         shell: pwsh
         run: |
           $files = @(
-            "${{ github.workspace }}\packages\opencode\dist\opencode-windows-arm64\bin\opencode.exe",
-            "${{ github.workspace }}\packages\opencode\dist\opencode-windows-x64\bin\opencode.exe",
-            "${{ github.workspace }}\packages\opencode\dist\opencode-windows-x64-baseline\bin\opencode.exe"
+            "${{ github.workspace }}\packages\cli\dist\cli-windows-arm64\bin\opencode.exe",
+            "${{ github.workspace }}\packages\cli\dist\cli-windows-x64\bin\opencode.exe",
+            "${{ github.workspace }}\packages\cli\dist\cli-windows-x64-baseline\bin\opencode.exe"
           )
 
           foreach ($file in $files) {
@@ -192,40 +337,17 @@ jobs:
             }
           }
 
-      - name: Repack Windows CLI archives
-        working-directory: packages/opencode/dist
-        shell: pwsh
-        run: |
-          Compress-Archive -Path "opencode-windows-arm64\bin\*" -DestinationPath "opencode-windows-arm64.zip" -Force
-          Compress-Archive -Path "opencode-windows-x64\bin\*" -DestinationPath "opencode-windows-x64.zip" -Force
-          Compress-Archive -Path "opencode-windows-x64-baseline\bin\*" -DestinationPath "opencode-windows-x64-baseline.zip" -Force
-
-      - name: Upload signed Windows CLI release assets
-        if: needs.version.outputs.release != ''
-        shell: pwsh
-        env:
-          GH_TOKEN: ${{ steps.committer.outputs.token }}
-        run: |
-          gh release upload "v${{ needs.version.outputs.version }}" `
-            "${{ github.workspace }}\packages\opencode\dist\opencode-windows-arm64.zip" `
-            "${{ github.workspace }}\packages\opencode\dist\opencode-windows-x64.zip" `
-            "${{ github.workspace }}\packages\opencode\dist\opencode-windows-x64-baseline.zip" `
-            --clobber `
-            --repo "${{ needs.version.outputs.repo }}"
-
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
-          name: opencode-cli-signed-windows
-          path: |
-            packages/opencode/dist/opencode-windows-arm64
-            packages/opencode/dist/opencode-windows-x64
-            packages/opencode/dist/opencode-windows-x64-baseline
+          name: opencode-preview-cli
+          path: packages/cli/dist/cli-*
+          if-no-files-found: error
 
   build-electron:
     needs:
-      - build-cli
       - version
-    if: github.repository == 'anomalyco/opencode' && github.ref_name != 'v2'
+      - sign-cli-windows
+    if: github.repository == 'anomalyco/opencode' && (github.ref_name != 'v2' || needs.version.outputs.release != '')
     continue-on-error: false
     env:
       AZURE_CLIENT_ID: ${{ secrets.AZURE_CLIENT_ID }}
@@ -263,6 +385,11 @@ jobs:
     steps:
       - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
 
+      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        with:
+          name: opencode-preview-cli
+          path: packages/cli/dist
+
       - uses: apple-actions/import-codesign-certs@8f3fb608891dd2244cdab3d69cd68c0d37a7fe93 # v2.0.0
         if: runner.os == 'macOS'
         with:
@@ -320,16 +447,15 @@ jobs:
         env:
           OPENCODE_VERSION: ${{ needs.version.outputs.version }}
           OPENCODE_CHANNEL: ${{ (github.ref_name == 'beta' && 'beta') || 'prod' }}
-          OPENCODE_CLI_ARTIFACT: ${{ (runner.os == 'Windows' && 'opencode-cli-windows') || 'opencode-cli' }}
-          RUST_TARGET: ${{ matrix.settings.target }}
-          GH_TOKEN: ${{ github.token }}
-          GITHUB_RUN_ID: ${{ github.run_id }}
+          OPENCODE_CLI_TARGET: ${{ matrix.settings.target }}
+          OPENCODE_CLI_DIST: ${{ github.workspace }}/packages/cli/dist
 
       - name: Build
         run: bun run build
         working-directory: packages/desktop
         env:
           NODE_OPTIONS: --max-old-space-size=4096
+          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
           OPENCODE_CHANNEL: ${{ (github.ref_name == 'beta' && 'beta') || 'prod' }}
           SENTRY_AUTH_TOKEN: ${{ secrets.SENTRY_AUTH_TOKEN }}
           SENTRY_ORG: ${{ vars.SENTRY_ORG }}
@@ -338,6 +464,8 @@ jobs:
           VITE_SENTRY_DSN: ${{ vars.WEB_SENTRY_DSN }}
           VITE_SENTRY_ENVIRONMENT: ${{ (github.ref_name == 'beta' && 'beta') || 'production' }}
           VITE_SENTRY_RELEASE: desktop@${{ needs.version.outputs.version }}
+          OPENCODE_CLI_TARGET: ${{ matrix.settings.target }}
+          OPENCODE_CLI_DIST: ${{ github.workspace }}/packages/cli/dist
 
       - name: Package
         if: needs.version.outputs.release
@@ -347,8 +475,7 @@ jobs:
         env:
           OPENCODE_CHANNEL: ${{ (github.ref_name == 'beta' && 'beta') || 'prod' }}
           GH_TOKEN: ${{ steps.committer.outputs.token }}
-          CSC_LINK: ${{ secrets.APPLE_CERTIFICATE }}
-          CSC_KEY_PASSWORD: ${{ secrets.APPLE_CERTIFICATE_PASSWORD }}
+          CSC_KEYCHAIN: build.keychain
           APPLE_API_KEY: ${{ runner.temp }}/apple-api-key.p8
           APPLE_API_KEY_ID: ${{ secrets.APPLE_API_KEY }}
           APPLE_API_ISSUER: ${{ secrets.APPLE_API_ISSUER }}
@@ -413,6 +540,9 @@ jobs:
     needs:
       - version
       - build-cli
+      - sign-cli-macos
+      - build-node-app-archive
+      - build-node-cli
       - sign-cli-windows
       - build-electron
     if: always() && !failure() && !cancelled()
@@ -441,17 +571,19 @@ jobs:
           registry-url: "https://registry.npmjs.org"
 
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        if: github.ref_name != 'v2' && github.ref_name != 'beta'
         with:
           name: opencode-cli
           path: packages/opencode/dist
 
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        if: github.ref_name != 'v2' && github.ref_name != 'beta'
         with:
           name: opencode-cli-windows
           path: packages/opencode/dist
 
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        if: github.ref_name != 'v2'
+        if: github.ref_name != 'v2' && github.ref_name != 'beta'
         with:
           name: opencode-cli-signed-windows
           path: packages/opencode/dist
@@ -461,6 +593,13 @@ jobs:
           name: opencode-preview-cli
           path: packages/cli/dist
 
+      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        if: needs.build-node-cli.result == 'success'
+        with:
+          pattern: opencode-node-cli-*
+          path: packages/cli/dist/node
+          merge-multiple: true
+
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
         if: needs.version.outputs.release
         with:
@@ -500,19 +639,6 @@ jobs:
           git config --global user.name "opencode"
           ssh-keyscan -H aur.archlinux.org >> ~/.ssh/known_hosts || true
 
-      - name: Upload desktop release assets
-        if: needs.version.outputs.release
-        env:
-          GH_TOKEN: ${{ steps.committer.outputs.token }}
-        run: |
-          shopt -s nullglob
-          files=(/tmp/desktop/*.{exe,blockmap,dmg,zip,AppImage,deb,rpm} /tmp/desktop/*.app.tar.gz)
-          if (( ${#files[@]} == 0 )); then
-            echo "No desktop release assets found"
-            exit 1
-          fi
-          gh release upload "v${{ needs.version.outputs.version }}" "${files[@]}" --clobber --repo "${{ needs.version.outputs.repo }}"
-
       - run: ./script/publish.ts
         env:
           OPENCODE_VERSION: ${{ needs.version.outputs.version }}
@@ -524,3 +650,6 @@ jobs:
           LATEST_YML_DIR: /tmp/latest-yml
           TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
           TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
+          OPENCODE_DESKTOP_DIST: /tmp/desktop
+          CLOUDFLARE_ACCOUNT_ID: 15d29c8639fd3733b1b5486a2acfd968
+          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}

.github/workflows/test.yml

@@ -22,6 +22,36 @@ env:
   FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true
 
 jobs:
+  affected:
+    name: affected packages
+    runs-on: blacksmith-4vcpu-ubuntu-2404
+    outputs:
+      app: ${{ steps.packages.outputs.app }}
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
+        with:
+          token: ${{ secrets.GITHUB_TOKEN }}
+          fetch-depth: 0
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
+        with:
+          bun-version-file: package.json
+
+      - name: Find affected packages
+        id: packages
+        env:
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            echo "app=true" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          bun x turbo@2.10.2 ls --affected --filter=@opencode/app --output=json > affected.json
+          bun -e 'const result = await Bun.file("affected.json").json(); console.log(`app=${result.packages.count > 0}`)' >> "$GITHUB_OUTPUT"
+
   unit:
     name: unit (${{ matrix.settings.name }})
     strategy:
@@ -41,6 +71,7 @@ jobs:
         uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
+          fetch-depth: 0
 
       - name: Setup Node
         uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
@@ -50,11 +81,25 @@ jobs:
       - name: Setup Bun
         uses: ./.github/actions/setup-bun
 
+      - name: Test Effect simplification rules
+        if: runner.os == 'Linux'
+        run: bun run test:effect-simplification-rules
+
+      - name: Check Effect simplifications
+        if: runner.os == 'Linux'
+        run: bun run lint:effect-simplifications
+
       - name: Configure git identity
         run: |
           git config --global user.email "bot@opencode.ai"
           git config --global user.name "opencode"
 
+      - name: Install ffmpeg
+        if: runner.os == 'Linux'
+        run: |
+          sudo apt-get update
+          sudo apt-get install --yes ffmpeg
+
       - name: Cache Turbo
         uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
         with:
@@ -66,18 +111,74 @@ jobs:
 
       - name: Run unit tests
         timeout-minutes: 20
-        run: GITHUB_ACTIONS=false bun turbo test
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            GITHUB_ACTIONS=false bun turbo test
+            exit 0
+          fi
+          GITHUB_ACTIONS=false bun turbo test --affected
         env:
           OPENCODE_EXPERIMENTAL_DISABLE_FILEWATCHER: ${{ runner.os == 'Windows' && 'true' || 'false' }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}
+
+      - name: Verify published codemode package
+        if: runner.os == 'Linux'
+        working-directory: packages/codemode
+        run: bun run script/publish.ts --dry-run
+
+      - name: Verify packed workerd SDK
+        if: runner.os == 'Linux'
+        timeout-minutes: 15
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            bun turbo verify:package --filter=@opencode/sdk
+            exit 0
+          fi
+          bun turbo verify:package --affected --filter=@opencode/sdk
+        env:
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}
+
+      - name: Verify compiled service lifecycle
+        if: always()
+        timeout-minutes: 10
+        working-directory: packages/cli
+        env:
+          NODE_OPTIONS: ${{ runner.os == 'Windows' && '--max-old-space-size=4096' || '' }}
+        run: |
+          bun run script/build.ts --single --skip-install
+          bun run script/service-smoke.ts
+
+      - name: Setup Node build runtime
+        if: always()
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
+        with:
+          node-version: "26.4.0"
+
+      - name: Verify Node build
+        if: always()
+        timeout-minutes: 15
+        working-directory: packages/cli
+        env:
+          NODE_OPTIONS: ${{ runner.os == 'Windows' && '--max-old-space-size=4096' || '' }}
+        run: |
+          bun run script/build-node.ts --single --skip-install --outdir=dist/node
+          bun run script/service-smoke.ts --node
 
       - name: Check generated client
         if: runner.os == 'Linux'
         working-directory: packages/client
         run: bun run check:generated
 
+      - name: Check generated documentation
+        if: runner.os == 'Linux'
+        working-directory: services/www
+        run: bun run check:generated
+
   e2e:
     name: e2e (${{ matrix.settings.name }})
-    if: github.ref_name != 'v2' && github.head_ref != 'v2'
+    needs: affected
     strategy:
       fail-fast: false
       matrix:
@@ -88,32 +189,38 @@ jobs:
             host: blacksmith-4vcpu-windows-2025
     runs-on: ${{ matrix.settings.host }}
     env:
+      E2E_ENABLED: ${{ needs.affected.outputs.app == 'true' && github.ref_name != 'v2' && github.head_ref != 'v2' }}
       PLAYWRIGHT_BROWSERS_PATH: ${{ github.workspace }}/.playwright-browsers
     defaults:
       run:
         shell: bash
     steps:
       - name: Checkout repository
+        if: env.E2E_ENABLED == 'true'
         uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Setup Node
+        if: env.E2E_ENABLED == 'true'
         uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
         with:
           # Playwright 1.59 hangs while extracting Chromium with Node 24.16.
           node-version: "24.15"
 
       - name: Setup Bun
+        if: env.E2E_ENABLED == 'true'
         uses: ./.github/actions/setup-bun
 
       - name: Read Playwright version
+        if: env.E2E_ENABLED == 'true'
         id: playwright-version
         run: |
           version=$(node -e 'console.log(require("./package.json").workspaces.catalog["@playwright/test"])')
           echo "version=$version" >> "$GITHUB_OUTPUT"
 
       - name: Cache Playwright browsers
+        if: env.E2E_ENABLED == 'true'
         id: playwright-cache
         uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
         with:
@@ -121,23 +228,30 @@ jobs:
           key: ${{ runner.os }}-${{ runner.arch }}-playwright-${{ steps.playwright-version.outputs.version }}-chromium
 
       - name: Install Playwright system dependencies
-        if: runner.os == 'Linux'
+        if: env.E2E_ENABLED == 'true' && runner.os == 'Linux'
         working-directory: packages/app
         run: bunx playwright install-deps chromium
 
       - name: Install Playwright browsers
-        if: steps.playwright-cache.outputs.cache-hit != 'true'
+        if: env.E2E_ENABLED == 'true' && steps.playwright-cache.outputs.cache-hit != 'true'
         working-directory: packages/app
         run: bunx playwright install chromium
 
-      - name: Run app e2e tests
-        run: bun --cwd packages/app test:e2e:local
+      - name: Run app e2e tests against production build
+        if: env.E2E_ENABLED == 'true'
+        run: bun --cwd packages/app test:e2e:built
         env:
           CI: true
         timeout-minutes: 30
 
+      - name: Verify service worker precaching and upgrades
+        if: env.E2E_ENABLED == 'true'
+        working-directory: packages/app
+        run: bunx playwright test --config e2e/service-worker/playwright.config.ts
+        timeout-minutes: 5
+
       - name: Upload Playwright artifacts
-        if: always()
+        if: always() && env.E2E_ENABLED == 'true'
         uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: playwright-${{ matrix.settings.name }}-${{ github.run_attempt }}

.github/workflows/typecheck.yml

@@ -1,21 +0,0 @@
-name: typecheck
-
-on:
-  push:
-    branches: [dev, v2]
-  pull_request:
-    branches: [dev, v2]
-  workflow_dispatch:
-
-jobs:
-  typecheck:
-    runs-on: blacksmith-4vcpu-ubuntu-2404
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
-
-      - name: Setup Bun
-        uses: ./.github/actions/setup-bun
-
-      - name: Run typecheck
-        run: bun typecheck

.gitignore

@@ -11,6 +11,7 @@ node_modules
 playground
 tmp
 dist
+dist-node
 ts-dist
 .turbo
 .typecheck-profiles
@@ -25,11 +26,13 @@ Session.vim
 a.out
 target
 .scripts
+.cache
 .direnv/
 
 # Local dev files
 opencode-dev
 UPCOMING_CHANGELOG.md
+RELEASE_REVIEW.md
 logs/
 *.bun-build
 tsconfig.tsbuildinfo

.husky/pre-push

@@ -17,4 +17,4 @@ if (process.versions.bun !== expectedBunVersion) {
   console.warn(`Warning: Bun version ${process.versions.bun} differs from expected ${expectedBunVersion}`);
 }
 '
-bun typecheck
+bun run check

.opencode/command/ai-deps.md

@@ -2,7 +2,7 @@
 description: "Bump AI sdk dependencies minor / patch versions only"
 ---
 
-Please read @package.json and @packages/opencode/package.json.
+Please read @package.json and @packages/core/package.json.
 
 Your job is to look into AI SDK dependencies, figure out if they have versions that can be upgraded (minor or patch versions ONLY no major ignore major changes).
 

.opencode/command/commit.md

@@ -6,15 +6,7 @@ subtask: true
 
 commit and push
 
-make sure it includes a prefix like
-docs:
-tui:
-core:
-ci:
-ignore:
-wip:
-
-For anything in the packages/web use the docs: prefix.
+Use `type(scope): summary` with one of these types: `feat`, `fix`, `docs`, `chore`, `refactor`, or `test`. The scope is optional.
 
 prefer to explain WHY something was done from an end user perspective instead of
 WHAT was done.

.opencode/command/rmslop.md

@@ -2,7 +2,7 @@
 description: Remove AI code slop
 ---
 
-Check the diff against dev, and remove all AI generated slop introduced in this branch.
+Check the diff against `origin/v2`, and remove all AI generated slop introduced in this branch.
 
 This includes:
 

.opencode/command/translate.md

@@ -1,6 +1,6 @@
 ---
 description: translate English to other languages
-model: opencode/claude-opus-4-8
+model: opencode/gpt-5.6-sol
 ---
 
 run git diff and translate changed english doc and UI copy files to other international languages. Translate all languages in parallel to save time.

.opencode/glossary/he.md

@@ -0,0 +1,48 @@
+# he Glossary
+
+## Sources
+
+- Hebrew Academy approved IT terminology: https://terms.hebrew-academy.org.il/Millonim/ShowMillon?KodMillon=192
+- Firefox Hebrew localization corpus: https://github.com/mozilla-l10n/firefox-l10n/tree/main/he
+- KDE Hebrew localization team and corpus: https://l10n.kde.org/team-infos.php?teamcode=he
+- Community-maintained VS Code Hebrew language pack: https://github.com/AMAARETS/vscode-language-pack-he
+- Microsoft Hebrew developer documentation for Git terminology: https://learn.microsoft.com/he-il/power-platform/alm/tutorials/github-actions-deploy
+- W3C guidance for bidirectional text: https://www.w3.org/International/articles/strings-and-bidi/
+
+## Do Not Translate (Locale Additions)
+
+- `OpenCode` (preserve casing in prose and UI copy)
+- `API`, `MCP`, `LSP`, `OAuth`, `Git`, model names, and provider names
+- Commands, flags, keyboard shortcuts, file paths, URLs, identifiers, hashes, and code literals
+- Keep `commit` and `diff` when they name the exact Git artifact or operation
+
+## Preferred Terms
+
+| English / Context | Preferred     | Notes                                                                     |
+| ----------------- | ------------- | ------------------------------------------------------------------------- |
+| session           | `הפעלה`       | Use `שיחה` only when the source specifically means a chat or conversation |
+| workspace         | `סביבת עבודה` |                                                                           |
+| terminal          | `מסוף`        | Prefer the established Hebrew term over transliteration                   |
+| command           | `פקודה`       |                                                                           |
+| provider          | `ספק`         | Use `ספק מודלים` where the bare noun is ambiguous                         |
+| model             | `מודל`        |                                                                           |
+| API key           | `מפתח API`    | Keep the acronym in Latin letters                                         |
+| plugin            | `תוסף`        |                                                                           |
+| repository        | `מאגר`        | Use `מאגר Git` where context is ambiguous                                 |
+| branch            | `ענף`         |                                                                           |
+| context           | `הקשר`        | Use `חלון הקשר` for context window                                        |
+| tokens            | `אסימונים`    |                                                                           |
+
+## Guidance
+
+- Prefer natural modern Israeli Hebrew over word-for-word translation or obscure coined terms.
+- Use short action verbs for controls and translate complete phrases in context.
+- Keep recognized developer acronyms and exact Git vocabulary in Latin script instead of phonetic transliteration.
+- Treat embedded code, paths, commands, shortcuts, hashes, model IDs, and other Latin technical artifacts as LTR content inside the RTL interface.
+- Keep recurring concepts consistent and do not collapse session, chat, run, and launch into one Hebrew term.
+
+## Avoid
+
+- Avoid transliterations such as `טרמינל`, `פלאגין`, and `קומנד` when `מסוף`, `תוסף`, and `פקודה` are clear.
+- Avoid translating `commit` as `התחייבות`.
+- Avoid inventing Hebrew expansions for `API`, `MCP`, or `LSP`.

.opencode/skills/effect/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: effect
-description: Work with Effect v4 / effect-smol TypeScript code in this repo
+description: Work with Effect v4 TypeScript code in this repo
 ---
 
 # Effect
@@ -9,10 +9,10 @@ This codebase uses Effect for typed, composable TypeScript services, schemas, an
 
 ## Source Of Truth
 
-Use the current Effect v4 / effect-smol source, not memory or older Effect v2/v3 examples.
+Use the current Effect v4 source, not memory or older Effect v2/v3 examples.
 
-1. If `.opencode/references/effect-smol` is missing, clone `https://github.com/Effect-TS/effect-smol` there. Do this in the project, not in the skill folder.
-2. Search `.opencode/references/effect-smol` for exact APIs, examples, tests, and naming patterns before answering or implementing Effect-specific code.
+1. If `.opencode/references/effect` is missing, clone `https://github.com/Effect-TS/effect` there. Do this in the project, not in the skill folder.
+2. Search `.opencode/references/effect` for exact APIs, examples, tests, and naming patterns before answering or implementing Effect-specific code.
 3. Also inspect existing repo code for local house style before introducing new patterns.
 4. Prefer answers and implementations backed by specific source files or nearby repo examples.
 
@@ -27,12 +27,12 @@ Use the current Effect v4 / effect-smol source, not memory or older Effect v2/v3
 - Keep layer composition explicit. Avoid broad hidden provisioning that makes missing dependencies hard to see.
 - In tests, prefer the repo's existing Effect test helpers and live tests for filesystem, git, child process, locks, or timing behavior.
 - Do not introduce `any`, non-null assertions, unchecked casts, or older Effect APIs just to satisfy types.
-- Do not answer from memory. Verify against `.opencode/references/effect-smol` or nearby code first.
+- Do not answer from memory. Verify against `.opencode/references/effect` or nearby code first.
 
 ## Testing Patterns
 
-- Use `testEffect(...)` from `packages/opencode/test/lib/effect.ts` for tests that exercise Effect services, layers, runtime context, scoped resources, or platform integrations.
+- Use `testEffect(...)` from `packages/core/test/lib/effect.ts` for tests that exercise Effect services, layers, runtime context, scoped resources, or platform integrations.
 - Use `it.live(...)` for filesystem, git repositories, HTTP servers, sockets, child processes, locks, real time, and other live platform behavior.
-- Run tests from package directories such as `packages/opencode`; never run package tests from the repo root.
+- Run tests from package directories such as `packages/core`; never run package tests from the repo root.
 - Prefer explicit test layers over ad hoc managed runtimes. Keep dependency provisioning visible in the test file.
 - Use scoped fixtures and finalizers for resources that must be cleaned up, including temporary directories, flags, databases, fibers, servers, and global state.

.opencode/skills/opencode-dev/SKILL.md

@@ -0,0 +1,231 @@
+---
+name: opencode-dev
+description: Use when interactively running, debugging, or verifying opencode's own V2 CLI/TUI or server during development in this repo — starting the dev TUI, driving it with termctrl, comparing V2 against the legacy TUI, hitting the V2 server/API directly, reading log files, or attaching Bun's inspector.
+---
+
+# Debugging opencode itself
+
+Workflow for interactively exercising the V2 CLI/TUI and server while developing in this repo. All commands below run from `packages/cli` unless noted otherwise.
+
+## Server/client model
+
+opencode V2 is a client/server system, not a single monolithic process:
+
+- **Server process** runs the Effect HTTP API (`packages/server`) and owns all domain state: sessions, database, plugins, permissions, Location services. It's started by the `serve` command (`packages/cli/src/commands/handlers/serve.ts`).
+- **TUI process** is a separate process that runs no application logic itself — it's an HTTP/SSE client of the server via the generated SDK (`createOpencodeClient` / `sdk.client.v2`).
+- **Discovery**: CLI processes find the shared server through a JSON registration file at `~/.local/state/opencode/service.json` (or `service-local.json` for the local/dev channel) containing `{id, version, url, pid}`. A separate password file under `~/.config/opencode/service.json` provides HTTP Basic auth. Before reusing a registration, the client calls `GET /health` to confirm the server is alive, authenticated, and version-compatible.
+- **Sharing**: because of this registration/health-check dance, many concurrent `opencode`/TUI invocations converge on one shared background daemon rather than each spawning their own. If no compatible healthy daemon is found, a new one is spawned detached (`serve --service`) and registers itself.
+- **`bun dev service start|status|stop|restart`** manages this shared background daemon's lifecycle directly — useful when you need to force a fresh server, confirm one is running, or kill a stuck one.
+- **Standalone mode** (`--standalone`) opts a single invocation out of the shared daemon: it spawns a private one-off `serve --stdio --port 0` child tied to that invocation's lifetime, with its own random password. Use this to isolate a debugging session from your other running opencode sessions.
+- Every log line is tagged `role=server` or `role=cli` and a per-process `run=<id>`, so you can distinguish server-side and client-side activity in one shared log file (see "Logs" below) even when both roles are interleaved from concurrent processes.
+
+## Starting the dev TUI
+
+- This package is the V2 CLI adapter. Run its `dev` script when testing the TUI; do not use the repository-root `bun dev`, which launches the legacy `packages/opencode` CLI.
+- Run commands from `packages/cli`. Use `bun dev` for most debugging so the TUI starts with a private V2 server.
+
+## Interactive debugging with termctrl
+
+- Use `termctrl` for interactive checks instead of starting the TUI as a blocking foreground process. It provides a real PTY, handles OpenTUI's host handshake, and can save reviewable screenshots.
+- Use a dedicated session name and do not reuse or kill an unrelated session.
+
+```bash
+termctrl start opencode-v2-dev --host opentui --cols 112 --rows 34 -- bun dev
+termctrl wait opencode-v2-dev "Ask anything" --timeout 20000
+termctrl show opencode-v2-dev
+```
+
+- Wait for visible text before interacting instead of relying on fixed sleeps. Use the text expected from the screen under test, such as `Ask anything` or `Connect a provider`.
+- Drive the running TUI with `termctrl send`. Prefix typed input with `text:` and send control keys separately so the interaction matches real terminal input.
+
+```bash
+termctrl send opencode-v2-dev 'text:example prompt' enter
+termctrl send opencode-v2-dev ctrl-c
+```
+
+- Use `termctrl show` after each meaningful interaction and inspect the full visible screen for rendering errors, stale state, error toasts, and unexpected exits.
+- Save PNG evidence for every user-visible bug and fix. Do not save text captures; inspect the rendered PNG. Write temporary captures outside the repository unless the artifact is intended to be committed.
+
+```bash
+termctrl save opencode-v2-dev --format png --out /tmp/opencode/v2-tui.png
+```
+
+- For resize-sensitive changes, resize the viewport, wait for the expected content, and capture the screen again:
+
+```bash
+termctrl resize opencode-v2-dev --cols 100 --rows 30
+termctrl show opencode-v2-dev
+```
+
+- Source changes may require restarting the process. Use `termctrl restart opencode-v2-dev` rather than assuming the running TUI reloaded the change.
+- To exercise background-service behavior, use `bun dev service start`, `bun dev service status`, and `bun dev service stop`.
+- Always clean up the Terminal Control session when the check is complete:
+
+```bash
+termctrl stop opencode-v2-dev
+```
+
+## Comparing V2 against the legacy TUI
+
+Run both versions in separate Terminal Control sessions and save PNG-only captures at equivalent states:
+
+```bash
+# From packages/cli: local V2 TUI
+termctrl start opencode-v2-dev --host opentui --cols 112 --rows 34 -- bun dev
+
+# Released legacy TUI behavior reference
+termctrl start opencode-legacy --host opentui --cols 112 --rows 34 -- bunx opencode-ai@latest
+
+termctrl save opencode-v2-dev --format png --out /tmp/opencode/v2.png
+termctrl save opencode-legacy --format png --out /tmp/opencode/legacy.png
+```
+
+- Use the same viewport and send equivalent inputs to both sessions before comparing screenshots. The released CLI is a behavioral reference, not a source of V2 API design; keep the local implementation on V2 endpoints.
+- Stop both sessions after comparison: `termctrl stop opencode-v2-dev` and `termctrl stop opencode-legacy`.
+
+## Server/API debugging
+
+- Use `bun dev api --help` from `packages/cli` to inspect the API debugging command. It sends one request to the V2 server using the same daemon discovery/auth path as the CLI.
+- Use `bun dev api` to introspect the server-side data backing the TUI. This is useful when debugging UI bugs: compare what the screen renders with the raw session, message, event, agent, or health data returned by the API to determine whether the bug is in the server state, the client data layer, or the TUI rendering.
+- `bun dev api` accepts either an OpenAPI operation ID or a raw HTTP method plus path:
+
+```bash
+bun dev api get /health
+bun dev api get /openapi.json
+bun dev api <operationId> --param key=value
+```
+
+- Pass JSON request bodies with `--data`/`-d`; the command sets `content-type: application/json` automatically unless you provide a header. Add extra headers with `--header`/`-H name:value`.
+- If no compatible background server is registered, `bun dev api` starts one through the daemon service. Use `bun dev service status`, `bun dev service restart`, and `bun dev service stop` when you need explicit lifecycle control.
+- Prefer raw method/path calls for quick server debugging and operation IDs when exercising documented OpenAPI routes with path or query parameters.
+
+## Auditing installed `opencode2` sessions
+
+Installed next-channel sessions normally use `~/.local/share/opencode/opencode-next.db` and `~/.local/share/opencode/log/opencode.log`; `OPENCODE_DB` can override the database. Before calling `opencode2 api`, inspect `~/.local/state/opencode/service.json` because the command may start a daemon when none is healthy.
+
+For a supplied `ses_...` ID, compare three sources:
+
+- `opencode2 api get /api/session/active` and the Session/message endpoints for live server state.
+- The database's ordered `event` rows for durable history.
+- `packages/tui/src/context/data.tsx` and the relevant route for client projection and rendering.
+
+Locate an uncertain database without modifying it:
+
+```bash
+SESSION=ses_...
+for db in ~/.local/share/opencode/*.db; do
+  sqlite3 "file:$db?mode=ro" "select 1 from session where id='$SESSION' limit 1" 2>/dev/null | grep -q 1 && printf '%s\n' "$db"
+done
+```
+
+## Logs
+
+- Log files live under `~/.local/share/opencode/log/`. In a local/dev checkout the active file is `opencode-local.log`; `opencode.log` is used for non-local (released) channel installs. Both are append-only, shared across every CLI and server process on the machine.
+- Each line is structured `key=value` text: `timestamp`, `level`, `run=<id>` (per-process run ID), `message`, and a `role=cli` or `role=server` tag. Use `run=` to isolate one process's activity and `role=` to separate client-side from server-side log lines, since a shared daemon interleaves many processes' output in one file.
+- Tail the live file while reproducing an issue instead of guessing from stale output:
+
+```bash
+tail -f ~/.local/share/opencode/log/opencode-local.log
+```
+
+- Filter to one run or role when the file is noisy:
+
+```bash
+grep 'run=8fc3b1d5' ~/.local/share/opencode/log/opencode-local.log
+grep 'role=server' ~/.local/share/opencode/log/opencode-local.log
+```
+
+- `OPENCODE_LOG_LEVEL` controls verbosity (default `INFO`); set it before starting `bun dev` or `serve` to get `DEBUG` output for a specific repro.
+- `OPENCODE_PRINT_LOGS=1` additionally tees log output to stderr of the process that emitted it, which is useful when a process fails before you'd think to check the shared log file.
+- `termctrl logs <session>` surfaces stdout/stderr for a Terminal Control session specifically (e.g. inspector output or startup failures before the TUI renderer starts) — use the log file above for anything emitted by a separate server/daemon process instead.
+
+## Heap snapshots
+
+The CLI installs a `SIGUSR1` listener on non-Windows processes in `packages/cli/src/heap.ts`. Use it to capture the installed `opencode2` server without restarting it or attaching an inspector.
+
+1. Find the processes and inspect their roles and memory:
+
+```bash
+pgrep -a -f 'opencode2\.exe|opencode2'
+ps -o pid,ppid,rss,vsz,lstart,etime,cmd -p <pid>,<pid>
+```
+
+2. Signal the process whose heap needs investigation. For shared-service memory, target the `opencode2.exe serve --service` child, not the short wrapper/TUI process:
+
+```bash
+kill -USR1 <server-pid>
+```
+
+3. Wait for `heap snapshot written` in the channel's log before opening the file. Snapshots are written to the same log directory as `heap-<pid>-<timestamp>.heapsnapshot`; writing a large heap can take several seconds and the file is incomplete until the completion message appears:
+
+```bash
+grep 'heap snapshot' ~/.local/share/opencode/log/opencode.log | tail
+find ~/.local/share/opencode/log -maxdepth 1 -name 'heap-<server-pid>-*.heapsnapshot' -printf '%T@ %s %p\n' | sort -nr | head
+```
+
+Use `opencode-local.log` instead for a local/dev channel process. The log's `path=` field is authoritative.
+
+4. Analyze the snapshot with Chrome DevTools, a V8 heap snapshot parser, or a temporary tool installed outside the repository. Start with the largest retained objects, dominators, object counts grouped by constructor/name, and retainer paths back to GC roots. Relate suspicious names and paths back to the source tree rather than treating large shallow allocations as leaks.
+
+For command-line analysis, install tooling under `/tmp/opencode`, not in the repository. For example, MemLab can rank dominators and trace a reported heap object ID back to a GC root:
+
+```bash
+npm install --prefix /tmp/opencode/heap-analysis @memlab/cli
+/tmp/opencode/heap-analysis/node_modules/.bin/memlab analyze object-size --snapshot <snapshot>
+/tmp/opencode/heap-analysis/node_modules/.bin/memlab analyze shape --snapshot <snapshot>
+/tmp/opencode/heap-analysis/node_modules/.bin/memlab trace --snapshot <snapshot> --node-id=<id>
+```
+
+A single snapshot explains what retains memory at one point in time, but does not by itself prove a leak. For leak confirmation, capture a baseline, perform a controlled repeated workload, allow idle cleanup/GC when possible, capture another snapshot, and compare growth and retainer paths. Also compare snapshot heap size with process RSS: a large difference can indicate native allocations, database mappings, allocator fragmentation, or other memory outside the JavaScript heap.
+
+```bash
+cat /proc/<pid>/smaps_rollup
+pmap -x <pid> | sort -k3 -nr | head -25
+```
+
+Heap serialization itself can temporarily increase RSS and allocator high-water marks, so record `ps`/`smaps_rollup` both before and after capture. Large anonymous mappings with a comparatively small live heap require native-allocation or allocator investigation; they cannot be explained from JavaScript retainer paths alone.
+
+## CPU profiles
+
+The CLI installs a `SIGPROF` listener on non-Windows processes in `packages/cli/src/cpu-profile.ts`. One signal starts a ten-second CPU profile and stops it automatically; additional signals are ignored while a profile is active. There is no CPU profile CLI flag or environment variable.
+
+1. Get the PID from the health endpoint. For shared-service performance, target the server PID returned here rather than the short wrapper or TUI process:
+
+```bash
+opencode2 api get /api/health
+```
+
+Use `bun dev api get /api/health` instead when targeting the local/dev channel.
+
+2. Start the capture:
+
+```bash
+kill -PROF <server-pid>
+```
+
+3. Wait for `CPU profile written` in the channel's log before opening the file. Profiles are written to the same log directory as `cpu-<pid>-<timestamp>.cpuprofile`; the log's `path=` field is authoritative:
+
+```bash
+grep 'CPU profile' ~/.local/share/opencode/log/opencode.log | tail
+find ~/.local/share/opencode/log -maxdepth 1 -name 'cpu-<server-pid>-*.cpuprofile' -printf '%T@ %s %p\n' | sort -nr | head
+```
+
+Use `opencode-local.log` for a local/dev process. Load the completed `.cpuprofile` in Chrome DevTools or another V8 CPU profile viewer and inspect the hottest functions, call stacks, and self time during the controlled workload.
+
+## Debugger
+
+- To debug the V2 CLI or TUI with Bun's inspector, launch the CLI entrypoint through Terminal Control with an inspector URL, then attach a debugger to that URL:
+
+```bash
+termctrl start opencode-v2-debug --host opentui --cols 112 --rows 34 -- \
+  bun run --inspect=ws://localhost:6499/ src/index.ts
+```
+
+- Use `--inspect-wait` or `--inspect-brk` when execution must pause until a debugger attaches.
+- Use `termctrl logs opencode-v2-debug` for inspector output or startup failures emitted before the TUI renderer starts. Use `termctrl show` for the visible full-screen TUI.
+
+## Verification
+
+- Run `bun typecheck` from `packages/cli` after CLI adapter changes.
+- Run `bun typecheck` and `bun test` from `packages/tui` after shared TUI changes. Do not run tests from the repository root.
+- Treat automated checks and Terminal Control smoke tests as complementary. For user-visible changes, verify initial render, the changed interaction, Ctrl-C exit behavior, and save a screenshot of the corrected state.

.opencode/skills/opencode-drive/SKILL.md

@@ -1,254 +0,0 @@
----
-name: opencode-drive
-description: Use when an agent needs drive OpenCode via a script or interact with an isolated instance
----
-
-# OpenCode Drive
-
-Use `opencode-drive` to launch an isolated OpenCode instance and control it via commands or a script.
-
-There are two modes. Always default to using a script unless specifically directed to be interactive (connect
-to an existing running instance, or start a new one, and make a few changes to the UI and read it, and iterate
-on changes).
-
-Scripts allow you to run a full walkthrough in one run. When the script is done opencode-drive exits,
-stops all processes, and cleans up all artifacts.
-
-# Prepare The Environment
-
-Use `init` when files must be added to the isolated home or project before OpenCode starts. It prints the artifact directory without launching OpenCode. A later `start` with the same name reuses it.
-
-```bash
-artifacts=$(opencode-drive init --name demo)
-cp -R ./fixtures/home/. "$artifacts/"
-cp -R ./fixtures/project/. "$artifacts/files/"
-opencode-drive start --name demo --dev ~/projects/opencode
-```
-
-The simulated project is under `$artifacts/files`. Running `start` without a prior `init` initializes the artifacts automatically.
-
-# Scripted usage
-
-You can write scripts that walk through entire flows, and gives you full access to controlling
-the backend too. See examples of the script API at the bottom of this file.
-
-After creating or editing a script, always typecheck it before running. Never skip this step:
-
-```bash
-opencode-drive check ./reproduce-stale-exploring-empty.ts
-```
-
-Run it by passing `--script` to start:
-
-```bash
-opencode-drive start --name auto-stop-reproduction --script ./reproduce-stale-exploring-empty.ts
-```
-
-It will output information about the run, including paths to log files which you can read
-to inspect what happened. If you need to dig into failures that aren't clear, read those log
-files. If the script is unsuccessful, automatically fix the script and run it again.
-
-Scripts use one typed definition object. `setup` runs before OpenCode starts,
-and `fs.writeFile` always writes inside the simulated project.
-
-You can read the full typed API here: https://raw.githubusercontent.com/jlongster/opencode-drive/refs/heads/main/src/script/types.ts
-
-```ts
-import { defineScript } from "opencode-drive"
-
-export default defineScript({
-  async setup({ fs, config }) {
-    config.autoupdate = false
-    await fs.writeFile("src/example.ts", "export const value = 1\n")
-  },
-
-  async run({ ui, llm }) {
-    await ui.submit("Open src/example.ts")
-    await llm.send(llm.text("The file exports `value`."))
-    await ui.waitFor("The file exports `value`.")
-  },
-})
-```
-
-`setup` receives the current OpenCode config object, which starts from the
-default drive config unless the prepared instance already has one. When a script
-needs custom config, mutate this `config` parameter instead of generating and
-writing a new config object from scratch, so the script keeps the default
-provider/model settings unless it intentionally changes them.
-
-Note that the simulated model is a GPT model type, and opencode uses the `patch` tool for working with files Do not use a `edit` or `write` tool to edit files.
-
-Use `launch: "manual"` when the script needs to launch the server and every TUI
-itself (this is extremely rare, do not use this unless explicitly asked). In this
-mode `ui` is typed as `null`; call `server.launch()` exactly
-once before launching clients. Each `clients.launch(name)` result provides the
-same UI methods as the automatic client. You can see an example of this API
-here: https://raw.githubusercontent.com/jlongster/opencode-drive/refs/heads/main/examples/multiple-clients.ts
-
-Use the exported `wait(milliseconds)` utility for an unconditional delay.
-
-`await llm.send(...)` waits for the next request and resolves after OpenCode
-acknowledges its complete response. `llm.queue(...)` declares responses in
-advance. Chunks may be built with `text`, `reasoning`, `toolCall`, `raw`,
-`finish`, and `disconnect`. A normal response receives `finish("stop")`
-automatically unless it yields or queues an explicit terminal event.
-
-`llm.text(text, { delay, chunkSize })` defaults to a 2 ms delay and a
-15-character target varied by plus or minus 5 per chunk.
-
-`llm.reasoning` accepts the same options, and `llm.pause(milliseconds)` adds a
-delay between any two outputs.
-
-Use `llm.serve` for an ongoing typed response generator:
-
-```ts
-llm.serve(async function* (request, index) {
-  yield llm.reasoning(`Handling request ${index + 1}`)
-  yield llm.text(`Received ${request.id}`)
-  yield llm.finish("stop")
-})
-```
-
-The backend connection, response cleanup, cancellation, and recording
-completion are automatic.
-
-You can see some example scripts here:
-
-- https://raw.githubusercontent.com/jlongster/opencode-drive/refs/heads/main/examples/simple.ts
-- https://raw.githubusercontent.com/jlongster/opencode-drive/refs/heads/main/examples/serve.ts
-
-## Prune
-
-- `prune` removes artifact directories. These are always cleaned up after running a script
-  successfully, but leftover on failed runs. Always call this if a script fails.
-
-```bash
-opencode-drive prune --name demo
-
-// --force cleans up all artifcat directories
-opencode-dirve prune --force
-```
-
-# Live interaction usage
-
-- Always give headless instances a unique `--name`. Visible instances may omit it.
-- A normal headless `start` detaches automatically and returns after the instance is ready.
-- Do not add `&`; the long-running owner already runs in the background.
-- Configure simulated model responses after startup when needed.
-- Send ordered UI commands with `send`.
-- Always stop the instance when finished.
-
-```bash
-opencode-drive start --name demo
-
-opencode-drive send --name demo \
-  --command.ui.type '{"text":"Explain this project"}' \
-  --command.ui.enter
-
-opencode-drive stop --name demo
-```
-
-## Send UI Commands
-
-- Every `send` opens a connection to the named instance, runs its commands in order, and exits.
-- Combine typing and Enter in one command when submitting a prompt.
-- JSON-valued commands require one JSON argument.
-- Multiple command flags execute from left to right.
-
-Commands:
-
-- `--command.ui.type <json>` types into the focused editor. Arguments: `text` string.
-- `--command.ui.press <json>` presses a key. Arguments: `key` string; optional `modifiers` object with boolean `ctrl`, `shift`, `meta`, `super`, or `hyper`.
-- `--command.ui.enter` presses Enter. Arguments: none.
-- `--command.ui.arrow <json>` presses an arrow key. Arguments: `direction` is `up`, `down`, `left`, or `right`.
-- `--command.ui.focus <json>` focuses an element. Arguments: `target` is the numeric element `num` returned by `ui.state`.
-- `--command.ui.click <json>` clicks an element. Arguments: numeric `target`, `x`, and `y`; use the element `num` returned by `ui.state` as `target`.
-- `--command.ui.state` prints focus and interactive element metadata as JSON. Arguments: none.
-- `--command.ui.matches <json>` prints whether literal, case-sensitive text appears on screen. Arguments: `text` string.
-
-```bash
-opencode-drive send --name demo \
-  --command.ui.type '{"text":"Find the relevant code and explain it"}' \
-  --command.ui.enter
-
-opencode-drive send --name demo \
-  --command.ui.press '{"key":"p","modifiers":{"ctrl":true}}'
-
-opencode-drive send --name demo \
-  --command.ui.arrow '{"direction":"down"}'
-
-opencode-drive send --name demo \
-  --command.ui.focus '{"target":12}'
-
-opencode-drive send --name demo \
-  --command.ui.click '{"target":12,"x":4,"y":1}'
-
-opencode-drive send --name demo \
-  --command.ui.matches '{"text":"OpenCode"}'
-```
-
-To read the UI state and see information about interactable elements, use the `ui.state` command:
-
-```bash
-opencode-drive send --name demo --command.ui.state
-```
-
-## Configure LLM Responses
-
-- `responses` controls what the LLM responds with
-- Only use this if you are wanting to reproduce an exact type of response
-- Defaults are `text,reasoning,diff,tool` with `write,apply_patch`.
-- Supported types are `text`, `reasoning`, `diff`, and `tool`.
-- `--tools` limits generated tool calls to names offered by OpenCode.
-
-```bash
-opencode-drive responses --name demo \
-  --types text,reasoning,diff,tool \
-  --tools write,apply_patch
-
-opencode-drive responses --name demo \
-  --types tool \
-  --tools read,glob,grep
-```
-
-## Inspect The UI
-
-- `ui.state` prints focus and interactive element metadata as JSON.
-- `ui.matches` checks for literal, case-sensitive screen text.
-- `screenshot` prints the generated image path.
-
-```bash
-opencode-drive screenshot --name demo
-```
-
-## Lifecycle
-
-- `stop` waits for recording export and owner cleanup before returning.
-
-```bash
-opencode-drive stop --name demo
-```
-
-# Record The UI
-
-- Start with `--record` to capture a headless instance from its first rendered frame.
-- `stop` finishes the recording, exports an MP4, and prints its path.
-
-```bash
-opencode-drive start --name demo --record
-
-opencode-drive send --name demo \
-  --command.ui.type '{"text":"Show me the current architecture"}' \
-  --command.ui.enter
-
-opencode-drive stop --name demo
-```
-
-# Artifacts dir
-
-- `dir` prints the artifact directory for the instance.
-
-```bash
-opencode-drive dir --name demo
-```
-

.opencode/skills/rtl-aware-development/SKILL.md

@@ -0,0 +1,63 @@
+---
+name: rtl-aware-development
+description: OpenCode Desktop should be RTL-aware. Use when implementing or reviewing RTL/LTR behavior in the web app, desktop app, CSS, menus, scrolling, resizing, icons, mixed-direction text, or Electron title bars.
+---
+
+# RTL-Aware Development
+
+Treat direction as independent from language. Test English in both directions as well as real RTL and mixed-script content.
+
+## Guidelines
+
+- Set `lang` and `dir` on the document, and propagate direction through component providers used by portaled menus and popovers. Do not change the selected locale merely to force RTL.
+- Keep DOM and focus order semantic. Flexbox and Grid already follow `dir`; do not add `row-reverse`, CSS `order`, or reversed markup just to mirror a layout.
+- Prefer logical CSS for semantic layout. Reserve physical coordinates for pointer positions, canvas geometry, native window controls, and other genuinely physical placement.
+
+```css
+/* Avoid */
+padding-left: 12px;
+right: 0;
+border-right: 1px solid;
+text-align: left;
+
+/* Prefer */
+padding-inline-start: 12px;
+inset-inline-end: 0;
+border-inline-end: 1px solid;
+text-align: start;
+```
+
+- Isolate mixed-direction text. Use `dir="auto"` or `<bdi>` for unknown text; keep code, URLs, IDs, and filesystem paths LTR without forcing the surrounding component LTR.
+
+```html
+<span class="file-row"><bdi dir="auto">README.md</bdi></span> <bdi dir="ltr"><code>C:\src\app.ts</code></bdi>
+```
+
+- Mirror directional meaning, not every image. Back/forward, previous/next, disclosure, indentation, and directional progress may need mirroring. Do not mirror brands, clocks, media controls, charts, or text. Reverse physical gradients, `translateX`, SVG transforms, and animation deltas explicitly.
+- Map interactions through direction. `clientX` remains physical; resizing a logical edge needs an RTL-aware delta. Logical previous/next keyboard controls may swap ArrowLeft/ArrowRight. Follow the relevant WAI-ARIA widget pattern.
+- Do not assume LTR scrolling. RTL `scrollLeft` can start at `0` and become negative. Prefer `scrollIntoView({ inline: "nearest" })` or a tested direction-normalizing helper.
+- For Electron title bars, prefer native caption controls and use `titleBarOverlay` plus `env(titlebar-area-*)` for the safe content rectangle. Keep Windows/macOS native-control avoidance and `trafficLightPosition` physical; keep app navigation inside that rectangle logical. Mark interactive titlebar children `app-region: no-drag`.
+- Verify behavior, not screenshots alone. Check computed styles, pseudo-element geometry, hit zones, focus order, keyboard behavior, submenu direction, zoom/scaling, and both LTR and RTL scroll endpoints.
+
+## Test Matrix
+
+- English + LTR
+- English + forced RTL
+- A real RTL locale + RTL
+- Mixed RTL/LTR content, long labels, numbers, code, and paths
+- Keyboard, pointer resize, scrolling, menus/submenus, and Electron titlebar controls in both directions
+
+## References
+
+- [RTL Styling 101, Ahmad Shadeed](https://rtlstyling.com/posts/rtl-styling/)
+- [CSS-Tricks: RTL Styling 101](https://css-tricks.com/rtl-styling-101/)
+- [CSS-Tricks: CSS Logical Properties and Values](https://css-tricks.com/css-logical-properties-and-values/)
+- [W3C: Structural markup and right-to-left text](https://www.w3.org/International/questions/qa-html-dir)
+- [W3C: Inline bidirectional markup](https://www.w3.org/International/articles/inline-bidi-markup/)
+- [MDN: CSS logical properties and values](https://developer.mozilla.org/en-US/docs/Web/CSS/Guides/Logical_properties_and_values)
+- [MDN: `dir`](https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Global_attributes/dir)
+- [MDN: `scrollLeft`](https://developer.mozilla.org/en-US/docs/Web/API/Element/scrollLeft)
+- [web.dev: Logical properties](https://web.dev/learn/css/logical-properties/)
+- [Electron: Custom title bar](https://www.electronjs.org/docs/latest/tutorial/custom-title-bar)
+- [WAI-ARIA: Window splitter pattern](https://www.w3.org/WAI/ARIA/apg/patterns/windowsplitter/)
+- [Kobalte: I18n Provider](https://kobalte.dev/docs/core/components/i18n-provider/)

.opencode/tool/github-pr-search.ts

@@ -1,5 +1,5 @@
 /// <reference path="../env.d.ts" />
-import { tool } from "@opencode-ai/plugin"
+import { tool } from "@opencode/plugin"
 async function githubFetch(endpoint: string, options: RequestInit = {}) {
   const response = await fetch(`https://api.github.com${endpoint}`, {
     ...options,

.opencode/tool/github-triage.ts

@@ -1,10 +1,10 @@
 /// <reference path="../env.d.ts" />
-import { tool } from "@opencode-ai/plugin"
+import { tool } from "@opencode/plugin"
 
 const TEAM = {
   tui: ["kommander", "simonklee"],
   desktop_web: ["Hona", "Brendonovich"],
-  core: ["jlongster", "rekram1-node", "nexxeln", "kitlangton", "starptech"],
+  core: ["jlongster", "rekram1-node", "neriousy", "nexxeln", "kitlangton"],
   inference: ["fwang", "MrMushrooooom", "starptech"],
   windows: ["Hona"],
 } as const

.opencode/tui.json

@@ -1,19 +1,3 @@
 {
-  "$schema": "https://opencode.ai/tui.json",
-  "plugin": [
-    [
-      "./plugins/tui-smoke.tsx",
-      {
-        "enabled": false,
-        "label": "workspace",
-        "keybinds": {
-          "smoke_modal": "ctrl+alt+m",
-          "smoke_screen": "ctrl+alt+o",
-          "smoke_screen_home": "escape,ctrl+shift+h",
-          "smoke_screen_modal": "ctrl+alt+m",
-          "smoke_dialog_close": "escape,q"
-        }
-      }
-    ]
-  ]
+  "$schema": "https://opencode.ai/tui.json"
 }

.oxlintrc.json

@@ -1,51 +1,22 @@
 {
   "$schema": "https://raw.githubusercontent.com/nicolo-ribaudo/oxc-project.github.io/refs/heads/json-schema/src/public/.oxlintrc.schema.json",
-  "options": {
-    "typeAware": true
-  },
   "categories": {
-    "suspicious": "warn"
+    "correctness": "off",
+    "suspicious": "off",
+    "pedantic": "off",
+    "perf": "off",
+    "style": "off",
+    "restriction": "off",
+    "nursery": "off"
   },
   "rules": {
-    "typescript/no-base-to-string": "warn",
-    // Effect uses `function*` with Effect.gen/Effect.fnUntraced that don't always yield
-    "require-yield": "off",
-    // SolidJS uses `let ref: T | undefined` for JSX ref bindings assigned at runtime
-    "no-unassigned-vars": "off",
-    // SolidJS tracks reactive deps by reading properties inside createEffect
-    "no-unused-expressions": "off",
-    // Intentional control char matching (ANSI escapes, null byte sanitization)
-    "no-control-regex": "off",
-    // SST and plugin tools require triple-slash references
-    "triple-slash-reference": "off",
-
-    // Suspicious category: suppress noisy rules
-    // Effect's nested function* closures inherently shadow outer scope
-    "no-shadow": "off",
-    // Namespace-heavy codebase makes this too noisy
-    "unicorn/consistent-function-scoping": "off",
-    // Opinionated — .sort()/.reverse() mutation is fine in this codebase
-    "unicorn/no-array-sort": "off",
-    "unicorn/no-array-reverse": "off",
-    // Not relevant — this isn't a DOM event handler codebase
-    "unicorn/prefer-add-event-listener": "off",
-    // Bundler handles module resolution
-    "unicorn/require-module-specifiers": "off",
-    // postMessage target origin not relevant for this codebase
-    "unicorn/require-post-message-target-origin": "off",
-    // Side-effectful constructors are intentional in some places
-    "no-new": "off",
-
-    // Type-aware: catch unhandled promises
-    "typescript/no-floating-promises": "warn",
-    // Warn when spreading non-plain objects (Headers, class instances, etc.)
-    "typescript/no-misused-spread": "warn"
-  },
-  "options": {
-    "typeAware": true
-  },
-  "options": {
-    "typeAware": true
+    "no-restricted-globals": [
+      "error",
+      {
+        "name": "Reflect",
+        "message": "Use typed property access or direct invocation. Suppress this rule only for genuine reflection."
+      }
+    ]
   },
   "ignorePatterns": ["**/node_modules", "**/dist", "**/.build", "**/.sst", "**/*.d.ts", "**/sdk.gen.ts"]
 }

AGENTS.md

@@ -1,9 +1,31 @@
-- To regenerate the legacy JavaScript SDK, run `./packages/sdk/js/script/build.ts`.
-- After changing the public Protocol or Server `HttpApi`, run `bun run generate` from `packages/client`. Do not edit `src/generated` or `src/generated-effect` directly.
-- Keep runtime dependencies directed from Schema to Core and Protocol, then from Core and Protocol to Server. Client runtime code may depend on Schema and Protocol but never Core or Server; `sdk-next` composes Client, Core, and Server.
-- Do not modify `packages/opencode` unless the user explicitly asks for V1 work. `packages/opencode` is the V1 implementation and is present for reference only. New implementation changes should land in the V2 package set: `packages/core`, `packages/cli`, `packages/server`, `packages/protocol`, `packages/schema`, and related generated client surfaces when required.
-- The default branch in this repo is `dev`.
-- Local `main` ref may not exist; use `dev` or `origin/dev` for diffs.
+- After changing the public Protocol or Server `HttpApi`, run `bun run generate` from `packages/client`. Do not edit generated client files directly.
+- Keep runtime dependencies directed from Schema to Core and Protocol, then from Core and Protocol to Server. Client runtime code may depend on Schema and Protocol but never Core or Server; `sdk` composes Client, Core, and Server.
+- Current implementation changes belong in `packages/core`, `packages/cli`, `packages/server`, `packages/protocol`, `packages/schema`, and related generated client surfaces when required.
+- This repository does not use Changesets. Do not add `.changeset` files; follow the existing release workflow instead.
+- The default branch in this repo is `v2`.
+- Default new branches and worktrees to `v2`, or `origin/v2` when the local `v2` ref is unavailable, and default pull requests to target `v2`. Use another base or target branch when the requester explicitly instructs it.
+- Local `main` ref may not exist; use `v2` or `origin/v2` for diffs.
+
+## Live V2 TUI Testing
+
+- Run `bun run dev:live` from a development worktree to test its TUI against the currently elected `opencode` background server and live sessions.
+- Pass a directory after the script when needed, for example `bun run dev:live /path/to/project`.
+- The script discovers the server with `opencode service status`, injects its private local credential from `opencode service get password`, and uses the `dev` TUI storage channel so tabs and other client-local state match the installed client.
+- Prefer `dev:live` over plain `bun run dev` for this workflow. An implicit managed-service connection may replace the live server when the worktree client version differs; explicit `--server` warns and continues without replacing it.
+
+## V2 TUI Stories
+
+- When a user asks for a TUI story, add a fixture-driven story under `packages/tui/src/feature-plugins/system/storybook` and register it in `index.tsx`.
+- Render the real production component rather than a visual copy. Keep submissions and other side effects local to the story so it is safe to explore repeatedly.
+- Expose the meaningful state dimensions through story keybindings and list them in `StoryFooter`; include a reset command when combinations can leave the fixture in a confusing state.
+- Run a specific story with `OPENCODE_STORY=<story-id> bun run dev:live` from the development worktree, and exercise narrow and wide terminal sizes when layout is relevant.
+
+## TUI Theme Tokens
+
+- Choose theme tokens by semantic role, not by their current color. Do not use raw `theme.hue` values or borrow an unrelated semantic token to achieve a preferred appearance.
+- Use `text.feedback` and `background.feedback` only for outcome or status feedback such as errors, warnings, success messages, and informational messages. Use `formfield` states for form-control text, ordinals, and selection markers, and `action` states for actions.
+- If the theme does not expose a token for the required semantic role, extend the theme schema, defaults, resolution, and types with that role before using it in a component. Do not repurpose the nearest-looking existing token.
+- When changing the public theme token surface, verify the built-in light and dark defaults and the custom-theme fallback path in addition to the affected TUI component.
 
 ## Branch Names
 
@@ -19,13 +41,12 @@ Valid types are `feat`, `fix`, `docs`, `chore`, `refactor`, and `test`. Scopes a
 
 Examples: `fix(tui): simplify thinking toggle styling`, `docs: update contributing guide`, `chore(sdk): regenerate types`.
 
-Never bypass Git hooks. Do not use `--no-verify` or otherwise disable, skip, or circumvent commit or push hooks. If a hook fails, fix the failure or stop and report it to the user.
-
 ## Style Guide
 
 ### General Principles
 
 - Keep things in one function unless composable or reusable
+- Validate unknown values once at the boundary that owns them. Pass typed values inward instead of repeating `typeof value === "object"` and property-existence checks. Do not defensively revalidate values already guaranteed by a schema, constructor, or internal type.
 - Do not extract single-use helpers preemptively. Inline the logic at the call site unless the helper is reused, hides a genuinely complex boundary, or has a clear independent name that improves the caller.
 - Before adding complexity for a speculative or vanishingly unlikely race or security edge case, explain the concrete failure mode, likelihood, and complexity cost to the user and get their buy-in. Do not silently expand scope for theoretical robustness.
 - Avoid `try`/`catch` where possible
@@ -63,8 +84,9 @@ const { a, b } = obj
 ### Imports
 
 - Never alias imports. Do not use `import { foo as bar } from "..."` or renamed imports like `resolve as pathResolve`.
+- Never use type-position `import("...")` references such as `Schema.declare<import("@opencode/plugin/effect/plugin").Plugin["effect"]>`. Only when two imports genuinely collide on a name and no other option exists, an aliased type import (`import type { Plugin as PluginDefinition } from "..."`) is permitted as a last resort — still strongly preferred not to.
 - Never use star imports. Do not use `import * as Foo from "..."` or `import type * as Foo from "..."`.
-- If a namespace-style value is needed, import the module's own exported namespace by name, for example `import { Project } from "@opencode-ai/core/project"`, then reference `Project.ID`.
+- If a namespace-style value is needed, import the module's own exported namespace by name, for example `import { Project } from "@opencode/core/project"`, then reference `Project.ID`.
 - Prefer dynamic imports for heavy modules that are only needed in selected code paths, especially in startup-sensitive entrypoints. Destructure dynamic import bindings near the top of the narrowest scope that needs them so they read like normal imports. Avoid inline chains such as `await import("./module").then((mod) => mod.value())` or `(await import("./module")).value()`. Keep branch-specific imports inside the branch that needs them to preserve lazy loading.
 
 ### Variables
@@ -146,23 +168,24 @@ const table = sqliteTable("session", {
 
 - Avoid mocks as much as possible, you shouldn't be using globalThis.\* at all unless it's the only option.
 - Test actual implementation, do not duplicate logic into tests
-- Tests cannot run from repo root (guard: `do-not-run-tests-from-root`); run from package dirs like `packages/opencode`.
+- Tests cannot run from repo root (guard: `do-not-run-tests-from-root`); run from package directories such as `packages/core`.
 
-## Type Checking
+## Checks
 
-- Always run `bun typecheck` from package directories (e.g., `packages/opencode`), never `tsc` directly.
+- Run `bun run check` from the repository root as the canonical full lint and type-check verification.
+- During focused iteration, run `bun typecheck` from the affected package directory (for example, `packages/core`). Never run `tsc` directly.
 
 ## V2 Session Core
 
 - Keep durable events minimal: record irreducible new facts and do not repeat state derivable by folding the ordered aggregate history. Enrich projections and read models with previous or derived state when consumers need self-contained views.
-- Keep durable prompt admission separate from model execution. `SessionV2.prompt(...)` admits one durable `session_pending` row before scheduling advisory `SessionExecution.wake(sessionID)` unless `resume: false` requests admit-only behavior. The serialized runner promotes admitted inputs into visible user messages at safe boundaries, consuming the pending row in the same event transaction; `session_pending` stores only unconsumed work.
-- Reusing a Session ID adopts the existing Session. Reusing a prompt message ID reconciles an exact retry only when Session, prompt, and delivery mode match; conflicting reuse fails. Retry of an already-promoted input reconciles against the projected message and the durable admitted event rather than a retained row.
+- Keep durable prompt admission separate from model execution. `Session.prompt(...)` publishes `session.inbox.enqueued`, whose projection inserts one durable `session_inbox` row, before scheduling advisory `SessionExecution.wake(sessionID)` unless `resume: false` requests admit-only behavior. Delivery publishes `session.inbox.delivered`; its projection consumes the inbox row and inserts the visible message in the same transaction. `session_inbox` stores only unconsumed work.
+- Reusing a Session ID adopts the existing Session. Reusing a user or synthetic inbox item ID is idempotent when Session and type match: the first admission wins and the retried payload, metadata, and delivery mode are ignored, whether the item is still pending or already delivered (reconciled from the projected message without retained enqueue history). Cross-Session or cross-type reuse fails. Control items keep their operation-specific conflict behavior.
 - Keep `SessionExecution` process-global and Session-ID based. Its local implementation owns the process-local Session coordinator and discovers placement through `SessionStore` plus `LocationServiceMap.get(session.location)` only when a drain starts; no layer should take a Session ID. V2 interruption targets the active process-local ownership chain for that Session; interruption of a known but idle or locally unowned Session is a no-op, while the public API rejects an unknown Session.
 - Keep `SessionRunner`, model resolution, tool registry, permissions, and filesystem Location-scoped. Omitted `Location.workspaceID` means implicit-local placement; explicit workspace identity remains reserved for future placement semantics.
-- Preserve one explicit `llm.stream(request)` call per Physical Attempt and reload projected history before durable continuation. Most Steps have one Physical Attempt; overflow-triggered compaction recovery may rebuild one Step for a second attempt. Do not bridge through legacy `SessionPrompt.loop(...)` or delegate orchestration to an in-memory tool loop.
-- Keep local Session drains process-local until clustering is implemented. `SessionRunCoordinator` joins explicit same-Session resumes, coalesces prompt wakeups, and allows different Sessions to run concurrently. Advisory wakes drain eligible durable inbox rows only; post-crash continuation recovery requires a separate explicit design before it may retry provider work. A drain has no durable identity or transcript boundary.
-- Keep delivery vocabulary explicit. Prompts steer by default and promote at the next safe step boundary while the current drain requires continuation. An explicit `queue` input remains pending until the Session would otherwise become idle; promote one queued input at that boundary, then reevaluate continuation before promoting another. Promoting any new user input resets the selected agent's step allowance; a batch of steers resets it once.
+- Preserve one explicit `llm.stream(request)` call per Physical Attempt and reload projected history before durable continuation. A logical Step may use generic pre-output retries, one full-context retry after continuation rejection, incomplete-stream continuation, or one overflow-compaction rebuild. Generic retries retain the logical step number and do not consume another agent-step allowance. Do not delegate orchestration to an in-memory tool loop.
+- Keep local Session drains process-local until clustering is implemented. `SessionRunCoordinator` joins explicit same-Session resumes, coalesces prompt wakeups, and allows different Sessions to run concurrently. A write-ahead execution claim marks a process-local busy period for restart recovery: terminal completion, failure, or user interruption releases it, while shutdown interruption and process death preserve it. Startup recovery resumes claimed top-level Sessions with durable per-execution attempt accounting. The claim is a recovery marker, not clustered ownership, fencing, or an exactly-once guarantee.
+- Keep delivery vocabulary explicit. Prompts steer by default. At safe step boundaries, steered compaction takes priority up to the first steered move control; other steers retain enqueue order. At an idle boundary, steers take priority; otherwise exactly one queued item delivers before the runner reevaluates continuation. Inbox items may be cancelled or changed between queue and steer before delivery. Promoting new user input resets the selected agent's step allowance; a batch of steers resets it once.
 - One step is one logical LLM call; its durable record covers only the model-visible span. Do not write "provider turn", and do not use bare "turn" for a single call: "turn" is reserved for the future assistant-turn unit containing all steps from prompt promotion until the session would go idle.
-- Keep EventV2 replay owner claims separate from clustered Session execution ownership.
+- Keep event replay ownership separate from clustered Session execution ownership.
 - Keep the Instructions algebra and built-ins in `src/instructions`; keep instruction producers with their observed domains, and keep Session History selection plus `InstructionState` and `InstructionEntry` persistence Session-owned. `InstructionDiscovery` observes ambient global and upward-project instructions. The runner composes built-ins, discovery, guidance, and entries explicitly in `loadInstructions`; there is no instruction registry.
-- `session.instructions.updated` stores only changed source keys and content hashes. Blob values live once in `instruction_blob`; `instruction_state` is a rebuildable fold cache, never primary state. Render initial instructions and chronological updates from values during request assembly. Completed compaction moves the instruction epoch; Session movement and committed revert clear it. Unavailable sources retain the last value and block only the initial complete delta.
+- `session.instructions.updated` stores changed source keys and content hashes and may freeze rendered chronological update text. Blob values live once in `instruction_blob`; the projected `instruction_state` row is the normal boundary-processing source of current and initial values. Request assembly renders the epoch baseline from stored values, while later frozen updates enter history as durable System messages. Completed compaction moves the instruction epoch; Session movement retains it so destination instruction changes are chronological, while committed revert clears it. Forks adopt the parent's newest instruction values even when copied message history ends at an earlier boundary. Unavailable sources retain the last value and block only the initial complete delta.

CONTEXT.md

@@ -1,245 +0,0 @@
-# OpenCode Session Runtime
-
-OpenCode sessions preserve durable conversational history while assembling the runtime context an agent needs to act correctly in its current environment.
-
-## Language
-
-**Model Context**:
-The complete model-visible input assembled for one **Step**, including system instructions, **Session History**, tool definitions, and step-local additions. **Instructions** are one component of Model Context, not a synonym for it.
-_Avoid_: System Context
-
-**Instructions**:
-The opaque algebra of independently refreshable typed instruction sources that render the durable instruction baseline and chronological updates shown to the model.
-_Avoid_: Model Context, System Context
-
-**Session History**:
-The projected chronological conversation selected for a **Step** after applying the active compaction boundary and interleaving derived **Instruction Updates** from the current **Instruction Epoch**.
-_Avoid_: Session Context
-
-**Instruction Source**:
-One independently read typed value within **Instructions**, represented by a stable namespaced key, canonical JSON codec, pure first/changed renderers, and an optional removal renderer.
-_Avoid_: Prompt fragment
-
-**InstructionEntry**:
-One API-managed, durable, per-Session instruction value. Its slash-free client key maps to the `api/<key>` **Instruction Source** key. Entries deliberately render to the model as mechanism-neutral `<context>` blocks: the model sees session context, not how it was attached.
-
-**InstructionDiscovery**:
-The Location-scoped service that observes ambient global and upward-project `AGENTS.md` files as one ordered aggregate **Instruction Source**.
-
-**Instruction State**:
-The Session-owned projection cache of one instruction log fold: epoch start, values at that start, current values, and the last folded sequence. It is rebuilt from durable events and never authors model-visible facts.
-
-**Instruction Update**:
-A durable `session.instructions.updated` value delta admitted at a **Safe Step Boundary**. Its model-visible System text is rendered from stored values at request assembly and is never persisted verbatim.
-_Avoid_: Correction, stored prose, raw text diff
-
-**Initial Instructions**:
-The deterministic instruction text rendered from values at the current **Instruction Epoch** start and sent as provider-cache prefix state until completed compaction moves the epoch or Session movement or committed revert resets it.
-_Avoid_: Live system prompt
-
-**Instruction Epoch**:
-The span between completed compactions. Its start is the last `session.compaction.ended` sequence, or the initial complete instruction delta when no prior epoch exists.
-
-**Instruction Values**:
-The key-to-hash map produced by folding instruction deltas in durable sequence order. Hash bodies live once in the content-addressed instruction blob store.
-
-**Unavailable Instruction Source**:
-An expected temporary inability to read an **Instruction Source** value; the runtime retains its prior effective value and emits no update, while an unavailable source blocks the initial complete delta.
-
-**Safe Step Boundary**:
-The point during Step preparation, after prior tool settlement and before durable input promotion, where instruction changes may be admitted chronologically.
-
-**Admitted Prompt**:
-A durable user input accepted into the Session inbox but not yet included in **Session History**.
-
-**Prompt Promotion**:
-The durable transition that removes an **Admitted Prompt** from pending input and appends its user message to **Session History**.
-
-**Step**:
-One logical LLM call spanning pre-flight instruction synchronization, input promotion, request build, and compaction check; the provider stream; and tool settlement.
-_Avoid_: provider turn, turn (unqualified)
-
-**Physical Attempt**:
-One actual provider request on the wire in service of a **Step**; most Steps have one Physical Attempt, while overflow-triggered compaction recovery may give one Step two.
-
-**Assistant Turn**:
-A reserved name for the not-yet-modeled unit containing all **Steps** from prompt promotion until the assistant yields the floor; do not reify it until something durable needs it.
-
-**Settlement**:
-The terminal transition for a unit of work: Step and tool settlement are durable, while drain and execution settlement are coordinator-observed.
-
-**Execution**:
-One session-scoped coordinator busy period from first wake until idle. An Execution is process-local coordination rather than a durable domain entity.
-
-**Session Drain**:
-One process-local execution span that promotes eligible input and runs required **Steps** until no immediate continuation remains. A Session Drain has no durable identity or transcript boundary.
-
-**Model Tool Output**:
-The bounded projection of a Core-executed tool result persisted in Session history and replayed to the model. A tool may shape this projection semantically, but the Tool Registry enforces the final size limit.
-
-**Managed Tool Output File**:
-A temporary file created under OpenCode's shared tool-output directory to retain complete output that was too large for Session history.
-
-**Model Request Options**:
-Provider-semantic model settings selected from the Catalog and active Session variant before the LLM protocol adapter encodes them for a provider request.
-_Avoid_: Request body, wire options
-
-**Generation Controls**:
-Provider-neutral sampling and output controls, partitioned from provider semantics and compatibility wire fields when model metadata enters the Catalog.
-
-**Native Continuation Metadata**:
-Opaque protocol-shaped data attached to assistant content and required to continue that content natively with a compatible model, such as a reasoning signature or provider-hosted item identifier.
-
-**PTY Environment**:
-The host-supplied environment overlay applied by the server when creating a PTY, observed for the request Location and resolved PTY working directory.
-
-**OpenCode Client**:
-The generated Promise and Effect APIs derived from the public `HttpApi`; **Embedded OpenCode** shares the Effect API through an in-memory `HttpClient` against the same router and handlers.
-_Avoid_: Remote client
-
-**SDK Contract IR**:
-The runtime-neutral compiled representation of the authoritative `HttpApi`, preserving encoded and decoded type projections plus transport metadata so independent SDK emitters can choose their public value model and runtime interpreter.
-
-**Embedded OpenCode**:
-A scoped in-process host that structurally extends the **OpenCode Client**, supplies an in-memory HTTP transport, and exposes additional same-process capabilities directly.
-_Avoid_: Local implementation
-
-**Page**:
-A bounded ordered result containing `items` and opaque `previous` and `next` cursor links for navigating the same query in either direction.
-_Avoid_: Response envelope
-
-## Relationships
-
-- **Instructions** is an opaque carrier composed from zero or more **Instruction Sources**.
-- **Model Context** is broader than **Instructions**. For each **Step**, the runner assembles the selected agent or provider system text, **Initial Instructions**, **Session History**, available tools, and step-local additions into one model request.
-- **Session History** persists conversational messages. The runner derives model-facing **Instruction Update** messages from value deltas and interleaves them by durable sequence; **Initial Instructions** remain separate provider-request state.
-- The runner explicitly loads and combines instruction built-ins, **InstructionDiscovery**, selected-agent skill guidance, reference guidance, MCP guidance, and **InstructionEntry** values. There is no instruction registry.
-- `Instructions.combine(...)` preserves caller order and rejects duplicate stable namespaced source keys. The runner loads its producers concurrently, then combines them in its fixed declared order.
-- Each **Instruction Source** read returns one coherent typed value, explicit removal, or temporary unavailability. `Instructions.make(...)` hides the value type so differently typed sources compose uniformly; its canonical codec defines storage and hash equivalence, while pure renderers produce first, changed, and optional removal text.
-- `Instructions.read(...)` reads every composed source concurrently and exactly once at the boundary. `Instructions.diff(...)` compares encoded-value hashes with current **Instruction Values** and returns one delta plus new blob bodies.
-- `Instructions.renderInitial(...)` renders values at the **Instruction Epoch** start. `Instructions.renderUpdate(...)` renders one hydrated delta against the values immediately before it.
-- A changed **Instruction Source** contributes its hash to one **Instruction Update**; explicit removal contributes the `"removed"` sentinel.
-- An **Instruction Update** persists only its value delta. Rendered text is derived during request assembly and excluded from compaction summaries.
-- The instruction blob insert, durable delta, and **Instruction State** advance commit atomically.
-- Changes from multiple **Instruction Sources** admitted at one safe boundary combine into one **Instruction Update**.
-- Instruction changes are sampled and admitted lazily at a **Safe Step Boundary**, never pushed asynchronously when their source changes.
-- At a **Safe Step Boundary**, prior tool results are already settled; instruction preparation completes before newly admitted user input promotes.
-- An **Admitted Prompt** is replayable pending input, not yet model-visible **Session History**.
-- **Prompt Promotion** atomically consumes the pending inbox entry and appends its model-visible user message.
-- Steering prompts promote at the next **Safe Step Boundary** while the current **Session Drain** still requires continuation. Promoting any newly admitted user input resets the selected agent's step allowance; multiple prompts promoted at one boundary reset it once.
-- A queued prompt does not promote while the current **Session Drain** requires continuation. The runner promotes one queued prompt when the Session would otherwise become idle, then reevaluates continuation before promoting another.
-- A **Session Drain** is process-local coordination rather than a durable domain entity. Durable recovery must reason from prompts, projected history, physical attempts, and tool state rather than inventing an enclosing execution identity.
-- An **Execution** contains one or more **Session Drains**; a **Session Drain** contains one reserved assistant-turn span at a time; that span contains **Steps**; and each **Step** contains one or more **Physical Attempts** plus any tool calls it requires.
-- A **Step** record covers only the model-visible span from first assistant output through tool settlement; pre-flight leaves no record, and one Step settles at most one record.
-- The first **Step** admits one complete delta and renders **Initial Instructions** without narrating that delta in history; an unavailable initial source blocks the Step instead of persisting incomplete values.
-- Instruction preparation precedes durable input promotion on every Step so an unavailable first baseline leaves pending input untouched and later updates enter history before newly promoted input.
-- Completed compaction moves the **Instruction Epoch** to the exact `session.compaction.ended` sequence and copies current hashes to the epoch's initial values. Earlier updates leave active model history while durable deltas remain.
-- A newly composed **Instruction Source** absent from current **Instruction Values** emits its first rendering once at the next **Safe Step Boundary**.
-- **Unavailable Instruction Source** uses stale-while-revalidate semantics and is distinct from a successfully loaded absence, which may emit removal text.
-- **InstructionDiscovery** observes ambient instructions as one ordered aggregate **Instruction Source**.
-- Ambient discovery reads global and upward-project `AGENTS.md` files and honors `OPENCODE_DISABLE_PROJECT_CONFIG` for project files.
-- After a successful internal file or directory read, nearby `AGENTS.md` files toward the Location root are injected once per Session as durable synthetic instruction messages.
-- **InstructionEntry** stores API-managed per-Session JSON values. Each entry contributes one `api/<key>` **Instruction Source**, so adding, replacing, or removing an entry is reconciled at the next **Safe Step Boundary**.
-- Location-scoped instruction producers naturally re-resolve when a moved Session next runs in its destination Location.
-- Moving a Session clears its **Instruction State**, so the destination must admit a complete delta before another prompt can promote. Committed revert does the same; replay derives both resets from their durable events.
-- Selected-agent available-skill guidance is an **Instruction Source** composed explicitly by the runner. It lists only names and descriptions permitted for that agent; skill bodies and locations are exposed only through the permission-checked `skill` tool.
-- The selected agent and model are sampled when a **Step** starts. Changes admitted after that boundary apply to the next Step and do not restart the current Step.
-- An agent switch that changes selected-agent guidance produces an **Instruction Update** while preserving the current baseline.
-- Local tool authorization and pending permission requests retain the effective agent of the **Step** that issued the call; a later agent switch cannot change that call's policy.
-- Instruction source changes never wake idle Sessions; the next naturally scheduled **Safe Step Boundary** loads and compares current values lazily.
-- Once admitted, an **Instruction Update** remains durable even if the following **Physical Attempt** fails and is replayed unchanged on retry.
-- **Instruction Updates** remain durable value history but are not `session_message` rows. Clients display changed keys rather than model-facing prose.
-- The date **Instruction Source** initially preserves host-local calendar-date behavior; a configured user timezone may replace that default later.
-- **Initial Instructions** are recomputed deterministically from durable values for every request; rendered bytes are not stored.
-- A model/provider switch preserves current **Instruction Values**, the **Instruction Epoch**, and chronological conversation history; the new selection applies to the next **Step**.
-- **Native Continuation Metadata** remains in durable history. Step projection includes it only for a successful exact originating provider/model match; failed Steps and incompatible models omit opaque metadata, while non-empty visible reasoning lowers to ordinary assistant text after a model switch. This conservative relation may widen only when recorded provider tests establish compatibility.
-- **Model Request Options** remain provider-semantic through Catalog resolution. The Session runner maps them into the LLM package's provider-option namespace; the selected protocol adapter alone owns provider wire encoding.
-- **Generation Controls**, protocol-semantic **Model Request Options**, and compatibility request body fields are separate Catalog domains. A shared ingestion adapter partitions legacy and models.dev AI-SDK-shaped options before routing.
-- The **PTY Environment** is a server concern rather than a Core PTY concern. PTY creation merges caller values, then the host overlay, then Core-forced terminal invariants such as `TERM` and `OPENCODE_TERMINAL`.
-- Networked and **Embedded OpenCode** use the same **OpenCode Client** and preserve the full HTTP encoding, routing, middleware, and decoding boundary; only the `HttpClient` transport differs.
-- The Effect-native network constructor obtains `HttpClient.HttpClient` from its environment so callers own transport selection, recording, tracing, retries, and tests. Convenience runtimes may provide a fetch transport separately.
-- Creating **Embedded OpenCode** is scoped. Closing its owning Scope releases the in-process server resources, database resources, registrations, and fibers.
-- **Embedded OpenCode** exposes shared client capabilities and embedded-only capabilities on one object; consumers do not navigate through a nested `.client` property.
-- The beta **OpenCode Client** currently uses plural consumer-facing capability groups such as `sessions`; whether the stable Session namespace should instead be singular `session` must be settled before stabilization. Internal server identifiers do not implicitly define public client names.
-- Server's concrete `HttpApi` is authoritative for shared **OpenCode Client** capabilities. Codegen compiles its Session group directly; the Effect runtime uses an equivalent Protocol-only projection so generated artifacts remain independent of Core and Server.
-- SDK generation reflects the public `HttpApi` once into an **SDK Contract IR**. Promise and Effect emitters share endpoint structure and transport metadata without being required to expose identical public values: an emitter may select encoded wire types, decoded domain types, compile-time brands, runtime validation, and its own execution abstraction independently.
-- The first Effect emitter is the rich projection: it exposes decoded Effect-native values, preserves brands and schema transformations, performs runtime schema decoding, and delegates transport interpretation to `HttpApiClient`. Lighter wire-shaped Effect output remains possible through another emitter policy rather than constraining the shared IR.
-- The rich Effect emitter regenerates private executable schemas when the **SDK Contract IR** proves that their transport semantics can be reproduced exactly. Contracts with authoritative custom transformations use the import-based Effect emitter against a Protocol-only client projection whose generated transport output is tested against Server's concrete API; the Promise emitter still derives zero-Effect structural wire types from the same IR.
-- `@opencode-ai/protocol` owns Session endpoint construction and middleware placement. Server supplies concrete middleware keys to produce the authoritative build-time API; the client projection supplies transport-only keys without importing Core or Server at runtime.
-- The first Promise emitter targets the same clean domain-oriented method organization rather than Hey API source compatibility. It returns unwrapped values directly, rejects declared and infrastructure failures, and begins with minimal client-level transport configuration; result wrappers, interceptors, and legacy generated signatures are outside the initial surface.
-- The first Promise emitter parses response syntax and trusts its generated structural types; it does not perform runtime structural validation. Malformed payload syntax fails, while a syntactically valid shape mismatch is not detected at the SDK boundary. Standalone validator generation remains an optional future emitter policy.
-- Declared Promise-client failures retain their tagged structural wire values and have generated type guards. Consumers do not depend on generated `Error` subclass identity, preserving discrimination across package copies and realms while remaining structurally aligned with Effect domain errors.
-- Promise-client infrastructure failures use one generated `ClientError` class with a structured reason such as transport failure, unexpected status, unsupported content type, or malformed response. Promise methods reject with either a tagged declared domain failure or `ClientError`, matching the Effect client's conceptual domain/infrastructure error division.
-- Promise methods accept a separate optional per-call transport-options argument containing `AbortSignal` and header overrides. Cancellation and transport metadata do not enter the domain input object; broader interceptor and response-mode APIs remain deferred.
-- Promise streaming methods return a lazy `AsyncIterable` directly rather than a Promise-wrapped stream object. Iteration opens the connection, `AbortSignal` cancels it, and ending iteration closes the underlying request; the Effect emitter analogously returns `Stream` directly.
-- Promise SSE connection establishment, declared HTTP failures, and infrastructure failures occur during `AsyncIterable` iteration, beginning with its first `next()` call, rather than during synchronous method construction.
-- Neither generated streaming runtime automatically reconnects after disconnection. Promise `AsyncIterable` and Effect `Stream` fail explicitly; live consumers refresh and resubscribe, while durable sequence-based resume remains explicit composition above the generated client.
-- Promise client construction is synchronous and network-free. It requires `baseUrl`, defaults to `globalThis.fetch`, accepts client-level headers, and merges them with per-call header overrides.
-- Effect client construction accepts an explicit `baseUrl` and obtains `HttpClient.HttpClient` from the Effect environment. It does not install fetch or duplicate per-call transport policy; callers transform/provide the client for headers, tracing, retries, recording, and tests, while fiber interruption owns cancellation.
-- Promise and Effect emitters each own their generated public type modules. The **SDK Contract IR**, not a physically shared generated type package, is the common source; this permits zero-Effect wire types and rich decoded Effect types to evolve independently.
-- Promise and Effect network clients ship from `@opencode-ai/client` behind isolated root and `/effect` exports. The root has no runtime path to Effect; `/effect` imports only Effect, Schema, and Protocol.
-- The Effect-native scoped host belongs to `@opencode-ai/sdk-next`, which will assume the existing `@opencode-ai/sdk` name after legacy consumers migrate. Client remains network-only and SDK depends one-way on Client.
-- SDK executes Server's assembled `HttpRouter` in memory. It opens no listener and performs no network I/O, while preserving Server routing, middleware, codecs, handlers, and errors.
-- The Effect Client and SDK re-export their decoded datatype facade from Schema so callers do not depend on internal package locations or Core's versioned names.
-- A capability intended for both networked and **Embedded OpenCode** belongs in the authoritative public `HttpApi`; embedded-only same-process capabilities extend **Embedded OpenCode** separately.
-- `sessions.log({ sessionID, after, follow })` is the public durable Session event stream. It verifies the Session, replays durable events after the optional aggregate sequence, optionally continues with newly committed durable events, excludes live-only fragments, and is transported as SSE in both networked and embedded modes.
-- `events.subscribe()` is a distinct public instance-wide live stream for Session and non-Session activity. It has no replay guarantee and includes connection, heartbeat, and instance-disposal lifecycle events; consumers recover from disconnection by refreshing authoritative state.
-- A Session ID is not an optional filter on `events.subscribe()`: instance-wide live events and durable Session events have different schemas, replay guarantees, cursors, lifecycle events, and failure behavior.
-- The initial common OpenCode Client does not expose server-global event aggregation. `events.subscribe()` is bounded to the connected OpenCode instance or workspace; any future cross-instance administrative stream requires a separately designed API.
-- `events.subscribe()` does not automatically reconnect after transport loss. The live-only stream fails with `ClientError`; consumers refresh authoritative state before explicitly opening a new subscription because events missed during disconnection cannot be replayed.
-- `sessions.log({ sessionID, after, follow })` returns the generated HTTP client's cold durable event stream and does not build reconnection policy into the endpoint or client constructor. Transport loss fails the stream with `ClientError`. Callers may compose an explicit resuming stream above it by retaining the last observed durable sequence and opening a new subscription with `after`; any reusable resume helper remains a separate API design question.
-- The stable `sessions.list(...)` design returns a **Page** in both networked and **Embedded OpenCode**; embedded execution does not define a separate unbounded array-returning list operation. The beta client currently preserves the existing HTTP `{ data, cursor }` envelope until emitter-level Page projection is implemented.
-- Session list cursors are opaque branded values carrying continuation query and ordering state. Consumers pass them back unchanged and do not inspect storage anchors or encoded filter fields.
-- A Session list continuation accepts only its opaque cursor. Scope, filters, ordering, and page size are fixed by the initial query and carried by that cursor.
-- `sessions.messages(...)` returns a **Page** and uses the same cursor discipline as `sessions.list(...)`: the initial request supplies `sessionID`, ordering, and page size; continuation supplies `sessionID` plus only an opaque branded message cursor carrying ordering, page size, direction, and message anchor. Using a cursor with another Session is invalid.
-- `sessions.message({ sessionID, messageID })` is a required resource lookup. An unknown Session fails with `SessionNotFoundError`; a known Session with an absent or differently owned message fails with `MessageNotFoundError` without disclosing cross-Session ownership. Absence is not represented as `undefined` across the public HTTP boundary.
-- `sessions.interrupt({ sessionID })` first verifies that the durable Session exists, failing with `SessionNotFoundError` otherwise. For a known Session, interruption is idempotent: idle, already-settled, or locally unowned execution is a no-op.
-- `sessions.active()` snapshots the current process's foreground Session drain registry as a record of Session IDs to `{ type: "running" }`. Missing IDs are inactive; background subagents and tasks do not make their parent Session active, and process restart clears the registry.
-- `sessions.context({ sessionID })` preserves the existing message-only operation. It returns projected **Session History**; it does not include or represent the complete **Model Context**, whose agent system text, **Initial Instructions**, tools, and step-local additions remain separate.
-- **Open question**: Should a future, separately named operation expose complete **Model Context**, including the instruction baseline, applied instruction metadata, tools, and step-local additions?
-- `sessions.prompt(...)` exposes `resume?: boolean`. Omitting it preserves durable admission followed by an advisory execution wake; `resume: false` requests durable admit-only behavior.
-- The public operation remains `sessions.prompt(...)`; `SessionPending.admit` is the internal primitive, while the public `Admission` result and `resume` option express its durable admission semantics.
-- `sessions.create(...)` accepts an optional `location`. Omission resolves through the connected OpenCode instance's default or current location; an explicit value selects a known location. Networked and embedded transports use the same handler semantics.
-- `sessions.switchAgent({ sessionID, agent })` is part of the common client alongside `sessions.switchModel(...)`. It affects subsequent Session activity and fails with `SessionNotFoundError` for an unknown Session.
-- The **Embedded OpenCode** Layer delegates to the same scoped creation path; it does not define a second implementation.
-- A **PTY Environment** adapter observes plugins in the request Location while passing the resolved PTY working directory to the hook; standalone servers use an empty adapter.
-- An **Instruction Update** lowers to the provider's native chronological instruction role when supported and to a wrapped chronological fallback otherwise.
-- When the aggregate discovered instruction set changes, its **Instruction Update** includes the complete current ordered set and supersedes the prior aggregate value; when no discovered instructions remain, the message states that previously loaded instructions no longer apply.
-- Ambient project instruction discovery honors `OPENCODE_DISABLE_PROJECT_CONFIG`; global instructions remain eligible.
-- Oversized textual **Model Tool Output** retains a bounded preview in Session history while its complete text moves to managed tool-output storage. Arbitrary structured-result size is a separate concern.
-- One tool settlement receives one aggregate textual limit, using the configured maximum lines or UTF-8 bytes, whichever is reached first. The limit is provider-independent; token pressure belongs to context assembly and compaction.
-- Generic truncation preserves the beginning and end of textual output. Tools may apply a more meaningful strategy before the Tool Registry enforces the final limit.
-- A truncated **Model Tool Output** identifies its complete text in the bounded model-visible preview. The Tool Registry also supplies managed paths as internal metadata to tool hooks; Session events do not expose a typed `outputPaths` field.
-- A **Managed Tool Output File** is temporary and may expire after its retention period. The bounded **Model Tool Output**, not the file, is the durable replayable record.
-- Failure to retain a **Managed Tool Output File** fails settlement operationally. The Session never publishes a successful result whose complete output was lost during generic bounding.
-- Once a tool operation succeeds, bounding its **Model Tool Output** and publishing its one durable settlement form an interruption-safe completion region. Raw oversized success is never published before a later correction.
-- When a structured-only result would exceed the **Model Tool Output** limit, its validated structured value remains unchanged for Session consumers while model replay uses a bounded textual JSON preview and optional managed output path.
-- Existing tool-managed output paths survive generic bounding. A fallback file retains exactly the complete projected text received by the Tool Registry and never claims to reconstruct output already discarded by tool-specific shaping.
-- **Managed Tool Output Files** use globally unique names in one shared flat directory. They receive no special filesystem authority; each tool applies its ordinary external-path policy.
-- Provider-executed tool results remain provider-native transcript facts outside generic Tool Registry bounding. Their context control requires provider-aware pruning or compaction because some providers require exact structured round-trip payloads.
-
-## Client contract architecture
-
-Semantic values that mean the same thing internally and publicly live in the lightweight Schema leaf. Core consumes Schema for domain behavior; Protocol composes Schema values into paths, payloads, envelopes, errors, cursors, and streams; Server imports both, hosts Protocol's exact groups, and owns protocol/domain adaptation. The root Promise client remains zero-Effect, `/effect` depends on Effect plus Schema and Protocol, and `@opencode-ai/sdk-next` composes the scoped in-process host above Client, Core, and Server.
-
-Shared public records are plain objects declared with `Schema.Struct`. A same-name inferred interface gives object records readable TypeScript signatures without constructors, prototypes, or nominal identity; unions retain explicit type aliases.
-
-Before stabilizing the client API:
-
-- Keep additional public schemas in Schema and additional network groups in Protocol; neither package may transitively load databases, Drizzle, Session execution, providers, watchers, native modules, or WASM.
-- Keep concrete Location middleware keys in Server while Protocol owns their placement. Client projections may supply transport-only keys, but must prove generated equivalence with Server's concrete API.
-- Project the existing list response envelope to the stable client **Page** shape and enforce separate initial-query and cursor-continuation inputs without changing the hosted V2 wire contract.
-- Settle the stable consumer namespace (`session` versus the current beta `sessions`) and use an explicit codegen annotation if the consumer name should differ from the server group identifier.
-- Preserve V2 route paths, operation IDs, codecs, errors, middleware behavior, and OpenAPI output while making this change.
-- Preserve browser-safe `@opencode-ai/client` and `@opencode-ai/client/effect` bundles through import-boundary tests.
-- Define embedded-host placement before supporting multiple hosts over one database. Hosts that share durable Session storage must also share process-local Session execution coordination, or each host must receive isolated storage explicitly.
-- Keep an embedded request scope alive until any streamed response body finishes. The initial non-streaming Session surface does not exercise this lifetime boundary; Session and instance event streams must do so before joining the embedded client.
-
-## Example dialogue
-
-> **Dev:** "The date changed while the session was active. Should the **Instruction Update** say what the old date was?"
-> **Domain expert:** "No. Emit the newly effective date so the agent can act on the current instructions."
-
-## Flagged ambiguities
-
-- Legacy `experimental.chat.system.transform` can mutate assembled system text arbitrarily, but V2 plugins do not yet expose an equivalent hook. Decide separately whether to port it, model dynamic uses as explicit **Instruction Sources**, or narrow its semantics.

CONTRIBUTING.md

@@ -1,299 +1,112 @@
 # Contributing to OpenCode
 
-We want to make it easy for you to contribute to OpenCode. Here are the most common type of changes that get merged:
+The changes most likely to be accepted are:
 
 - Bug fixes
-- Additional LSPs / Formatters
-- Improvements to LLM performance
-- Support for new providers
-- Fixes for environment-specific quirks
+- Additional LSPs and formatters
+- LLM performance improvements
+- Environment-specific fixes
 - Missing standard behavior
 - Documentation improvements
 
-However, any UI or core product feature must go through a design review with the core team before implementation.
+UI and core product features require design review before implementation. If you are unsure whether a change fits, ask a maintainer or choose an issue labeled [`help wanted`](https://github.com/anomalyco/opencode/issues?q=is%3Aissue%20state%3Aopen%20label%3Ahelp-wanted), [`good first issue`](https://github.com/anomalyco/opencode/issues?q=is%3Aissue%20state%3Aopen%20label%3A%22good%20first%20issue%22), [`bug`](https://github.com/anomalyco/opencode/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug), or [`perf`](https://github.com/anomalyco/opencode/issues?q=is%3Aopen%20is%3Aissue%20label%3A%22perf%22).
 
-If you are unsure if a PR would be accepted, feel free to ask a maintainer or look for issues with any of the following labels:
-
-- [`help wanted`](https://github.com/anomalyco/opencode/issues?q=is%3Aissue%20state%3Aopen%20label%3Ahelp-wanted)
-- [`good first issue`](https://github.com/anomalyco/opencode/issues?q=is%3Aissue%20state%3Aopen%20label%3A%22good%20first%20issue%22)
-- [`bug`](https://github.com/anomalyco/opencode/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug)
-- [`perf`](https://github.com/anomalyco/opencode/issues?q=is%3Aopen%20is%3Aissue%20label%3A%22perf%22)
+Want to take on an issue? Leave a comment and a maintainer may assign it unless it is already being worked on.
 
 > [!NOTE]
 > PRs that ignore these guardrails will likely be closed.
 
-Want to take on an issue? Leave a comment and a maintainer may assign it to you unless it is something we are already working on.
+## Adding Providers
 
-## Adding New Providers
+New providers should rarely require OpenCode changes. Add the provider to [models.dev](https://github.com/anomalyco/models.dev) first.
 
-New providers shouldn't require many if ANY code changes, but if you want to add support for a new provider first make a PR to:
-https://github.com/anomalyco/models.dev
+## Development
 
-## Developing OpenCode
-
-- Requirements: Bun 1.3+
-- Install dependencies and start the dev server from the repo root:
-
-  ```bash
-  bun install
-  bun dev
-  ```
-
-### Running against a different directory
-
-By default, `bun dev` runs OpenCode in the `packages/opencode` directory. To run it against a different directory or repository:
+OpenCode requires Bun 1.3 or newer. From the repository root:
 
 ```bash
-bun dev <directory>
+bun install
+bun dev [directory]
 ```
 
-To run OpenCode in the root of the opencode repo itself:
+`bun dev` runs the V2 CLI and TUI. Pass a directory to open another project, or `.` to open this repository.
+
+To test a development TUI against your installed OpenCode V2 background service and live sessions:
 
 ```bash
-bun dev .
+bun run dev:live [directory]
 ```
 
-### Building a "localcode"
-
-To compile a standalone executable:
+For web development, run the backend and app in separate terminals. Other interfaces have root scripts:
 
 ```bash
-./packages/opencode/script/build.ts --single
+bun dev serve --port 4096
+bun run dev:web
+bun run dev:desktop
+bun run dev:www
 ```
 
-Then run it with:
+### Packages
+
+- `packages/schema`: shared wire and storage contracts
+- `packages/core`: domain behavior and persistence
+- `packages/protocol`: public API definitions
+- `packages/server`: HTTP server and runtime composition
+- `packages/client`: generated TypeScript clients
+- `packages/cli`: command-line entrypoint and service lifecycle
+- `packages/tui`: terminal interface
+- `packages/app`: shared web interface
+- `packages/desktop`: Electron desktop application
+- `packages/plugin`: plugin API
+
+### Verification
+
+Run typechecks, and tests where defined, from the affected package rather than the repository root:
 
 ```bash
-./packages/opencode/dist/opencode-<platform>/bin/opencode
+cd packages/core
+bun run test
+bun typecheck
 ```
 
-Replace `<platform>` with your platform (e.g., `darwin-arm64`, `linux-x64`).
+Follow package-specific instructions in nearby `AGENTS.md` files. After changing the public Protocol or Server `HttpApi`, run `bun run generate` from `packages/client`; never edit generated client files directly.
 
-- Core pieces:
-  - `packages/opencode`: OpenCode core business logic & server.
-  - `packages/opencode/src/cli/cmd/tui/`: The TUI code, written in SolidJS with [opentui](https://github.com/sst/opentui)
-  - `packages/app`: The shared web UI components, written in SolidJS
-  - `packages/desktop`: The native desktop app, built with Electron (wraps `packages/app`)
-  - `packages/plugin`: Source for `@opencode-ai/plugin`
+Follow the repository [style guide](./AGENTS.md).
 
-### Understanding bun dev vs opencode
+## Pull Requests
 
-During development, `bun dev` is the local equivalent of the built `opencode` command. Both run the same CLI interface:
+### Link Issues When Required
 
-```bash
-# Development (from project root)
-bun dev --help           # Show all available commands
-bun dev serve            # Start headless API server
-bun dev web              # Start server + open web interface
-bun dev <directory>      # Start TUI in specific directory
+Bug fixes, chores, and tests must reference an existing issue. Documentation, refactor, and feature PRs are exempt from the automated linked-issue check. When required, use `Fixes #123` or `Closes #123` in the PR description.
 
-# Production
-opencode --help          # Show all available commands
-opencode serve           # Start headless API server
-opencode web             # Start server + open web interface
-opencode <directory>     # Start TUI in specific directory
-```
+Before implementing new functionality, open a feature request describing the problem, why it belongs in OpenCode, and your proposed approach if you have one. Wait for design approval before opening the implementation PR.
 
-### Running the API Server
+Base branches on `v2`, not `dev`, and complete the provided pull request template.
 
-To start the OpenCode headless API server:
+### Keep It Focused
 
-```bash
-bun dev serve
-```
+- Keep PRs small and focused.
+- Explain the problem and why the change fixes it.
+- Check whether the functionality already exists.
+- For UI changes, include before-and-after screenshots or video.
+- For logic changes, explain what you tested and how a reviewer can verify it.
 
-This starts the headless server on port 4096 by default. You can specify a different port:
+### Keep It Brief
 
-```bash
-bun dev serve --port 8080
-```
+Long, AI-generated PR descriptions and issues may be ignored. Write a short explanation in your own words. If the change cannot be explained briefly, the PR may be too large.
 
-### Running the Web App
+### Use Conventional Titles
 
-To test UI changes during development:
-
-1. **First, start the OpenCode server** (see [Running the API Server](#running-the-api-server) section above)
-2. **Then run the web app:**
-
-```bash
-bun run --cwd packages/app dev
-```
-
-This starts a local dev server at http://localhost:5173 (or similar port shown in output). Most UI changes can be tested here, but the server must be running for full functionality.
-
-### Running the Desktop App
-
-The desktop app is an Electron application that wraps the web UI.
-
-To run the desktop app in development:
-
-```bash
-bun run --cwd packages/desktop dev
-```
-
-To create a production build and package the app:
-
-```bash
-bun run --cwd packages/desktop build
-bun run --cwd packages/desktop package
-```
-
-> [!NOTE]
-> If you make changes to the API or SDK (e.g. `packages/opencode/src/server/server.ts`), run `./script/generate.ts` to regenerate the SDK and related files.
-
-Please try to follow the [style guide](./AGENTS.md)
-
-### Setting up a Debugger
-
-Bun debugging is currently rough around the edges. We hope this guide helps you get set up and avoid some pain points.
-
-The most reliable way to debug OpenCode is to run it manually in a terminal via `bun run --inspect=<url> dev ...` and attach
-your debugger via that URL. Other methods can result in breakpoints being mapped incorrectly, at least in VSCode (YMMV).
-
-Caveats:
-
-- If you want to run the OpenCode TUI and have breakpoints triggered in the server code, you might need to run `bun dev spawn` instead of
-  the usual `bun dev`. This is because `bun dev` runs the server in a worker thread and breakpoints might not work there.
-- If `spawn` does not work for you, you can debug the server separately:
-  - Debug server: `bun run --inspect=ws://localhost:6499/ --cwd packages/opencode ./src/index.ts serve --port 4096`,
-    then attach TUI with `opencode attach http://localhost:4096`
-  - Debug TUI: `bun run --inspect=ws://localhost:6499/ --cwd packages/opencode --conditions=browser ./src/index.ts`
-
-Other tips and tricks:
-
-- You might want to use `--inspect-wait` or `--inspect-brk` instead of `--inspect`, depending on your workflow
-- Specifying `--inspect=ws://localhost:6499/` on every invocation can be tiresome, you may want to `export BUN_OPTIONS=--inspect=ws://localhost:6499/` instead
-
-#### VSCode Setup
-
-If you use VSCode, you can use our example configurations [.vscode/settings.example.json](.vscode/settings.example.json) and [.vscode/launch.example.json](.vscode/launch.example.json).
-
-Some debug methods that can be problematic:
-
-- Debug configurations with `"request": "launch"` can have breakpoints incorrectly mapped and thus unusable
-- The same problem arises when running OpenCode in the VSCode `JavaScript Debug Terminal`
-
-With that said, you may want to try these methods, as they might work for you.
-
-## Pull Request Expectations
-
-### Issue First Policy
-
-**All PRs must reference an existing issue.** Before opening a PR, open an issue describing the bug or feature. This helps maintainers triage and prevents duplicate work. PRs without a linked issue may be closed without review.
-
-- Use `Fixes #123` or `Closes #123` in your PR description to link the issue
-- For small fixes, a brief issue is fine - just enough context for maintainers to understand the problem
-
-### General Requirements
-
-- Keep pull requests small and focused
-- Explain the issue and why your change fixes it
-- Before adding new functionality, ensure it doesn't already exist elsewhere in the codebase
-
-### UI Changes
-
-If your PR includes UI changes, please include screenshots or videos showing the before and after. This helps maintainers review faster and gives you quicker feedback.
-
-### Logic Changes
-
-For non-UI changes (bug fixes, new features, refactors), explain **how you verified it works**:
-
-- What did you test?
-- How can a reviewer reproduce/confirm the fix?
-
-### No AI-Generated Walls of Text
-
-Long, AI-generated PR descriptions and issues are not acceptable and may be ignored. Respect the maintainers' time:
-
-- Write short, focused descriptions
-- Explain what changed and why in your own words
-- If you can't explain it briefly, your PR might be too large
-
-### PR Titles
-
-PR titles should follow conventional commit standards:
-
-- `feat:` new feature or functionality
-- `fix:` bug fix
-- `docs:` documentation or README changes
-- `chore:` maintenance tasks, dependency updates, etc.
-- `refactor:` code refactoring without changing behavior
-- `test:` adding or updating tests
-
-You can optionally include a scope to indicate which package is affected:
-
-- `feat(app):` feature in the app package
-- `fix(desktop):` bug fix in the desktop package
-- `chore(opencode):` maintenance in the opencode package
+Use `type(scope): summary`. Supported types are `feat`, `fix`, `docs`, `chore`, `refactor`, and `test`. The scope is optional.
 
 Examples:
 
-- `docs: update contributing guidelines`
-- `fix: resolve crash on startup`
-- `feat: add dark mode support`
-- `feat(app): add dark mode support`
-- `fix(desktop): resolve crash on startup`
-- `chore: bump dependency versions`
+- `docs: update contributing guide`
+- `fix(tui): restore scroll position`
+- `feat(app): add workspace search`
 
-### Style Preferences
+## Issues
 
-These are not strictly enforced, they are just general guidelines:
+Bug reports and feature requests must use their issue templates. Blank issues are not allowed; ask support and how-to questions in the [Discord community](https://discord.gg/opencode).
 
-- **Functions:** Keep logic within a single function unless breaking it out adds clear reuse or composition benefits.
-- **Destructuring:** Do not do unnecessary destructuring of variables.
-- **Control flow:** Avoid `else` statements.
-- **Error handling:** Prefer `.catch(...)` instead of `try`/`catch` when possible.
-- **Types:** Reach for precise types and avoid `any`.
-- **Variables:** Stick to immutable patterns and avoid `let`.
-- **Naming:** Choose concise single-word identifiers when they remain descriptive.
-- **Runtime APIs:** Use Bun helpers such as `Bun.file()` when they fit the use case.
-
-## Feature Requests
-
-For net-new functionality, start with a design conversation. Open an issue describing the problem, your proposed approach (optional), and why it belongs in OpenCode. The core team will help decide whether it should move forward; please wait for that approval instead of opening a feature PR directly.
-
-## Trust & Vouch System
-
-This project uses [vouch](https://github.com/mitchellh/vouch) to manage contributor trust. The vouch list is maintained in [`.github/VOUCHED.td`](.github/VOUCHED.td).
-
-### How it works
-
-- **Vouched users** are explicitly trusted contributors.
-- **Denounced users** are explicitly blocked. Issues and pull requests from denounced users are automatically closed. If you have been denounced, you can request to be unvouched by reaching out to a maintainer on [Discord](https://opencode.ai/discord)
-- **Everyone else** can participate normally — you don't need to be vouched to open issues or PRs.
-
-### For maintainers
-
-Collaborators with write access can manage the vouch list by commenting on any issue:
-
-- `vouch` — vouch for the issue author
-- `vouch @username` — vouch for a specific user
-- `denounce` — denounce the issue author
-- `denounce @username` — denounce a specific user
-- `denounce @username <reason>` — denounce with a reason
-- `unvouch` / `unvouch @username` — remove someone from the list
-
-Changes are committed automatically to `.github/VOUCHED.td`.
-
-### Denouncement policy
-
-Denouncement is reserved for users who repeatedly submit low-quality AI-generated contributions, spam, or otherwise act in bad faith. It is not used for disagreements or honest mistakes.
-
-## Issue Requirements
-
-All issues **must** use one of our issue templates:
-
-- **Bug report** — for reporting bugs (requires a description)
-- **Feature request** — for suggesting enhancements (requires verification checkbox and description)
-- **Question** — for asking questions (requires the question)
-
-Blank issues are not allowed. When a new issue is opened, an automated check verifies that it follows a template and meets our contributing guidelines. If an issue doesn't meet the requirements, you'll receive a comment explaining what needs to be fixed and have **2 hours** to edit the issue. After that, it will be automatically closed.
-
-Issues may be flagged for:
-
-- Not using a template
-- Required fields left empty or filled with placeholder text
-- AI-generated walls of text
-- Missing meaningful content
-
-If you believe your issue was incorrectly flagged, let a maintainer know.
+Automated checks flag missing templates, placeholder text, AI-generated walls of text, and missing meaningful content. You have two hours to correct a flagged issue before it closes automatically. Ask a maintainer if an issue was flagged incorrectly.

V2_HTTP_API_AUDIT.md

@@ -0,0 +1,279 @@
+# V2 HTTP API audit checklist
+
+**Source:** `packages/protocol/openapi.json`  
+**Current endpoint count:** 139
+**Last regenerated:** 2026-09-13
+
+## How to use this checklist
+
+Review endpoints in document order. For each endpoint, select one disposition and capture rationale or follow-up work in Notes. Mark **Reviewed** only after the disposition is agreed.
+
+### Review criteria
+
+- Resource and operation naming
+- HTTP method and idempotency
+- Request parameters and location scope
+- Response shape and error taxonomy
+- Authentication and authorization
+- Current production consumers
+- Stability level: public, experimental, or internal
+- Whether the generated client API is intuitive
+
+### Disposition legend
+
+- **Keep:** ship unchanged as a supported V2 API
+- **Change:** retain after a defined contract change
+- **Remove:** exclude from the official V2 API
+- **Experimental-only:** retain outside the stable API commitment
+
+## Progress
+
+- [x] Group 1: Foundation and placement (4)
+- [ ] Group 2: Configuration and capability catalogs (16)
+- [ ] Group 3: Credentials, integrations, MCP, and web search (22)
+- [ ] Group 4: Session lifecycle (12)
+- [ ] Group 5: Session execution and inputs (11)
+- [ ] Group 6: Session history and recovery (13)
+- [ ] Group 7: Inbox, permissions, and forms (19)
+- [ ] Group 8: Filesystem, worktrees, and VCS (12)
+- [ ] Group 9: PTYs, persistent terminals, and shells (24)
+- [ ] Group 10: Events, RPC, and experimental operations (6)
+
+## Resolved during audit
+
+### [x] `POST /api/plugin/await-activation`
+
+- **Decision:** Remove
+- **Notes:** Activation timing is an internal server concern. Catalog reads remain non-blocking.
+
+### [x] Location response wrappers
+
+- **Decision:** Reduce generic endpoint response locations to `{ directory }`.
+- **Notes:** Full project metadata remains available from `GET /api/location`; no consumers used it from wrapped responses.
+
+### [x] `GET /api/health` and `GET /api/server`
+
+- **Decision:** Merge and rename
+- **Replacement:** `GET /api/status` with operation ID `server.status`.
+- **Notes:** Returns `version`, `pid`, and connection `urls`; readiness is conveyed by HTTP status.
+
+### [x] `GET /api/project/current`
+
+- **Decision:** Remove
+- **Replacement:** `GET /api/location`, using `project` from the response.
+- **Notes:** The endpoint duplicated `Location.Info.project`; production callers were migrated.
+
+### [x] `POST /api/workspace` and `DELETE /api/workspace/{workspaceID}`
+
+- **Decision:** Remove
+- **Notes:** Provider-backed workspaces are not part of the V2 HTTP contract and can be introduced later. Core and the embedded SDK retain internal workspace support.
+
+## Group 1: Foundation and placement
+
+**Endpoints:** 4
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [x] 001–002 | `GET` | `/api/status` | `server.status` | Keep | Replaces the former health and server endpoints. |
+| [x] 003 | `GET` | `/api/location` | `location.get` | Keep | Workspace selectors and response fields removed until workspace support ships. |
+| [x] 004 | `GET` | `/api/project` | `project.list` | Keep | Removed unused `time.initialized`; the database column remains for migration data. |
+| [x] 005 | `PATCH` | `/api/project/{projectID}` | `project.update` | Keep | Request and response accepted as-is. |
+
+## Group 2: Configuration and capability catalogs
+
+**Endpoints:** 16
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [x] 008 | `GET` | `/api/agent` | `agent.list` | Keep | Request and response accepted as-is. |
+| [x] 009 | `GET` | `/api/agent/{agentID}` | `agent.get` | Keep | Request, response, and not-found error accepted as-is. |
+| [x] 010 | `GET` | `/api/plugin` | `plugin.list` | Keep | Request and response accepted as-is. |
+| [x] 012 | `POST` | `/api/plugin/check` | `plugin.check` | Keep | Request and response accepted as-is. |
+| [x] 013 | `POST` | `/api/plugin/update` | `plugin.update` | Keep | Request and errors accepted as-is. |
+| [x] 014 | `GET` | `/api/model` | `model.list` | Keep | Request and response accepted as-is. |
+| [x] 015 | `GET` | `/api/model/default` | `model.default` | Keep | Request and nullable response accepted as-is. |
+| [x] 016 | `GET` | `/api/provider` | `provider.list` | Keep | Request and response accepted as-is. |
+| [x] 017 | `GET` | `/api/provider/{providerID}` | `provider.get` | Keep | Request, response, and not-found error accepted as-is. |
+| [x] 018 | `GET` | `/api/command` | `command.list` | Keep | Request and response accepted as-is. |
+| [x] 019 | `GET` | `/api/skill` | `skill.list` | Keep | Renamed `location` to `path`; removed the skill-specific `slash` flag and slash-command behavior. |
+| [x] 020 | `GET` | `/api/reference` | `reference.list` | Keep | Removed duplicate `description` and `hidden` fields from nested `source`. |
+| [x] 021 | `GET` | `/api/config` | `config.get` | Keep | Compatibility entries removed; response now contains only documents and OpenCode directories. |
+| [x] 022 | `GET` | `/api/config/preferences` | `config.preferences` | Remove | Redundant special projection of global config. |
+| [x] 023 | `PATCH` | `/api/config/preferences` | `config.updatePreferences` | Remove | Redundant field-specific config mutation API. |
+| [ ] 024 | `GET` | `/api/config/shell` | `config.shells` |  |  |
+| [x] 024a | `PATCH` | `/api/experimental/config` | `experimental.config.update` | Change | Experimental global config mutation; initially accepts only `shell`. |
+
+## Group 3: Credentials, integrations, MCP, and web search
+
+**Endpoints:** 22
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [x] 025 | `GET` | `/api/integration` | `integration.list` | Keep | Full integration inventory is consumed by authentication and integration-selection clients. |
+| [x] 026 | `GET` | `/api/integration/{integrationID}` | `integration.get` | Change | Missing integration now returns typed `404` instead of optional data. |
+| [x] | `POST` | `/api/experimental/integration/wellknown` | `experimental.integration.wellknown.add` | Experimental-only | Retained outside the stable API commitment. |
+| [ ] 028 | `POST` | `/api/integration/{integrationID}/connect/key` | `integration.connect.key` |  |  |
+| [ ] 029 | `POST` | `/api/integration/{integrationID}/connect/oauth` | `integration.oauth.connect` |  |  |
+| [ ] 030 | `GET` | `/api/integration/{integrationID}/connect/oauth/{attemptID}` | `integration.oauth.status` |  |  |
+| [ ] 031 | `DELETE` | `/api/integration/{integrationID}/connect/oauth/{attemptID}` | `integration.oauth.cancel` |  |  |
+| [ ] 032 | `POST` | `/api/integration/{integrationID}/connect/oauth/{attemptID}/complete` | `integration.oauth.complete` |  |  |
+| [ ] 033 | `POST` | `/api/integration/{integrationID}/connect/command` | `integration.command.connect` |  |  |
+| [ ] 034 | `GET` | `/api/integration/{integrationID}/connect/command/{attemptID}` | `integration.command.status` |  |  |
+| [ ] 035 | `DELETE` | `/api/integration/{integrationID}/connect/command/{attemptID}` | `integration.command.cancel` |  |  |
+| [x] 036 | `GET` | `/api/mcp` | `mcp.list` | Keep | MCP inventory and connection status retained. |
+| [x] 037 | `PUT` | `/api/experimental/mcp/{server}` | `experimental.mcp.add` | Experimental-only | Runtime-only MCP override; does not persist configuration. |
+| [x] 038 | `DELETE` | `/api/experimental/mcp/{server}` | `experimental.mcp.remove` | Experimental-only | Runtime removal override; missing server returns `404`. |
+| [x] 039 | `POST` | `/api/experimental/mcp/{server}/connect` | `experimental.mcp.connect` | Experimental-only | Runtime connection override retained outside the stable API. |
+| [x] 040 | `POST` | `/api/experimental/mcp/{server}/disconnect` | `experimental.mcp.disconnect` | Experimental-only | Runtime disconnection override retained outside the stable API. |
+| [ ] 041 | `GET` | `/api/mcp/resource` | `mcp.resource.catalog` |  | Deferred for later review. |
+| [x] 042 | `PATCH` | `/api/credential/{credentialID}` | `credential.update` | Change | Removed redundant location query; credentials and events are global. |
+| [x] 043 | `DELETE` | `/api/credential/{credentialID}` | `credential.remove` | Change | Removed redundant location query; credentials and events are global. |
+| [x] 044 | `POST` | `/api/credential/{credentialID}/activate` | `credential.activate` | Change | Removed redundant location query; credentials and events are global. |
+| [x] 045 | `GET` | `/api/websearch/provider` | `websearch.providers` | Keep | Provider availability remains location-scoped; singular resource path retained. |
+| [x] 046 | `POST` | `/api/websearch` | `websearch.query` | Keep | Unknown provider remains an invalid request; published time documented as Unix epoch milliseconds. |
+
+## Group 4: Session lifecycle
+
+**Endpoints:** 12
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [x] 047 | `GET` | `/api/session` | `session.list` | Keep | Existing filtering, ordering, and cursor contract retained for now. |
+| [x] 048 | `POST` | `/api/session` | `session.create` | Keep | Existing creation contract retained; model reference includes optional variant. |
+| [x] 049 | `GET` | `/api/experimental/session/stats` | `experimental.session.stats` | Experimental-only | Session analytics retained outside the stable API commitment. |
+| [x] 050 | `GET` | `/api/session/active` | `session.active` | Keep | Status record retained for future active-state expansion. |
+| [x] 051 | `GET` | `/api/session/{sessionID}` | `session.get` | Keep | Specific session read and typed `404` retained. |
+| [x] 052 | `DELETE` | `/api/session/{sessionID}` | `session.remove` | Keep | Session and child deletion with typed `404` retained. |
+| [x] 053 | `POST` | `/api/session/{sessionID}/fork` | `session.fork` | Change | Request now accepts optional branded `before` message ID; omission copies full history. |
+| [x] 054 | `POST` | `/api/session/{sessionID}/agent` | `session.switchAgent` | Keep | Subsequent-execution agent selection retained. |
+| [x] 055 | `POST` | `/api/session/{sessionID}/model` | `session.switchModel` | Keep | Subsequent-execution model and optional variant selection retained. |
+| [x] 056 | `PATCH` | `/api/session/{sessionID}` | `session.rename` | Change | Title-only rename now uses the session resource path. |
+| [x] 057 | `POST` | `/api/session/{sessionID}/move` | `session.move` | Change | Removed inaccurate local-change transfer claim; delivery behavior retained. |
+| [x] 058 | `POST` | `/api/session/{sessionID}/background` | `session.background` | Keep | Backgroundable foreground tools transition to background observation; idle requests remain no-ops. |
+
+## Group 5: Session execution and inputs
+
+**Endpoints:** 11
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [x] 059 | `POST` | `/api/session/{sessionID}/prompt` | `session.prompt` | Keep | Durable admission, delivery mode, and admit-only resume control retained. |
+| [x] 060 | `POST` | `/api/session/{sessionID}/command` | `session.command` | Change | Renamed request field from `command` to `name`; `204` retained. |
+| [x] 061 | `POST` | `/api/experimental/session/{sessionID}/skill` | `experimental.session.skill` | Experimental-only | Skill ID is now the `id` field; standalone activation remains experimental. |
+| [x] 062 | `POST` | `/api/session/{sessionID}/synthetic` | `session.synthetic` | Keep | Durable synthetic admission and delivery controls retained. |
+| [x] 063 | `POST` | `/api/session/{sessionID}/shell` | `session.shell` | Change | Caller ID is now the optimistic shell message ID; server derives its event ID. |
+| [x] 064 | `POST` | `/api/session/{sessionID}/compact` | `session.compact` | Keep | Durable compaction admission and delivery controls retained. |
+| [x] 065 | `POST` | `/api/experimental/session/{sessionID}/wait` | `experimental.session.wait` | Experimental-only | Race-free idle barrier retained outside the stable API. |
+| [x] 066 | `POST` | `/api/session/{sessionID}/generate` | `session.generate` | Keep | Transient generation from session context retained. |
+| [x] 067 | `POST` | `/api/session/{sessionID}/interrupt` | `session.interrupt` | Change | Renamed `continue` to `resume` across public and internal interruption APIs. |
+| [x] 068 | `PUT` | `/api/experimental/session/{sessionID}/environment` | `experimental.session.environment` | Experimental-only | Process-local environment replacement retained outside the stable API. |
+| [x] 069 | `POST` | `/api/session/{sessionID}/view` | `session.view` | Change | Idle watermark now uses the standard epoch-millisecond timestamp schema. |
+
+## Group 6: Session history and recovery
+
+**Endpoints:** 13
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [x] 070 | `POST` | `/api/experimental/session/import` | `experimental.session.import` | Experimental-only | Existing projected transcript import contract retained outside the stable API. |
+| [x] 071 | `GET` | `/api/experimental/session/{sessionID}/export` | `experimental.session.export` | Experimental-only | Existing projected transcript export contract retained outside the stable API. |
+| [x] 072 | `POST` | `/api/session/{sessionID}/revert/stage` | `session.revert.stage` | Keep | Existing staged history and optional file restoration behavior retained. |
+| [x] 073 | `DELETE` | `/api/session/{sessionID}/revert` | `session.revert.clear` | Change | Clearing staged revert now deletes the session revert resource. |
+| [x] 074 | `POST` | `/api/session/{sessionID}/revert/commit` | `session.revert.commit` | Keep | Explicit staged-revert commit action retained. |
+| [x] 075 | `GET` | `/api/session/{sessionID}/context` | `session.context` | Keep | Active model-context projection retained. |
+| [x] 076 | `GET` | `/api/session/{sessionID}/diff` | `session.diff` | Keep | Turn-range structured diff contract retained. |
+| [x] 077 | `GET` | `/api/experimental/session/{sessionID}/instructions/entries` | `experimental.session.instructions.entry.list` | Experimental-only | API-managed durable context entries retained outside the stable API. |
+| [x] 078 | `PUT` | `/api/experimental/session/{sessionID}/instructions/entries/{key}` | `experimental.session.instructions.entry.put` | Experimental-only | API-managed durable context entries retained outside the stable API. |
+| [x] 079 | `DELETE` | `/api/experimental/session/{sessionID}/instructions/entries/{key}` | `experimental.session.instructions.entry.remove` | Experimental-only | API-managed durable context entries retained outside the stable API. |
+| [x] | `GET` | `/api/experimental/session/{sessionID}/log` | `session.log` | Experimental-only | Retained outside the stable API commitment. |
+| [x] 081 | `GET` | `/api/session/{sessionID}/message/{messageID}` | `session.message.get` | Change | Normalized specific-message operation ID. |
+| [x] 082 | `GET` | `/api/session/{sessionID}/message` | `session.message.list` | Change | Normalized session-scoped message-list operation ID. |
+
+## Group 7: Inbox, permissions, and forms
+
+**Endpoints:** 19
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [x] 083 | `GET` | `/api/session/{sessionID}/inbox` | `session.inbox.list` | Change | Inbox timestamps now use the standard nested `time.created` shape. |
+| [ ] 084 | `DELETE` | `/api/session/{sessionID}/inbox/{inboxID}` | `session.inbox.cancel` |  |  |
+| [ ] 085 | `POST` | `/api/session/{sessionID}/inbox/{inboxID}/steer` | `session.inbox.steer` |  |  |
+| [ ] 086 | `POST` | `/api/session/{sessionID}/inbox/{inboxID}/queue` | `session.inbox.queue` |  |  |
+| [ ] 087 | `GET` | `/api/form/request` | `form.request.list` |  |  |
+| [ ] 088 | `GET` | `/api/session/{sessionID}/form` | `session.form.list` |  |  |
+| [ ] 089 | `POST` | `/api/session/{sessionID}/form` | `session.form.create` |  |  |
+| [ ] 090 | `GET` | `/api/session/{sessionID}/form/{formID}` | `session.form.get` |  |  |
+| [ ] 091 | `GET` | `/api/session/{sessionID}/form/{formID}/state` | `session.form.state` |  |  |
+| [ ] 092 | `POST` | `/api/session/{sessionID}/form/{formID}/reply` | `session.form.reply` |  |  |
+| [ ] 093 | `POST` | `/api/session/{sessionID}/form/{formID}/cancel` | `session.form.cancel` |  |  |
+| [ ] 094 | `GET` | `/api/permission/request` | `permission.request.list` |  |  |
+| [ ] 095 | `GET` | `/api/permission/saved` | `permission.saved.list` |  |  |
+| [ ] 096 | `DELETE` | `/api/permission/saved/{id}` | `permission.saved.remove` |  |  |
+| [ ] 097 | `POST` | `/api/session/{sessionID}/permission` | `session.permission.create` |  |  |
+| [ ] 098 | `GET` | `/api/session/{sessionID}/permission` | `session.permission.list` |  |  |
+| [ ] 099 | `GET` | `/api/session/{sessionID}/permission/{requestID}` | `session.permission.get` |  |  |
+| [ ] 100 | `POST` | `/api/session/{sessionID}/permission/{requestID}/reply` | `session.permission.reply` |  |  |
+| [ ] 101 | `PUT` | `/api/session/{sessionID}/permission/rules` | `session.permission.rules` |  |  |
+
+## Group 8: Filesystem, worktrees, and VCS
+
+**Endpoints:** 12
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [ ] 102 | `GET` | `/api/fs/read/*` | `fs.read` |  |  |
+| [ ] 103 | `GET` | `/api/fs/list` | `fs.list` |  |  |
+| [ ] 104 | `GET` | `/api/fs/find` | `fs.find` |  |  |
+| [ ] 105 | `GET` | `/api/worktree` | `worktree.list` |  |  |
+| [ ] 106 | `POST` | `/api/worktree` | `worktree.create` |  |  |
+| [ ] 107 | `DELETE` | `/api/worktree` | `worktree.remove` |  |  |
+| [ ] 108 | `POST` | `/api/worktree/refresh` | `worktree.refresh` |  |  |
+| [ ] 109 | `GET` | `/api/vcs` | `vcs.get` |  |  |
+| [ ] 110 | `GET` | `/api/vcs/base` | `vcs.base` |  |  |
+| [ ] 111 | `GET` | `/api/vcs/status` | `vcs.status` |  |  |
+| [ ] 112 | `GET` | `/api/vcs/branches` | `vcs.branches` |  |  |
+| [ ] 113 | `GET` | `/api/vcs/diff` | `vcs.diff` |  |  |
+
+## Group 9: PTYs, persistent terminals, and shells
+
+**Endpoints:** 24
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [ ] 114 | `GET` | `/api/pty` | `pty.list` |  |  |
+| [ ] 115 | `POST` | `/api/pty` | `pty.create` |  |  |
+| [ ] 116 | `GET` | `/api/pty/{ptyID}` | `pty.get` |  |  |
+| [ ] 117 | `PUT` | `/api/pty/{ptyID}` | `pty.update` |  |  |
+| [ ] 118 | `DELETE` | `/api/pty/{ptyID}` | `pty.remove` |  |  |
+| [ ] 119 | `POST` | `/api/pty/{ptyID}/connect-token` | `pty.connect.token` |  |  |
+| [ ] 120 | `GET` | `/api/pty/{ptyID}/connect` | `pty.connect` |  |  |
+| [x] | `GET` | `/api/experimental/session/{sessionID}/terminal/read` | `server.experimental.persistentPty.read` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `GET` | `/api/experimental/session/{sessionID}/terminal` | `server.experimental.persistentPty.list` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `POST` | `/api/experimental/session/{sessionID}/terminal` | `server.experimental.persistentPty.create` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `POST` | `/api/experimental/persistent-pty/shutdown` | `server.experimental.persistentPty.shutdown` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `POST` | `/api/experimental/persistent-pty/handoff` | `server.experimental.persistentPty.handoff` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `GET` | `/api/experimental/persistent-pty/{ptyID}` | `server.experimental.persistentPty.get` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `PUT` | `/api/experimental/persistent-pty/{ptyID}` | `server.experimental.persistentPty.update` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `DELETE` | `/api/experimental/persistent-pty/{ptyID}` | `server.experimental.persistentPty.remove` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `GET` | `/api/experimental/persistent-pty/{ptyID}/snapshot` | `server.experimental.persistentPty.snapshot` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `POST` | `/api/experimental/persistent-pty/{ptyID}/connect-token` | `server.experimental.persistentPty.connectToken` | Experimental-only | Retained outside the stable API commitment. |
+| [x] | `GET` | `/api/experimental/persistent-pty/{ptyID}/connect` | `persistentPty.connect` | Experimental-only | Retained outside the stable API commitment. |
+| [ ] 132 | `GET` | `/api/shell` | `shell.list` |  |  |
+| [ ] 133 | `POST` | `/api/shell` | `shell.create` |  |  |
+| [ ] 134 | `GET` | `/api/shell/{id}` | `shell.get` |  |  |
+| [ ] 135 | `DELETE` | `/api/shell/{id}` | `shell.remove` |  |  |
+| [ ] 136 | `PATCH` | `/api/shell/{id}/timeout` | `shell.timeout` |  |  |
+| [ ] 137 | `GET` | `/api/shell/{id}/output` | `shell.output` |  |  |
+
+## Group 10: Events, RPC, and experimental operations
+
+**Endpoints:** 6
+
+| Done | Method | Path | Operation ID | Decision | Notes |
+|---|---|---|---|---|---|
+| [ ] 138 | `POST` | `/api/generate` | `generate.text` |  |  |
+| [ ] 139 | `POST` | `/api/rpc/{rpcID}/{method}` | `rpc.call` |  |  |
+| [ ] 140 | `GET` | `/api/event` | `event.subscribe` |  |  |
+| [ ] 141 | `GET` | `/api/debug/location` | `debug.location.list` |  |  |
+| [ ] 142 | `DELETE` | `/api/debug/location` | `debug.location.evict` |  |  |
+| [x] | `GET` | `/api/experimental/migration/v1` | `experimental.migration.v1.status` | Experimental-only | Retained outside the stable API commitment. |

bunfig.toml

@@ -2,7 +2,7 @@
 exact = true
 # Only install newly resolved package versions published at least 3 days ago.
 minimumReleaseAge = 259200
-minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@opentui/core", "@opentui/core-darwin-arm64", "@opentui/core-darwin-x64", "@opentui/core-linux-arm64", "@opentui/core-linux-arm64-musl", "@opentui/core-linux-x64", "@opentui/core-linux-x64-musl", "@opentui/core-win32-arm64", "@opentui/core-win32-x64", "@opentui/keymap", "@opentui/solid", "opentui-spinner", "gitlab-ai-provider", "opencode-gitlab-auth", "@ff-labs/fff-node", "@ff-labs/fff-bun", "@ff-labs/fff-bin-darwin-arm64", "@ff-labs/fff-bin-darwin-x64", "@ff-labs/fff-bin-linux-arm64-gnu", "@ff-labs/fff-bin-linux-arm64-musl", "@ff-labs/fff-bin-linux-x64-gnu", "@ff-labs/fff-bin-linux-x64-musl", "@ff-labs/fff-bin-win32-arm64", "@ff-labs/fff-bin-win32-x64", "@pierre/diffs", "@pierre/theming", "app-builder-lib", "dmg-builder", "electron-builder", "electron-publish"]
+minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@brendonovich/vite-plugin-opencode", "@opencode/sdk", "@opencode-ai/pty", "@opencode-ai/pty-darwin-arm64", "@opencode-ai/pty-darwin-x64", "@opencode-ai/pty-linux-arm64-gnu", "@opencode-ai/pty-linux-arm64-musl", "@opencode-ai/pty-linux-x64-gnu", "@opencode-ai/pty-linux-x64-musl", "@opentui/core", "@opentui/core-darwin-arm64", "@opentui/core-darwin-x64", "@opentui/core-linux-arm64", "@opentui/core-linux-arm64-musl", "@opentui/core-linux-x64", "@opentui/core-linux-x64-musl", "@opentui/core-win32-arm64", "@opentui/core-win32-x64", "@opentui/keymap", "@opentui/solid", "opentui-spinner", "gitlab-ai-provider", "opencode-gitlab-auth", "@ff-labs/fff-node", "@ff-labs/fff-bun", "@ff-labs/fff-bin-darwin-arm64", "@ff-labs/fff-bin-darwin-x64", "@ff-labs/fff-bin-linux-arm64-gnu", "@ff-labs/fff-bin-linux-arm64-musl", "@ff-labs/fff-bin-linux-x64-gnu", "@ff-labs/fff-bin-linux-x64-musl", "@ff-labs/fff-bin-win32-arm64", "@ff-labs/fff-bin-win32-x64", "@pierre/diffs", "@pierre/theming", "app-builder-lib", "dmg-builder", "electron", "electron-builder", "electron-publish", "blume", "mermaid"]
 
 [test]
 root = "./do-not-run-tests-from-root"

github/package.json

@@ -15,6 +15,6 @@
     "@actions/github": "6.0.1",
     "@octokit/graphql": "9.0.1",
     "@octokit/rest": "catalog:",
-    "@opencode-ai/sdk": "workspace:*"
+    "@opencode-ai/sdk": "1.18.5"
   }
 }

infra/app.ts

@@ -1,4 +1,5 @@
 import { domain } from "./stage"
+import { createWebApp } from "./webapp"
 
 const GITHUB_APP_ID = new sst.Secret("GITHUB_APP_ID")
 const GITHUB_APP_PRIVATE_KEY = new sst.Secret("GITHUB_APP_PRIVATE_KEY")
@@ -59,11 +60,4 @@ new sst.cloudflare.x.Astro("Web", {
   },
 })
 
-new sst.cloudflare.StaticSite("WebApp", {
-  domain: "app." + domain,
-  path: "packages/app",
-  build: {
-    command: "bun turbo build",
-    output: "./dist",
-  },
-})
+createWebApp("app." + domain)

infra/console.ts

@@ -235,6 +235,7 @@ const bucketNew = new sst.cloudflare.Bucket("ZenDataNew")
 const DISCORD_INCIDENT_WEBHOOK_URL = new sst.Secret("DISCORD_INCIDENT_WEBHOOK_URL")
 const AWS_SES_ACCESS_KEY_ID = new sst.Secret("AWS_SES_ACCESS_KEY_ID")
 const AWS_SES_SECRET_ACCESS_KEY = new sst.Secret("AWS_SES_SECRET_ACCESS_KEY")
+const ENTERPRISE_SALES_INBOX_EMAIL = new sst.Secret("ENTERPRISE_SALES_INBOX_EMAIL")
 
 const SALESFORCE_CLIENT_ID = new sst.Secret("SALESFORCE_CLIENT_ID")
 const SALESFORCE_CLIENT_SECRET = new sst.Secret("SALESFORCE_CLIENT_SECRET")
@@ -263,6 +264,7 @@ new sst.cloudflare.x.SolidStart("Console", {
     EMAILOCTOPUS_API_KEY,
     AWS_SES_ACCESS_KEY_ID,
     AWS_SES_SECRET_ACCESS_KEY,
+    ENTERPRISE_SALES_INBOX_EMAIL,
     SALESFORCE_CLIENT_ID,
     SALESFORCE_CLIENT_SECRET,
     SALESFORCE_INSTANCE_URL,
@@ -288,8 +290,12 @@ new sst.cloudflare.x.SolidStart("Console", {
     server: {
       placement: { region: "aws:us-east-2" },
       transform: {
-        worker: {
-          tailConsumers: [{ service: logProcessor.nodes.worker.scriptName }],
+        worker: (args) => {
+          args.compatibilityFlags = $resolve(args.compatibilityFlags).apply((flags) => [
+            ...(flags ?? []),
+            "global_fetch_strictly_public",
+          ])
+          args.tailConsumers = [{ service: logProcessor.nodes.worker.scriptName }]
         },
       },
     },

infra/stage.ts

@@ -8,6 +8,25 @@ export const zoneID = "430ba34c138cfb5360826c4909f99be8"
 export const awsStage = $app.stage === "production" ? "production" : "dev"
 export const deployAws = $app.stage === awsStage
 
+if ($app.stage === "production") {
+  new cloudflare.DnsRecord("TrustCenter", {
+    zoneId: zoneID,
+    name: "trust.opencode.ai",
+    type: "CNAME",
+    content: "3a69a5bb27875189.vercel-dns-016.com",
+    proxied: false,
+    ttl: 60,
+  })
+
+  new cloudflare.DnsRecord("TrustCenterVerification", {
+    zoneId: zoneID,
+    name: "opencode.ai",
+    type: "TXT",
+    content: "compai-domain-verification=org_6993a99c6200a2d642bb115d",
+    ttl: 60,
+  })
+}
+
 new cloudflare.RegionalHostname("RegionalHostname", {
   hostname: domain,
   regionKey: "us",

infra/webapp.ts

@@ -0,0 +1,19 @@
+export function createWebApp(domain: string) {
+  return new sst.cloudflare.StaticSite("WebApp", {
+    domain,
+    path: "packages/app",
+    environment:
+      $app.stage === "beta"
+        ? {
+            OPENCODE_CHANNEL: "beta",
+            VITE_OPENCODE_SERVER_MODE: "none",
+            VITE_SENTRY_ENVIRONMENT: "beta",
+          }
+        : undefined,
+    build: {
+      // Preserve Sentry credentials and run source-map uploads on every deployment.
+      command: "bun run build",
+      output: "./dist",
+    },
+  })
+}

install

@@ -1,6 +1,7 @@
 #!/usr/bin/env bash
 set -euo pipefail
 APP=opencode
+SOURCE_APP=opencode
 
 MUTED='\033[0;2m'
 RED='\033[0;31m'
@@ -15,13 +16,13 @@ Usage: install.sh [options]
 
 Options:
     -h, --help              Display this help message
-    -v, --version <version> Install a specific version (e.g., 1.0.180)
+    -v, --version <version> Install a specific version (e.g., 0.0.0-beta-17236)
     -b, --binary <path>     Install from a local binary instead of downloading
         --no-modify-path    Don't modify shell config files (.zshrc, .bashrc, etc.)
 
 Examples:
-    curl -fsSL https://opencode.ai/install | bash
-    curl -fsSL https://opencode.ai/install | bash -s -- --version 1.0.180
+    curl -fsSL https://opencode.ai/v2/install | bash
+    curl -fsSL https://opencode.ai/v2/install | bash -s -- --version 0.0.0-beta-17236
     ./install --binary /path/to/opencode
 EOF
 }
@@ -109,11 +110,6 @@ else
         ;;
     esac
 
-    archive_ext=".zip"
-    if [ "$os" = "linux" ]; then
-      archive_ext=".tar.gz"
-    fi
-
     is_musl=false
     if [ "$os" = "linux" ]; then
       if [ -f /etc/alpine-release ]; then
@@ -165,42 +161,50 @@ else
       target="$target-musl"
     fi
 
-    filename="$APP-$target$archive_ext"
-
-
-    if [ "$os" = "linux" ]; then
-        if ! command -v tar >/dev/null 2>&1; then
-             echo -e "${RED}Error: 'tar' is required but not installed.${NC}"
-             exit 1
-        fi
-    else
-        if ! command -v unzip >/dev/null 2>&1; then
-            echo -e "${RED}Error: 'unzip' is required but not installed.${NC}"
-            exit 1
-        fi
+    if ! command -v tar >/dev/null 2>&1; then
+        echo -e "${RED}Error: 'tar' is required but not installed.${NC}"
+        exit 1
     fi
 
+    package_scope="@opencode"
     if [ -z "$requested_version" ]; then
-        url="https://github.com/anomalyco/opencode/releases/latest/download/$filename"
-        specific_version=$(curl -s https://api.github.com/repos/anomalyco/opencode/releases/latest | sed -n 's/.*"tag_name": *"v\([^"]*\)".*/\1/p')
+        metadata=$(curl -fsSL https://opencode.ai/update/api/beta/cli/npm || true)
+        specific_version=$(echo "$metadata" | sed -n 's/.*"version":"\([^"]*\)".*/\1/p')
+        package=$(echo "$metadata" | sed -n 's/.*"package":"\([^"]*\)".*/\1/p')
 
-        if [[ $? -ne 0 || -z "$specific_version" ]]; then
+        if [ -z "$specific_version" ] || [ -z "$package" ]; then
             echo -e "${RED}Failed to fetch version information${NC}"
             exit 1
         fi
+        package_scope="${package%/cli}"
     else
         # Strip leading 'v' if present
         requested_version="${requested_version#v}"
-        url="https://github.com/anomalyco/opencode/releases/download/v${requested_version}/$filename"
         specific_version=$requested_version
+    fi
 
-        # Verify the release exists before downloading
-        http_status=$(curl -sI -o /dev/null -w "%{http_code}" "https://github.com/anomalyco/opencode/releases/tag/v${requested_version}")
-        if [ "$http_status" = "404" ]; then
-            echo -e "${RED}Error: Release v${requested_version} not found${NC}"
-            echo -e "${MUTED}Available releases: https://github.com/anomalyco/opencode/releases${NC}"
-            exit 1
-        fi
+    package_name="$package_scope/cli-$target"
+    http_status=$(curl -s -o /dev/null -w "%{http_code}" "https://registry.npmjs.org/$package_scope%2fcli-$target/$specific_version" || true)
+    # Older clients install the minimum release before they can migrate package names.
+    if [ "$http_status" = "404" ] && [ -n "$requested_version" ]; then
+        package_name="@opencode-ai/cli-$target"
+        http_status=$(curl -s -o /dev/null -w "%{http_code}" "https://registry.npmjs.org/@opencode-ai%2fcli-$target/$specific_version" || true)
+    fi
+    if [ "$http_status" = "404" ]; then
+        echo -e "${RED}Error: Version ${specific_version} is not available for $target${NC}"
+        echo -e "${MUTED}Available versions: https://www.npmjs.com/package/$package_name?activeTab=versions${NC}"
+        exit 1
+    fi
+    if [ "$http_status" != "200" ]; then
+        echo -e "${RED}Failed to fetch package information${NC}"
+        exit 1
+    fi
+
+    filename="cli-$target-$specific_version.tgz"
+    url="https://registry.npmjs.org/$package_name/-/$filename"
+    binary_name="$SOURCE_APP"
+    if [ "$os" = "windows" ]; then
+        binary_name="$SOURCE_APP.exe"
     fi
 fi
 
@@ -219,18 +223,15 @@ print_message() {
 }
 
 check_version() {
-    if command -v opencode >/dev/null 2>&1; then
-        opencode_path=$(which opencode)
+    if command -v "$APP" >/dev/null 2>&1; then
+        opencode_path=$(which "$APP")
 
         ## Check the installed version
-        installed_version=$(opencode --version 2>/dev/null || echo "")
+        installed_version=$("$APP" --version 2>/dev/null || echo "")
+        installed_version="${installed_version##* }"
+        installed_version="${installed_version#v}"
 
-        if [[ "$installed_version" != "$specific_version" ]]; then
-            print_message info "${MUTED}Installed version: ${NC}$installed_version."
-        else
-            print_message info "${MUTED}Version ${NC}$specific_version${MUTED} already installed"
-            exit 0
-        fi
+        print_message info "${MUTED}Installed version: ${NC}$installed_version."
     fi
 }
 
@@ -287,7 +288,7 @@ download_with_progress() {
     trap "trap - RETURN; rm -f \"$tracefile\"; printf '\033[?25h' >&4; exec 4>&-" RETURN
 
     (
-        curl --trace-ascii "$tracefile" -s -L -o "$output" "$url"
+        curl --trace-ascii "$tracefile" -fsL -o "$output" "$url"
     ) &
     local curl_pid=$!
 
@@ -325,30 +326,56 @@ download_with_progress() {
 }
 
 download_and_install() {
-    print_message info "\n${MUTED}Installing ${NC}opencode ${MUTED}version: ${NC}$specific_version"
+    print_message info "\n${MUTED}Installing ${NC}$APP ${MUTED}version: ${NC}$specific_version"
     local tmp_dir="${TMPDIR:-/tmp}/opencode_install_$$"
     mkdir -p "$tmp_dir"
 
     if [[ "$os" == "windows" ]] || ! [ -t 2 ] || ! download_with_progress "$url" "$tmp_dir/$filename"; then
         # Fallback to standard curl on Windows, non-TTY environments, or if custom progress fails
-        curl -# -L -o "$tmp_dir/$filename" "$url"
+        curl -f -# -L -o "$tmp_dir/$filename" "$url"
     fi
 
-    if [ "$os" = "linux" ]; then
-        tar -xzf "$tmp_dir/$filename" -C "$tmp_dir"
-    else
-        unzip -q "$tmp_dir/$filename" -d "$tmp_dir"
+    tar -xzf "$tmp_dir/$filename" -C "$tmp_dir"
+    local installed_binary="$APP"
+    if [ "$os" = "windows" ]; then
+        installed_binary="$APP.exe"
     fi
-
-    mv "$tmp_dir/opencode" "$INSTALL_DIR"
-    chmod 755 "${INSTALL_DIR}/opencode"
+    mv "$tmp_dir/package/bin/$binary_name" "$INSTALL_DIR/$installed_binary"
+    chmod 755 "$INSTALL_DIR/$installed_binary"
     rm -rf "$tmp_dir"
 }
 
 install_from_binary() {
-    print_message info "\n${MUTED}Installing ${NC}opencode ${MUTED}from: ${NC}$binary_path"
-    cp "$binary_path" "${INSTALL_DIR}/opencode"
-    chmod 755 "${INSTALL_DIR}/opencode"
+    print_message info "\n${MUTED}Installing ${NC}$APP ${MUTED}from: ${NC}$binary_path"
+    local installed_binary="$APP"
+    case "$(uname -s)" in
+        MINGW*|MSYS*|CYGWIN*) installed_binary="$APP.exe" ;;
+    esac
+    cp "$binary_path" "$INSTALL_DIR/$installed_binary"
+    chmod 755 "$INSTALL_DIR/$installed_binary"
+}
+
+install_legacy_shim() {
+    local shim_os="${os:-}"
+    if [[ -z "$shim_os" ]]; then
+        case "$(uname -s)" in
+            MINGW*|MSYS*|CYGWIN*) shim_os="windows" ;;
+        esac
+    fi
+    rm -f "$INSTALL_DIR/opencode2" "$INSTALL_DIR/opencode2.exe" "$INSTALL_DIR/opencode2.cmd"
+    if [[ "$shim_os" == "windows" ]]; then
+        cat > "$INSTALL_DIR/opencode2.cmd" <<'EOF'
+@echo off
+"%~dp0opencode.exe" %*
+exit /b %errorlevel%
+EOF
+        return
+    fi
+    cat > "$INSTALL_DIR/opencode2" <<'EOF'
+#!/bin/sh
+exec "$(dirname "$0")/opencode" "$@"
+EOF
+    chmod 755 "$INSTALL_DIR/opencode2"
 }
 
 if [ -n "$binary_path" ]; then
@@ -357,6 +384,7 @@ else
     check_version
     download_and_install
 fi
+install_legacy_shim
 
 
 add_to_path() {
@@ -455,6 +483,6 @@ echo -e ""
 echo -e "cd <project>  ${MUTED}# Open directory${NC}"
 echo -e "opencode      ${MUTED}# Run command${NC}"
 echo -e ""
-echo -e "${MUTED}For more information visit ${NC}https://opencode.ai/docs"
+echo -e "${MUTED}For more information visit ${NC}https://opencode.ai/v2/docs"
 echo -e ""
 echo -e ""

nix/desktop.nix

@@ -4,14 +4,16 @@
   bun,
   nodejs,
   darwin,
-  electron_41,
+  callPackage,
   makeWrapper,
   writableTmpDirAsHomeHook,
   autoPatchelfHook,
+  copyDesktopItems,
+  makeDesktopItem,
   opencode,
 }:
 let
-  electron = electron_41;
+  electron = callPackage ./electron.nix { };
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "opencode-desktop";
@@ -27,9 +29,12 @@ stdenv.mkDerivation (finalAttrs: {
     nodejs
     makeWrapper
     writableTmpDirAsHomeHook
-  ] ++ lib.optionals stdenv.hostPlatform.isLinux [
+  ]
+  ++ lib.optionals stdenv.hostPlatform.isLinux [
     autoPatchelfHook
-  ] ++ lib.optionals stdenv.hostPlatform.isDarwin [
+    copyDesktopItems
+  ]
+  ++ lib.optionals stdenv.hostPlatform.isDarwin [
     # Ad-hoc sign the .app: --config.mac.identity=null below skips signing.
     darwin.autoSignDarwinBinariesHook
   ];
@@ -38,20 +43,35 @@ stdenv.mkDerivation (finalAttrs: {
     (lib.getLib stdenv.cc.cc)
   ];
 
+  desktopItems = lib.optional stdenv.hostPlatform.isLinux (makeDesktopItem {
+    name = "ai.opencode.desktop";
+    desktopName = "OpenCode";
+    exec = "opencode-desktop %U";
+    icon = "ai.opencode.desktop";
+    # Electron derives X11 WM_CLASS from app.name.
+    startupWMClass = "OpenCode";
+    categories = [ "Development" ];
+  });
+
   env = opencode.env // {
     ELECTRON_SKIP_BINARY_DOWNLOAD = "1";
   };
 
-  # https://github.com/electron/electron/issues/31121
-  # mac builds use a .app bundle which doesnt have this issue
-  postPatch = lib.optionalString stdenv.isLinux ''
-    BASE_PATH=packages/desktop
-    FILES=(src/main/windows.ts)
-    for file in "''${FILES[@]}"; do
-      substituteInPlace $BASE_PATH/$file \
+  postPatch =
+    # NOTE: Relax Bun version check to be a warning instead of an error
+    ''
+      substituteInPlace packages/script/src/index.ts \
+        --replace-fail 'throw new Error(`This script requires bun@''${expectedBunVersionRange}' \
+                       'console.warn(`Warning: This script requires bun@''${expectedBunVersionRange}'
+    ''
+    # https://github.com/electron/electron/issues/31121
+    # mac builds use a .app bundle which doesnt have this issue
+    + lib.optionalString stdenv.isLinux ''
+      substituteInPlace \
+        packages/desktop/src/main/windows/appearance.ts \
+        packages/desktop/src/main/service/desktop-cli.ts \
         --replace-fail "process.resourcesPath" "'$out/opt/opencode-desktop/resources'"
-    done
-  '';
+    '';
 
   preBuild = ''
     cp -r "${electron.dist}" $HOME/.electron-dist
@@ -67,6 +87,11 @@ stdenv.mkDerivation (finalAttrs: {
 
     cd packages/desktop
 
+    export OPENCODE_CLI_DIST="$TMPDIR/desktop-cli"
+    cli_package=$(bun -e 'import { getCurrentCli } from "./scripts/utils.ts"; console.log(getCurrentCli().package.replace("@opencode/", ""))')
+    mkdir -p "$OPENCODE_CLI_DIST/$cli_package/bin"
+    cp ${lib.getExe opencode} "$OPENCODE_CLI_DIST/$cli_package/bin/opencode"
+
     bun run build
     npx electron-builder --dir \
       --config electron-builder.config.ts \
@@ -76,27 +101,38 @@ stdenv.mkDerivation (finalAttrs: {
     runHook postBuild
   '';
 
-  installPhase =
-    ''
-      runHook preInstall
-    ''
-    + lib.optionalString stdenv.hostPlatform.isDarwin ''
-      mkdir -p $out/Applications
-      mv dist/mac*/*.app $out/Applications
-      makeWrapper "$out/Applications/OpenCode.app/Contents/MacOS/OpenCode" $out/bin/opencode-desktop
-    ''
-    + lib.optionalString stdenv.hostPlatform.isLinux ''
-      mkdir -p $out/opt/opencode-desktop
-      cp -r dist/linux*-unpacked/{resources,LICENSE*} $out/opt/opencode-desktop
-      makeWrapper ${lib.getExe electron} $out/bin/opencode-desktop \
-        --inherit-argv0 \
-        --set ELECTRON_FORCE_IS_PACKAGED 1 \
-        --add-flags $out/opt/opencode-desktop/resources/app.asar \
-        --add-flags "\''${NIXOS_OZONE_WL:+\''${WAYLAND_DISPLAY:+--ozone-platform-hint=auto --enable-features=WaylandWindowDecorations --enable-wayland-ime=true}}"
-    ''
-    + ''
-      runHook postInstall
-    '';
+  installPhase = ''
+    runHook preInstall
+  ''
+  + lib.optionalString stdenv.hostPlatform.isDarwin ''
+    mkdir -p $out/Applications
+    mv dist/mac*/*.app $out/Applications
+    makeWrapper "$out/Applications/OpenCode.app/Contents/MacOS/OpenCode" $out/bin/opencode-desktop
+  ''
+  + lib.optionalString stdenv.hostPlatform.isLinux ''
+    mkdir -p $out/opt/opencode-desktop
+    cp -r dist/linux*-unpacked/{resources,LICENSE*} $out/opt/opencode-desktop
+    install -Dm644 resources/icons/32x32.png \
+      "$out/share/icons/hicolor/32x32/apps/ai.opencode.desktop.png"
+    install -Dm644 resources/icons/64x64.png \
+      "$out/share/icons/hicolor/64x64/apps/ai.opencode.desktop.png"
+    install -Dm644 resources/icons/128x128.png \
+      "$out/share/icons/hicolor/128x128/apps/ai.opencode.desktop.png"
+    install -Dm644 resources/icons/128x128@2x.png \
+      "$out/share/icons/hicolor/256x256/apps/ai.opencode.desktop.png"
+    install -Dm644 resources/icons/icon.png \
+      "$out/share/icons/hicolor/512x512/apps/ai.opencode.desktop.png"
+    install -Dm644 resources/ai.opencode.desktop.metainfo.xml \
+      "$out/share/metainfo/ai.opencode.desktop.metainfo.xml"
+    makeWrapper ${lib.getExe electron} $out/bin/opencode-desktop \
+     --inherit-argv0 \
+     --set ELECTRON_FORCE_IS_PACKAGED 1 \
+     --add-flags $out/opt/opencode-desktop/resources/app.asar \
+     --add-flags "\''${NIXOS_OZONE_WL:+\''${WAYLAND_DISPLAY:+--ozone-platform-hint=auto --enable-features=WaylandWindowDecorations --enable-wayland-ime=true}}"
+  ''
+  + ''
+    runHook postInstall
+  '';
 
   autoPatchelfIgnoreMissingDeps = [
     "libc.musl-x86_64.so.1"

nix/electron.nix

@@ -0,0 +1,13 @@
+{ callPackage, path }:
+let
+  version = (builtins.fromJSON (builtins.readFile ../packages/desktop/package.json)).devDependencies.electron;
+in
+(callPackage (path + "/pkgs/development/tools/electron/binary/generic.nix") { }) version {
+  # Electron 42.10.1 SHASUMS256.txt; update with the desktop package version.
+  aarch64-linux = "20e68d6c4e47f3ebf59de7c6b1f8b8bec6a6ebda6a451132f9b465f3f13ce467";
+  x86_64-linux = "2452b27112d92387471fa2488aafac85d79ea3f2ee1216c0abd5150d6c12362b";
+  aarch64-darwin = "ac7194a3dfd81930ba35355c01620262c1254752859b42dcb8f4b9e4d174a871";
+  x86_64-darwin = "4489aba55477a0082266cb690db1c829503ba3338048599d8fd243953df37dab";
+  # fetchzip hashes the unpacked headers, not the release tarball.
+  headers = "sha256-4eUy3BZVvxTl7KUOsxio7769lL6ag/ecbeK+qLURWMI=";
+}

nix/hashes.json

@@ -1,8 +1,8 @@
 {
   "nodeModules": {
-    "x86_64-linux": "sha256-JTtn+wXTXg+yklvIMDLcGFaYhTU6ZrCgKT9JTNEQ3gA=",
-    "aarch64-linux": "sha256-gXU6zyhvAZrZirkL/PlHdkHtEof/7PVSPCaE34Jnd4U=",
-    "aarch64-darwin": "sha256-Q0oTG3uzOlD/X2kJingLle529lKFoTpyCW2rHXOZ6iE=",
-    "x86_64-darwin": "sha256-LINvKHxPibTlJeNzfACQx0x+Yj5oROT6Du3I5AtqqXk="
+    "x86_64-linux": "sha256-E5T4o3wNivOg8q4wRV7yE4CUNaUq4QNPxZsnvLaGUcg=",
+    "aarch64-linux": "sha256-xQQi7LgxInZQVCznASfa0Pm+cNGBo5j4Tfp0E/bKnNw=",
+    "aarch64-darwin": "sha256-qxb371dCf7WG09VvQE+HZ/x3EURpFIr11bdnQwGTHhw=",
+    "x86_64-darwin": "sha256-xi1qiYr41hzbgAQl+I3xdtnkiPx2SIcA/0OMpN78qXw="
   }
 }

nix/node_modules.nix

@@ -10,7 +10,7 @@
     ]).nodeModules.${stdenvNoCC.hostPlatform.system},
 }:
 let
-  packageJson = lib.pipe ../packages/opencode/package.json [
+  packageJson = lib.pipe ../packages/cli/package.json [
     builtins.readFile
     builtins.fromJSON
   ];
@@ -27,11 +27,12 @@ stdenvNoCC.mkDerivation {
     fileset = lib.fileset.intersection (lib.fileset.fromSource (lib.sources.cleanSource ../.)) (
       lib.fileset.unions [
         ../packages
+        ../services
         ../bun.lock
         ../package.json
         ../patches
         ../install # required by desktop build (cli.rs include_str!)
-        ../.github/TEAM_MEMBERS # required by @opencode-ai/script
+        ../.github/TEAM_MEMBERS # required by @opencode/script
       ]
     );
   };
@@ -52,7 +53,7 @@ stdenvNoCC.mkDerivation {
       --cpu="${bunCpu}" \
       --os="${bunOs}" \
       --filter '!./' \
-      --filter './packages/opencode' \
+      --filter './packages/cli' \
       --filter './packages/desktop' \
       --filter './packages/app' \
       --frozen-lockfile \

nix/opencode.nix

@@ -8,6 +8,7 @@
   makeBinaryWrapper,
   models-dev,
   ripgrep,
+  wayland,
   installShellFiles,
   versionCheckHook,
   writableTmpDirAsHomeHook,
@@ -48,13 +49,13 @@ stdenvNoCC.mkDerivation (finalAttrs: {
   env.OPENCODE_DISABLE_MODELS_FETCH = true;
   env.OPENCODE_VERSION = finalAttrs.version;
   env.OPENCODE_CHANNEL = "prod";
+  env.NODE_OPTIONS = "--max-old-space-size=4096";
 
   buildPhase = ''
     runHook preBuild
 
-    cd ./packages/opencode
+    cd ./packages/cli
     bun --bun ./script/build.ts --single --skip-install
-    bun --bun ./script/schema.ts schema.json
 
     runHook postBuild
   '';
@@ -62,9 +63,9 @@ stdenvNoCC.mkDerivation (finalAttrs: {
   installPhase = ''
     runHook preInstall
 
-    install -Dm755 dist/opencode-*/bin/opencode $out/bin/opencode
-    install -Dm644 schema.json $out/share/opencode/schema.json
+    install -Dm755 dist/cli-*/bin/opencode $out/bin/opencode
 
+    # OpenTUI dlopens Wayland for clipboard images.
     wrapProgram $out/bin/opencode \
       --prefix PATH : ${
         lib.makeBinPath (
@@ -74,7 +75,11 @@ stdenvNoCC.mkDerivation (finalAttrs: {
           # bun runs sysctl to detect if running on rosetta2
           ++ lib.optional stdenvNoCC.hostPlatform.isDarwin sysctl
         )
-      }
+      } ${lib.optionalString stdenvNoCC.hostPlatform.isLinux ''
+        --prefix LD_LIBRARY_PATH : ${lib.makeLibraryPath [ wayland ]}
+      ''}
+
+    ln -s opencode $out/bin/opencode2
 
     runHook postInstall
   '';
@@ -84,6 +89,10 @@ stdenvNoCC.mkDerivation (finalAttrs: {
     installShellCompletion --cmd opencode \
       --bash <($out/bin/opencode completion) \
       --zsh <(SHELL=/bin/zsh $out/bin/opencode completion)
+
+    installShellCompletion --cmd opencode2 \
+      --bash <($out/bin/opencode2 completion) \
+      --zsh <(SHELL=/bin/zsh $out/bin/opencode2 completion)
   '';
 
   nativeInstallCheckInputs = [
@@ -95,7 +104,6 @@ stdenvNoCC.mkDerivation (finalAttrs: {
   versionCheckProgramArg = "--version";
 
   passthru = {
-    jsonschema = "${placeholder "out"}/share/opencode/schema.json";
     env = finalAttrs.env;
   };
 

package.json

@@ -2,21 +2,29 @@
   "$schema": "https://json.schemastore.org/package.json",
   "name": "opencode",
   "description": "AI-powered development tool",
-  "version": "0.0.0",
+  "version": "2.0.3",
   "private": true,
   "type": "module",
-  "packageManager": "bun@1.3.14",
+  "packageManager": "bun@1.4.2",
   "scripts": {
-    "dev": "bun run --cwd packages/cli --conditions=browser src/index.ts",
+    "dev": "bun run --cwd packages/cli src/index.ts",
+    "dev:live": "sh -c 'OPENCODE_TUI_CHANNEL=dev OPENCODE_PASSWORD=\"$(opencode service get password)\" exec bun run dev \"$@\" --server \"$(opencode service status)\"' --",
+    "dev:vite": "bun run --cwd packages/cli --conditions=browser dev/vite.ts",
+    "dev:vite:live": "sh -c 'OPENCODE_TUI_CHANNEL=dev OPENCODE_PASSWORD=\"$(opencode service get password)\" exec bun run dev:vite \"$@\" --server \"$(opencode service status)\"' --",
     "dev:desktop": "bun --cwd packages/desktop dev",
     "dev:web": "bun --cwd packages/app dev",
     "dev:console": "ulimit -n 10240 2>/dev/null; bun run --cwd packages/console/app dev",
     "dev:stats": "bun sst shell --stage=production -- bun run --cwd packages/stats/app dev",
+    "dev:www": "bun run --cwd services/www dev",
     "dev:storybook": "bun --cwd packages/storybook storybook",
+    "bench:devex": "bun run --cwd packages/app test:bench:devex",
     "lint": "oxlint",
-    "lint:effect-patterns": "ast-grep scan -c script/ast-grep/sgconfig.yml packages/core/src packages/server/src packages/protocol/src packages/cli/src",
+    "lint:effect-patterns": "ast-grep scan -c script/ast-grep/sgconfig.yml packages/util/src packages/core/src packages/server/src packages/protocol/src packages/cli/src",
+    "lint:effect-simplifications": "ast-grep scan -c script/ast-grep/effect-simplifications/sgconfig.yml --off=unused-suppression packages",
     "test:lint-rules": "ast-grep test -c script/ast-grep/sgconfig.yml",
+    "test:effect-simplification-rules": "ast-grep test -c script/ast-grep/effect-simplifications/sgconfig.yml",
     "typecheck": "bun turbo typecheck --concurrency=3",
+    "check": "bun run lint && bun run typecheck",
     "typecheck:profile": "bun script/profile-typecheck.ts",
     "typecheck:profile:packages": "bun script/profile-typecheck-packages.ts",
     "upgrade-opentui": "bun run script/upgrade-opentui.ts",
@@ -24,34 +32,34 @@
     "prepare": "husky",
     "random": "echo 'Random script'",
     "sso": "aws sso login --sso-session=opencode --no-browser",
-    "translate:app": "bun run script/translate-app.ts",
     "test": "echo 'do not run tests from root' && exit 1"
   },
   "workspaces": {
     "packages": [
       "packages/*",
+      "services/*",
       "packages/console/*",
-      "packages/stats/*",
-      "packages/sdk/js",
-      "packages/slack"
+      "packages/stats/*"
     ],
     "catalog": {
-      "@effect/opentelemetry": "4.0.0-beta.83",
-      "@effect/platform-node": "4.0.0-beta.83",
-      "@effect/sql-sqlite-bun": "4.0.0-beta.83",
+      "@effect/opentelemetry": "4.0.0-rc.112",
+      "@effect/platform-node": "4.0.0-rc.112",
+      "@effect/platform-node-shared": "4.0.0-rc.112",
+      "@effect/sql-sqlite-bun": "4.0.0-rc.112",
       "@npmcli/arborist": "9.4.0",
-      "@types/bun": "1.3.13",
+      "@types/bun": "1.4.0",
       "@types/cross-spawn": "6.0.6",
       "@octokit/rest": "22.0.0",
       "@hono/standard-validator": "0.2.0",
       "@hono/zod-validator": "0.4.2",
-      "@opentui/core": "0.4.3",
-      "@opentui/keymap": "0.4.3",
-      "@opentui/solid": "0.4.3",
-      "@tanstack/solid-virtual": "3.13.28",
-      "@shikijs/stream": "4.2.0",
+      "@opentui/core": "0.5.10",
+      "@opentui/keymap": "0.5.10",
+      "@opentui/solid": "0.5.10",
+      "@tanstack/solid-virtual": "3.13.37",
+      "@shikijs/stream": "4.4.3",
+      "@standard-schema/spec": "1.1.0",
       "ulid": "3.0.1",
-      "@kobalte/core": "0.13.11",
+      "@kobalte/core": "0.13.13",
       "@corvu/drawer": "0.2.4",
       "@types/luxon": "3.7.1",
       "@types/node": "24.12.2",
@@ -62,22 +70,26 @@
       "@openauthjs/openauth": "0.0.0-20250322224806",
       "@pierre/diffs": "1.2.10",
       "opentui-spinner": "0.0.7",
-      "@solid-primitives/storage": "4.3.3",
+      "@solid-primitives/event-listener": "2.4.6",
+      "@solid-primitives/media": "2.3.6",
+      "@solid-primitives/resize-observer": "2.2.0",
+      "@solid-primitives/storage": "4.4.0",
       "@tailwindcss/vite": "4.1.11",
-      "diff": "8.0.2",
-      "dompurify": "3.3.1",
-      "drizzle-kit": "1.0.0-rc.2",
-      "drizzle-orm": "1.0.0-rc.2",
-      "effect": "4.0.0-beta.83",
+      "diff": "8.0.4",
+      "dompurify": "3.4.14",
+      "drizzle-kit": "1.0.0-rc.5-ab785fc",
+      "drizzle-orm": "1.0.0-rc.5-169397b",
+      "effect": "4.0.0-rc.112",
       "ai": "6.0.168",
       "cross-spawn": "7.0.6",
       "hono": "4.10.7",
       "hono-openapi": "1.1.2",
       "fuzzysort": "3.1.0",
+      "get-east-asian-width": "1.6.0",
       "luxon": "3.6.1",
-      "marked": "17.0.1",
+      "marked": "18.0.11",
       "marked-shiki": "1.2.1",
-      "remend": "1.3.0",
+      "remend": "1.3.1",
       "@playwright/test": "1.59.1",
       "semver": "7.7.4",
       "typescript": "5.8.2",
@@ -85,16 +97,18 @@
       "zod": "4.1.8",
       "remeda": "2.26.0",
       "sst": "4.13.1",
-      "shiki": "4.2.0",
+      "shiki": "4.4.3",
       "solid-list": "0.3.0",
+      "string-width": "7.2.0",
       "tailwindcss": "4.1.11",
-      "vite": "7.1.4",
+      "vite": "7.3.6",
       "@solidjs/meta": "0.29.4",
-      "@solidjs/router": "0.15.4",
+      "@solidjs/router": "1.0.0",
       "@solidjs/start": "https://pkg.pr.new/@solidjs/start@dfb2020",
-      "@sentry/solid": "10.36.0",
-      "@sentry/vite-plugin": "4.6.0",
-      "solid-js": "1.9.10",
+      "@sentry/solid": "10.71.0",
+      "@sentry/vite-plugin": "5.4.0",
+      "solid-js": "1.9.15",
+      "solid-sonner": "0.3.1",
       "vite-plugin-solid": "2.11.10",
       "@lydell/node-pty": "1.2.0-beta.12"
     }
@@ -102,6 +116,9 @@
   "devDependencies": {
     "@actions/artifact": "5.0.1",
     "@ast-grep/cli": "0.44.0",
+    "@opencode/client": "workspace:*",
+    "@types/react": "19.2.17",
+    "@types/react-dom": "19.2.3",
     "@tsconfig/bun": "catalog:",
     "@types/mime-types": "3.0.1",
     "@typescript/native-preview": "catalog:",
@@ -116,9 +133,8 @@
   },
   "dependencies": {
     "@aws-sdk/client-s3": "3.933.0",
-    "@opencode-ai/plugin": "workspace:*",
-    "@opencode-ai/script": "workspace:*",
-    "@opencode-ai/sdk": "workspace:*",
+    "@opencode/plugin": "workspace:*",
+    "@opencode/script": "workspace:*",
     "heap-snapshot-toolkit": "1.1.3",
     "typescript": "catalog:"
   },
@@ -135,33 +151,35 @@
     "esbuild",
     "node-pty",
     "protobufjs",
-    "tree-sitter",
-    "tree-sitter-bash",
-    "tree-sitter-powershell",
-    "web-tree-sitter",
     "electron"
   ],
   "overrides": {
     "@opentui/core": "catalog:",
     "@opentui/keymap": "catalog:",
     "@opentui/solid": "catalog:",
+    "@effect/platform-node-shared": "catalog:",
     "@types/bun": "catalog:",
-    "@types/node": "catalog:"
+    "@types/node": "catalog:",
+    "bun-types": "1.4.2",
+    "effect": "catalog:",
+    "solid-js": "catalog:"
   },
   "patchedDependencies": {
-    "@ff-labs/fff-bun@0.9.3": "patches/@ff-labs%2Ffff-bun@0.9.3.patch",
+    "@ai-sdk/openai-compatible@2.0.41": "patches/@ai-sdk%2Fopenai-compatible@2.0.41.patch",
+    "@dnd-kit/dom@0.5.0": "patches/@dnd-kit%2Fdom@0.5.0.patch",
     "@npmcli/agent@4.0.2": "patches/@npmcli%2Fagent@4.0.2.patch",
     "@silvia-odwyer/photon-node@0.3.4": "patches/@silvia-odwyer%2Fphoton-node@0.3.4.patch",
     "@standard-community/standard-openapi@0.2.9": "patches/@standard-community%2Fstandard-openapi@0.2.9.patch",
-    "solid-js@1.9.10": "patches/solid-js@1.9.10.patch",
-    "@ai-sdk/xai@3.0.102": "patches/@ai-sdk%2Fxai@3.0.102.patch",
+    "solid-js@1.9.15": "patches/solid-js@1.9.15.patch",
+    "@ai-sdk/mistral@3.0.51": "patches/@ai-sdk%2Fmistral@3.0.51.patch",
     "gcp-metadata@8.1.2": "patches/gcp-metadata@8.1.2.patch",
     "pacote@21.5.0": "patches/pacote@21.5.0.patch",
     "@ai-sdk/google@3.0.73": "patches/@ai-sdk%2Fgoogle@3.0.73.patch",
-    "@tanstack/solid-virtual@3.13.28": "patches/@tanstack%2Fsolid-virtual@3.13.28.patch",
     "@pierre/trees@1.0.0-beta.4": "patches/@pierre%2Ftrees@1.0.0-beta.4.patch",
-    "@modelcontextprotocol/sdk@1.29.0": "patches/@modelcontextprotocol%2Fsdk@1.29.0.patch",
-    "@tanstack/virtual-core@3.17.0": "patches/@tanstack%2Fvirtual-core@3.17.0.patch",
-    "effect@4.0.0-beta.83": "patches/effect@4.0.0-beta.83.patch"
+    "@tanstack/virtual-core@3.17.8": "patches/@tanstack%2Fvirtual-core@3.17.8.patch",
+    "@ff-labs/fff-bun@0.10.5": "patches/@ff-labs%2Ffff-bun@0.10.5.patch",
+    "ghostty-web@github:anomalyco/ghostty-web#83c0a07": "patches/ghostty-web@0.3.0.patch",
+    "vite@8.2.2": "patches/vite@8.2.2.patch",
+    "@modelcontextprotocol/client@2.0.0": "patches/@modelcontextprotocol%2Fclient@2.0.0.patch"
   }
 }

packages/ai/AGENTS.md

@@ -0,0 +1,317 @@
+# AI Package Guide
+
+## Effect
+
+- Prefer `HttpClient.HttpClient` / `HttpClientResponse.HttpClientResponse` over web `fetch` / `Response` at package boundaries.
+- Use `Stream.Stream` for streaming data flow. Avoid ad hoc async generators or manual web reader loops unless an Effect `Stream` API cannot model the behavior.
+- Use Effect Schema codecs for JSON encode/decode (`Schema.fromJsonString(...)`) instead of direct `JSON.parse` / `JSON.stringify` in implementation code.
+- In `Effect.gen`, yield yieldable errors directly (`return yield* new MyError(...)`) instead of `Effect.fail(new MyError(...))`.
+- Use `Effect.void` instead of `Effect.succeed(undefined)` when the successful value is intentionally void.
+
+## Conventions
+
+Per-type constructors live on the type, not as top-level re-exports. Use `Message.system(...)`, `Message.user(...)`, `Message.assistant(...)`, `Message.tool(...)`, `Message.media(...)`, `LanguageModel.make(...)`, `ToolDefinition.make(...)`, `ToolCallPart.make(...)`, `ToolResultPart.make(...)`, `ToolChoice.make(...)`, `ToolChoice.named(...)`, `SystemPart.make(...)`, and `GenerationOptions.make(...)` directly. The top-level `LLM` namespace is reserved for request-shaped call APIs: `LLM.request`, `LLM.generate`, `LLM.stream`, and `LLM.generateObject`. Use `LLMRequest.update(...)` when deriving canonical request data; do not add a duplicate `LLM.updateRequest(...)` path. Two ways to construct the same thing is one too many.
+
+Modality namespaces mirror `LLM` exactly: `Image.request`, `Image.generate`, `Image.stream` (later `Video`, `Speech`, `Transcription`). Common request fields (`images`, `mask`, `n`, `size`, `aspectRatio`, `seed`, `format`) lower natively or fail with a typed `AIError`; provider-native controls always live under `providerOptions`, never under a modality-specific `options` key.
+
+Media payloads are always `Media.Asset` (`src/media.ts`). Construct them with `Media.bytes`, `Media.base64`, `Media.url`, `Media.ref`, `Media.fromDataUrl`, or `Media.file`; never introduce a parallel `data: string | Uint8Array` shape. `MediaPart.media`, `ImageRequest.images`/`mask`, `ImageResponse.images`, and the `media` `LLMEvent` all share it. Protocols branch on `asset.source.type` and `asset.kind` and use `ProviderShared.inlineMedia` / `requireInlineMedia` / `mediaUrl` / `MediaInput.refID` rather than re-deriving base64 or URL handling.
+
+`schema/messages.ts → media.ts → route/executor-service.ts` is an accepted runtime dependency from the schema layer on the executor service tag: `Media.Asset.bytes()` must be able to download `url` sources, and the tag lives in that leaf module precisely so the schema barrel never imports the executor implementation (which imports the schema barrel back). Do not move the tag into `route/executor.ts` or import `route/executor.ts` from `src/schema/*` or `src/media.ts`.
+
+Nothing in `src/*` except `src/promise.ts` may know about Promises. `@opencode/ai/promise` (`AI.make({ layer? })`, default `ai`) is the single Promise/`AsyncIterable` surface for LLM and media; it runs the Effect APIs in one `ManagedRuntime` and rethrows `AIError` unchanged.
+
+- Keep provider-defined string enums forward-compatible. Expose known values for autocomplete while accepting future values with `Known | (string & {})`; use `Schema.String` at runtime unless rejecting unknown values is required for correctness.
+- Order reasoning-effort values from lowest to highest: `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`. Provider-specific subsets follow the same relative order in types, schemas, option lists, and tests.
+
+## Tests
+
+- Use `testEffect(...)` from `test/lib/effect.ts` for tests requiring Effect layers.
+- Keep provider tests fixture-first. Live provider calls must stay behind `RECORD=true` and required API-key checks.
+
+## Errors
+
+- `AIError` wraps a union of tagged reason errors. It stores only `reason`, derives `message` from the reason, and exposes the reason as its `cause`.
+- Each reason owns its readable `message`, category-specific fields, and optional `body`, `http`, and underlying exception in `cause`.
+- `reason.body` is the sole original-response or triggering-event payload field. Preserve original text before schema decoding removes fields; do not replace the complete event with only its nested error.
+- `reason.http` describes an observed HTTP response with required `url`, `status`, and response `headers`. Do not invent status codes or derive a separate request ID from headers.
+- Reclassification and transport recovery must preserve the reason's body, HTTP context, and underlying cause. Error `message` and `cause` are non-enumerable: copy them explicitly when constructing an enriched reason with its constructor or `AIErrorReason.make`.
+
+## Architecture
+
+This package is an Effect Schema-first LLM core. The Schema classes in `src/schema/` are the canonical runtime data model. Convenience functions in `src/llm.ts` are thin constructors that return those same Schema class instances; they should improve callsites without creating a second model.
+
+Session integration lives in `packages/core/src/session`: `runner/llm.ts` owns orchestration, `model-request.ts` lowers Session state into `LLMRequest`, and `model-transport.ts` selects transport behavior.
+
+Keep this package independent of Session concerns. Session auth, permissions, plugins, telemetry headers, and runtime selection belong in Core.
+
+### Request Flow
+
+The intended callsite is:
+
+```ts
+const request = LLM.request({
+  model: OpenAI.configure({ apiKey }).responses("gpt-4o-mini"),
+  system: "You are concise.",
+  prompt: "Say hello.",
+})
+
+const response = yield * LLMClient.generate(request)
+```
+
+`LLM.request(...)` builds an `LLMRequest`. `LLMClient.generate(...)` reads the executable route carried by `request.model.route`, builds the provider-native body, asks the route's transport for a real `HttpClientRequest.HttpClientRequest`, sends it through `RequestExecutor.Service`, parses the provider stream into common `LLMEvent`s, and finally returns an `LLMResponse`.
+
+Use `LLMClient.stream(request)` when callers want incremental `LLMEvent`s. Use `LLMClient.generate(request)` when callers want those same events collected into an `LLMResponse`.
+
+Filter or narrow `LLMEvent` streams with `LLMEvent.is.*` (camelCase guards, e.g. `events.filter(LLMEvent.is.toolCall)`). The kebab-case `LLMEvent.guards["tool-call"]` form also works but prefer `is.*` in new code.
+
+### Routes
+
+A route is the runnable composition of four orthogonal pieces:
+
+- **`Protocol`** (`src/route/protocol.ts`) — semantic API contract. Owns request body construction (`body.from`), the body schema (`body.schema`), the streaming-event schema (`stream.event`), and the event-to-`LLMEvent` state machine (`stream.step`). `Route.make(...)` validates and JSON-encodes the body from `body.schema` and decodes frames with `stream.event`. Examples: `OpenAIChat.protocol`, `OpenResponses.protocol`, `OpenAIResponses.protocol`, `AnthropicMessages.protocol`, `Gemini.protocol`, `BedrockConverse.protocol`.
+- **`Endpoint`** (`src/route/endpoint.ts`) — URL construction. The host, path, and route query live on the endpoint. `Endpoint.path("/chat/completions", { baseURL })` is the common case; pass a function for paths that embed the model id or a body field (e.g. `Endpoint.path(({ body }) => `/model/${body.modelId}/converse-stream`)`).
+- **`Auth`** (`src/route/auth.ts`) — per-request transport authentication. Provider facades configure credentials onto the route before model selection, usually via `Auth.bearer(apiKey)` or `Auth.header(name, apiKey)`. Routes that need per-request signing (Bedrock SigV4, future Vertex IAM, Azure AAD) implement `Auth` as a function that signs the body and merges signed headers into the result.
+- **`Framing`** (`src/route/framing.ts`) — bytes → frames. SSE (`Framing.sse`) is shared; Bedrock keeps its AWS event-stream framing as a typed `Framing<object>` value alongside its protocol.
+
+Compose them via `Route.make(...)`:
+
+```ts
+export const route = Route.make({
+  id: "openai-chat",
+  provider: "openai",
+  protocol: OpenAIChat.protocol,
+  endpoint: Endpoint.path("/chat/completions", {
+    baseURL: "https://api.openai.com/v1",
+  }),
+  auth: Auth.bearer(Auth.config("OPENAI_API_KEY")),
+  framing: Framing.sse,
+})
+```
+
+Route defaults are request-shaping defaults such as `headers`, `limits`, `generation`, `providerOptions`, and `http`. Endpoint host/query belongs on the route endpoint. Selected `LanguageModel` values carry only model id, provider id, and the configured route value. Model capability/catalog metadata lives outside this package; protocol support is enforced by request lowering and typed `AIError`s.
+
+The four-axis decomposition is the reason DeepSeek, TogetherAI, Cerebras, Baseten, Fireworks, and DeepInfra all reuse `OpenAIChat.protocol` verbatim — each provider owns a small `Route.make(...)` composition instead of a protocol clone. Bug fixes in one protocol propagate to every consumer of that protocol in a single commit.
+
+When a provider supports multiple physical transports, selection remains execution policy below its semantic route. `OpenResponsesChannel.transport(...)` owns the provider-neutral Responses WebSocket concept: it prepares one final request, executes HTTP by default, strips WebSocket-disallowed fields, and passes a generic channel exchange to a per-call `WebSocketChannelExecutor` when supplied. Provider-specific Responses routes opt in with handshake and connection-age policy. `Route.streamPrepared` owns decoding and acknowledges channel completion only after successful full consumption.
+
+### Media Routes
+
+Media does not fit the SSE-frames-to-event-state-machine LLM route. `MediaRoute.make(...)` (`src/route/media.ts`) composes a `MediaProtocol` kind with `Endpoint` and `Auth` and owns the transport plumbing: `http` option merging, URL/query rendering, auth headers, JSON vs multipart encoding, and handing the response back to the protocol. `MediaProtocol.inline` (`src/route/media-protocol.ts`) is `body.from(request)` plus `response.decode(response, context)`; use `MediaProtocol.decodeJson` / `text` / `bytes` so decode failures retain the raw body and HTTP context. `MediaProtocol.JobProtocol` is the queued shape and `Job` (`src/job.ts`) is the provider-neutral handle over a `JobRoute` (`status`, `result`, `cancel`, `pollHint`); the first video route implements it. Image protocol files follow the same section order as LLM protocols and declare unsupported common fields once through `MediaInput.rejectUnsupported`.
+
+### URL Construction
+
+`Endpoint` owns `{ baseURL, path, query }`. Each protocol route includes a canonical endpoint when the provider has one (e.g. `https://api.openai.com/v1`); provider helpers override endpoint fields by configuring the route before selecting a model. Generic OpenAI-compatible routes have no canonical URL and require configuration before execution.
+
+For providers where the URL is derived from typed inputs (Azure resource name, Bedrock region), the provider helper configures the route endpoint before calling `.model(...)`. Use `AtLeastOne<T>` from `route/auth-options.ts` for inputs that accept either of two derivation paths (Azure: `resourceName` or `baseURL`).
+
+### Provider Facades
+
+Provider-facing APIs are configured facades over route values. Endpoint/auth/resource/API-version setup happens before model selection, and model selectors accept only a model or deployment id. Every facade is wrapped in `ModelRef.facade(...)`, which makes it callable: `openai("gpt-4o-mini")` returns a `ModelRef` with lazy per-modality routes derived from the facade's own selectors (`model` → `llm`, `image` → `image`). `LLM.request` and `Image.request` resolve their route from the ref at request time, so one value names a model for every modality while named selectors stay the single source of truth:
+
+```ts
+const openai = OpenAI.configure({ apiKey, baseURL })
+const ref = openai("gpt-4o-mini") // ModelRef: LLM.request → .responses, Image.request → .image
+const model = openai.responses("gpt-4o-mini")
+
+const azure = Azure.configure({ resourceName, apiKey, apiVersion: "v1" })
+const deployment = azure.responses("my-deployment")
+
+const gateway = CloudflareAIGateway.configure({ accountId, gatewayId, gatewayApiKey, apiKey })
+const proxied = gateway.model("openai/gpt-4o-mini")
+```
+
+Keep provider facades small and explicit:
+
+- Use branded `ProviderID.make(...)` and `ModelID.make(...)` where ids are constructed directly.
+- Use `model` for the default API path and named methods for provider-native alternatives such as OpenAI `responses` and `chat`.
+- Put provider-specific setup on `.configure(...)`; do not add `model(id, overrides)` as a duplicate construction path.
+- Export lower-level `routes` arrays separately only when advanced internal wiring needs them.
+- Prefer `apiKey` as provider-specific sugar and `auth` as the explicit override; keep them mutually exclusive in provider option types with `ProviderAuthOption`.
+- Resolve `apiKey` → `Auth` with `AuthOptions.bearer(options, "<PROVIDER>_API_KEY")` (it honors an explicit `auth` override and falls back to `Auth.config(envVar)` so missing keys surface a typed `Authentication` error rather than a runtime crash).
+- Use separate top-level facades for products with different required setup, such as `CloudflareAIGateway` and `CloudflareWorkersAI`.
+- Give every named provider its own file and top-level export. Keep its endpoint, auth defaults, and route setup in that file. Compose shared protocols directly; do not nest named provider presets under generic compatible facades or keep their endpoints in a shared provider profile registry.
+
+`Provider.make(...)` remains available for simple static provider definitions, but new built-in providers should prefer plain configured facades unless a helper removes real duplication without adding runtime behavior.
+
+### Provider Package Entrypoints
+
+Catalog-selected native providers use package-like export paths from `@opencode/ai`. They are internal entrypoints in one npm package, not separately published provider packages. Every entrypoint implements `ProviderPackage.Definition` and exposes `model(modelID, settings)`, where settings are one flat serializable object: the connection keys the entrypoint declares (`apiKey`, `baseURL`, `region`, …), the common `headers` and `body` overlays, and the protocol's request options (`reasoningEffort`, `thinking`, …) side by side. Each entrypoint destructures its own connection keys and passes the rest to the route as `providerOptions`; there is no nested `providerOptions` at the entrypoint.
+
+```ts
+import { model } from "@opencode/ai/providers/openai/responses"
+
+const selected = model("gpt-5", {
+  apiKey,
+  reasoningEffort: "high",
+})
+```
+
+Keep semantic APIs as separate entrypoints, such as OpenAI `chat` and `responses`. Transport is execution policy: OpenAI Responses uses HTTP by default and may receive a per-call WebSocket channel executor through `StreamOptions` without changing model or route identity.
+
+Do not expose `Route` in provider package settings. Route composition stays an implementation detail behind `model(...)`.
+
+The dependency arrow points down: `providers/*.ts` files import protocol routes and auth-option utilities; protocol modules import `endpoint`, `auth`, `framing`, and transport pieces. Protocols do not import provider facades. Lower-level modules know nothing about provider catalog metadata. `OpenAIResponses` composes the provider-neutral `OpenResponses` protocol; the baseline never imports the OpenAI extension.
+
+### Shared protocol helpers
+
+`ProviderShared` exports a small toolkit used inside protocol implementations to keep them focused on provider-native shapes:
+
+- `joinText(parts)` — joins an array of `TextPart` (or anything with a `.text`) with newlines. Use this anywhere a protocol flattens text content into a single string for a provider field.
+- `parseToolInput(route, name, raw)` — Schema-decodes a tool-call argument string with the canonical "Invalid JSON input for `<route>` tool call `<name>`" error message. Treats empty input as `{}`.
+- `parseJson(route, raw, message)` — generic JSON-via-Schema decode for non-tool bodies.
+- `eventError(route, message, ...)` — typed `InvalidProviderOutput` constructor for stream-time decode failures.
+- `validateWith(decoder)` — maps Schema decode errors to `InvalidRequest`. `Route.make(...)` uses this for body validation; lower-level routes can reuse it.
+- `matchToolChoice(provider, choice, branches)` — branches over `LLMRequest["toolChoice"]` for provider-specific lowering.
+
+If you find yourself copying a 3-to-5-line snippet between two protocols, lift it into `ProviderShared` next to these helpers rather than duplicating.
+
+### Chronological System Updates
+
+`LLMRequest.system` is the initial privileged prompt that applies ahead of the conversation. `Message.system(...)` is a separate, provider-neutral chronological operator update inside `LLMRequest.messages`; it applies only from its position in history onward and accepts text content only.
+
+Native chronological system messages are route/model-specific. Open Responses lowers them to standard `developer` messages, while Anthropic Messages lowers them to native system messages for Claude Opus 4.8 (`claude-opus-4-8`). Other routes and models intentionally lower the update in place into ordinary user-compatible text using this stable escaped representation:
+
+```text
+<system-update>
+...
+</system-update>
+```
+
+The wrapped-user fallback preserves ordering while visibly lowering authority. Never silently pass a raw chronological `role: "system"` through a route that might reject it. Do not insert raw retrieved documents, tool output, or web content into privileged chronological system updates; keep untrusted content in ordinary user/tool channels.
+
+### Effort Updates
+
+`Message.effort({ effort, previous })` is a chronological "reasoning effort changed here" marker (`undefined` means the model default). Changing a top-level effort invalidates the whole provider prompt cache, so protocols with a native per-message update (`Protocol.supportsEffortUpdates`) keep the top-level effort at the first marker's `previous` and lower each marker in place: Anthropic Messages emits an empty `role: "system"` message with `output_config.effort` plus the `mid-conversation-output-config-2026-07-01` beta, and OpenAI Responses emits `configuration_update` items. `applyEffortUpdates` runs in `prepareRequest` and strips the markers for every other route, so a protocol without support keeps today's plain top-level behaviour. When the last marker disagrees with the effort the request asks for (reverted or forked history), `resolveEffortUpdates` strips the markers and falls back to a plain top-level change.
+
+### Tools
+
+Tool loops are represented in common messages and events:
+
+```ts
+const call = ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } })
+const result = Message.tool({ id: "call_1", name: "lookup", result: { forecast: "sunny" } })
+
+const followUp = LLM.request({
+  model,
+  messages: [Message.user("Weather?"), Message.assistant([call]), result],
+})
+```
+
+Routes lower these into provider-native assistant tool-call messages and tool-result messages. Streaming providers should emit `tool-input-delta` events while arguments arrive, then a final `tool-call` event with parsed input.
+
+### Tool dispatch
+
+`LLM.stream(request)` and `LLM.generate(request)` each run exactly one model call. Add tool schemas to `request.tools` with `Tool.toDefinitions(tools)`. When a caller wants the package's typed one-call execution behavior, pass each canonical local `tool-call` event to `ToolRuntime.dispatch(tools, call)`.
+
+```ts
+const get_weather = Tool.make({
+  description: "Get current weather for a city",
+  parameters: Schema.Struct({ city: Schema.String }),
+  success: Schema.Struct({ temperature: Schema.Number, condition: Schema.String }),
+  execute: (input) =>
+    Effect.gen(function* () {
+      const data = yield* WeatherApi.fetch(input.city)
+      return { temperature: data.temp, condition: data.cond }
+    }),
+})
+
+const tools = { get_weather, get_time, ... }
+const events = yield* LLM.stream(
+  LLMRequest.update(request, { tools: Tool.toDefinitions(tools) }),
+).pipe(Stream.runCollect)
+
+const call = Array.from(events).find(LLMEvent.is.toolCall)
+if (call && !call.providerExecuted) {
+  const dispatched = yield* ToolRuntime.dispatch(tools, call)
+  // Persist call + dispatched.result, then construct the next request explicitly.
+}
+```
+
+The dispatcher:
+
+- On `tool-call`: looks up the named tool, decodes input against `parameters` Schema, dispatches to the typed `execute`, encodes the result against `success` Schema, and returns canonical `tool-result` events.
+- Does not stream providers, construct Session events, schedule fibers, append history, count steps, or continue model rounds.
+- Leaves persistence and continuation to the enclosing product flow.
+
+Handler dependencies (services, permissions, plugin hooks, abort handling) are closed over by the consumer at tool-construction time. Build the tools record inside an `Effect.gen` once and reuse it across many dispatches.
+
+Errors must be expressed as `ToolFailure`. The runtime catches it and emits a `tool-error` event, then a `tool-result` of `type: "error"`, so the model can self-correct on the next step. Anything that is not a `ToolFailure` is treated as a defect and fails the stream. Three recoverable error paths produce `tool-error` events:
+
+- The model called an unknown tool name.
+- Input failed the `parameters` Schema.
+- The handler returned a `ToolFailure`.
+
+Provider-defined / hosted tools (Anthropic `web_search` / `code_execution` / `web_fetch`, OpenAI Responses `web_search_call` / `file_search_call` / `code_interpreter_call` / `mcp_call` / `image_generation_call` / `computer_use_call`) pass through the runtime untouched:
+
+- Routes surface the model's call as a `tool-call` event with `providerExecuted: true`, and the provider's result as a matching `tool-result` event with `providerExecuted: true`.
+- Callers detect `providerExecuted` on `tool-call` and **skip local dispatch** — no handler is invoked and no `tool-error` is raised for "unknown tool". The provider already executed it.
+- Callers that continue should retain both events in explicit history when the protocol requires it. Anthropic encodes them back as `server_tool_use` + `web_search_tool_result` (or `code_execution_tool_result` / `web_fetch_tool_result`) blocks; OpenAI Responses callers typically use `previous_response_id` instead of resending hosted-tool items.
+
+Add provider-defined tools to `request.tools` (no runtime entry needed). The matching route must know how to lower the tool definition into the provider-native shape; right now Anthropic accepts `web_search` / `code_execution` / `web_fetch` and OpenAI Responses accepts the hosted tool names listed above.
+
+## Protocol File Style
+
+Protocol files should look self-similar. Provider quirks belong behind named helpers so a new route can be reviewed by comparing the same sections across files.
+
+### Section order
+
+Use this order for every protocol module:
+
+1. Public model input
+2. Request body schema
+3. Streaming event schema
+4. Parser state
+5. Request body construction (`fromRequest`)
+6. Stream parsing (`step` and per-event handlers)
+7. Protocol and route
+8. Protocol route export
+
+### Rules
+
+- Keep protocol files focused on the protocol. Move provider-specific projection, signing, media normalization, or other bulky transformations into `src/protocols/utils/*`.
+- Use `Effect.fn("Provider.fromRequest")` for request body construction entrypoints. Use `Effect.fn(...)` for event handlers that yield effects; keep purely synchronous handlers as plain functions returning a `StepResult` that the dispatcher lifts via `Effect.succeed(...)`.
+- Parser state owns terminal information. The state machine records finish reason, usage, and pending tool calls; emit one terminal `finish` event (or `provider-error`) for each completed response. If a provider splits reason and usage across events, merge them in parser state before flushing.
+- Emit exactly one terminal `finish` event for a completed response, normally after a matching `step-finish`. Use `stream.terminal` to stop reading when the provider has a completion sentinel; use `stream.onHalt` when the final event must be flushed after the framed stream ends.
+- Use shared helpers for repeated protocol policy such as text joining, usage totals, JSON parsing, and tool-call accumulation. `ToolStream` (`protocols/utils/tool-stream.ts`) accumulates streamed tool-call arguments uniformly.
+- Make intentional provider differences explicit in helper names or comments. If two protocol files differ visually, the reason should be obvious from the names.
+- Prefer dispatched per-event handlers (`onMessageStart`, `onContentBlockDelta`, ...) called from a small top-level `step` switch over a long if-chain. The dispatcher keeps the event surface visible at a glance.
+- Keep tests in the same conceptual order as the protocol: basic prepare, tools prepare, unsupported lowering, text/usage parsing, tool streaming, finish reasons, provider errors.
+
+### Review checklist
+
+- Can the file be skimmed side-by-side with `openai-chat.ts` without hunting for equivalent sections?
+- Are provider quirks named, isolated, and covered by focused tests?
+- Does request body construction validate unsupported common content at the protocol boundary?
+- Does stream parsing emit stable common events without leaking provider event order to callers?
+- Does `toolChoice: "none"` behavior read as intentional?
+
+## Recording Tests
+
+Recorded tests use one cassette file per scenario. A cassette holds an ordered array of `{ request, response }` interactions, so multi-step flows (tool loops, retries, polling) record into a single file. Use `recordedTests({ prefix, requires })` and let the helper derive cassette names from test names:
+
+```ts
+const recorded = recordedTests({ prefix: "openai-chat", requires: ["OPENAI_API_KEY"] })
+
+recorded.effect("streams text", () =>
+  Effect.gen(function* () {
+    // test body
+  }),
+)
+```
+
+Replay is the default. `RECORD=true` records fresh cassettes locally and requires the listed env vars; unset `CI` before recording because CI always forces replay. Cassettes are written as pretty-printed JSON so multi-interaction diffs stay reviewable.
+
+Pass `provider`, `protocol`, and optional `tags` to `recordedTests(...)` / `recorded.effect.with(...)` so cassettes carry searchable metadata. Use recorded-test filters to replay or record a narrow subset without rewriting a whole file:
+
+- `RECORDED_PROVIDER=openai` matches tests tagged with `provider:openai`; comma-separated values are allowed.
+- `RECORDED_PREFIX=openai-chat` matches cassette groups by `recordedTests({ prefix })`; comma-separated values are allowed.
+- `RECORDED_TAGS=tool` requires all listed tags to be present, e.g. `RECORDED_TAGS=provider:togetherai,tool`.
+- `RECORDED_TEST="streams text"` matches by test name, kebab-case test id, or cassette path.
+
+Filters apply in replay and record mode. Combine them with `RECORD=true` when refreshing only one provider or scenario.
+
+**Binary response bodies.** Most providers stream text (SSE, JSON). The recorder treats known textual media types (`text/*`, JSON/XML structured types, JavaScript, forms, YAML, and SVG) as text and stores every other response as base64 with `bodyEncoding: "base64"`. This preserves binary formats such as AWS event-stream frames without a lossy UTF-8 round trip.
+
+**Matching strategy.** A runtime request atomically claims the first unused recorded interaction that matches its method, URL, allow-listed headers, and canonical JSON body. Distinct requests may replay in any order or concurrently. Repeated identical requests consume their matching responses in cassette order, preserving deterministic retry and polling behavior. `scriptedResponses` (in `test/lib/http.ts`) is the deterministic counterpart for tests that don't need a live provider; it scripts response bodies in order without reading from disk.
+
+Do not blanket re-record an entire test file when adding one cassette. `RECORD=true` rewrites every recorded case that runs, and provider streams contain volatile IDs, timestamps, fingerprints, and obfuscation fields. Prefer deleting the one cassette you intend to refresh, or run a focused test pattern that only registers the scenario you want to record. Keep stable existing cassettes unchanged unless their request shape or expected behavior changed.

packages/ai/README.md

@@ -0,0 +1,932 @@
+# @opencode/ai
+
+Schema-first language model and image-generation APIs built with Effect.
+
+```ts
+import { Effect, Layer } from "effect"
+import { LLM, LLMClient } from "@opencode/ai"
+import { RequestExecutor } from "@opencode/ai/route"
+import { OpenAI } from "@opencode/ai/providers"
+
+const openai = OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY })
+
+const request = LLM.request({
+  model: openai("gpt-4o-mini"), // the request namespace picks the route; `.responses(...)` / `.chat(...)` stay explicit
+  system: "You are concise.",
+  prompt: "Say hello in one short sentence.",
+  generation: { maxTokens: 40 },
+})
+
+const program = Effect.gen(function* () {
+  const response = yield* LLMClient.generate(request)
+  console.log(response.text)
+})
+
+const llmLayer = LLMClient.layer.pipe(Layer.provide(RequestExecutor.fetchLayer))
+
+await Effect.runPromise(program.pipe(Effect.provide(llmLayer)))
+```
+
+Run `LLMClient.stream(request)` instead of `generate` when you want incremental `LLMEvent`s. The event stream is provider-neutral — same shape across OpenAI Chat, OpenAI Responses, Anthropic Messages, Gemini, Bedrock Converse, and any OpenAI-compatible deployment.
+
+The same configured facade names image models. `Image.request` resolves the provider's image route from the ref and
+returns `Media.Asset`s with lazily decoded bytes:
+
+```ts
+import { NodeFileSystem } from "@effect/platform-node"
+import { Image, ImageClient, Media } from "@opencode/ai"
+
+const image = Effect.gen(function* () {
+  const response = yield* Image.generate({
+    model: openai("gpt-image-2"),
+    prompt: "A robot tending a rooftop garden",
+    size: "1024x1024",
+    providerOptions: { quality: "high" }, // typed per image model
+  })
+  yield* Media.write(response.image, "./garden.png")
+})
+
+// `asset.bytes()` / `Media.write` also need the executor, so merge it into the environment instead of hiding it.
+const imageLayer = ImageClient.layer.pipe(Layer.provideMerge(RequestExecutor.fetchLayer))
+
+await Effect.runPromise(image.pipe(Effect.provide(imageLayer), Effect.provide(NodeFileSystem.layer)))
+```
+
+Prefer promises? `@opencode/ai/promise` exposes the same LLM and image APIs over one managed runtime:
+
+```ts
+import { AI } from "@opencode/ai/promise"
+
+const ai = AI.make()
+const text = await ai.llm.generate({ model: openai("gpt-4o-mini"), prompt: "Say hello." })
+const generated = await ai.image.generate({ model: openai("gpt-image-2"), prompt: "A lighthouse" })
+for await (const event of ai.llm.stream({ model: openai("gpt-4o-mini"), prompt: "Stream hello." })) {
+  // LLMEvent
+}
+await ai.dispose()
+```
+
+## Alibaba Cloud Model Studio
+
+`Alibaba` provides standard Model Studio inference. Configure a region explicitly, then select
+Chat Completions (`.model` or `.chat`), Anthropic-compatible Messages (`.messages`), or OpenAI-compatible
+Responses (`.responses`). These routes use HTTP/SSE.
+
+```ts
+import { LLM } from "@opencode/ai"
+import { Alibaba } from "@opencode/ai/providers"
+
+const alibaba = Alibaba.configure({
+  region: "ap-southeast-1", // Singapore
+  apiKey: process.env.DASHSCOPE_API_KEY,
+  // workspaceID: "llm-your-workspace", // use a workspace-dedicated endpoint
+})
+
+const request = LLM.request({
+  model: alibaba.model("qwen3.8-max"),
+  prompt: "Explain this design.",
+  providerOptions: { reasoningEffort: "medium" },
+})
+```
+
+### Regions and credentials
+
+| Region              | `region`         | Shared host when `workspaceID` is omitted |
+| ------------------- | ---------------- | ----------------------------------------- |
+| Singapore           | `ap-southeast-1` | `dashscope-intl.aliyuncs.com`             |
+| China (Beijing)     | `cn-beijing`     | `dashscope.aliyuncs.com`                  |
+| China (Hong Kong)   | `cn-hongkong`    | `cn-hongkong.dashscope.aliyuncs.com`      |
+| US (Virginia)       | `us-east-1`      | `dashscope-us.aliyuncs.com`               |
+| Germany (Frankfurt) | `eu-central-1`   | Supply `workspaceID` or `baseURL`         |
+| Japan (Tokyo)       | `ap-northeast-1` | Supply `workspaceID` or `baseURL`         |
+
+With `workspaceID`, the host is `{workspaceID}.{region}.maas.aliyuncs.com`. A complete `baseURL`
+overrides regional setup, including the API prefix: `/compatible-mode/v1` for Chat/Responses,
+or `/apps/anthropic/v1` for Messages. The selector appends its operation path.
+
+Keys and model availability are region-specific. Auth resolves from explicit `auth` or `apiKey`,
+then `DASHSCOPE_API_KEY`, then `ALIBABA_API_KEY`.
+
+The access region and inference scope differ: Virginia's `-us` model IDs request US-only inference;
+some regions select scope through their workspace. Model IDs pass through unchanged.
+Alibaba's [regional guide](https://www.alibabacloud.com/help/en/model-studio/regions) and
+[base URL table](https://www.alibabacloud.com/help/en/model-studio/base-url) disagree about Virginia's
+shared host; the entry above follows the base URL table. Dedicated hosts can be copied from the console.
+
+### Native options
+
+- **Chat:** `reasoningEffort` → `reasoning_effort`, `enableThinking` → `enable_thinking`,
+  `thinkingBudget` → `thinking_budget`, and `preserveThinking` → `preserve_thinking`.
+  Replay complete `response.message` values to retain `reasoning_content` separately from answer text.
+  Qwen 3.8 defaults to preserving thinking; older models have different defaults.
+  Additional options include `toolStream`, `parallelToolCalls`, `repetitionPenalty`, `responseFormat`,
+  `enableSearch`, and native `searchOptions`. `generation.topK` lowers to `top_k`.
+  `clearThinking` is a hosted GLM control, and `thinking.type` is available for hosted MiniMax models.
+- **Messages:** `effort` → `output_config.effort`. `thinking.type` accepts enabled/disabled with an
+  optional `budgetTokens` (or native `budget_tokens`). `outputConfig.format` accepts a JSON schema.
+  Model Studio's empty thinking signatures are accepted; supplied signatures are replayed unchanged.
+- **Responses:** `reasoningEffort` → `reasoning.effort`, plus `enableThinking`, `store`,
+  `previousResponseId`, and `conversation`. Omitted `store` retains the API's default (`true`);
+  set it to `false` for client-managed history. `previousResponseId` requires a stored response.
+  Hosted tools are `Alibaba.webSearch()`, `Alibaba.webExtractor()`, and `Alibaba.codeInterpreter()`.
+  Web extraction is used together with web search. Hosted calls/results carry `providerExecuted: true`.
+
+Omitted options preserve provider defaults. Effort values pass through unchanged and accept future
+strings. Qwen 3.8 Chat rejects requests combining a thinking budget with effort.
+
+Package entrypoints are `@opencode/ai/providers/alibaba`, `alibaba/chat`, `alibaba/messages`,
+and `alibaba/responses`. Live recordings cover all three APIs in Singapore; regional URL construction
+is unit-tested for all six regions.
+
+## Z.AI
+
+`ZAI` uses the standard API. Chat Completions is the default language-model API;
+the existing `.image(...)` selector provides image generation.
+
+```ts
+import { LLM } from "@opencode/ai"
+import { ZAI, ZAICodingPlan } from "@opencode/ai/providers"
+
+const zai = ZAI.configure({ apiKey: process.env.ZAI_API_KEY })
+const request = LLM.request({
+  model: zai.model("glm-5.3"), // also zai.chat("glm-5.3")
+  prompt: "Explain this design.",
+  providerOptions: {
+    reasoningEffort: "high",
+    thinking: { type: "enabled", clear_thinking: false },
+  },
+})
+
+const coding = ZAICodingPlan.configure({ apiKey: process.env.ZAI_API_KEY })
+const messages = LLM.request({
+  model: coding.messages("glm-5.3"),
+  prompt: "Explain this design.",
+  providerOptions: { effort: "high" },
+})
+```
+
+The products have distinct provider identities and endpoints:
+
+| Provider                            | Selector                    | Default base URL                      |
+| ----------------------------------- | --------------------------- | ------------------------------------- |
+| `ZAI` (`zai`)                       | `.model`, `.chat`, `.image` | `https://api.z.ai/api/paas/v4`        |
+| `ZAICodingPlan` (`zai-coding-plan`) | `.model`, `.chat`           | `https://api.z.ai/api/coding/paas/v4` |
+| `ZAICodingPlan`                     | `.messages`                 | `https://api.z.ai/api/anthropic/v1`   |
+| `ZAICodingPlan`                     | `.responses`                | `https://api.z.ai/api/v1`             |
+
+Both read `ZAI_API_KEY` when `apiKey` is omitted and support an explicit `auth` override.
+Coding Plan requires an active subscription. `baseURL` overrides the selected API's
+complete base, including its version prefix. Language-model routes use HTTP/SSE.
+
+Options retain the selected API's native semantics:
+
+- Chat `reasoningEffort` lowers to `reasoning_effort`; Responses lowers it to `reasoning.effort`.
+  Messages `effort` lowers to `output_config.effort`. Omission preserves provider defaults.
+- Chat `thinking` passes `type` and `clear_thinking` through unchanged. Set
+  `clear_thinking: false` and replay complete `response.message` values to preserve reasoning
+  across user messages and tool loops. The standard API defaults to clearing historical thinking;
+  Coding Plan documents preservation by default.
+- Messages accepts `thinking: { type: "enabled" | "adaptive" | "disabled" }` without requiring
+  an Anthropic token budget. Coding Plan documents a disabled toggle as low-effort thinking
+  for GLM-5.3, with explicit effort taking precedence.
+- Chat also offers `toolStream`, `doSample`, `responseFormat`, `requestID`, and `userID`.
+  Tool-argument streaming is enabled when tools are present on GLM-4.6/4.7/5.x;
+  `toolStream: false` explicitly disables it. Older model families omit the opt-in.
+- Effort and thinking values remain forward-compatible strings. Their meaning is model-specific:
+  GLM-5.3 accepts `low`, `high`, and `max` effort and rejects disabled thinking with HTTP 400;
+  the direct GLM-5.2 recordings returned reasoning even with `none` and `minimal` effort,
+  whereas explicit `thinking.type: "disabled"` disabled it on GLM-5.2 and GLM-4.7.
+
+Standard API recordings cover GLM-5.3 efforts and a full preserved-reasoning tool loop with
+a subsequent user follow-up, GLM-5.2 efforts, older-model thinking toggles, GLM-4.5 tool calls,
+GLM-5.3-Flash image input, and JSON output. Coding Plan has unit coverage for routing,
+request options, and reasoning replay; successful live recordings are pending.
+
+Package entrypoints are `@opencode/ai/providers/zai`, `zai/chat`, `zai-coding-plan`,
+`zai-coding-plan/chat`, `zai-coding-plan/messages`, and `zai-coding-plan/responses`.
+
+## Moonshot
+
+Moonshot defaults to Chat Completions, with Messages and Responses selectors for Kimi K3:
+
+```ts
+import { LLM } from "@opencode/ai"
+import { Moonshot } from "@opencode/ai/providers"
+
+const moonshot = Moonshot.configure({ apiKey: process.env.MOONSHOT_API_KEY })
+
+const request = LLM.request({
+  model: moonshot.model("kimi-k3"), // also moonshot.chat("kimi-k3")
+  prompt: "Explain the tradeoffs in this design.",
+  providerOptions: { reasoningEffort: "high" },
+})
+
+const messages = LLM.request({
+  model: moonshot.messages("kimi-k3"),
+  prompt: "Explain the tradeoffs in this design.",
+  providerOptions: { effort: "high" },
+})
+
+const responses = LLM.request({
+  model: moonshot.responses("kimi-k3"),
+  prompt: "Explain the tradeoffs in this design.",
+  providerOptions: { reasoningEffort: "high" },
+})
+```
+
+When `apiKey` is omitted, authentication reads `MOONSHOT_API_KEY`, then `MOONSHOTAI_API_KEY`.
+Chat and Responses use `https://api.moonshot.ai/v1`; Messages uses
+`https://api.moonshot.ai/anthropic/v1`. `baseURL` overrides the selected API's complete base,
+including the version prefix, for regional endpoints or gateways. Each endpoint requires its own valid credentials.
+All three routes use HTTP/SSE.
+
+Reasoning options stay native to the selected API and model:
+
+| Model/API                   | Provider options                                                                        |
+| --------------------------- | --------------------------------------------------------------------------------------- |
+| K3 Chat / Responses         | `reasoningEffort: "low" \| "high" \| "max"`; default is `max`                           |
+| K3 Messages                 | `effort: "low" \| "high" \| "max"`; default is `max`                                    |
+| K2.6 Chat                   | `thinking: { type: "enabled" \| "disabled", keep?: "all" \| null }`; default is enabled |
+| K2.7 Code / high-speed Chat | Omit `thinking` to use always-on, preserved reasoning                                   |
+
+Omitting options preserves the model's defaults. K3 uses effort rather than the K2.x `thinking`
+parameter. Known effort values have autocomplete while future strings remain accepted.
+For K2.6, `thinking.keep: "all"` enables preservation of reasoning across user messages.
+K3 and both K2.7 Code variants always preserve reasoning. Continue with the returned
+`response.message` and matching tool results so reasoning content and any Messages signatures are retained.
+Leave sampling options such as `temperature` unset to use these models' fixed defaults.
+
+The recorded suite covers all three K3 APIs, default and explicit efforts, K2.6 thinking modes,
+both K2.7 Code variants, generated tool loops with a subsequent user follow-up, required/disabled
+tool choice, image-byte input, and native structured output through `http.body` overlays.
+K3 Chat and Messages accept required and disabled tool choice. Responses supports automatic tool
+choice only; explicit `required` and `none` produce a provider `InvalidRequest` error, also covered by recordings.
+The provider targets the Moonshot Open Platform; Kimi Code is a separate product and endpoint.
+
+Package entrypoints are `@opencode/ai/providers/moonshot`, `moonshot/chat`, `moonshot/messages`,
+and `moonshot/responses`; each exports `model(modelID, settings)`.
+
+## MiniMax
+
+MiniMax defaults to its Messages API and reads `MINIMAX_API_KEY` when `apiKey` is omitted:
+
+```ts
+import { Effect, Layer } from "effect"
+import { LLM, LLMClient } from "@opencode/ai"
+import { MiniMax } from "@opencode/ai/providers"
+import { RequestExecutor } from "@opencode/ai/route"
+
+const minimax = MiniMax.configure({ apiKey: process.env.MINIMAX_API_KEY })
+const request = LLM.request({
+  model: minimax.model("MiniMax-M3"), // also minimax.messages("MiniMax-M3")
+  prompt: "What is 173 multiplied by 219?",
+  providerOptions: { thinking: { type: "adaptive" } },
+  generation: { maxTokens: 1536 },
+})
+
+const layer = LLMClient.layer.pipe(Layer.provide(RequestExecutor.fetchLayer))
+const response = await Effect.runPromise(LLMClient.generate(request).pipe(Effect.provide(layer)))
+console.log(response.text)
+```
+
+Select `minimax.chat("MiniMax-M3")` or `minimax.responses("MiniMax-M3")` for MiniMax's native Chat Completions
+and Responses APIs. The matching package entrypoints are `@opencode/ai/providers/minimax/messages`,
+`@opencode/ai/providers/minimax/chat`, and `@opencode/ai/providers/minimax/responses`.
+
+- **Messages:** M3 thinking defaults off. Set `thinking: { type: "adaptive" }` to enable it or
+  `thinking: { type: "disabled" }` to disable it.
+- **Chat:** M3 thinking defaults on and uses the same `thinking` control. The provider enables `reasoning_split`
+  by default so reasoning is separate from answer text; `reasoningSplit: false` selects native `<think>`-tagged text.
+- **Responses:** M3 reasoning defaults off. `reasoningEffort: "none"` disables it; `"minimal"`, `"low"`,
+  `"medium"`, and `"high"` enable reasoning without changing its depth.
+
+M2.x models always think, even when a disabling option is supplied. For tool continuations, retain the complete
+`response.message` in history before adding `Message.tool(...)` results; this preserves reasoning and any signatures.
+
+The default API bases are `https://api.minimax.io/anthropic/v1` for Messages and `https://api.minimax.io/v1` for
+Chat and Responses. `configure({ baseURL })` replaces the selected API's base, including its version prefix.
+
+## Meta
+
+Use Meta's direct [Model API](https://dev.meta.ai/docs/overview) with `META_API_KEY`:
+
+```ts
+import { Meta } from "@opencode/ai/providers"
+
+const meta = Meta.configure() // or Meta.configure({ apiKey })
+const request = LLM.request({
+  model: meta.responses("muse-spark-1.3"), // meta.model(...) also selects Responses
+  prompt: "What is 173 multiplied by 219? Reply with the integer.",
+  providerOptions: { reasoningEffort: "low" },
+  generation: { maxTokens: 1024 },
+})
+```
+
+`meta.chat("muse-spark-1.3")` selects Chat Completions; `meta.messages("muse-spark-1.3")` selects
+the Anthropic-compatible Messages API. All use `https://api.meta.ai/v1`. The package entrypoints
+`@opencode/ai/providers/meta/responses`, `meta/chat`, and `meta/messages` expose `model(modelID, settings)`.
+
+[Muse Spark](https://dev.meta.ai/docs/models) supports `minimal`, `low`, `medium`, `high`, and
+`xhigh` reasoning effort; standard-tier 1.3 also supports `max`. Omitting effort uses the model's
+default. Muse Spark always reasons and rejects `none`. The output-token budget includes private reasoning.
+
+Responses defaults to `store: false` and `include: ["reasoning.encrypted_content"]`. Preserve
+`response.message` along with matching `Message.tool(...)` results in subsequent requests to replay
+reasoning through tool loops. Optional `reasoningSummary: "auto"` requests a readable summary.
+For server-managed history, override `store: true, include: []` and send the response ID through
+`http: { body: { previous_response_id: responseID } }` with only the new input.
+Chat Completions redacts private reasoning and cannot carry it between calls.
+Responses and Chat support only `toolChoice: "auto"` (the default). Messages also accepts `"none"`;
+its documented forced `"any"` choice currently returns HTTP 400. Messages defaults to adaptive thinking
+with `display: "omitted"`, preserving encrypted `redacted_thinking` in `response.message`. Use
+`providerOptions: { effort: "low" }` for depth or `thinking: { type: "enabled", budgetTokens: 1024 }`
+for budget compatibility (with `generation.maxTokens > 1024`).
+
+Add `tools: [Meta.webSearch()]` to a Spark Responses or Messages request for hosted web search.
+Responses exposes hosted results and URL citations in text-part `providerMetadata.meta.annotations`.
+To include search result lists, set `include: ["reasoning.encrypted_content", "web_search_call.results"]`.
+Messages exposes hosted search calls; the recorded Messages API stream does not supply structured
+citations or separate result blocks. Retain `response.message` for either API's continuation.
+
+Use `Image.generate` for one-off generation or editing:
+
+```ts
+import { Image, Media } from "@opencode/ai"
+
+const generation = Image.generate({
+  model: meta("muse-image-1.0"),
+  prompt: "A flat black square on a white background.",
+  n: 1,
+  providerOptions: { reasoningStrength: "low" },
+})
+
+const edit = Image.generate({
+  model: meta("muse-image-1.0"),
+  prompt: "Make the square purple.",
+  images: [Media.bytes(imageBytes, "image/webp")],
+  format: "png",
+  providerOptions: { reasoningStrength: "low" },
+})
+```
+
+The default image format is WEBP; `format` also accepts PNG/JPEG and `responseFormat: "url"`
+returns a signed URL. `size` is an aspect-ratio hint. For conversational images, select
+`meta.responses("muse-image-1.0")` with `tools: [Meta.imageGeneration({ reasoningStrength: "low" })]`.
+Generated images are provider-executed tool results with file content. Retain `response.message` to
+replay the signed image handle on the next request. Muse Image accepts only the `image_generation` tool.
+
+Meta Responses is explicitly HTTP/SSE-only and does not use WebSockets, even when a caller supplies
+`StreamOptions.webSocket`. The public `/v1/responses` endpoint rejects WebSocket upgrades with HTTP 405 (`Allow: POST`).
+
+## Image generation
+
+Use `Image.generate` with an image model for direct asset generation. `Image.request` mirrors `LLM.request`: the
+model is a callable-facade ref (`openai("gpt-image-2")`) or an explicit `.image(...)` model, common fields
+(`images`, `mask`, `n`, `size`, `aspectRatio`, `seed`, `format`) lower natively or fail typed, and
+`providerOptions` is inferred from the selected model:
+
+```ts
+import { Image, Media } from "@opencode/ai"
+import { OpenAI } from "@opencode/ai/providers"
+
+const openai = OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY })
+
+const program = Effect.gen(function* () {
+  const response = yield* Image.generate({
+    model: openai("gpt-image-2"),
+    prompt: "A robot tending a rooftop garden",
+    n: 2,
+    size: "1024x1024",
+    format: "webp",
+    providerOptions: {
+      quality: "high", // inferred from the OpenAI image model
+      future_option: true, // unknown native options pass through unchanged
+    },
+  })
+
+  return response.images // Media.Asset[] with owned bytes or a provider URL
+})
+```
+
+`Media.Asset` is the one asset type shared by image requests, image responses, LLM messages, and tool results.
+`asset.source` is the serializable `Media.Source` (`bytes`, `base64`, `url`, or `ref`); `asset.bytes()`,
+`asset.base64()`, and `asset.dataUrl()` decode or download lazily and cache; `asset.materialize()` pulls a `url`
+asset into owned bytes before the provider URL expires. Construct assets with `Media.bytes`, `Media.base64`,
+`Media.url`, `Media.ref(provider, id)`, `Media.fromDataUrl`, or `Media.file(path)`.
+
+Pass ordered image inputs to the same method for editing, composition, or image-conditioned generation:
+
+```ts
+const response =
+  yield *
+  Image.generate({
+    model,
+    prompt: "Combine these product photos into one studio scene",
+    images: [
+      Media.bytes(firstBytes, "image/png"),
+      Media.url("https://example.com/second.webp"),
+      Media.ref("openai", "file_123"),
+    ],
+    providerOptions,
+    http,
+  })
+```
+
+`Media.ref(provider, id)` represents provider file handles such as OpenAI file IDs or Gemini Files URIs; routes
+only forward refs that belong to their own provider. Raw strings are not accepted as image inputs, avoiding
+ambiguity between base64, URLs, and provider IDs. Empty or omitted `images` uses text-to-image generation; a
+non-empty array selects the provider's edit behavior without enforcing provider image-count limits locally. OpenAI
+uses multipart for byte/data-URL edits and its JSON reference body for URL or file-ID edits. The common `mask`
+field selects inpainting; routes that cannot honor it fail with `UnsupportedOperation`:
+
+```ts
+yield *
+  Image.generate({
+    model: openai("gpt-image-2"),
+    prompt,
+    images: [Media.bytes(sourceBytes, "image/png")],
+    mask: Media.bytes(maskBytes, "image/png"),
+  })
+```
+
+On multipart requests, `http.body` can override option fields but not structural `model`, `prompt`, `image[]`,
+or `mask` fields, and the transport owns the multipart `Content-Type` boundary. For JSON requests, `http.body`
+remains the final raw-native overlay. Gemini does not fetch public HTTP URLs, and hosted Z.ai image generation does
+not accept image inputs. These cases fail with a typed `AIError` before network I/O.
+
+Provider-native image options belong to each request. Raw `http.body` fields have final precedence over them:
+
+```ts
+yield *
+  Image.generate({
+    model: openai("gpt-image-2"),
+    prompt,
+    providerOptions: { quality: "medium" },
+    http,
+  })
+```
+
+xAI image models use the same request API with xAI-native controls:
+
+```ts
+yield *
+  Image.generate({
+    model: XAI.configure({ apiKey })("any-model-id"),
+    prompt,
+    n: 2,
+    aspectRatio: "16:9",
+    providerOptions: {
+      resolution: "1k",
+      responseFormat: "b64_json",
+      future_option: true,
+    },
+    http,
+  })
+```
+
+Google's current Gemini image models use the same direct API:
+
+```ts
+import { Google } from "@opencode/ai/providers"
+
+const googleProgram = Effect.gen(function* () {
+  const response = yield* Image.generate({
+    model: Google.configure({ apiKey })("any-model-id"),
+    prompt: "A robot tending a rooftop garden",
+    aspectRatio: "16:9",
+    seed: 42,
+    providerOptions: {
+      imageSize: "2K",
+      thinkingLevel: "HIGH",
+      includeThoughts: true,
+      futureOption: true,
+    },
+    http,
+  })
+
+  return response.images
+})
+```
+
+Google image options are request-scoped and inferred from the selected model. Known fields autocomplete while
+future string values and arbitrary native Gemini `generationConfig` fields remain available. Native fields override
+their mapped aliases, and `http.body` is the final deep overlay. The selected model ID is sent to Gemini
+`generateContent` without a local allowlist.
+
+Z.ai image models infer open Z.ai-native options from the selected model:
+
+```ts
+yield *
+  Image.generate({
+    model: ZAI.configure({ apiKey })("any-model-id"),
+    prompt,
+    providerOptions: {
+      quality: "hd",
+      userID: "user-123",
+      future_option: true,
+    },
+    http,
+  })
+```
+
+Z.ai does not include trustworthy MIME metadata for output URLs, so generated images use
+`application/octet-stream` until materialized. Output URLs expire after 30 days; call `asset.materialize()` and
+persist the bytes promptly if they must remain available.
+
+Conversational image generation remains part of the LLM interaction. OpenAI Responses exposes it through its hosted image tool:
+
+```ts
+const program = Effect.gen(function* () {
+  const response = yield* LLM.generate(
+    LLM.request({
+      model: OpenAI.configure({ apiKey }).responses("gpt-5"),
+      prompt: "Design a solarpunk rooftop garden, then show me.",
+      tools: [OpenAI.imageGeneration({ quality: "high" })],
+    }),
+  )
+
+  return response.message
+})
+```
+
+The hosted result is represented as a provider-executed tool call and tool result, and the generated image is also emitted as a first-class `media` `LLMEvent` (`response.message` then carries a `media` part). Gemini image-capable models emit the same `media` event for inline image output. Retaining `response.message` preserves the generated image for continuation on both routes.
+
+## Public API
+
+- **`LLM.request({...})`** — build a provider-neutral `LLMRequest`. Accepts ergonomic inputs (`system: string`, `prompt: string`) that normalize into the canonical Schema classes.
+- **`LLM.generate` / `LLM.stream`** — re-exported from `LLMClient` for one-import use.
+- **`Message.user(...)` / `Message.assistant(...)` / `Message.tool(...)`** — message constructors from the canonical schema model.
+- **`LanguageModel.make(...)` / `ToolCallPart.make(...)` / `ToolResultPart.make(...)` / `ToolDefinition.make(...)`** — model and tool-related constructors from the canonical schema model.
+- **`LLMEvent.is.*`** — typed guards (`is.textDelta`, `is.toolCall`, `is.finish`, …) for filtering streams.
+- **`Image.request` / `Image.generate` / `Image.stream`** — generate images through a provider-neutral image request and response model.
+- **`ImageClient`** — Effect service and layer for image execution, parallel to `LLMClient`.
+- **`Media`** — the shared asset type (`Media.Asset`, `Media.Source`) and constructors used by messages, tool results, and media requests.
+- **`ModelRef`** — the value returned by calling a configured facade (`openai("gpt-5")`); each request namespace resolves its own route from it.
+- **`Job`** — provider-neutral async job handle (`await`, `refresh`, `cancel`, `events`) used by queued media routes.
+- **`@opencode/ai/promise`** — `AI.make({ layer? })` and a default `ai` client exposing `llm` and `image` as Promise / `AsyncIterable` APIs.
+
+## Testing
+
+Use the deterministic test client from `@opencode/ai/testing` to script provider-neutral responses and inspect
+the requests sent by code under test:
+
+```ts
+import { Effect } from "effect"
+import { TestLLM } from "@opencode/ai/testing"
+
+const programWithTestClient = Effect.gen(function* () {
+  const test = yield* TestLLM.Test
+  yield* test.push(TestLLM.text("Hello from the test model", "text-1"))
+  const result = yield* program
+  console.log(yield* test.requests())
+  return result
+}).pipe(Effect.provide(TestLLM.testLayer()))
+```
+
+`testLayer()` provides the same object under `LLMClient.Service` and `TestLLM.Test`. Production consumes the
+normal client; tests use the additional controls. Each layer build has fresh state.
+
+- `test.push(...)` queues one-shot responses in execution order. Each argument is one response.
+- `test.always(response)` installs a repeatable fallback. The layer's `fallback` option sets its initial value.
+- `test.serve(request => response)` installs a request-dependent fallback. `always` and `serve` replace each
+  other without changing queued replies; queued replies take precedence.
+- `test.requests()` returns an array snapshot. `transformRequest` changes only the recorded observation;
+  `serve` receives the original canonical request.
+- `test.wait(count)` waits for request arrivals, not output or completion, and supports concurrent waiters.
+- `test.gate()` returns a scoped gate with countable `started` notifications and a `release` Effect. Release
+  unblocks all requests captured by that gate; closing its scope also releases it. Effect-aware test runners
+  already provide Scope.
+
+Constructing `stream()` or `generate()` does not record a request, invoke a responder, or consume a script.
+Each execution does. An exhausted queue without a fallback defects immediately rather than waiting for a
+future reply.
+
+Generation responses remain canonical event arrays or arbitrary `Stream<LLMEvent, AIError>` values. The client consumes
+supplied streams directly, preserving failure identity, finalizers, incomplete output, and post-finish tails;
+it does not repair or truncate them.
+
+For explicit compaction, script a `CompactionResponse` through `push`, `always`, or `serve`. Its `replacement` contains the next context window, including retained user messages. The client returns that result and usage directly, with the same lazy request recording and gates. Generation and compaction reject fixtures for the wrong operation instead of converting between response shapes.
+
+For `compact(request, { mechanism: "trigger" })`, script a `CompactionCheckpointResponse` instead. It carries `checkpoint`, `responseID`, and optional `usage`. Endpoint and trigger calls reject each other's fixtures; both share the same queue, gates, lazy recording, and fallback controls.
+
+The published legacy `Service`, `layer`, `clientLayer`, and module-level controls remain available as adapters
+over the same implementation, including the legacy live `requests` array. New tests should use `Test` and
+`testLayer`.
+
+## Provider compaction
+
+Compaction is opt-in. The package supports automatic compaction in OpenAI/Azure Responses and Anthropic Messages (including Claude on Vertex), and explicit compaction calls in OpenAI/Azure/xAI Responses. Model and deployment support still depends on the provider.
+
+This is different from prompt caching, server-side history storage, or truncation. Compaction returns provider-owned context that must be replayed to continue the conversation.
+
+### Explicit compaction
+
+`LLMClient.compact(request)` (equivalently, `{ mechanism: "endpoint" }`) is the caller-controlled operation for OpenAI, Azure, and xAI Responses. It performs exactly one HTTP call to `/responses/compact`, using the selected route's endpoint, credentials, query, and HTTP middleware. It returns a `CompactionResponse` with `replacement: Message[]` and optional `usage`, not a normal generation response. This mechanism does not accept a WebSocket executor.
+
+Prefer this operation, where supported, when the application owns compaction policy and durable context updates.
+
+```ts
+const result = yield * LLMClient.compact(request)
+const next = LLMRequest.update(request, {
+  messages: result.replacement,
+})
+const response = yield * LLMClient.generate(next)
+```
+
+`replacement` replaces the complete input window. Do not append it to the original transcript or extract only the encrypted item: the provider may retain additional messages in its output. Retained user and assistant messages remain ordinary messages with typed text, media, or reasoning parts, in their original order. Provider-specific message IDs, status, and phase use `providerMetadata`, not a raw output array hidden in an assistant message. Unsupported returned item types fail explicitly.
+
+The selected model carries explicit-compaction capability through request construction and updates. Calls using unsupported routes fail type checking. When the model is selected dynamically, narrow the request with `LLMClient.canCompact(request)` before calling `LLMClient.compact`; a model or route switch does not inherit the old capability. Runtime validation still rejects unsupported calls from untyped consumers. Capability describes the route's API, not whether every model or custom deployment supports the operation.
+
+Generation-only body overlays such as `stream` and `store` are not sent to the compact endpoint. Supported compact controls such as service tier and prompt-cache settings preserve request defaults and HTTP-overlay precedence. Retained image and file detail settings survive serialization and replay.
+
+The input must still fit the model's context window. Explicit compaction is not an overflow-recovery operation. Anthropic does not expose this operation in this package; its in-band compaction remains available below. Compatible routes do not inherit an explicit compact endpoint simply because they use a Responses protocol.
+
+### Streamed checkpoint compaction
+
+OpenAI Responses also exposes a separate, explicitly selected mechanism:
+
+```ts
+const result =
+  yield *
+  LLMClient.compact(request, {
+    mechanism: "trigger",
+    webSocket, // Optional: without it, the request uses HTTP/SSE.
+  })
+
+result.checkpoint // Successful encrypted CompactionPart.
+result.responseID
+result.usage
+```
+
+This appends a native `compaction_trigger` control item to the full input and sends a normal Responses request, with tools and instructions retained, `stream: true`, `store: false`, and parallel tool calls enabled. It removes normal-answer text/output-format controls, forced tool choices, output-token/tool-call limits, and automatic `context_management`. Body overlays cannot replace `input` or supply `previous_response_id`/`conversation`; the complete canonical history is required for safe stateless replay. Request metadata, auth, headers, query parameters, service tier, and supported prompt-cache settings are preserved.
+
+Only a successful `response.completed` with a response ID and exactly one logical encrypted checkpoint succeeds. Repeated item events are correlated by ID/output slot, including ID-less checkpoints. Other output is ignored, not returned as assistant text or dispatched as tools. Failed, incomplete, malformed, and interrupted responses return errors rather than partial checkpoints.
+
+The result is **not a replacement window**. The caller selects retained history, combines it with `result.checkpoint`, and durably installs it before continuing. The operation does not choose a retention budget, prune messages, or modify the original request.
+
+The supplied WebSocket executor can reuse a compatible append baseline for the compaction request. On completion the protocol supplies no continuation checkpoint, clearing the old baseline so the next generation sends the newly installed window in full. Validation occurs before transport completion is acknowledged. There is no operation-level retry or fallback to `/responses/compact`; existing safe transport fallback may use SSE, with full history and no connection-local response ID.
+
+Trigger support is separate from endpoint support. Only the OpenAI Responses route advertises it; Azure, xAI, Chat, and compatible Responses routes do not inherit it. Untyped calls still fail before sending: missing route capabilities return `UnsupportedOperation`, while unknown mechanism names and invalid inputs return `InvalidRequest`. Dynamic callers must narrow for the selected mechanism:
+
+```ts
+if (LLMClient.canCompact(request, { mechanism: "trigger" })) {
+  const result = yield * LLMClient.compact(request, { mechanism: "trigger" })
+}
+```
+
+This capability describes protocol implementation, **not universal availability on OpenAI API deployments**. The host application owns subscription/deployment eligibility, OAuth, endpoint selection, and deployment-specific headers. Local protocol/socket tests do not establish live provider support.
+
+### Advanced: in-band compaction
+
+`providerOptions.contextManagement` lets the provider decide when to compact during an ordinary `generate` or `stream` call. This is an advanced option for callers that own persistence and recovery: persist the complete assistant message, including its checkpoint, before continuing. Enabling the option does not provide durable checkpoint storage, interruption recovery, or model-switch policy. Keep the prior context until a successful checkpoint has been persisted.
+
+Inside an `Effect.gen`, enable OpenAI compaction with typed provider options:
+
+```ts
+import { LLM, LLMClient, LLMRequest, Message } from "@opencode/ai"
+import { OpenAI } from "@opencode/ai/providers"
+
+const request = LLM.request({
+  model: OpenAI.configure({ apiKey }).responses("gpt-5.3-codex"),
+  messages,
+  providerOptions: {
+    contextManagement: [{ type: "compaction", compactThreshold: 200_000 }],
+  },
+})
+const response = yield * LLMClient.generate(request)
+const next = LLMRequest.update(request, {
+  messages: [...request.messages, response.message, Message.user("Continue")],
+})
+```
+
+`store: false` remains the default. Keep the entire `response.message`, not just `response.text`. Compaction events become ordered `CompactionPart`s alongside text and reasoning. The conversation contains everything needed to continue; there is no separate replay object or hidden provider transcript.
+
+A compaction part has `provider` and exactly one representation: `encrypted` for Responses, or `text` for Anthropic. Responses also preserves the optional checkpoint `id`. These fields survive message serialization without becoming visible assistant text. Sending a checkpoint to another provider or an incompatible API fails rather than silently losing context.
+
+```ts
+import { CompactionPart, ProviderID } from "@opencode/ai"
+
+CompactionPart.make({ provider: ProviderID.make("openai"), id: "cmp_123", encrypted: "..." })
+CompactionPart.make({ provider: ProviderID.make("anthropic"), text: "Summary of the conversation..." })
+```
+
+For Anthropic, use:
+
+```ts
+providerOptions: {
+  contextManagement: {
+    edits: [{
+      type: "compact_20260112",
+      trigger: { type: "input_tokens", value: 150_000 },
+      pauseAfterCompaction: true,
+      instructions: "Summarize the task and decisions. Do not call tools while summarizing.",
+    }],
+  },
+}
+```
+
+- The trigger is optional (provider default: 150,000 tokens), with a minimum of 50,000.
+- Custom instructions replace Anthropic's default summarization instructions.
+- The route adds `compact-2026-01-12` to existing beta headers, including when replaying a checkpoint without enabling new compactions.
+- A pause is exposed as `response.finishReason.raw === "compaction"`. It occurs only if the threshold triggers compaction: `pauseAfterCompaction` does not mean "compact now". The caller explicitly issues the next request; the package never automatically resumes.
+- Anthropic can return a compaction block with `content: null` when summarization fails. This becomes a compaction part with `text: null`, which is **not** a successful replacement for prior history. The package never prunes history automatically.
+- `Usage` totals include all reported Anthropic `usage.iterations`, including compaction. `contextTokens` separately reports the final message iteration's inclusive input size, when available. A compaction-only pause does not report a post-compaction context size. Raw iteration usage remains in `providerMetadata`.
+
+### Recording tests
+
+Tests cover serialized round trips, real local HTTP plus a tool loop, WebSocket recovery, provider errors, malformed blocks, and usage accounting. Live provider tests are gated by `RECORD=true` and the relevant API keys:
+
+```sh
+# Run from packages/ai. Only records the selected new cassette group.
+RECORD=true RECORDED_PREFIX=openai-compaction bun test test/provider/compaction.recorded.test.ts
+RECORD=true RECORDED_PREFIX=xai-compaction bun test test/provider/compaction.recorded.test.ts
+RECORD=true RECORDED_PREFIX=anthropic-compaction bun test test/provider/compaction.recorded.test.ts
+```
+
+Provider references: [OpenAI](https://developers.openai.com/api/docs/guides/compaction), [Azure](https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/responses#server-side-compaction), [Anthropic](https://platform.claude.com/docs/en/build-with-claude/compaction), [xAI](https://docs.x.ai/developers/advanced-api-usage/context-compaction).
+
+## Caching
+
+Prompt caching is **on by default**. Every `LLMRequest` resolves to `cache: "auto"` unless the caller opts out with `cache: "none"`. Each protocol translates `CacheHint`s to its wire format (`cache_control` on Anthropic, `cachePoint` on Bedrock; OpenAI and Gemini do implicit caching server-side and don't need inline markers — auto is a no-op there).
+
+### Auto placement
+
+`"auto"` places up to four breakpoints — the last tool definition, the first system part, the last system part when distinct, and the final message boundary. These expose successively larger reusable prefixes for tool definitions, system instructions, and the active conversation. The rolling final-message boundary advances on every request so recent conversation prefixes remain reusable during tool loops.
+
+Tools precede every system and conversation block in the provider prefix, so tool definitions must remain byte-stable and deterministically ordered for downstream breakpoints to remain reusable.
+
+Requests below a provider's minimum cacheable size simply do not produce a reusable cache entry.
+
+### Opting out
+
+```ts
+LLM.request({
+  model,
+  system,
+  prompt: "one-off question",
+  cache: "none",
+})
+```
+
+### Granular policy
+
+```ts
+cache: {
+  tools?: boolean,
+  system?: boolean,
+  messages?: "latest-user-message" | "latest-assistant" | { tail: number },
+  ttlSeconds?: number,         // ≥ 3600 → 1h on Anthropic/Bedrock; else 5m
+}
+```
+
+### Manual hints
+
+Inline `CacheHint` on any text / system / tool / tool-result part overrides automatic placement. The auto policy preserves manual hints, counts them against Anthropic and Bedrock's four-breakpoint limit, and only fills the remaining slots.
+
+```ts
+LLM.request({
+  model,
+  system: [
+    { type: "text", text: "stable system prompt", cache: { type: "ephemeral" } },
+  ],
+  ...
+})
+```
+
+### Provider behavior table
+
+| Protocol                | `cache: "auto"`                                                           |
+| ----------------------- | ------------------------------------------------------------------------- |
+| Anthropic Messages      | emits up to 4 `cache_control` markers (4-breakpoint cap enforced)         |
+| Bedrock Converse        | emits up to 4 `cachePoint` blocks (4-breakpoint cap enforced)             |
+| OpenRouter              | emits up to 4 `cache_control` markers                                     |
+| OpenAI Chat / Responses | no-op (implicit caching above 1024 tokens)                                |
+| Gemini                  | no-op (implicit caching on 2.5+; explicit `CachedContent` is out-of-band) |
+
+Normalized cache usage is read back into `response.usage.cacheReadInputTokens` and `cacheWriteInputTokens` across every provider.
+
+## Providers
+
+Provider facades configure endpoint/auth/deployment details first, then expose model selectors that take only a model or deployment id. The selected model carries the executable route value used at runtime.
+
+```ts
+import { OpenAI, CloudflareAIGateway } from "@opencode/ai/providers"
+
+const openai = OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY }).responses("gpt-4o-mini")
+const gateway = CloudflareAIGateway.configure({
+  accountId: process.env.CLOUDFLARE_ACCOUNT_ID,
+  gatewayApiKey: process.env.CLOUDFLARE_API_TOKEN,
+}).model("workers-ai/@cf/meta/llama-3.1-8b-instruct")
+```
+
+Included LLM providers: OpenAI, Anthropic, Google (Gemini), Google Vertex, Amazon Bedrock, Azure OpenAI, Baseten, Cerebras, Cloudflare AI Gateway, Cloudflare Workers AI, DeepInfra, DeepSeek, Fireworks, Groq, Mistral, OpenRouter, TogetherAI, and xAI. Z.ai currently exposes image generation. Generic Chat Completions, Responses, and Anthropic Messages-compatible entrypoints support custom endpoints.
+
+Each named provider owns its module, endpoint, authentication, and route setup. Providers with the same wire format compose the shared protocol directly:
+
+```ts
+import { DeepSeek, Fireworks } from "@opencode/ai/providers"
+
+const deepseek = DeepSeek.configure({ apiKey }).model("deepseek-chat")
+const fireworks = Fireworks.configure({ apiKey }).model("accounts/fireworks/models/my-model")
+```
+
+The former `OpenAICompatible.baseten`, `.cerebras`, `.deepinfra`, `.deepseek`, `.fireworks`, `.groq`, and `.togetherai` presets are replaced by the top-level `Baseten`, `Cerebras`, `DeepInfra`, `DeepSeek`, `Fireworks`, `Groq`, and `TogetherAI` exports. Use `CloudflareAIGateway` and `CloudflareWorkersAI` directly; each has its own module. `OpenAICompatible` configures generic endpoints with an explicit `baseURL`.
+
+### Provider entrypoints
+
+Provider modules are available through dedicated exports from `@opencode/ai`. Each LLM entrypoint exports `model(modelID, settings)`, where `settings` contains provider configuration plus common `headers` and `body` overlays.
+
+```ts
+import { model } from "@opencode/ai/providers/openai/responses"
+
+const selected = model("gpt-5", {
+  apiKey: process.env.OPENAI_API_KEY,
+  headers: { "x-application": "example" },
+})
+```
+
+APIs have separate entrypoints:
+
+- `@opencode/ai/providers/openai/chat`
+- `@opencode/ai/providers/openai/responses`
+- `@opencode/ai/providers/openai-compatible/responses`
+- `@opencode/ai/providers/anthropic-compatible`
+- `@opencode/ai/providers/google-vertex/gemini`
+- `@opencode/ai/providers/google-vertex/chat`
+- `@opencode/ai/providers/google-vertex/responses`
+- `@opencode/ai/providers/google-vertex/messages`
+
+OpenAI Responses has one semantic route and uses HTTP by default. Advanced callers may supply a per-call WebSocket channel executor through `StreamOptions`; transport policy does not change provider settings, model identity, or route identity. The provider-neutral Open Responses implementation owns the reusable WebSocket request and event contract, while each provider opts in with its own handshake and connection policy. Azure follows the same Chat/Responses split at `providers/azure/chat` and `providers/azure/responses`. Generic OpenAI-compatible Chat remains at `providers/openai-compatible`; the Responses adapter at `providers/openai-compatible/responses` uses the provider-neutral Open Responses protocol. OpenAI Responses extends that baseline with OpenAI tools, event variants, metadata, and defaults. Generic Anthropic Messages-compatible providers use `providers/anthropic-compatible`, which the named Anthropic provider composes. Google Gemini and Amazon Bedrock expose their single native API through their existing provider paths.
+
+Vertex Gemini, Vertex Chat, Vertex Responses, and Vertex Messages are separate API entrypoints. All accept `project`, `location`, and an optional `accessToken`; when no explicit token or auth override is supplied they lazily use Google Application Default Credentials. Vertex Gemini instead selects express mode when `apiKey` or `GOOGLE_VERTEX_API_KEY` is present. Vertex Chat targets MaaS models through the OpenAI-compatible Chat Completions endpoint, while Vertex Responses targets Grok models and defaults `store` to `false` as required by Vertex. `providers/google-vertex` remains the default alias for `providers/google-vertex/gemini`.
+
+Tuned Vertex Gemini deployments use model ids shaped like `endpoints/1234567890` and require OAuth or ADC; Vertex express-mode API keys support publisher models only.
+
+```ts
+import { model } from "@opencode/ai/providers/google-vertex/gemini"
+
+model("gemini-3.5-flash", { project: "my-project", location: "global" })
+```
+
+```ts
+import { model } from "@opencode/ai/providers/google-vertex/chat"
+
+model("deepseek-ai/deepseek-v3.2-maas", { project: "my-project", location: "global" })
+```
+
+```ts
+import { model } from "@opencode/ai/providers/google-vertex/responses"
+
+model("xai/grok-4.20-reasoning", { project: "my-project", location: "global" })
+```
+
+```ts
+import { model } from "@opencode/ai/providers/google-vertex/messages"
+
+model("claude-sonnet-4-6", { project: "my-project", location: "global" })
+```
+
+Additional provider entrypoints include:
+
+- `@opencode/ai/providers/baseten`
+- `@opencode/ai/providers/deepseek`
+- `@opencode/ai/providers/fireworks`
+- `@opencode/ai/providers/cloudflare-ai-gateway`
+- `@opencode/ai/providers/cloudflare-workers-ai`
+
+## Provider options & HTTP overlays
+
+Request options in order of stability:
+
+1. **`generation`** — portable knobs (`maxTokens`, `temperature`, `topP`, `topK`, penalties, seed, stop).
+2. **`promptCacheKey`** — stable cache affinity lowered by every protocol that supports it.
+3. **`providerOptions: { ... }`** — flat options inferred from the selected model (OpenAI `store`, Anthropic `thinking`, Gemini `thinkingConfig`, OpenRouter routing).
+4. **`http: { body, headers, query }`** — last-resort serializable overlays merged into the final HTTP request. Reach for this only when a stable typed path doesn't yet exist.
+
+Route/provider defaults are overridden by request-level values for each axis.
+
+The selected model supplies the provider-specific option type, so per-request overrides stay flat while the canonical runtime request remains provider-neutral:
+
+```ts
+LLM.request({
+  model,
+  prompt,
+  providerOptions: {
+    reasoningEffort: "low",
+  },
+})
+```
+
+## Routes
+
+Compose a route with `Route.make({ protocol, endpoint, auth, framing, ... })`. The route owns endpoint/auth/framing and the protocol owns body construction plus stream parsing. Transports receive the route's endpoint and auth when preparing requests. Unsupported request shapes fail during protocol lowering.
+
+## Effect
+
+This package is built on Effect. Public methods return `Effect` or `Stream`; provide `LLMClient.layer` for LLM dispatch and `ImageClient.layer` for image dispatch, then import the provider/protocol modules for the routes you use. The example at `example/tutorial.ts` is a runnable walkthrough.
+
+## See also
+
+- `AGENTS.md` — architecture, route construction, contributor guide
+- `example/tutorial.ts` — runnable end-to-end walkthrough
+- `test/provider/*.test.ts` — fixture-first protocol tests; `*.recorded.test.ts` files cover live cassettes

packages/ai/docs/media-design.md

@@ -0,0 +1,342 @@
+# Media generation in `@opencode/ai` — public API direction
+
+Status: proposal. Branch `media-support`.
+
+## Goal
+
+`@opencode/ai` becomes the one package you reach for to generate anything: text, images, video, speech, transcripts, and later music and realtime. The LLM surface already exists and is shaped by three constraints: Effect-first, used by OpenCode Core, usable externally. Media has a different priority order: **external DX first**, Effect and Promise as peers, Core as one consumer among many.
+
+The design below is derived from a survey of the raw provider APIs (OpenAI, Gemini/Veo/Imagen, xAI, Stability, BFL, fal, Replicate, Runway, Luma, Kling, MiniMax, ElevenLabs, Deepgram, Cartesia, AssemblyAI, Lyria) and from the Vercel AI SDK v7 (`generateImage`, `generateSpeech`, `transcribe`, `experimental_generateVideo`, `ImageModelV4`/`SpeechModelV4`/`TranscriptionModelV4`/`Experimental_VideoModelV4`).
+
+## What the survey forces
+
+1. **Three execution shapes, everywhere.** Inline sync (OpenAI images, all TTS, Gemini), async job with polling or webhook (every video provider, BFL, fal, Replicate, AssemblyAI), and bidirectional streams (ElevenLabs/Cartesia/Deepgram WS, realtime). Video has no sync provider at all.
+2. **Output is never just bytes.** base64, signed URLs with TTLs from 10 minutes (BFL) to 2 days (Veo), URLs that need auth plus redirect (Veo), separate download endpoints (Sora `/content?variant=`), raw bodies (Stability, TTS). Multi-output is the norm.
+3. **Inputs have roles.** First/last frame, mask, style/subject reference, source video for edit/extend, reference audio, prior generation id, provider-side file handles (`file_id`, `gs://`, `runway://`, `mm_file://`).
+4. **Partial streaming is modality-specific.** Images: a few whole partial frames. Audio: ordered chunks plus timestamp events. Jobs: status/progress/logs. Video: none.
+5. **Usage is a union**: tokens, seconds, characters (often only in headers), credits, compute time.
+6. **Moderation can be partial success** (Veo strips audio but returns video). Deprecations are constant (Sora API shuts down 2026-09-24, Imagen on Gemini API 2026-08-17).
+
+## Where Vercel is weak and we should not be
+
+- No streaming TTS at all.
+- Video job handle is experimental and only `start`/`getStatus`; polling loop is inside `generateVideo` with an injectable `delay`.
+- Unsupported inputs become silent `warnings` arrays, so a request can succeed while dropping your mask.
+- `n` is fanned out into hidden parallel calls (`maxImagesPerCall`), which obscures cost and idempotency.
+- Each modality has its own bespoke result type; the file abstraction is a lazy base64/bytes pair with no URL, expiry, or provider ref.
+- Effect's own `unstable/ai` has no media generation. Nothing in the Effect ecosystem owns this.
+
+## Design principles
+
+- **Same shape as LLM.** `X.request(...)` → Schema class; `X.generate(request)` / `X.stream(request)`; `XClient.Service` + `layer`; typed `AIError`. If you know `LLM`, you know `Video`.
+- **Execution shape is route policy, not API shape.** `Image.generate` returns an image whether the provider is inline or queued. Job control is available uniformly when you want it.
+- **Errors, not warnings.** Unsupported common fields fail at the protocol boundary with a typed `AIError`, as the LLM routes do today. Provider-side partial results (filtered audio, moderated sample) surface as `notices` on the response, never as silent drops.
+- **One asset type in, one asset type out**, shared with LLM messages and tool results.
+- **Typed per-model options**, no hidden fan-out, no implicit retries that spend money.
+- **Promise API is one mechanism for the whole package**, not a media-only wrapper.
+- **The request namespace names the modality; the model does not repeat it.** `Image.request({ model: openai("gpt-image-2") })`, not `openai.image("gpt-image-2")`.
+
+## Public API
+
+### Model selection — `ModelRef`
+
+Today a model value is built as `OpenAI.configure({ apiKey }).responses("gpt-5")` or `.image("gpt-image-2")`: `configure` fixes credentials, endpoint, and defaults; the selector fixes which of the provider's APIs to hit and binds the typed `providerOptions` generic. The selector exists because OpenAI has two LLM APIs. For media that is the exception, and the request namespace already names the modality, so repeating it in the model is ceremony.
+
+A configured facade is callable and returns a `ModelRef`. Each request namespace resolves its own route from the ref.
+
+```ts
+import { OpenAI, Google } from "@opencode/ai/providers"
+
+const openai = OpenAI.configure({ apiKey })      // OpenAI(...) alone uses env auth (OPENAI_API_KEY)
+
+LLM.request({ model: openai("gpt-5"), prompt })                 // → routes.llm   (default: responses)
+Image.request({ model: openai("gpt-image-2"), prompt })         // → routes.image
+Video.request({ model: openai("sora-2"), prompt })              // → routes.video
+Speech.request({ model: openai("gpt-4o-mini-tts"), text })      // → routes.speech
+Transcription.request({ model: openai("gpt-4o-transcribe"), audio })
+
+// Explicit selectors remain only where a provider has two APIs for one modality.
+LLM.request({ model: openai.chat("gpt-4o"), prompt })
+Image.request({ model: google.imagen("imagen-4.0-generate-001"), prompt })   // default is Gemini-native image
+```
+
+Mechanics:
+
+- `ModelRef<Routes>` carries `{ id, provider, routes }` where `Routes = { llm?: Route<…>; image?: ImageRoute<Opts>; video?: VideoRoute<Opts>; speech?: …; transcription?: … }`. Routes are lazy; unused protocols are not constructed.
+- `X.request<M extends XModel | ModelRef<{ x: XRoute<any> }>>` infers `providerOptions` from `M["routes"]["x"]`. Passing a ref whose provider has no `x` route is a compile error. Model-id validity stays a runtime provider error, as today.
+- Explicit selectors (`openai.chat`, `google.imagen`) return the concrete `XModel` directly, exactly as `.responses(...)` does now. `.model(id)` stays as an alias of the callable for LLM compatibility.
+- Provider package entrypoints keep `model(modelID, settings)` and gain the same resolution: `@opencode/ai/providers/openai` `model(...)` returns a ref; `@opencode/ai/providers/openai/responses` returns the concrete LLM model.
+- One default per modality per provider is part of the facade definition (OpenAI image → Images API, Google image → Gemini-native since Imagen on the Gemini API shuts down 2026-08-17).
+
+This applies to LLM in the same pass so the package has one way to name a model.
+
+### `Media` — the asset type
+
+Replaces `MediaPart.data: string | Uint8Array`, `ImageInput`, `GeneratedImage`, and aligns `Tool.FileContent`.
+
+```ts
+import { Media } from "@opencode/ai"
+
+Media.Source =
+  | { type: "bytes";  data: Uint8Array; mediaType: string }
+  | { type: "base64"; data: string;     mediaType: string }
+  | { type: "url";    url: string; mediaType?: string; expiresAt?: number; headers?: Record<string, string> }
+  | { type: "ref";    provider: ProviderID; id: string; mediaType?: string }   // file_id, gs://, runway://, prior generation
+
+class Media.Asset {
+  readonly source: Media.Source
+  readonly mediaType: string                    // always resolved (sniffed when the provider omits it)
+  readonly kind: "image" | "video" | "audio" | "document" | "other"
+  readonly info?: { width?; height?; durationSeconds?; sampleRate?; channels?; encoding?; format? }
+  readonly expiresAt?: number
+  readonly providerMetadata?: ProviderMetadata
+
+  bytes(): Effect<Uint8Array, AIError, RequestExecutor.Service>   // downloads/decodes lazily, cached
+  base64(): Effect<string, AIError, RequestExecutor.Service>
+  dataUrl(): Effect<string, AIError, RequestExecutor.Service>
+  materialize(): Effect<Media.Asset, AIError, RequestExecutor.Service>  // url/ref → bytes, before the URL dies
+}
+
+Media.bytes(data, mediaType?)      Media.base64(data, mediaType?)
+Media.url(url, options?)           Media.ref(provider, id)
+Media.file(path)                   // Bun/Node: reads + sniffs; Effect FileSystem variant for layers
+Media.write(asset, path)           // convenience, uses FileSystem
+```
+
+Raw-PCM outputs (Gemini TTS, Cartesia raw, Deepgram WS) carry `info.encoding/sampleRate/channels` because there is no container header.
+
+### Modality namespaces
+
+Each namespace mirrors `LLM` exactly.
+
+```ts
+import { Image, Video, Speech, Transcription } from "@opencode/ai"
+import { OpenAI, Google, ElevenLabs, Fal } from "@opencode/ai/providers"
+```
+
+#### Image
+
+```ts
+const request = Image.request({
+  model: openai("gpt-image-2"),
+  prompt: "A robot tending a rooftop garden",
+  images: [Media.file("./ref.png")],           // references / edit sources
+  mask: Media.file("./mask.png"),
+  n: 2,
+  size: "1536x1024",                            // or aspectRatio: "3:2"
+  seed: 7,
+  format: "webp",
+  providerOptions: { quality: "high", background: "transparent" },   // typed per model
+})
+
+const response = yield* Image.generate(request)   // ImageResponse
+response.image                                    // Media.Asset (first)
+response.images                                   // Media.Asset[]
+response.usage                                    // Usage union (see below)
+response.notices                                  // moderation / partial-result notices
+
+yield* Image.stream(request)                      // Stream<ImageEvent>
+// ImageEvent: job-queued | job-progress | image-partial { index, image } | image { index, image } | finish { usage }
+```
+
+Editing is not a separate function; `images`/`mask` on the request select the edit path in the route (OpenAI `/images/edits`, Gemini multimodal parts, xAI `/images/edits`). Routes that cannot honor `mask` fail with `Unsupported`.
+
+#### Video
+
+```ts
+const request = Video.request({
+  model: google("veo-3.1-generate-preview"),
+  prompt: "Panning wide shot of a calico kitten sleeping in the sunshine",
+  frames: { first: Media.file("./start.png"), last: Media.file("./end.png") },
+  references: [Media.url("https://…/style.png")],
+  video: Media.ref("openai", "video_123"),      // edit / extend / remix source
+  durationSeconds: 8,
+  aspectRatio: "16:9",
+  resolution: "1080p",
+  audio: true,
+  n: 1,
+  providerOptions: { personGeneration: "dont_allow", negativePrompt: "text, watermark" },
+})
+
+// Simple: wait for it.
+const response = yield* Video.generate(request, { poll: { interval: "10 seconds", timeout: "10 minutes" } })
+response.video                                     // Media.Asset (url with expiresAt, or bytes when the route downloads)
+yield* response.video.materialize()                // pull bytes before the URL expires
+
+// Explicit job control.
+const job = yield* Video.start(request)            // Job<VideoResponse>
+job.id; job.status; job.progress; job.token        // token is serializable JSON
+yield* job.await({ poll })                         // VideoResponse
+yield* job.cancel()
+
+// Resume from another process.
+const resumed = yield* Video.resume(model, token)  // Job<VideoResponse>
+
+// Progress as a stream.
+yield* Video.stream(request)                       // Stream<VideoEvent>: job-queued { position } | job-progress { progress, logs } | video { index, video } | finish
+```
+
+Webhooks: `Video.complete(model, token, webhook)` finishes a job from a webhook payload without polling. Token shape is route-owned and opaque (Veo operation name, fal `response_url`, Runway task id).
+
+#### Speech (TTS)
+
+```ts
+const request = Speech.request({
+  model: elevenlabs("eleven_v3"),
+  text: "Hello from OpenCode.",
+  voice: "JBFqnCBsd6RMkjVDRZzb",                  // name, uuid, or { id } — provider-normalized
+  format: "mp3",                                   // mp3 | wav | pcm | opus | aac | flac | (string & {})
+  speed: 1.0,
+  language: "en",
+  instructions: "Warm, unhurried.",
+  providerOptions: { stability: 0.5 },
+})
+
+const response = yield* Speech.generate(request)   // SpeechResponse: audio: Media.Asset, timestamps?, usage
+yield* Speech.stream(request)                      // Stream<SpeechEvent>: audio-delta { chunk } | timestamps { words } | finish
+```
+
+Streaming TTS is first-class on day one: OpenAI `stream_format: sse`, ElevenLabs `/stream`, Cartesia SSE, Deepgram chunked. Input-streaming TTS (WS, text arrives incrementally) is a later `Speech.session(...)` scoped resource, not part of `generate`.
+
+#### Transcription (STT)
+
+```ts
+const request = Transcription.request({
+  model: openai("gpt-4o-transcribe"),
+  audio: Media.file("./call.wav"),
+  language: "en",
+  prompt: "Names: Shoubhit, OpenCode.",
+  timestamps: "word",                              // none | segment | word
+  diarize: true,
+  providerOptions: { chunkingStrategy: "auto" },
+})
+
+const response = yield* Transcription.generate(request)
+response.text; response.segments; response.words; response.language; response.durationSeconds
+yield* Transcription.stream(request)               // Stream<TranscriptionEvent>: text-delta | segment | finish
+```
+
+Realtime STT over WebSocket is the same future `session` shape as input-streaming TTS.
+
+### `Job` — shared async execution
+
+```ts
+class Job<Response> {
+  readonly id: string
+  readonly model: MediaModel
+  readonly token: unknown                          // route-owned serializable JSON
+  readonly status: "queued" | "running" | "completed" | "failed" | "cancelled" | "expired"
+  readonly progress?: number                       // 0..1, normalized
+  readonly position?: number
+  readonly expiresAt?: number
+  refresh(): Effect<Job<Response>, AIError>
+  await(options?: { poll?: Poll }): Effect<Response, AIError>
+  cancel(): Effect<void, AIError>
+  events(options?): Stream<JobEvent, AIError>
+}
+
+Poll = { interval?: Duration; timeout?: Duration; schedule?: Schedule }   // route may override from provider hints (`openai-poll-after-ms`)
+```
+
+`Job` is not video-specific. Image routes on BFL, fal, and Replicate are jobs; `Image.start` exists for them. A route declares itself `inline` or `job`; `generate` on a job route is `start` then `await`.
+
+### Usage
+
+```ts
+Usage =
+  | { type: "tokens"; input; output; total; details? }
+  | { type: "seconds"; seconds }
+  | { type: "characters"; characters }
+  | { type: "credits"; credits }
+  | { type: "compute"; seconds }
+```
+
+Header-only usage (ElevenLabs `character-cost`, Deepgram `dg-char-count`) is lifted into `usage` by the route.
+
+### Promise API — `@opencode/ai/promise`
+
+Mirrors the `packages/plugin/src/effect` and `packages/plugin/src/promise` split that already exists in this repo. One mechanism for LLM and media.
+
+```ts
+import { AI } from "@opencode/ai/promise"
+
+const ai = AI.make()                                // ManagedRuntime over RequestExecutor.fetchLayer + all clients
+// AI.make({ layer }) to inject a custom executor / recorder / middleware
+
+const image = await ai.image.generate({ model, prompt })
+await image.image.bytes()
+
+for await (const event of ai.speech.stream({ model, text, voice })) { … }
+
+const job = await ai.video.start({ model, prompt })
+const video = await job.await({ poll: { interval: 10_000 }, signal })
+const resumed = ai.video.resume(model, JSON.parse(saved))
+
+const text = await ai.llm.generate({ model, prompt })   // closes today's gap: LLM has no promise API either
+for await (const event of ai.llm.stream(request)) { … }
+
+await ai.dispose()
+```
+
+Streams become `AsyncIterable` via `Stream.toAsyncIterable`. `AIError` is thrown as-is. `AbortSignal` maps to interruption. Nothing in `src/*` except this entrypoint knows about promises.
+
+### Providers
+
+Existing facades gain media routes behind the callable `ModelRef`; the modality routes each facade provides:
+
+| Facade | llm | image | video | speech | transcription | other |
+|---|---|---|---|---|---|---|
+| `OpenAI` | responses (default), chat | Images API | Sora (deprecated 2026-09-24) | ✓ | ✓ | |
+| `Google` | Gemini | Gemini-native (default), `imagen` | Veo | Gemini TTS | Gemini transcribe | |
+| `XAI` | ✓ | ✓ | ✓ | | | |
+| `ElevenLabs` | | | | ✓ | Scribe | soundEffect, music |
+| `Fal` | | ✓ | ✓ | | | |
+| `Replicate`, `Runway`, `Luma`, `Kling`, `MiniMax`, `Deepgram`, `Cartesia`, `AssemblyAI`, `BlackForestLabs`, `Stability` | | per provider | | | | |
+
+New facades follow the existing one-file-per-provider rule. Package entrypoints: `@opencode/ai/providers/openai` `model(id, settings)` returns the ref; modality-specific entrypoints such as `@opencode/ai/providers/openai/images` return the concrete model.
+
+`ImageModel<Options>` already gives typed `providerOptions` per model; `VideoModel`, `SpeechModel`, `TranscriptionModel` follow the same generic and `ModelRef` infers through to them. A shared `MediaModel` union is what `Job` and the promise client key on.
+
+### Routes and protocols
+
+Media does not fit the LLM four-axis route (SSE frames → event state machine) except for streaming TTS/STT. Reuse `Endpoint`, `Auth`, `Framing`, `RequestExecutor`, and add media protocol kinds:
+
+- `MediaProtocol.inline` — `body.from(request)` (JSON, multipart, or query), `response.decode(response)` (JSON, or binary body → `Media.Asset`).
+- `MediaProtocol.job` — `start`, `status`, `result`, `cancel`, optional `download`, `pollHint`, `token` schema.
+- `MediaProtocol.stream` — framing + `step` state machine emitting modality events, same discipline as LLM protocols.
+
+`Route.make` for media composes one protocol kind with endpoint/auth. The existing `ImageRoute { generate(request, execute) }` is the ad-hoc version of `inline` and gets folded in.
+
+### LLM integration
+
+- `MediaPart` becomes `{ type: "media"; media: Media.Asset; … }` so protocols branch on `kind` and can pass `url`/`ref` sources through natively (OpenAI `image_url`, Gemini `fileData`).
+- New `LLMEvent`s: `media { media: Media.Asset }` so Gemini inline image output is first-class instead of dropped. OpenAI Responses `image_generation_call` keeps its single carrier — the provider-executed `tool-result` with `file` content — because Core consumes hosted tool-result content today and has no `media` event handling yet; it switches to the `media` carrier when Core adopts the event, so the image is never emitted twice.
+- `Message.assistant([...])` accepts media parts; Gemini multi-turn image editing replays them.
+- `Tool.FileContent` aligns with `Media.Source`.
+
+## Decisions
+
+All settled:
+
+1. **Callable facades + `ModelRef`** replace per-modality selectors as the primary way to name a model, for LLM and media alike. Explicit selectors stay only for providers with two APIs in one modality.
+2. **`providerOptions` everywhere** (rename current `Image.options`) for consistency with LLM.
+3. **No hidden `n` fan-out.** `n` lowers natively; routes that cannot do `n > 1` fail typed. Callers use `Effect.all` / `Promise.all` explicitly.
+4. **Errors over warnings** for unsupported common fields; `notices` for provider-side partial results only.
+5. **`Media.Asset` is a class** (lazy bytes, cached) with `Media.Source` as the serializable Schema for wire/persistence. `Asset.from(source)` / `asset.source` round-trip losslessly. Same pattern as `LanguageModel` today.
+6. **Promise entrypoint**: `@opencode/ai/promise` exporting `AI.make(options?: { layer? })` plus a module-level default `ai` for scripts, covering LLM too.
+7. **Modality set for v1**: `Image`, `Video`, `Speech`, `Transcription`. `Music`/`SoundEffect` and `session` (bidirectional WS, realtime) are designed-for but deferred.
+8. **Sora is skipped** (API shuts down 2026-09-24). Video launches with Veo, xAI, fal, Runway.
+
+## Build order
+
+Foundation + Image ship together as the reference implementation, serially. Video, Speech, and Transcription then proceed in parallel on separate branches. Image jobs and partial streaming come last, after Video has hardened `Job`.
+
+## Phasing
+
+1. **Foundation** — `ModelRef` + callable facades (LLM included, `.responses`/`.chat`/`.model` kept), `Media`, `Job`, `Poll`, `Usage` union, `MediaProtocol` kinds, `@opencode/ai/promise` with `llm` + `image`. Port the five existing image protocols onto it. Unify `MediaPart` and add the `media` LLM event (fixes Gemini image output being dropped).
+2. **Video** — Veo, xAI, fal, Runway first. Then Luma, Kling, MiniMax, Replicate.
+3. **Speech + Transcription** — OpenAI, ElevenLabs, Gemini TTS, Deepgram, Cartesia, AssemblyAI. Streaming TTS from the start.
+4. **Image jobs and partials** — BFL, fal, Replicate, Stability; OpenAI `partial_images` streaming.
+5. **Later** — ElevenLabs music/SFX, Lyria, `Speech.session` / `Transcription.session`, realtime.
+
+Core adoption (session attachments beyond png/jpeg/gif/webp/pdf, image-generation tool, TUI rendering) comes after phase 1 and is a Core concern.

packages/ai/example/tutorial.ts

@@ -0,0 +1,264 @@
+import { Config, Effect, Formatter, Layer, Schema, Stream } from "effect"
+import { NodeFileSystem } from "@effect/platform-node"
+import {
+  Image,
+  ImageClient,
+  LLM,
+  LLMClient,
+  LLMRequest,
+  Media,
+  Message,
+  ProviderID,
+  Tool,
+  ToolRuntime,
+} from "@opencode/ai"
+import { Route, Auth, Endpoint, Framing, Protocol, RequestExecutor } from "@opencode/ai/route"
+import { OpenAI } from "@opencode/ai/providers"
+
+/**
+ * A runnable walkthrough of the LLM package use-site API.
+ *
+ * Run from `packages/ai` with an OpenAI key in the environment:
+ *
+ *   OPENAI_API_KEY=... bun example/tutorial.ts
+ *
+ * The file is intentionally written as a normal TypeScript program. You can
+ * hover imports and local values to see how the public API is typed.
+ */
+
+const apiKey = Config.redacted("OPENAI_API_KEY")
+
+// 1. Configure a provider. The configured facade records provider identity,
+// deployment options, authentication, and defaults. Calling it with a model id
+// returns a `ModelRef`: `LLM.request` resolves the default LLM route
+// (Responses for OpenAI) and `Image.request` resolves the image route, so one
+// value names a model for every modality. `.responses(...)` / `.chat(...)`
+// remain available when you need a specific API.
+const openai = OpenAI.configure({
+  apiKey,
+  generation: { maxTokens: 160 },
+  providerOptions: {
+    store: false,
+  },
+})
+const model = openai("gpt-4o-mini")
+
+// 2. Build a provider-neutral request. This is useful when reusing one request
+// across generate and stream examples.
+//
+// Options can live on both the configured route/provider facade and the request:
+//
+//   - `generation`: common controls such as max tokens, temperature, topP/topK,
+//     penalties, seed, and stop sequences.
+//   - `promptCacheKey`: stable cache affinity for protocols that support it.
+//   - `providerOptions`: model-typed provider-native behavior. For example,
+//     OpenAI store behavior, Anthropic thinking, Gemini thinking config, or
+//     OpenRouter routing/reasoning.
+//   - `http`: last-resort serializable overlays for final request body, headers,
+//     and query params. Prefer typed `providerOptions` when a field is stable.
+//
+// Route/provider options are defaults. Request options override them for this call.
+const request = LLM.request({
+  model,
+  system: "You are concise and practical.",
+  prompt: "Tell me a joke",
+  generation: { maxTokens: 80, temperature: 0.7 },
+  promptCacheKey: "tutorial-joke",
+})
+
+// 3. `generate` sends the request and collects the event stream into one
+// response object. `response.text` is the collected text output.
+const generateOnce = Effect.gen(function* () {
+  const response = yield* LLM.generate(request)
+
+  console.log("\n== generate ==")
+  console.log("generated text:", response.text)
+  console.log("usage", Formatter.formatJson(response.usage, { space: 2 }))
+})
+
+// 4. `stream` exposes provider output as common `LLMEvent`s for UIs that want
+// incremental text, reasoning, tool input, usage, or finish events.
+const streamText = LLM.stream(request).pipe(
+  Stream.tap((event) =>
+    Effect.sync(() => {
+      if (event.type === "text-delta") process.stdout.write(`\ntext: ${event.text}`)
+      if (event.type === "finish")
+        process.stdout.write(
+          `\nfinish: ${event.reason.normalized}${event.reason.raw ? ` (${event.reason.raw})` : ""}\n`,
+        )
+    }),
+  ),
+  Stream.runDrain,
+)
+
+// 5. Tools are typed with Effect Schema. Provider turns remain explicit:
+// advertise definitions on the request, stream one turn, dispatch local calls,
+// then persist/build follow-up history in the enclosing product flow.
+const tools = {
+  get_weather: Tool.make({
+    description: "Get current weather for a city.",
+    parameters: Schema.Struct({ city: Schema.String }),
+    success: Schema.Struct({ forecast: Schema.String }),
+    execute: (input) => Effect.succeed({ forecast: `${input.city}: sunny, 72F` }),
+  }),
+}
+
+const streamWithTools = Effect.gen(function* () {
+  const request = LLM.request({
+    model,
+    prompt: "Use get_weather for San Francisco, then answer in one sentence.",
+    generation: { maxTokens: 80, temperature: 0 },
+    tools: Tool.toDefinitions(tools),
+  })
+  const events = Array.from(yield* LLM.stream(request).pipe(Stream.runCollect))
+  for (const event of events) {
+    if (event.type === "tool-call") console.log("tool call", event.name, event.input)
+    if (event.type === "text-delta") process.stdout.write(event.text)
+    if (event.type !== "tool-call" || event.providerExecuted) continue
+    const dispatched = yield* ToolRuntime.dispatch(tools, event)
+    console.log("tool result", event.name, dispatched.result)
+
+    // A durable agent would persist these messages before starting another
+    // raw model turn. This tutorial keeps the boundary visible instead.
+    const followUp = LLMRequest.update(request, {
+      messages: [
+        ...request.messages,
+        Message.assistant([event]),
+        Message.tool({ ...event, result: dispatched.result }),
+      ],
+    })
+    console.log("follow-up history messages:", followUp.messages.length)
+  }
+})
+
+// 6. `generateObject` is the structured-output helper. It forces a synthetic
+// tool call internally, so the same call site works across providers instead of
+// depending on provider-specific JSON mode flags.
+const WeatherReport = Schema.Struct({
+  city: Schema.String,
+  forecast: Schema.String,
+  highFahrenheit: Schema.Number,
+})
+
+const generateStructuredObject = Effect.gen(function* () {
+  const response = yield* LLM.generateObject({
+    model,
+    system: "Return only structured weather data.",
+    prompt: "Give me today's weather for San Francisco.",
+    schema: WeatherReport,
+    generation: { maxTokens: 120, temperature: 0 },
+  })
+
+  console.log("\n== generateObject ==")
+  console.log(Formatter.formatJson(response.object, { space: 2 }))
+})
+
+// If the shape is only known at runtime, pass raw JSON Schema instead. The
+// `.object` type is `unknown`; callers that need static types should validate it.
+const generateDynamicObject = LLM.generateObject({
+  model,
+  prompt: "Extract the city and forecast from: San Francisco is sunny.",
+  jsonSchema: {
+    type: "object",
+    properties: {
+      city: { type: "string" },
+      forecast: { type: "string" },
+    },
+    required: ["city", "forecast"],
+  },
+})
+
+// -----------------------------------------------------------------------------
+// Part 2: provider composition with a fake provider
+// -----------------------------------------------------------------------------
+
+// A protocol is the provider-native API shape: common request -> body, response
+// frames -> common events. This fake one turns text prompts into a JSON body
+// and treats every SSE frame as output text.
+const FakeBody = Schema.Struct({
+  model: Schema.String,
+  input: Schema.String,
+})
+type FakeBody = Schema.Schema.Type<typeof FakeBody>
+
+const FakeProtocol = Protocol.make<FakeBody, string, string, void>({
+  // Protocol ids are open strings, so external packages can define their own
+  // protocols without changing this package.
+  id: "fake-echo",
+  body: {
+    schema: FakeBody,
+    from: (request) =>
+      Effect.succeed({
+        model: request.model.id,
+        input: request.messages
+          .flatMap((message) => message.content)
+          .filter((part) => part.type === "text")
+          .map((part) => part.text)
+          .join("\n"),
+      }),
+  },
+  stream: {
+    event: Schema.String,
+    initial: () => undefined,
+    step: (_, frame) => Effect.succeed([undefined, [{ type: "text-delta", id: "text-0", text: frame }]] as const),
+    onHalt: () => [{ type: "finish", reason: { normalized: "stop" } }],
+  },
+})
+
+// A route is the runnable binding for that protocol. It adds the deployment
+// axes that the protocol deliberately does not know: URL, auth, and framing.
+const FakeAdapter = Route.make({
+  id: "fake-echo",
+  provider: "fake-echo",
+  protocol: FakeProtocol,
+  endpoint: Endpoint.path("/v1/echo", { baseURL: "https://fake.local" }),
+  auth: Auth.passthrough,
+  framing: Framing.sse,
+})
+
+// A provider module exports a configured facade. Configuration happens before
+// model selection; model selectors accept ids only.
+const FakeEcho = {
+  id: ProviderID.make("fake-echo"),
+  configure: () => ({
+    id: ProviderID.make("fake-echo"),
+    model: (id: string) => FakeAdapter.model({ id }),
+  }),
+}
+
+// 8. Image generation uses the same facade and the same request/generate shape.
+// `response.image` is a `Media.Asset`: bytes decode lazily and are cached, and
+// `Media.write` persists them through the Effect `FileSystem`.
+const generateImage = Effect.gen(function* () {
+  const response = yield* Image.generate({
+    model: openai("gpt-image-1-mini"),
+    prompt: "A flat black circle centered on a plain white background.",
+    size: "1024x1024",
+    format: "jpeg",
+    providerOptions: { quality: "low" },
+  })
+
+  console.log("\n== image ==")
+  console.log("media type:", response.image.mediaType)
+  console.log("bytes:", (yield* response.image.bytes()).byteLength)
+  console.log("usage", Formatter.formatJson(response.usage, { space: 2 }))
+  yield* Media.write(response.image, "tutorial-image.jpg").pipe(Effect.provide(NodeFileSystem.layer))
+})
+
+// Provide the LLM runtime and the HTTP request executor once. Keep one path
+// enabled at a time so the tutorial can demonstrate generate, stream, or
+// tool-loop behavior without spending tokens on every example.
+const requestExecutorLayer = RequestExecutor.fetchLayer
+const llmClientLayer = LLMClient.layer.pipe(Layer.provide(requestExecutorLayer))
+const imageClientLayer = ImageClient.layer.pipe(Layer.provide(requestExecutorLayer))
+
+const program = Effect.gen(function* () {
+  // yield* generateOnce
+  // yield* streamText
+  // yield* generateStructuredObject
+  // yield* generateDynamicObject.pipe(Effect.andThen((response) => Effect.sync(() => console.log(response.object))))
+  // yield* generateImage
+  yield* streamWithTools
+}).pipe(Effect.provide(Layer.mergeAll(requestExecutorLayer, llmClientLayer, imageClientLayer)))
+
+Effect.runPromise(program)

packages/ai/package.json

@@ -0,0 +1,40 @@
+{
+  "$schema": "https://json.schemastore.org/package.json",
+  "version": "2.0.3",
+  "name": "@opencode/ai",
+  "type": "module",
+  "license": "MIT",
+  "scripts": {
+    "setup:recording-env": "bun run script/setup-recording-env.ts",
+    "test": "bun test --timeout 30000 --only-failures",
+    "typecheck": "tsgo --noEmit && tsgo --noEmit -p tsconfig.types.json",
+    "build": "tsc -p tsconfig.build.json"
+  },
+  "files": [
+    "dist"
+  ],
+  "exports": {
+    ".": "./src/index.ts",
+    "./promise": "./src/promise.ts",
+    "./testing": "./src/testing.ts",
+    "./*": "./src/*.ts"
+  },
+  "devDependencies": {
+    "@clack/prompts": "1.0.0-alpha.1",
+    "@effect/platform-node": "catalog:",
+    "@opencode/http-recorder": "workspace:*",
+    "@tsconfig/bun": "catalog:",
+    "@types/bun": "catalog:",
+    "@typescript/native-preview": "catalog:",
+    "typescript": "catalog:"
+  },
+  "dependencies": {
+    "@aws-sdk/credential-providers": "3.1057.0",
+    "@smithy/eventstream-codec": "4.2.14",
+    "@smithy/util-utf8": "4.2.2",
+    "@opencode/schema": "workspace:*",
+    "aws4fetch": "1.0.20",
+    "effect": "catalog:",
+    "google-auth-library": "10.5.0"
+  }
+}

packages/ai/script/publish.ts

@@ -0,0 +1,38 @@
+#!/usr/bin/env bun
+import { Script } from "@opencode/script"
+import { $ } from "bun"
+import { fileURLToPath } from "url"
+
+const dir = fileURLToPath(new URL("..", import.meta.url))
+process.chdir(dir)
+
+async function published(name: string, version: string) {
+  return (await $`npm view ${name}@${version} version`.nothrow()).exitCode === 0
+}
+
+await $`bun run build`
+const originalText = await Bun.file("package.json").text()
+const pkg = JSON.parse(originalText) as {
+  name: string
+  version: string
+  exports: Record<string, string>
+}
+if (await published(pkg.name, pkg.version)) {
+  console.log(`already published ${pkg.name}@${pkg.version}`)
+} else {
+  for (const [key, value] of Object.entries(pkg.exports)) {
+    const file = value.replace("./src/", "./dist/").replace(".ts", "")
+    // @ts-ignore
+    pkg.exports[key] = {
+      import: file + ".js",
+      types: file + ".d.ts",
+    }
+  }
+  await Bun.write("package.json", JSON.stringify(pkg, null, 2))
+  try {
+    await $`bun pm pack`
+    await $`npm publish *.tgz --tag ${Script.channel} --access public`
+  } finally {
+    await Bun.write("package.json", originalText)
+  }
+}

packages/ai/script/recording-cost-report.ts

@@ -2,7 +2,7 @@ import * as fs from "node:fs/promises"
 import * as path from "node:path"
 
 const RECORDINGS_DIR = path.resolve(import.meta.dir, "..", "test", "fixtures", "recordings")
-const MODELS_DEV_URL = "https://models.dev/api.json"
+const MODELS_DEV_URL = "https://models.opencode.ai/api.json"
 
 type JsonRecord = Record<string, unknown>
 

packages/ai/script/setup-recording-env.ts

@@ -7,7 +7,8 @@ import { AwsV4Signer } from "aws4fetch"
 import { Config, ConfigProvider, Effect, FileSystem, PlatformError, Redacted } from "effect"
 import { FetchHttpClient, HttpClient, HttpClientRequest, type HttpClientResponse } from "effect/unstable/http"
 import * as ProviderShared from "../src/protocols/shared"
-import * as Cloudflare from "../src/providers/cloudflare"
+import { CloudflareAIGateway } from "../src/providers/cloudflare-ai-gateway.js"
+import { CloudflareWorkersAI } from "../src/providers/cloudflare-workers-ai.js"
 
 type Provider = {
   readonly id: string
@@ -120,11 +121,11 @@ const PROVIDERS: ReadonlyArray<Provider> = [
     ],
     validate: (env) =>
       validateChat({
-        url: `${Cloudflare.aiGatewayBaseURL({
+        url: `${CloudflareAIGateway.baseURL({
           accountId: env.CLOUDFLARE_ACCOUNT_ID,
           gatewayId: env.CLOUDFLARE_GATEWAY_ID || undefined,
         })}/chat/completions`,
-        token: Redacted.make(envValue(env, Cloudflare.aiGatewayAuthEnvVars)),
+        token: Redacted.make(envValue(env, CloudflareAIGateway.authEnvVars)),
         tokenHeader: "cf-aig-authorization",
         model: "workers-ai/@cf/meta/llama-3.1-8b-instruct",
       }),
@@ -140,8 +141,8 @@ const PROVIDERS: ReadonlyArray<Provider> = [
     ],
     validate: (env) =>
       validateChat({
-        url: `${Cloudflare.workersAIBaseURL({ accountId: env.CLOUDFLARE_ACCOUNT_ID })}/chat/completions`,
-        token: Redacted.make(envValue(env, Cloudflare.workersAIAuthEnvVars)),
+        url: `${CloudflareWorkersAI.baseURL({ accountId: env.CLOUDFLARE_ACCOUNT_ID })}/chat/completions`,
+        token: Redacted.make(envValue(env, CloudflareWorkersAI.authEnvVars)),
         model: "@cf/meta/llama-3.1-8b-instruct",
       }),
   },
@@ -157,9 +158,21 @@ const PROVIDERS: ReadonlyArray<Provider> = [
     id: "togetherai",
     label: "TogetherAI",
     tier: "compatible",
-    note: "Existing OpenAI-compatible text/tool recorded tests",
-    vars: [{ name: "TOGETHER_AI_API_KEY" }],
-    validate: (env) => validateBearer("https://api.together.xyz/v1/models", Redacted.make(env.TOGETHER_AI_API_KEY)),
+    note: "Native Together AI text/tool recorded tests",
+    vars: [{ name: "TOGETHER_API_KEY" }],
+    validate: (env) => validateBearer("https://api.together.xyz/v1/models", Redacted.make(env.TOGETHER_API_KEY)),
+  },
+  {
+    id: "minimax",
+    label: "MiniMax",
+    tier: "compatible",
+    note: "Anthropic-compatible Messages text/tool recorded tests",
+    vars: [{ name: "MINIMAX_API_KEY" }],
+    validate: (env) =>
+      HttpClientRequest.get("https://api.minimax.io/anthropic/v1/models").pipe(
+        HttpClientRequest.setHeader("x-api-key", Redacted.value(Redacted.make(env.MINIMAX_API_KEY))),
+        executeRequest,
+      ),
   },
   {
     id: "mistral",
@@ -188,8 +201,8 @@ const PROVIDERS: ReadonlyArray<Provider> = [
   {
     id: "cerebras",
     label: "Cerebras",
-    tier: "optional",
-    note: "OpenAI-compatible bridge",
+    tier: "compatible",
+    note: "Native Cerebras text/tool/tool-loop recorded tests",
     vars: [{ name: "CEREBRAS_API_KEY" }],
     validate: (env) => validateBearer("https://api.cerebras.ai/v1/models", Redacted.make(env.CEREBRAS_API_KEY)),
   },
@@ -358,7 +371,7 @@ const responseError = Effect.fn("RecordingEnv.responseError")(function* (
   response: HttpClientResponse.HttpClientResponse,
 ) {
   if (response.status >= 200 && response.status < 300) return undefined
-  const body = yield* response.text.pipe(Effect.catch(() => Effect.succeed("")))
+  const body = yield* response.text.pipe(Effect.orElseSucceed(() => ""))
   return `${response.status}${body ? `: ${body.slice(0, 180)}` : ""}`
 })
 

packages/ai/src/cache-policy.ts

@@ -0,0 +1,165 @@
+// Apply an `LLMRequest.cache` policy by injecting `CacheHint`s onto the parts
+// the policy designates. Runs once at compile time, before the per-protocol
+// body builder, so the existing inline-hint lowering path handles the rest.
+//
+// The default `"auto"` shape places breakpoints at the last tool definition,
+// the first and last distinct system parts, and the conversation tail. This
+// exposes reusable tool, base-agent, project, and session prefixes while
+// advancing the tail after each tool result keeps recent conversation prefixes
+// reusable during long agent runs.
+//
+// Manual `cache: CacheHint` placements on individual parts are preserved and
+// count against the four-breakpoint budget; auto only fills remaining slots.
+import { CacheHint, type CachePolicy, type CachePolicyObject } from "./schema/options.js"
+import { LLMRequest, Message, ToolDefinition, type ContentPart, type ToolEntry } from "./schema/messages.js"
+import { effortUpdate } from "./effort-updates.js"
+
+const AUTO: CachePolicyObject = {
+  tools: true,
+  system: true,
+  messages: { tail: 1 },
+}
+
+const NONE: CachePolicyObject = {}
+const BREAKPOINT_CAP = 4
+
+// Resolution rules:
+//   - undefined   → "auto" — caching is on by default.
+//   - "auto"      → tools + first/last system + final message boundary.
+//   - "none"      → no auto placement; manual `CacheHint`s still flow.
+//   - object form → exactly what the caller asked for.
+const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
+  if (policy === undefined || policy === "auto") return AUTO
+  if (policy === "none") return NONE
+  return policy
+}
+
+// Protocols whose wire format ignores inline cache markers (OpenAI's implicit
+// prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
+// whole policy pass for these — emitting hints would be harmless but pointless.
+const RESPECTS_INLINE_HINTS = new Set([
+  "anthropic-messages",
+  "google-vertex-messages",
+  "bedrock-converse",
+  "openrouter",
+])
+
+const makeHint = (ttlSeconds: number | undefined): CacheHint =>
+  ttlSeconds !== undefined ? new CacheHint({ type: "ephemeral", ttlSeconds }) : new CacheHint({ type: "ephemeral" })
+
+interface Budget {
+  remaining: number
+}
+
+const markLastTool = (tools: ReadonlyArray<ToolEntry>, hint: CacheHint, budget: Budget): ReadonlyArray<ToolEntry> => {
+  const target = tools.at(-1)
+  if (target === undefined) return tools
+  if (target.type === "namespace") {
+    const nested = markLastTool(target.tools, hint, budget)
+    return nested === target.tools ? tools : [...tools.slice(0, -1), { ...target, tools: nested }]
+  }
+  if (target.cache || budget.remaining === 0) return tools
+  budget.remaining -= 1
+  return [...tools.slice(0, -1), new ToolDefinition({ ...target, cache: hint })]
+}
+
+const countToolHints = (tools: ReadonlyArray<ToolEntry>): number =>
+  tools.reduce(
+    (count, tool) => count + (tool.type === "tool" ? (tool.cache === undefined ? 0 : 1) : countToolHints(tool.tools)),
+    0,
+  )
+
+const markSystemBoundaries = (system: LLMRequest["system"], hint: CacheHint, budget: Budget): LLMRequest["system"] => {
+  if (system.length === 0) return system
+  let changed = false
+  const next = system.map((part, index) => {
+    if ((index !== 0 && index !== system.length - 1) || part.cache || budget.remaining === 0) return part
+    budget.remaining -= 1
+    changed = true
+    return { ...part, cache: hint }
+  })
+  return changed ? next : system
+}
+
+const lastIndexOfRole = (messages: ReadonlyArray<Message>, role: Message["role"]): number =>
+  messages.findLastIndex((m) => m.role === role)
+
+// Mark the last text part of `messages[index]`. If no text part exists, mark
+// the last content part regardless of type — that's the breakpoint position
+// in tool-result-only messages too.
+const markMessageAt = (
+  messages: ReadonlyArray<Message>,
+  index: number,
+  hint: CacheHint,
+  budget: Budget,
+): ReadonlyArray<Message> => {
+  if (index < 0 || index >= messages.length) return messages
+  const target = messages[index]!
+  if (target.content.length === 0) return messages
+  const lastTextIndex = target.content.findLastIndex((part) => part.type === "text")
+  const markAt = lastTextIndex >= 0 ? lastTextIndex : target.content.length - 1
+  const existing = target.content[markAt]!
+  if (("cache" in existing && existing.cache) || budget.remaining === 0) return messages
+  budget.remaining -= 1
+  const nextContent = target.content.map((part, i) => (i === markAt ? ({ ...part, cache: hint } as ContentPart) : part))
+  const next = new Message({ ...target, content: nextContent })
+  // Single pass over `messages`, substituting the one updated entry. Long
+  // conversations call this on every request, so avoid `.map()` here — its
+  // closure dispatch and identity copies show up in profiling.
+  const result = messages.slice()
+  result[index] = next
+  return result
+}
+
+const markMessages = (
+  messages: ReadonlyArray<Message>,
+  strategy: NonNullable<CachePolicyObject["messages"]>,
+  hint: CacheHint,
+  budget: Budget,
+): ReadonlyArray<Message> => {
+  if (messages.length === 0) return messages
+  if (strategy === "latest-user-message")
+    return markMessageAt(messages, lastIndexOfRole(messages, "user"), hint, budget)
+  if (strategy === "latest-assistant")
+    return markMessageAt(messages, lastIndexOfRole(messages, "assistant"), hint, budget)
+  let start = messages.length
+  let remaining = strategy.tail
+  while (remaining > 0 && start > 0) {
+    start -= 1
+    if (effortUpdate(messages[start]!) === undefined) remaining -= 1
+  }
+  let next = messages
+  for (let i = start; i < messages.length; i++)
+    if (effortUpdate(messages[i]!) === undefined) next = markMessageAt(next, i, hint, budget)
+  return next
+}
+
+const countHints = (request: LLMRequest) =>
+  countToolHints(request.tools) +
+  request.system.reduce((count, part) => count + (part.cache === undefined ? 0 : 1), 0) +
+  request.messages.reduce(
+    (count, message) =>
+      count +
+      message.content.reduce(
+        (contentCount, part) => contentCount + ("cache" in part && part.cache !== undefined ? 1 : 0),
+        0,
+      ),
+    0,
+  )
+
+export const applyCachePolicy = (request: LLMRequest): LLMRequest => {
+  if (!RESPECTS_INLINE_HINTS.has(request.model.route.id)) return request
+  if (request.model.route.id === "openrouter" && (request.cache === undefined || request.cache === "auto"))
+    return request
+  const policy = resolve(request.cache)
+  if (!policy.tools && !policy.system && !policy.messages) return request
+
+  const hint = makeHint(policy.ttlSeconds)
+  const budget = { remaining: Math.max(0, BREAKPOINT_CAP - countHints(request)) }
+  const tools = policy.tools ? markLastTool(request.tools, hint, budget) : request.tools
+  const system = policy.system ? markSystemBoundaries(request.system, hint, budget) : request.system
+  const messages = policy.messages ? markMessages(request.messages, policy.messages, hint, budget) : request.messages
+
+  if (tools === request.tools && system === request.system && messages === request.messages) return request
+  return LLMRequest.update(request, { tools, system, messages })
+}

packages/ai/src/effort-updates.ts

@@ -0,0 +1,25 @@
+// Changing the top-level reasoning effort invalidates the provider prompt cache. Protocols with a native
+// per-message update keep it frozen and lower `Message.effort(...)` markers instead; other routes strip them.
+import { LLMRequest, type EffortPart, type Message } from "./schema/messages.js"
+
+export const effortUpdate = (message: Message): EffortPart | undefined => {
+  if (message.role !== "system" || message.content.length !== 1) return undefined
+  const part = message.content[0]
+  return part.type === "effort" ? part : undefined
+}
+
+export const stripEffortUpdates = (request: LLMRequest) => {
+  const messages = request.messages.filter((message) => effortUpdate(message) === undefined)
+  return messages.length === request.messages.length ? request : LLMRequest.update(request, { messages })
+}
+
+export const applyEffortUpdates = (request: LLMRequest): LLMRequest =>
+  request.model.route.supportsEffortUpdates?.(request) ? request : stripEffortUpdates(request)
+
+// Reverted or forked history can leave the last marker disagreeing with the requested effort.
+export const resolveEffortUpdates = (request: LLMRequest, current: string | undefined) => {
+  const updates = request.messages.flatMap((message) => effortUpdate(message) ?? [])
+  if (updates.length === 0) return { request, effort: current }
+  if (updates.at(-1)?.effort !== current) return { request: stripEffortUpdates(request), effort: current }
+  return { request, effort: updates[0]?.previous }
+}

packages/ai/src/image-client.ts

@@ -0,0 +1,61 @@
+import { Context, Effect, Layer, Stream } from "effect"
+import { RequestExecutor } from "./route/executor.js"
+import type { AIError } from "./schema/index.js"
+import {
+  responseEvents,
+  type ImageEvent,
+  type ImageOptions,
+  type ImageRequestFor,
+  type ImageResponse,
+} from "./image.js"
+
+export interface Interface {
+  readonly generate: <Options extends ImageOptions>(
+    request: ImageRequestFor<Options>,
+  ) => Effect.Effect<ImageResponse, AIError>
+  readonly stream: <Options extends ImageOptions>(
+    request: ImageRequestFor<Options>,
+  ) => Stream.Stream<ImageEvent, AIError>
+}
+
+export class Service extends Context.Service<Service, Interface>()("@opencode/ImageClient") {}
+
+export const generate = <Options extends ImageOptions>(
+  request: ImageRequestFor<Options>,
+): Effect.Effect<ImageResponse, AIError, Service> =>
+  Effect.gen(function* () {
+    const client = yield* Service
+    return yield* client.generate(request)
+  })
+
+export const stream = <Options extends ImageOptions>(
+  request: ImageRequestFor<Options>,
+): Stream.Stream<ImageEvent, AIError, Service> =>
+  Stream.unwrap(
+    Effect.gen(function* () {
+      const client = yield* Service
+      return client.stream(request)
+    }),
+  )
+
+export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer.effect(
+  Service,
+  Effect.gen(function* () {
+    const executor = yield* RequestExecutor.Service
+    const generate = <Options extends ImageOptions>(request: ImageRequestFor<Options>) =>
+      request.model.route.generate(request, executor.execute)
+    return Service.of({
+      generate,
+      // Inline routes have no partial frames yet; the stream is the completed response expanded into events.
+      stream: (request) =>
+        Stream.unwrap(generate(request).pipe(Effect.map((response) => Stream.fromIterable(responseEvents(response))))),
+    })
+  }),
+)
+
+export const ImageClient = {
+  Service,
+  layer,
+  generate,
+  stream,
+} as const

packages/ai/src/image.ts

@@ -0,0 +1,272 @@
+import { Effect, Schema, Stream } from "effect"
+import { Media } from "./media.js"
+import { ModelRef, type ResolveImageModel } from "./model-ref.js"
+import { Endpoint } from "./route/endpoint.js"
+import { MediaRoute } from "./route/media.js"
+import type { MediaProtocol } from "./route/media-protocol.js"
+import {
+  AIError,
+  HttpOptions,
+  InvalidRequestError,
+  MediaUsage,
+  ModelID,
+  ProviderID,
+  ProviderMetadata,
+  UnsupportedOperationError,
+} from "./schema/index.js"
+import { ImageClient, Service } from "./image-client.js"
+
+// ---------------------------------------------------------------------------
+// Model
+// ---------------------------------------------------------------------------
+
+export type ImageOptions = Record<string, unknown>
+
+export type ImageRoute<Options extends ImageOptions = ImageOptions> = MediaRoute.Route<
+  ImageRequestFor<Options>,
+  ImageResponse
+>
+
+export class ImageModel<Options extends ImageOptions = ImageOptions> {
+  declare protected readonly _Options: (options: Options) => Options
+  readonly id: ModelID
+  readonly provider: ProviderID
+  readonly route: ImageRoute<Options>
+  readonly http?: HttpOptions
+
+  constructor(input: ImageModel.Input<Options>) {
+    this.id = input.id
+    this.provider = input.provider
+    this.route = input.route
+    this.http = input.http
+  }
+
+  static make<Options extends ImageOptions = ImageOptions>(input: ImageModel.MakeInput<Options>) {
+    return new ImageModel<Options>({
+      id: ModelID.make(input.id),
+      provider: ProviderID.make(input.provider),
+      route: input.route,
+      http: input.http,
+    })
+  }
+
+  /** Compose an inline image protocol with its canonical path into a model for one deployment. */
+  static fromRoute<Options extends ImageOptions = ImageOptions>(
+    route: ImageModel.RouteInput<Options>,
+    input: MediaRoute.ModelInput,
+  ) {
+    return ImageModel.make<Options>({
+      id: input.id,
+      provider: route.provider,
+      http: input.http,
+      route: MediaRoute.make({
+        id: route.id,
+        provider: route.provider,
+        protocol: route.protocol,
+        endpoint: Endpoint.path(route.path, { baseURL: input.baseURL ?? route.baseURL }),
+        auth: input.auth,
+        headers: input.headers,
+      }),
+    })
+  }
+}
+
+export namespace ImageModel {
+  export interface Input<Options extends ImageOptions = ImageOptions> {
+    readonly id: ModelID
+    readonly provider: ProviderID
+    readonly route: ImageRoute<Options>
+    readonly http?: HttpOptions
+  }
+
+  export interface MakeInput<Options extends ImageOptions = ImageOptions>
+    extends Omit<Input<Options>, "id" | "provider"> {
+    readonly id: string | ModelID
+    readonly provider: string | ProviderID
+  }
+
+  export interface RouteInput<Options extends ImageOptions = ImageOptions> {
+    readonly id: string
+    readonly provider: string | ProviderID
+    readonly protocol: MediaProtocol.Inline<ImageRequestFor<Options>, ImageResponse>
+    readonly path: Endpoint.EndpointPart<MediaProtocol.Body, ImageRequestFor<Options>>
+    /** Canonical base URL; `ModelInput.baseURL` overrides it per deployment. */
+    readonly baseURL?: string
+  }
+}
+
+export const ImageModelSchema = Schema.declare((value): value is ImageModel => value instanceof ImageModel, {
+  expected: "Image.Model",
+})
+
+// ---------------------------------------------------------------------------
+// Request
+// ---------------------------------------------------------------------------
+
+export type ImageSize = `${number}x${number}`
+export const ImageSize = Schema.declare<ImageSize>(
+  (value): value is ImageSize => typeof value === "string" && /^\d+x\d+$/.test(value),
+  { title: "ImageSize" },
+)
+
+export type ImageAspectRatio = `${number}:${number}`
+export const ImageAspectRatio = Schema.declare<ImageAspectRatio>(
+  (value): value is ImageAspectRatio => typeof value === "string" && /^\d+(?:\.\d+)?:\d+(?:\.\d+)?$/.test(value),
+  { title: "ImageAspectRatio" },
+)
+
+export type ImageFormat = "png" | "jpeg" | "webp" | (string & {})
+
+export class ImageRequest extends Schema.Class<ImageRequest>("Image.Request")({
+  model: ImageModelSchema,
+  prompt: Schema.String,
+  /** Edit sources or style/subject references, in order. */
+  images: Schema.optional(Schema.Array(Media.AssetSchema)),
+  /** Inpainting mask; routes that cannot honor it fail with `UnsupportedOperation`. */
+  mask: Schema.optional(Media.AssetSchema),
+  n: Schema.optional(Schema.Int),
+  size: Schema.optional(ImageSize),
+  aspectRatio: Schema.optional(ImageAspectRatio),
+  seed: Schema.optional(Schema.Number),
+  format: Schema.optional(Schema.String),
+  providerOptions: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+  http: Schema.optional(HttpOptions),
+}) {
+  declare protected readonly _ImageRequest: void
+}
+
+export type ImageRequestFor<Options extends ImageOptions = ImageOptions> = Omit<
+  ImageRequest,
+  "model" | "providerOptions"
+> & {
+  readonly model: ImageModel<Options>
+  readonly providerOptions?: Options
+}
+
+export type ImageModelInput = ImageModel | ModelRef.WithImage
+
+export type ImageModelOptions<Model> = ResolveImageModel<Model> extends ImageModel<infer Options> ? Options : never
+
+export type ImageRequestInput<Model extends ImageModelInput = ImageModelInput> = Omit<
+  ConstructorParameters<typeof ImageRequest>[0],
+  "model" | "providerOptions" | "http"
+> & {
+  readonly model: Model
+  readonly format?: ImageFormat
+  readonly providerOptions?: NoInfer<ImageModelOptions<Model>>
+  readonly http?: HttpOptions.Input
+}
+
+// ---------------------------------------------------------------------------
+// Response and events
+// ---------------------------------------------------------------------------
+
+export class ImageResponse extends Schema.Class<ImageResponse>("Image.Response")({
+  images: Schema.Array(Media.AssetSchema),
+  usage: Schema.optional(MediaUsage),
+  notices: Schema.optional(Schema.Array(Media.Notice)),
+  providerMetadata: Schema.optional(ProviderMetadata),
+}) {
+  get image() {
+    return this.images[0]
+  }
+}
+
+export const ImageOutputEvent = Schema.Struct({
+  type: Schema.tag("image"),
+  index: Schema.Number,
+  image: Media.AssetSchema,
+}).annotate({ identifier: "Image.Event.Image" })
+
+export const ImageFinishEvent = Schema.Struct({
+  type: Schema.tag("finish"),
+  usage: Schema.optional(MediaUsage),
+  notices: Schema.optional(Schema.Array(Media.Notice)),
+  providerMetadata: Schema.optional(ProviderMetadata),
+}).annotate({ identifier: "Image.Event.Finish" })
+
+const imageEventTagged = Schema.Union([ImageOutputEvent, ImageFinishEvent]).pipe(Schema.toTaggedUnion("type"))
+export const ImageEvent = Object.assign(imageEventTagged, {
+  is: {
+    image: imageEventTagged.guards.image,
+    finish: imageEventTagged.guards.finish,
+  },
+})
+export type ImageEvent = Schema.Schema.Type<typeof imageEventTagged>
+
+/** Inline routes produce every image at once; expand the response into the streaming event shape. */
+export const responseEvents = (response: ImageResponse): ReadonlyArray<ImageEvent> => [
+  ...response.images.map((image, index) => ImageOutputEvent.make({ index, image })),
+  ImageFinishEvent.make({
+    usage: response.usage,
+    notices: response.notices,
+    providerMetadata: response.providerMetadata,
+  }),
+]
+
+// ---------------------------------------------------------------------------
+// Request-shaped call API
+// ---------------------------------------------------------------------------
+
+const resolveModel = (model: ImageModelInput): ImageModel => {
+  if (model instanceof ImageModel) return model
+  // Typed callers cannot reach the throw; it guards JS callers passing a ref without an image selector.
+  const selected = model.facade.image?.(model.id)
+  if (selected) return selected
+  throw new AIError({
+    reason: new UnsupportedOperationError({
+      operation: "image",
+      provider: model.provider,
+      message: `${model.provider} does not expose an image route for ${model.id}`,
+    }),
+  })
+}
+
+export function request<const Model extends ImageModelInput>(
+  input: ImageRequestInput<Model>,
+): ImageRequestFor<ImageModelOptions<Model>>
+export function request(input: ImageRequest): ImageRequest
+export function request(input: ImageRequest | ImageRequestInput) {
+  if (input instanceof ImageRequest) return input
+  return new ImageRequest({
+    ...input,
+    model: resolveModel(input.model),
+    http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+  })
+}
+
+const requestEffect = (input: ImageRequest | ImageRequestInput) =>
+  Effect.try({
+    try: () => request(input),
+    catch: (error) =>
+      error instanceof AIError
+        ? error
+        : new AIError({
+            reason: new InvalidRequestError({
+              message: error instanceof Error ? error.message : String(error),
+              cause: error,
+            }),
+          }),
+  })
+
+export function generate<const Model extends ImageModelInput>(
+  input: ImageRequestInput<Model>,
+): Effect.Effect<ImageResponse, AIError, Service>
+export function generate(input: ImageRequest): Effect.Effect<ImageResponse, AIError, Service>
+export function generate(input: ImageRequest | ImageRequestInput) {
+  return requestEffect(input).pipe(Effect.flatMap((request) => ImageClient.generate(request)))
+}
+
+export function stream<const Model extends ImageModelInput>(
+  input: ImageRequestInput<Model>,
+): Stream.Stream<ImageEvent, AIError, Service>
+export function stream(input: ImageRequest): Stream.Stream<ImageEvent, AIError, Service>
+export function stream(input: ImageRequest | ImageRequestInput) {
+  return Stream.unwrap(requestEffect(input).pipe(Effect.map((request) => ImageClient.stream(request))))
+}
+
+export const Image = {
+  request,
+  generate,
+  stream,
+} as const

packages/ai/src/index.ts

@@ -0,0 +1,63 @@
+export { LLMClient } from "./route/client.js"
+export { ImageClient } from "./image-client.js"
+export { Auth } from "./route/auth.js"
+export { Provider } from "./provider.js"
+export { ProviderPackage } from "./provider-package.js"
+export { isContextOverflow, isContextOverflowFailure } from "./provider-error.js"
+export type {
+  RouteLanguageModelInput,
+  RouteRoutedLanguageModelInput,
+  Interface as LLMClientShape,
+  Service as LLMClientService,
+} from "./route/client.js"
+export * from "./schema/index.js"
+export {
+  ImageAspectRatio,
+  ImageEvent,
+  ImageModel,
+  ImageModelSchema,
+  ImageRequest,
+  ImageResponse,
+  ImageSize,
+} from "./image.js"
+export type {
+  ImageFormat,
+  ImageModelInput,
+  ImageModelOptions,
+  ImageOptions,
+  ImageRequestFor,
+  ImageRequestInput,
+  ImageRoute,
+} from "./image.js"
+export { Image } from "./image.js"
+export { Media } from "./media.js"
+export { ModelRef } from "./model-ref.js"
+export type { ResolveImageModel, ResolveLanguageModel } from "./model-ref.js"
+export { Job } from "./job.js"
+export type { Event as JobEvent, Poll, Route as JobRoute, Snapshot as JobSnapshot, Status as JobStatus } from "./job.js"
+export { Tool, ToolFailure, toDefinitions } from "./tool.js"
+export { ToolRuntime } from "./tool-runtime.js"
+export type { DispatchResult as ToolDispatchResult, ToolSettlement } from "./tool-runtime.js"
+export type {
+  AnyExecutableTool,
+  AnyTool,
+  ExecutableTool,
+  ExecutableTools,
+  Definition as ToolShape,
+  ToolExecute,
+  ToolExecuteContext,
+  ToolModelOutputInput,
+  Tools,
+  ToolSchema,
+  ToolToModelOutput,
+} from "./tool.js"
+export * as LLM from "./llm.js"
+export type {
+  Definition as ProviderDefinition,
+  LanguageModelFactory as ProviderLanguageModelFactory,
+  LanguageModelOptions as ProviderLanguageModelOptions,
+} from "./provider.js"
+export type {
+  Definition as ProviderPackageDefinition,
+  Settings as ProviderPackageSettings,
+} from "./provider-package.js"

packages/ai/src/job.ts

@@ -0,0 +1,137 @@
+import { Duration, Effect, Schedule, Schema, Stream } from "effect"
+import { AIError, TimeoutError } from "./schema/errors.js"
+
+export const Status = Schema.Literals(["queued", "running", "completed", "failed", "cancelled", "expired"])
+export type Status = Schema.Schema.Type<typeof Status>
+
+/** Provider-neutral view of one job observation. */
+export interface Snapshot {
+  readonly id: string
+  readonly status: Status
+  /** Normalized 0..1 when the provider reports progress. */
+  readonly progress?: number
+  readonly position?: number
+  readonly expiresAt?: number
+}
+
+/**
+ * Route-owned job operations. `token` is the route's serializable handle (operation name, task id, response URL)
+ * so a job can be resumed from another process; its shape is opaque to `Job`.
+ */
+export interface Route<Response> {
+  readonly status: (token: unknown) => Effect.Effect<Snapshot, AIError>
+  readonly result: (token: unknown) => Effect.Effect<Response, AIError>
+  readonly cancel?: (token: unknown) => Effect.Effect<void, AIError>
+  /** Provider polling hint (e.g. `openai-poll-after-ms`) that overrides the default interval for the next poll. */
+  readonly pollHint?: (snapshot: Snapshot) => Duration.Duration | undefined
+}
+
+export interface Poll {
+  readonly interval?: Duration.Input
+  readonly timeout?: Duration.Input
+  /** Full override of the polling schedule; `interval` and `pollHint` are ignored when supplied. */
+  readonly schedule?: Schedule.Schedule<unknown, Snapshot>
+}
+
+export const DEFAULT_POLL_INTERVAL = Duration.seconds(5)
+export const DEFAULT_POLL_TIMEOUT = Duration.minutes(10)
+
+export type Event =
+  | { readonly type: "job-queued"; readonly id: string; readonly position?: number }
+  | { readonly type: "job-progress"; readonly id: string; readonly progress?: number }
+  | { readonly type: "job-finished"; readonly id: string; readonly status: Status }
+
+const TERMINAL: ReadonlySet<Status> = new Set(["completed", "failed", "cancelled", "expired"])
+
+export class Job<Response> {
+  readonly id: string
+  readonly status: Status
+  readonly progress?: number
+  readonly position?: number
+  readonly expiresAt?: number
+
+  constructor(
+    readonly route: Route<Response>,
+    readonly token: unknown,
+    snapshot: Snapshot,
+  ) {
+    this.id = snapshot.id
+    this.status = snapshot.status
+    this.progress = snapshot.progress
+    this.position = snapshot.position
+    this.expiresAt = snapshot.expiresAt
+  }
+
+  get snapshot(): Snapshot {
+    return {
+      id: this.id,
+      status: this.status,
+      progress: this.progress,
+      position: this.position,
+      expiresAt: this.expiresAt,
+    }
+  }
+
+  get terminal() {
+    return TERMINAL.has(this.status)
+  }
+
+  refresh(): Effect.Effect<Job<Response>, AIError> {
+    return this.route.status(this.token).pipe(Effect.map((snapshot) => new Job(this.route, this.token, snapshot)))
+  }
+
+  /** Poll until the job reaches a terminal status, then fetch the result. Fails with a `Timeout` reason on deadline. */
+  await(options?: { readonly poll?: Poll }): Effect.Effect<Response, AIError> {
+    const timeout = Duration.fromInputUnsafe(options?.poll?.timeout ?? DEFAULT_POLL_TIMEOUT)
+    const settled = this.terminal ? Effect.succeed(this) : this.poll(options?.poll)
+    return settled.pipe(
+      // Non-completed terminal states also go through `result` so the route can surface its provider failure body.
+      Effect.flatMap((job) => job.route.result(job.token)),
+      Effect.timeoutOrElse({
+        duration: timeout,
+        orElse: () =>
+          new AIError({
+            reason: new TimeoutError({
+              message: `Job ${this.id} did not finish within ${Duration.format(timeout)}`,
+              timeoutMs: Duration.toMillis(timeout),
+            }),
+          }),
+      }),
+    )
+  }
+
+  cancel(): Effect.Effect<void, AIError> {
+    return this.route.cancel?.(this.token) ?? Effect.void
+  }
+
+  /** Status observations as a stream, ending after the first terminal observation. */
+  events(options?: { readonly poll?: Poll }): Stream.Stream<Event, AIError> {
+    const observations = this.terminal
+      ? Stream.make(this)
+      : Stream.fromEffectSchedule(this.refresh(), this.schedule(options?.poll)).pipe(
+          Stream.takeUntil((job) => job.terminal),
+        )
+    return observations.pipe(
+      Stream.map((job): Event => {
+        if (job.terminal) return { type: "job-finished", id: job.id, status: job.status }
+        if (job.status === "queued") return { type: "job-queued", id: job.id, position: job.position }
+        return { type: "job-progress", id: job.id, progress: job.progress }
+      }),
+    )
+  }
+
+  private poll(poll: Poll | undefined) {
+    return this.refresh().pipe(Effect.repeat({ schedule: this.schedule(poll), until: (job) => job.terminal }))
+  }
+
+  private schedule(poll: Poll | undefined): Schedule.Schedule<unknown, Job<Response>> {
+    if (poll?.schedule) return poll.schedule.pipe(Schedule.setInputType<Job<Response>>())
+    const interval = poll?.interval ?? DEFAULT_POLL_INTERVAL
+    const pollHint = this.route.pollHint
+    const spaced = Schedule.spaced(interval).pipe(Schedule.setInputType<Job<Response>>())
+    if (!pollHint) return spaced
+    return spaced.pipe(
+      Schedule.modifyDelay((metadata) => Effect.succeed(pollHint(metadata.input.snapshot) ?? interval)),
+    )
+  }
+}

packages/ai/src/llm.ts

@@ -0,0 +1,187 @@
+import { Effect, JsonSchema, Schema } from "effect"
+import { LLMClient, Service } from "./route/client.js"
+import {
+  GenerationOptions,
+  HttpOptions,
+  InvalidProviderOutputError,
+  AIError,
+  LLMEvent,
+  LLMRequest,
+  LLMResponse,
+  Message,
+  LanguageModel,
+  SystemPart,
+  ToolChoice,
+  ToolEntry,
+  type ContentPart,
+  type LanguageModelProviderOptions,
+  type ToolEntryInput,
+} from "./schema/index.js"
+import { make as makeTool, toDefinitions, type ToolSchema } from "./tool.js"
+import { ModelRef, type ResolveLanguageModel } from "./model-ref.js"
+
+/** A concrete language model or a callable-facade ref whose provider exposes an LLM route. */
+export type LanguageModelInput = LanguageModel | ModelRef
+
+/** Input accepted by `LLM.request`, normalized into the canonical `LLMRequest` class. */
+export type RequestInput<SelectedModel extends LanguageModelInput = LanguageModelInput> = Omit<
+  ConstructorParameters<typeof LLMRequest>[0],
+  "model" | "system" | "messages" | "tools" | "toolChoice" | "generation" | "http" | "providerOptions"
+> & {
+  readonly model: SelectedModel
+  readonly system?: string | SystemPart | ReadonlyArray<SystemPart>
+  readonly prompt?: string | ContentPart | ReadonlyArray<ContentPart>
+  readonly messages?: ReadonlyArray<Message | Message.Input>
+  readonly tools?: ReadonlyArray<ToolEntryInput>
+  readonly toolChoice?: ToolChoice.Input
+  readonly generation?: GenerationOptions.Input
+  readonly providerOptions?: NoInfer<LanguageModelProviderOptions<ResolveLanguageModel<SelectedModel>>>
+  readonly http?: HttpOptions.Input
+}
+
+export const generate = LLMClient.generate
+
+export const stream = LLMClient.stream
+
+// The stored request keeps the concrete model; refs are resolved once here so nothing downstream knows about them.
+const resolveModel = (model: LanguageModelInput) => (model instanceof ModelRef ? model.facade.model(model.id) : model)
+
+export const request = <const SelectedModel extends LanguageModelInput>(
+  input: RequestInput<SelectedModel>,
+): LLMRequest<ResolveLanguageModel<SelectedModel>> => {
+  const {
+    model,
+    system: requestSystem,
+    prompt,
+    messages,
+    tools,
+    toolChoice: requestToolChoice,
+    generation: requestGeneration,
+    providerOptions: requestProviderOptions,
+    http: requestHttp,
+    ...rest
+  } = input
+  return new LLMRequest<ResolveLanguageModel<SelectedModel>>({
+    ...rest,
+    model: resolveModel(model) as ResolveLanguageModel<SelectedModel>,
+    system: SystemPart.content(requestSystem),
+    messages: [...(messages?.map(Message.make) ?? []), ...(prompt === undefined ? [] : [Message.user(prompt)])],
+    tools: tools?.map(ToolEntry.make) ?? [],
+    toolChoice: requestToolChoice ? ToolChoice.make(requestToolChoice) : undefined,
+    generation: requestGeneration === undefined ? undefined : GenerationOptions.make(requestGeneration),
+    providerOptions: requestProviderOptions,
+    http: requestHttp === undefined ? undefined : HttpOptions.make(requestHttp),
+  })
+}
+
+const GENERATE_OBJECT_TOOL_NAME = "generate_object"
+
+const GENERATE_OBJECT_TOOL_DESCRIPTION = "Return the structured result by calling this tool."
+
+type GenerateObjectBase<SelectedModel extends LanguageModelInput = LanguageModelInput> = Omit<
+  RequestInput<SelectedModel>,
+  "tools" | "toolChoice"
+>
+
+export class GenerateObjectResponse<T> {
+  constructor(
+    readonly object: T,
+    readonly response: LLMResponse,
+  ) {}
+
+  get events() {
+    return this.response.events
+  }
+
+  get usage() {
+    return this.response.usage
+  }
+}
+
+export interface GenerateObjectOptions<
+  S extends ToolSchema<any>,
+  SelectedModel extends LanguageModelInput = LanguageModelInput,
+> extends GenerateObjectBase<SelectedModel> {
+  readonly schema: S
+}
+
+export interface GenerateObjectDynamicOptions<SelectedModel extends LanguageModelInput = LanguageModelInput>
+  extends GenerateObjectBase<SelectedModel> {
+  /** Raw JSON Schema object describing the expected output shape. */
+  readonly jsonSchema: JsonSchema.JsonSchema
+}
+
+const runGenerateObject = Effect.fn("LLM.generateObject")(function* (
+  options: GenerateObjectBase,
+  tool: ReturnType<typeof makeTool>,
+) {
+  const baseRequest = request(options)
+  const generateRequest = LLMRequest.update(baseRequest, {
+    tools: toDefinitions({ [GENERATE_OBJECT_TOOL_NAME]: tool }),
+    toolChoice: ToolChoice.named(GENERATE_OBJECT_TOOL_NAME),
+  })
+  const response = yield* LLMClient.generate(generateRequest)
+  const call = response.toolCalls.find(
+    (event) => LLMEvent.is.toolCall(event) && event.name === GENERATE_OBJECT_TOOL_NAME,
+  )
+  if (!call || !LLMEvent.is.toolCall(call))
+    return yield* new AIError({
+      reason: new InvalidProviderOutputError({
+        message: `generateObject: model did not call the forced \`${GENERATE_OBJECT_TOOL_NAME}\` tool`,
+      }),
+    })
+  const object = yield* tool._decode(call.input).pipe(
+    Effect.mapError(
+      (error) =>
+        new AIError({
+          reason: new InvalidProviderOutputError({
+            message: `generateObject: tool input failed schema decode: ${error.message}`,
+            cause: error,
+          }),
+        }),
+    ),
+  )
+  return new GenerateObjectResponse(object, response)
+})
+
+/**
+ * Run a model and decode its output against `schema`. Works on every protocol
+ * because it forces a synthetic tool call internally — provider-native JSON
+ * modes are intentionally avoided so behaviour is uniform.
+ *
+ * Two input modes:
+ *
+ * 1. `schema: EffectSchema<T>` — `.object` is decoded and typed as `T`.
+ *    Decode failures surface as `AIError`.
+ * 2. `jsonSchema: JsonSchema.JsonSchema` — `.object` is `unknown`. Use when
+ *    the schema is only available at runtime (MCP, plugin manifests). Caller validates.
+ */
+export function generateObject<const SelectedModel extends LanguageModelInput, S extends ToolSchema<any>>(
+  options: GenerateObjectOptions<S, SelectedModel>,
+): Effect.Effect<GenerateObjectResponse<Schema.Schema.Type<S>>, AIError, Service>
+export function generateObject<const SelectedModel extends LanguageModelInput>(
+  options: GenerateObjectDynamicOptions<SelectedModel>,
+): Effect.Effect<GenerateObjectResponse<unknown>, AIError, Service>
+export function generateObject(options: GenerateObjectOptions<ToolSchema<any>> | GenerateObjectDynamicOptions) {
+  if ("schema" in options) {
+    const { schema, ...rest } = options
+    return runGenerateObject(
+      rest,
+      makeTool({
+        description: GENERATE_OBJECT_TOOL_DESCRIPTION,
+        parameters: schema,
+        success: Schema.Unknown as ToolSchema<unknown>,
+        execute: () => Effect.void,
+      }),
+    )
+  }
+  const { jsonSchema, ...rest } = options
+  return runGenerateObject(
+    rest,
+    makeTool({
+      description: GENERATE_OBJECT_TOOL_DESCRIPTION,
+      jsonSchema,
+      execute: () => Effect.void,
+    }),
+  )
+}

packages/ai/src/media.ts

@@ -0,0 +1,307 @@
+export * as Media from "./media.js"
+
+import { Effect, Encoding, FileSystem, Schema, SchemaGetter } from "effect"
+import { HttpClientRequest } from "effect/unstable/http"
+import { ProviderID } from "./schema/ids.js"
+import { AIError, HttpContext, InvalidProviderOutputError, InvalidRequestError } from "./schema/errors.js"
+import { ProviderMetadata } from "./schema/options.js"
+import { Service } from "./route/executor-service.js"
+import { detectMediaType, extensionMediaType } from "./utils/media-type.js"
+
+export { detectMediaType } from "./utils/media-type.js"
+
+const OCTET_STREAM = "application/octet-stream"
+
+// ---------------------------------------------------------------------------
+// Source — the serializable wire/persistence form of a media asset
+// ---------------------------------------------------------------------------
+
+const BytesSource = Schema.Struct({
+  type: Schema.Literal("bytes"),
+  data: Schema.Uint8Array,
+  mediaType: Schema.String,
+})
+
+const Base64Source = Schema.Struct({
+  type: Schema.Literal("base64"),
+  data: Schema.String,
+  mediaType: Schema.String,
+})
+
+const UrlSource = Schema.Struct({
+  type: Schema.Literal("url"),
+  url: Schema.String,
+  mediaType: Schema.optional(Schema.String),
+  /** Epoch milliseconds after which the provider no longer serves the URL. */
+  expiresAt: Schema.optional(Schema.Number),
+  /** Headers required to fetch the URL, such as provider auth for Veo downloads. */
+  headers: Schema.optional(Schema.Record(Schema.String, Schema.String)),
+})
+
+/** A provider-side handle: OpenAI `file_id`, Gemini file URI, `gs://`, `runway://`, or a prior generation id. */
+const RefSource = Schema.Struct({
+  type: Schema.Literal("ref"),
+  provider: ProviderID,
+  id: Schema.String,
+  mediaType: Schema.optional(Schema.String),
+})
+
+export const Source = Schema.Union([BytesSource, Base64Source, UrlSource, RefSource])
+  .pipe(Schema.toTaggedUnion("type"))
+  .annotate({ identifier: "Media.Source" })
+export type Source = Schema.Schema.Type<typeof Source>
+
+// ---------------------------------------------------------------------------
+// Kind, Info, Notice
+// ---------------------------------------------------------------------------
+
+export const Kind = Schema.Literals(["image", "video", "audio", "document", "other"])
+export type Kind = Schema.Schema.Type<typeof Kind>
+
+export const kindOf = (mediaType: string): Kind => {
+  const lower = mediaType.toLowerCase()
+  if (lower.startsWith("image/")) return "image"
+  if (lower.startsWith("video/")) return "video"
+  if (lower.startsWith("audio/")) return "audio"
+  if (lower === "application/pdf" || lower.startsWith("text/")) return "document"
+  return "other"
+}
+
+/** Container-independent facts about the payload; raw PCM audio relies on these because it has no header. */
+export const Info = Schema.Struct({
+  width: Schema.optional(Schema.Number),
+  height: Schema.optional(Schema.Number),
+  durationSeconds: Schema.optional(Schema.Number),
+  sampleRate: Schema.optional(Schema.Number),
+  channels: Schema.optional(Schema.Number),
+  encoding: Schema.optional(Schema.String),
+  format: Schema.optional(Schema.String),
+}).annotate({ identifier: "Media.Info" })
+export type Info = Schema.Schema.Type<typeof Info>
+
+/** A provider-side partial result such as stripped audio or a moderated sample; never a silent drop. */
+export const Notice = Schema.Struct({
+  type: Schema.Literals(["moderated", "filtered", "other"]),
+  message: Schema.String,
+  providerMetadata: Schema.optional(ProviderMetadata),
+}).annotate({ identifier: "Media.Notice" })
+export type Notice = Schema.Schema.Type<typeof Notice>
+
+// ---------------------------------------------------------------------------
+// Asset
+// ---------------------------------------------------------------------------
+
+const invalid = (message: string, cause?: unknown) =>
+  new AIError({ reason: new InvalidRequestError({ message, cause }) })
+
+/** Synchronous view of an inline payload; `undefined` for `url` and `ref` sources, which carry no local bytes. */
+export interface Inline {
+  readonly mime: string
+  readonly base64: string
+  readonly dataUrl: string
+}
+
+export class Asset {
+  readonly source: Source
+  /** Derived from the source: declared type, sniffed magic bytes, then `application/octet-stream`. */
+  readonly mediaType: string
+  readonly kind: Kind
+  readonly info?: Info
+  /** Epoch milliseconds after which a `url` source stops resolving. */
+  readonly expiresAt?: number
+  readonly providerMetadata?: ProviderMetadata
+
+  // Derived payload forms are cached on the instance because every protocol lowering re-reads the same payload. The
+  // cache is check-then-set (concurrent first reads of a `url` source may both download) and is never observable
+  // through `source`, so round-tripping through `Media.from(asset.source)` stays lossless.
+  #bytes: Uint8Array | undefined
+  #base64: string | undefined
+
+  constructor(input: Asset.Input) {
+    this.source = input.source
+    this.mediaType =
+      input.source.mediaType ??
+      (input.source.type === "bytes" ? detectMediaType(input.source.data) : undefined) ??
+      OCTET_STREAM
+    this.kind = kindOf(this.mediaType)
+    this.info = input.info
+    this.expiresAt = input.source.type === "url" ? input.source.expiresAt : undefined
+    this.providerMetadata = input.providerMetadata
+  }
+
+  /** Inline payload without effects, for protocols that embed base64 or data URLs directly. */
+  inline(): Inline | undefined {
+    const source = this.source
+    if (source.type !== "bytes" && source.type !== "base64") return undefined
+    const base64 = source.type === "base64" ? source.data : (this.#base64 ??= Encoding.encodeBase64(source.data))
+    const mime = this.mediaType.toLowerCase()
+    return { mime, base64, dataUrl: `data:${mime};base64,${base64}` }
+  }
+
+  /** Decoded payload; downloads `url` sources through the request executor and caches the result. */
+  bytes(): Effect.Effect<Uint8Array, AIError, Service> {
+    return Effect.suspend(() => {
+      const source = this.source
+      if (source.type === "bytes") return Effect.succeed(source.data)
+      if (this.#bytes !== undefined) return Effect.succeed(this.#bytes)
+      if (source.type === "ref")
+        return Effect.fail(invalid(`Cannot materialize provider ref ${source.provider}:${source.id}`))
+      const decoded =
+        source.type === "base64"
+          ? Effect.fromResult(Encoding.decodeBase64(source.data)).pipe(
+              Effect.mapError((cause) => invalid(`Media asset contains invalid base64 data`, cause)),
+            )
+          : download(source)
+      return decoded.pipe(Effect.tap((data) => Effect.sync(() => (this.#bytes = data))))
+    })
+  }
+
+  base64(): Effect.Effect<string, AIError, Service> {
+    return Effect.suspend(() => {
+      const source = this.source
+      if (source.type === "base64") return Effect.succeed(source.data)
+      if (this.#base64 !== undefined) return Effect.succeed(this.#base64)
+      return this.bytes().pipe(Effect.map((data) => (this.#base64 = Encoding.encodeBase64(data))))
+    })
+  }
+
+  dataUrl(): Effect.Effect<string, AIError, Service> {
+    return this.base64().pipe(Effect.map((data) => `data:${this.mediaType};base64,${data}`))
+  }
+
+  /**
+   * The `AssetEncoded` JSON form with `bytes` sources as base64, matching `Schema.toCodecJson(AssetSchema)`, so a
+   * plain `JSON.stringify` of messages or events stays lossless and decodes back through the JSON codec.
+   */
+  toJSON() {
+    const source = this.source
+    return {
+      source: source.type === "bytes" ? { ...source, data: Encoding.encodeBase64(source.data) } : source,
+      info: this.info,
+      providerMetadata: this.providerMetadata,
+    }
+  }
+
+  /** Pull `url` sources into owned bytes before the URL expires. Inline sources return themselves. */
+  materialize(): Effect.Effect<Asset, AIError, Service> {
+    if (this.source.type === "bytes" || this.source.type === "base64") return Effect.succeed(this)
+    return this.bytes().pipe(
+      Effect.map((data) =>
+        bytes(data, this.source.mediaType, { info: this.info, providerMetadata: this.providerMetadata }),
+      ),
+    )
+  }
+}
+
+export namespace Asset {
+  export interface Input {
+    readonly source: Source
+    readonly info?: Info
+    readonly providerMetadata?: ProviderMetadata
+  }
+}
+
+/** JSON form of an asset: the serializable `Source` plus caller-supplied metadata. `bytes` sources encode as base64. */
+export const AssetEncoded = Schema.Struct({
+  source: Source,
+  info: Schema.optional(Info),
+  providerMetadata: Schema.optional(ProviderMetadata),
+}).annotate({ identifier: "Media.AssetEncoded" })
+
+const encodeAsset = (asset: Asset): typeof AssetEncoded.Type => ({
+  source: asset.source,
+  info: asset.info,
+  providerMetadata: asset.providerMetadata,
+})
+
+const AssetInstance = Schema.declare((value): value is Asset => value instanceof Asset, {
+  expected: "Media.Asset",
+})
+
+/** `Asset` in the type domain and `AssetEncoded` on the wire, so messages and events holding assets serialize. */
+export const AssetSchema = AssetEncoded.pipe(
+  Schema.decodeTo(AssetInstance, {
+    decode: SchemaGetter.transform((encoded) => new Asset(encoded)),
+    encode: SchemaGetter.transform(encodeAsset),
+  }),
+)
+
+const download = Effect.fn("Media.download")(function* (source: Extract<Source, { readonly type: "url" }>) {
+  const executor = yield* Service
+  const response = yield* executor.execute(
+    HttpClientRequest.get(source.url).pipe(HttpClientRequest.setHeaders(source.headers ?? {})),
+  )
+  const buffer = yield* response.arrayBuffer.pipe(
+    Effect.mapError(
+      (cause) =>
+        new AIError({
+          reason: new InvalidProviderOutputError({
+            message: `Failed to read media from ${source.url}`,
+            http: new HttpContext({ url: response.request.url, status: response.status, headers: response.headers }),
+            cause,
+          }),
+        }),
+    ),
+  )
+  return new Uint8Array(buffer)
+})
+
+// ---------------------------------------------------------------------------
+// Constructors
+// ---------------------------------------------------------------------------
+
+export type AssetOptions = Omit<Asset.Input, "source">
+
+export const from = (source: Source, options?: AssetOptions) => new Asset({ ...options, source })
+
+export const bytes = (data: Uint8Array, mediaType?: string, options?: AssetOptions) =>
+  from({ type: "bytes", data, mediaType: mediaType ?? detectMediaType(data) ?? OCTET_STREAM }, options)
+
+export const base64 = (data: string, mediaType: string, options?: AssetOptions) =>
+  from({ type: "base64", data, mediaType }, options)
+
+export const url = (
+  value: string,
+  options?: AssetOptions & Omit<Extract<Source, { readonly type: "url" }>, "type" | "url">,
+) => {
+  const { mediaType, expiresAt, headers, ...rest } = options ?? {}
+  return from({ type: "url", url: value, mediaType, expiresAt, headers }, rest)
+}
+
+export const ref = (provider: string | ProviderID, id: string, mediaType?: string, options?: AssetOptions) =>
+  from({ type: "ref", provider: ProviderID.make(provider), id, mediaType }, options)
+
+const DATA_URL = /^data:([^;,]+)(?:;[^,]*)*;base64,(.*)$/s
+
+/** Parse a `data:<mime>;base64,<data>` URL, or `undefined` when the value is not a base64 data URL. */
+export const parseDataUrl = (value: string, options?: AssetOptions) => {
+  const match = DATA_URL.exec(value)
+  return match === null ? undefined : base64(match[2], match[1], options)
+}
+
+/** Parse a `data:<mime>;base64,<data>` URL. Malformed input throws a typed `AIError` because constructors are sync. */
+export const fromDataUrl = (dataUrl: string, options?: AssetOptions) => {
+  const asset = parseDataUrl(dataUrl, options)
+  if (asset === undefined) throw invalid("Media data URLs must contain a MIME type and base64 data")
+  return asset
+}
+
+/** Read a file through `FileSystem` and sniff its media type from magic bytes, then the extension. */
+export const file = (path: string, options?: AssetOptions): Effect.Effect<Asset, AIError, FileSystem.FileSystem> =>
+  Effect.gen(function* () {
+    const fs = yield* FileSystem.FileSystem
+    const data = yield* fs
+      .readFile(path)
+      .pipe(Effect.mapError((cause) => invalid(`Failed to read media file ${path}`, cause)))
+    return bytes(data, detectMediaType(data) ?? extensionMediaType(path), options)
+  })
+
+/** Materialize an asset and write its bytes through `FileSystem`. */
+export const write = (asset: Asset, path: string): Effect.Effect<void, AIError, FileSystem.FileSystem | Service> =>
+  Effect.gen(function* () {
+    const fs = yield* FileSystem.FileSystem
+    const data = yield* asset.bytes()
+    yield* fs
+      .writeFile(path, data)
+      .pipe(Effect.mapError((cause) => invalid(`Failed to write media file ${path}`, cause)))
+  })

packages/ai/src/model-ref.ts

@@ -0,0 +1,54 @@
+import { ModelID, type ProviderID } from "./schema/ids.js"
+import type { LanguageModel } from "./schema/options.js"
+import type { ImageModel } from "./image.js"
+
+/**
+ * A provider-scoped model name bound to its configured facade. Callable facades return one of these so
+ * `LLM.request({ model: openai("gpt-5") })` and `Image.request({ model: openai("gpt-image-2") })` name a model once
+ * and let the request namespace pick the selector (`facade.model`, `facade.image`, …) on demand.
+ */
+export class ModelRef<S extends ModelRef.Selectors = ModelRef.Selectors> {
+  constructor(
+    readonly id: ModelID,
+    readonly facade: S,
+  ) {}
+
+  get provider() {
+    return this.facade.id
+  }
+}
+
+export namespace ModelRef {
+  /** The named selectors a configured provider facade exposes; `model` is the default LLM route. */
+  export interface Selectors {
+    readonly id: ProviderID
+    // oxlint-disable-next-line typescript-eslint/no-explicit-any
+    readonly model: (id: ModelID) => LanguageModel<any, any>
+    // oxlint-disable-next-line typescript-eslint/no-explicit-any
+    readonly image?: (id: ModelID) => ImageModel<any>
+  }
+
+  /** A ref whose provider exposes an image route. */
+  // oxlint-disable-next-line typescript-eslint/no-explicit-any
+  export type WithImage = ModelRef<Selectors & { readonly image: (id: ModelID) => ImageModel<any> }>
+
+  /** Make a configured facade callable: `openai("gpt-5")` returns a `ModelRef` over the facade's own selectors. */
+  export const facade = <const S extends Selectors>(selectors: S) =>
+    Object.assign((id: string | ModelID) => new ModelRef(ModelID.make(id), selectors), selectors)
+}
+
+/** The concrete `LanguageModel` a request will run against: either the model itself or the ref's `model` selector. */
+// oxlint-disable-next-line typescript-eslint/no-explicit-any
+export type ResolveLanguageModel<Model> =
+  Model extends LanguageModel<any, any> ? Model : Model extends ModelRef<infer S> ? ReturnType<S["model"]> : never
+
+/** The concrete `ImageModel` a request will run against: either the model itself or the ref's `image` selector. */
+// oxlint-disable-next-line typescript-eslint/no-explicit-any
+export type ResolveImageModel<Model> =
+  Model extends ImageModel<any>
+    ? Model
+    : Model extends ModelRef<infer S>
+      ? S extends { readonly image: (id: ModelID) => infer Selected }
+        ? Selected
+        : never
+      : never

packages/ai/src/promise.ts

@@ -0,0 +1,100 @@
+import { Effect, Layer, ManagedRuntime, Stream } from "effect"
+import { Image, ImageRequest, type ImageModelInput, type ImageRequestInput } from "./image.js"
+import { ImageClient } from "./image-client.js"
+import { LLM } from "./index.js"
+import { LLMClient } from "./route/client.js"
+import { RequestExecutor } from "./route/executor.js"
+import { LLMRequest } from "./schema/index.js"
+import type { LanguageModelInput, RequestInput } from "./llm.js"
+
+/**
+ * Promise-first entrypoint for scripts and non-Effect callers. One `ManagedRuntime` hosts the LLM and image clients
+ * over a request executor; every method runs the corresponding Effect API and rethrows `AIError` unchanged.
+ */
+export interface Options {
+  /** Executor layer; defaults to `RequestExecutor.fetchLayer`. Inject a recorder or middleware here. */
+  readonly layer?: Layer.Layer<RequestExecutor.Service>
+}
+
+export interface RunOptions {
+  readonly signal?: AbortSignal
+}
+
+export type Services =
+  | Layer.Success<typeof LLMClient.layer>
+  | Layer.Success<typeof ImageClient.layer>
+  | RequestExecutor.Service
+
+const abortEffect = (signal: AbortSignal | undefined) =>
+  signal === undefined
+    ? Effect.never
+    : Effect.callback<void>((resume) => {
+        if (signal.aborted) {
+          resume(Effect.void)
+          return
+        }
+        const onAbort = () => resume(Effect.void)
+        signal.addEventListener("abort", onAbort, { once: true })
+        return Effect.sync(() => signal.removeEventListener("abort", onAbort))
+      })
+
+export const make = (options: Options = {}) => {
+  const runtime = ManagedRuntime.make(
+    Layer.mergeAll(LLMClient.layer, ImageClient.layer).pipe(
+      Layer.provideMerge(options.layer ?? RequestExecutor.fetchLayer),
+    ),
+  )
+
+  /** Run any package Effect (for example `asset.bytes()`) inside this runtime. */
+  const run = <A, E>(effect: Effect.Effect<A, E, Services>, options?: RunOptions) =>
+    runtime.runPromise(effect, { signal: options?.signal })
+
+  const iterate = <A, E>(stream: Stream.Stream<A, E, Services>, options?: RunOptions): AsyncIterable<A> =>
+    Stream.toAsyncIterable(
+      Stream.unwrap(
+        runtime.contextEffect.pipe(
+          Effect.map(
+            (context): Stream.Stream<A, E> =>
+              stream.pipe(Stream.interruptWhen(abortEffect(options?.signal)), Stream.provideContext(context)),
+          ),
+        ),
+      ),
+    )
+
+  // The typed `generate`/`stream` overloads take a concrete input or a request, not the union; normalize once here.
+  const llmRequest = (input: RequestInput | LLMRequest) => (input instanceof LLMRequest ? input : LLM.request(input))
+  const imageRequest = (input: ImageRequestInput | ImageRequest) =>
+    input instanceof ImageRequest ? input : Image.request(input)
+
+  return {
+    run,
+    llm: {
+      request: LLM.request,
+      generate: <const Model extends LanguageModelInput>(
+        input: RequestInput<Model> | LLMRequest,
+        options?: RunOptions,
+      ) => run(LLM.generate(llmRequest(input)), options),
+      stream: <const Model extends LanguageModelInput>(input: RequestInput<Model> | LLMRequest, options?: RunOptions) =>
+        iterate(LLM.stream(llmRequest(input)), options),
+    },
+    image: {
+      request: Image.request,
+      generate: <const Model extends ImageModelInput>(
+        input: ImageRequestInput<Model> | ImageRequest,
+        options?: RunOptions,
+      ) => run(Image.generate(imageRequest(input)), options),
+      stream: <const Model extends ImageModelInput>(
+        input: ImageRequestInput<Model> | ImageRequest,
+        options?: RunOptions,
+      ) => iterate(Image.stream(imageRequest(input)), options),
+    },
+    dispose: () => runtime.dispose(),
+  }
+}
+
+export type Client = ReturnType<typeof make>
+
+/** Default client over `RequestExecutor.fetchLayer` for scripts; the runtime builds its layer on first use. */
+export const ai = make()
+
+export * as AI from "./promise.js"

packages/ai/src/protocols.ts

@@ -0,0 +1 @@
+export * from "./protocols/index.js"

packages/ai/src/protocols/alibaba-chat.ts

@@ -0,0 +1,92 @@
+import { Effect, Schema } from "effect"
+import { Protocol } from "../route/protocol.js"
+import type { LanguageModelCompatibility } from "../schema/index.js"
+import { OpenAIChat } from "./openai-chat.js"
+import { JsonObject, ProviderShared } from "./shared.js"
+import { OpenResponsesOptions } from "./utils/open-responses-options.js"
+
+export type ReasoningEffort = OpenResponsesOptions.ReasoningEffort
+
+const Options = Schema.Struct({
+  reasoningEffort: OpenResponsesOptions.Options.fields.reasoningEffort,
+  enableThinking: Schema.optional(Schema.Boolean),
+  thinkingBudget: Schema.optional(Schema.Int),
+  preserveThinking: Schema.optional(Schema.Boolean),
+  clearThinking: Schema.optional(Schema.Boolean),
+  thinking: Schema.optional(
+    Schema.Struct({
+      type: Schema.declare<"adaptive" | "disabled" | (string & {})>(Schema.is(Schema.String)),
+    }),
+  ),
+  toolStream: Schema.optional(Schema.Boolean),
+  parallelToolCalls: OpenResponsesOptions.Options.fields.parallelToolCalls,
+  repetitionPenalty: Schema.optional(Schema.Number),
+  responseFormat: Schema.optional(
+    Schema.Struct({
+      type: Schema.declare<"text" | "json_object" | "json_schema" | (string & {})>(Schema.is(Schema.String)),
+      json_schema: Schema.optional(JsonObject),
+    }),
+  ),
+  enableSearch: Schema.optional(Schema.Boolean),
+  searchOptions: Schema.optional(
+    Schema.Struct({
+      forced_search: Schema.optional(Schema.Boolean),
+      search_strategy: Schema.optional(
+        Schema.declare<"turbo" | "max" | "agent" | "agent_max" | (string & {})>(Schema.is(Schema.String)),
+      ),
+      enable_search_extension: Schema.optional(Schema.Boolean),
+    }),
+  ),
+})
+export type OptionsInput = typeof Options.Type
+
+export const compatibility = {
+  maxTokensField: "max_completion_tokens",
+  supportsStore: false,
+  supportsStrictMode: false,
+  reasoningField: "reasoning_content",
+  zaiToolStream: false,
+} satisfies LanguageModelCompatibility
+
+export const protocol = Protocol.make({
+  id: "alibaba-chat",
+  body: {
+    schema: Schema.Struct({
+      ...OpenAIChat.bodyFields,
+      enable_thinking: Options.fields.enableThinking,
+      thinking_budget: Options.fields.thinkingBudget,
+      preserve_thinking: Options.fields.preserveThinking,
+      clear_thinking: Options.fields.clearThinking,
+      thinking: Options.fields.thinking,
+      parallel_tool_calls: Options.fields.parallelToolCalls,
+      repetition_penalty: Options.fields.repetitionPenalty,
+      top_k: Schema.optional(Schema.Int),
+      response_format: Options.fields.responseFormat,
+      enable_search: Options.fields.enableSearch,
+      search_options: Options.fields.searchOptions,
+    }),
+    from: Effect.fn("AlibabaChat.fromRequest")(function* (req) {
+      const opts = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(req.providerOptions ?? {})
+      return {
+        ...(yield* OpenAIChat.protocol.body.from(req)),
+        enable_thinking: opts.enableThinking,
+        thinking_budget: opts.thinkingBudget,
+        preserve_thinking: opts.preserveThinking,
+        clear_thinking: opts.clearThinking,
+        thinking: opts.thinking,
+        tool_stream: opts.toolStream,
+        parallel_tool_calls:
+          opts.parallelToolCalls ??
+          (req.toolChoice?.disableParallelToolUse === undefined ? undefined : !req.toolChoice.disableParallelToolUse),
+        repetition_penalty: opts.repetitionPenalty,
+        top_k: req.generation?.topK,
+        response_format: opts.responseFormat,
+        enable_search: opts.enableSearch,
+        search_options: opts.searchOptions,
+      }
+    }),
+  },
+  stream: OpenAIChat.protocol.stream,
+})
+
+export * as AlibabaChat from "./alibaba-chat.js"

packages/ai/src/protocols/alibaba-messages.ts

@@ -0,0 +1,48 @@
+import { Effect, Schema } from "effect"
+import { Protocol } from "../route/protocol.js"
+import { LLMRequest } from "../schema/index.js"
+import { AnthropicMessages } from "./anthropic-messages.js"
+import { ProviderShared } from "./shared.js"
+import { OpenResponsesOptions } from "./utils/open-responses-options.js"
+
+const Options = Schema.Struct({
+  effort: Schema.optional(OpenResponsesOptions.ReasoningEffort),
+  thinking: Schema.optional(
+    Schema.Struct({
+      type: Schema.declare<"enabled" | "disabled" | (string & {})>(Schema.is(Schema.String)),
+      budgetTokens: Schema.optional(Schema.Int),
+      budget_tokens: Schema.optional(Schema.Int),
+    }),
+  ),
+})
+export type OptionsInput = typeof Options.Type & Pick<AnthropicMessages.OptionsInput, "outputConfig">
+export const protocol = Protocol.make({
+  id: "alibaba-messages",
+  body: {
+    schema: Schema.Struct({
+      ...AnthropicMessages.AnthropicMessagesBody.fields,
+      thinking: Schema.optional(Schema.Struct({ type: Schema.String, budget_tokens: Schema.optional(Schema.Int) })),
+    }),
+    from: Effect.fn("AlibabaMessages.fromRequest")(function* (req) {
+      const opts = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(req.providerOptions ?? {})
+      // Model Studio accepts enabled thinking without Anthropic's mandatory token budget.
+      return {
+        ...(yield* AnthropicMessages.protocol.body.from(
+          LLMRequest.update(req, {
+            providerOptions: { ...req.providerOptions, thinking: undefined },
+          }),
+        )),
+        thinking:
+          opts.thinking === undefined
+            ? undefined
+            : {
+                type: opts.thinking.type,
+                budget_tokens: opts.thinking.budgetTokens ?? opts.thinking.budget_tokens,
+              },
+      }
+    }),
+  },
+  stream: AnthropicMessages.protocol.stream,
+})
+
+export * as AlibabaMessages from "./alibaba-messages.js"

packages/ai/src/protocols/alibaba-responses.ts

@@ -0,0 +1,98 @@
+import { Effect, Schema } from "effect"
+import { Protocol } from "../route/protocol.js"
+import { OpenResponses } from "./open-responses.js"
+import { JsonObject, optionalArray, ProviderShared } from "./shared.js"
+import { OpenResponsesOptions } from "./utils/open-responses-options.js"
+import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
+
+const Options = Schema.Struct({
+  reasoningEffort: OpenResponsesOptions.Options.fields.reasoningEffort,
+  enableThinking: Schema.optional(Schema.Boolean),
+  store: OpenResponsesOptions.Options.fields.store,
+  previousResponseId: Schema.optional(Schema.String),
+  conversation: Schema.optional(Schema.String),
+})
+export type OptionsInput = typeof Options.Type
+const NativeTool = Schema.Struct({ type: Schema.Literals(["web_search", "web_extractor", "code_interpreter"]) })
+const WebExtractorItem = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.Literal("web_extractor_call"),
+    id: Schema.String,
+    urls: Schema.optional(Schema.Array(Schema.String)),
+    goal: Schema.optional(Schema.String),
+  }),
+  [JsonObject],
+)
+const Body = Schema.Struct({
+  ...OpenResponses.coreFields,
+  input: Schema.Array(Schema.Union([OpenResponses.InputItem, WebExtractorItem])),
+  tools: optionalArray(Schema.Union([OpenResponses.Tool, NativeTool])),
+  enable_thinking: Options.fields.enableThinking,
+  previous_response_id: Options.fields.previousResponseId,
+  conversation: Options.fields.conversation,
+  stream: Schema.Literal(true),
+})
+const adapter = {
+  id: "alibaba-responses",
+  name: "Alibaba Responses",
+  nativeTool: (native) => ProviderShared.validateWith(Schema.decodeUnknownEffect(NativeTool))(native.alibaba),
+  restoreHostedToolItem: (item: unknown) => (Schema.is(WebExtractorItem)(item) ? item : undefined),
+} satisfies OpenResponses.ProviderAdapter
+
+const tools = {
+  web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
+  code_interpreter_call: { name: "code_interpreter", input: (item) => ({ code: item.code }) },
+} satisfies ResponsesHostedTools.Definitions
+
+export const protocol = Protocol.make({
+  id: adapter.id,
+  body: {
+    schema: Body,
+    from: Effect.fn("AlibabaResponses.fromRequest")(function* (req) {
+      const opts = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(req.providerOptions ?? {})
+      const body = yield* OpenResponses.fromRequestWithAdapter(req, adapter)
+      const choice = body.tool_choice
+      return yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Body))({
+        ...body,
+        enable_thinking: opts.enableThinking,
+        previous_response_id: opts.previousResponseId,
+        conversation: opts.conversation,
+        // Model Studio expresses named selection through allowed_tools.
+        tool_choice:
+          typeof choice === "object" && choice.type === "function"
+            ? { type: "allowed_tools" as const, mode: "required" as const, tools: [choice] }
+            : choice,
+      })
+    }),
+  },
+  stream: {
+    event: OpenResponses.protocol.stream.event,
+    initial: (req) => OpenResponses.initial(req, adapter),
+    step: (state, input) =>
+      Effect.gen(function* () {
+        const event = OpenResponses.normalize(state, input)
+        if (event.type !== "response.output_item.done" || !event.item) return yield* OpenResponses.step(state, event)
+        if (event.item.type === "web_extractor_call") {
+          const item = yield* Schema.decodeUnknownEffect(WebExtractorItem)(event.item).pipe(
+            Effect.mapError((cause) =>
+              ProviderShared.eventError(
+                adapter.id,
+                "Alibaba returned an invalid web extraction item",
+                ProviderShared.encodeJson(event),
+                cause,
+              ),
+            ),
+          )
+          return yield* ResponsesHostedTools.onDone(state, item, {
+            web_extractor_call: { name: "web_extractor", input: () => ({ urls: item.urls, goal: item.goal }) },
+          })
+        }
+        if (ResponsesHostedTools.isItem(event.item, tools))
+          return yield* ResponsesHostedTools.onDone(state, event.item, tools)
+        return yield* OpenResponses.step(state, event)
+      }),
+    terminal: OpenResponses.terminal,
+  },
+})
+
+export * as AlibabaResponses from "./alibaba-responses.js"

packages/ai/src/protocols/bedrock-converse.ts

@@ -0,0 +1,798 @@
+import { Effect, Encoding, Schema } from "effect"
+import { Route } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Protocol } from "../route/protocol.js"
+import {
+  AIError,
+  LLMEvent,
+  Usage,
+  type CacheHint,
+  type FinishReason,
+  type FinishReasonDetails,
+  type JsonSchema,
+  type LLMRequest,
+  type LanguageModelToolSchemaCompatibility,
+  type ProviderMetadata,
+  type ReasoningPart,
+  type ToolCallPart,
+  type ToolDefinition,
+  type ToolResultPart,
+} from "../schema/index.js"
+import { BedrockEventStream } from "./bedrock-event-stream.js"
+import { classifyProviderFailure } from "../provider-error.js"
+import { JsonObject, optionalArray, ProviderShared } from "./shared.js"
+import { BedrockAuth } from "./utils/bedrock-auth.js"
+import { BedrockCache } from "./utils/bedrock-cache.js"
+import { BedrockMedia } from "./utils/bedrock-media.js"
+import { Lifecycle } from "./utils/lifecycle.js"
+import { MistralToolID } from "./utils/mistral-tool-id.js"
+import { ToolSchemaProjection } from "./utils/tool-schema.js"
+import { ToolStream } from "./utils/tool-stream.js"
+
+const ADAPTER = "bedrock-converse"
+
+export type { Credentials as BedrockCredentials } from "./utils/bedrock-auth.js"
+
+// =============================================================================
+// Request Body Schema
+// =============================================================================
+const BedrockTextBlock = Schema.Struct({
+  text: Schema.String,
+})
+type BedrockTextBlock = Schema.Schema.Type<typeof BedrockTextBlock>
+
+const BedrockToolUseBlock = Schema.Struct({
+  toolUse: Schema.Struct({
+    toolUseId: Schema.String,
+    name: Schema.String,
+    input: Schema.Unknown,
+  }),
+})
+type BedrockToolUseBlock = Schema.Schema.Type<typeof BedrockToolUseBlock>
+
+const BedrockToolResultContentItem = Schema.Union([
+  Schema.Struct({ text: Schema.String }),
+  Schema.Struct({ json: Schema.Unknown }),
+  BedrockMedia.ImageBlock,
+  BedrockMedia.DocumentBlock,
+])
+
+const BedrockToolResultBlock = Schema.Struct({
+  toolResult: Schema.Struct({
+    toolUseId: Schema.String,
+    content: Schema.Array(BedrockToolResultContentItem),
+    status: Schema.optional(Schema.Literals(["success", "error"])),
+  }),
+})
+type BedrockToolResultBlock = Schema.Schema.Type<typeof BedrockToolResultBlock>
+
+const BedrockReasoningBlock = Schema.Struct({
+  reasoningContent: Schema.Union([
+    Schema.Struct({
+      reasoningText: Schema.Struct({
+        text: Schema.String,
+        signature: Schema.optional(Schema.String),
+      }),
+    }),
+    Schema.Struct({ redactedContent: Schema.String }),
+  ]),
+})
+
+const BedrockUserBlock = Schema.Union([
+  BedrockTextBlock,
+  BedrockMedia.ImageBlock,
+  BedrockMedia.DocumentBlock,
+  BedrockToolResultBlock,
+  BedrockCache.CachePointBlock,
+])
+type BedrockUserBlock = Schema.Schema.Type<typeof BedrockUserBlock>
+
+const BedrockAssistantBlock = Schema.Union([
+  BedrockTextBlock,
+  BedrockReasoningBlock,
+  BedrockToolUseBlock,
+  BedrockCache.CachePointBlock,
+])
+type BedrockAssistantBlock = Schema.Schema.Type<typeof BedrockAssistantBlock>
+
+const BedrockMessage = Schema.Union([
+  Schema.Struct({ role: Schema.Literal("user"), content: Schema.Array(BedrockUserBlock) }),
+  Schema.Struct({ role: Schema.Literal("assistant"), content: Schema.Array(BedrockAssistantBlock) }),
+]).pipe(Schema.toTaggedUnion("role"))
+type BedrockMessage = Schema.Schema.Type<typeof BedrockMessage>
+
+const BedrockSystemBlock = Schema.Union([BedrockTextBlock, BedrockCache.CachePointBlock])
+type BedrockSystemBlock = Schema.Schema.Type<typeof BedrockSystemBlock>
+
+const BedrockToolSpec = Schema.Struct({
+  toolSpec: Schema.Struct({
+    name: Schema.String,
+    description: Schema.optional(Schema.String),
+    inputSchema: Schema.Struct({
+      json: JsonObject,
+    }),
+  }),
+})
+type BedrockToolSpec = Schema.Schema.Type<typeof BedrockToolSpec>
+
+const BedrockTool = Schema.Union([BedrockToolSpec, BedrockCache.CachePointBlock])
+type BedrockTool = Schema.Schema.Type<typeof BedrockTool>
+
+const BedrockToolChoice = Schema.Union([
+  Schema.Struct({ auto: Schema.Struct({}) }),
+  Schema.Struct({ any: Schema.Struct({}) }),
+  Schema.Struct({ tool: Schema.Struct({ name: Schema.String }) }),
+])
+
+const BedrockBodyFields = {
+  modelId: Schema.String,
+  messages: Schema.Array(BedrockMessage),
+  system: optionalArray(BedrockSystemBlock),
+  inferenceConfig: Schema.optional(
+    Schema.Struct({
+      maxTokens: Schema.optional(Schema.Number),
+      temperature: Schema.optional(Schema.Number),
+      topP: Schema.optional(Schema.Number),
+      stopSequences: optionalArray(Schema.String),
+    }),
+  ),
+  toolConfig: Schema.optional(
+    Schema.Struct({
+      tools: Schema.Array(BedrockTool),
+      toolChoice: Schema.optional(BedrockToolChoice),
+    }),
+  ),
+  additionalModelRequestFields: Schema.optional(JsonObject),
+}
+const BedrockConverseBody = Schema.Struct(BedrockBodyFields)
+export type BedrockConverseBody = Schema.Schema.Type<typeof BedrockConverseBody>
+
+const BedrockUsageSchema = Schema.Struct({
+  inputTokens: Schema.optional(Schema.Number),
+  outputTokens: Schema.optional(Schema.Number),
+  totalTokens: Schema.optional(Schema.Number),
+  cacheReadInputTokens: Schema.optional(Schema.Number),
+  cacheWriteInputTokens: Schema.optional(Schema.Number),
+})
+type BedrockUsageSchema = Schema.Schema.Type<typeof BedrockUsageSchema>
+
+const BedrockStreamException = Schema.Struct({
+  message: Schema.optional(Schema.String),
+  originalMessage: Schema.optional(Schema.String),
+  originalStatusCode: Schema.optional(Schema.Number),
+})
+
+// Streaming event shape — the AWS event stream wraps each JSON payload by its
+// `:event-type` header (e.g. `messageStart`, `contentBlockDelta`). We
+// reconstruct that wrapping in `decodeFrames` below so the event schema can
+// stay a plain discriminated record.
+const BedrockEvent = Schema.Struct({
+  messageStart: Schema.optional(Schema.Struct({ role: Schema.String })),
+  contentBlockStart: Schema.optional(
+    Schema.Struct({
+      contentBlockIndex: Schema.Number,
+      start: Schema.optional(
+        Schema.Struct({
+          toolUse: Schema.optional(Schema.Struct({ toolUseId: Schema.String, name: Schema.String })),
+        }),
+      ),
+    }),
+  ),
+  contentBlockDelta: Schema.optional(
+    Schema.Struct({
+      contentBlockIndex: Schema.Number,
+      delta: Schema.optional(
+        Schema.Struct({
+          text: Schema.optional(Schema.String),
+          toolUse: Schema.optional(Schema.Struct({ input: Schema.String })),
+          reasoningContent: Schema.optional(
+            Schema.Struct({
+              text: Schema.optional(Schema.String),
+              signature: Schema.optional(Schema.String),
+              // Blob fields in Bedrock's JSON event stream are base64 strings.
+              redactedContent: Schema.optional(Schema.String),
+              // Vercel's Bedrock provider exposes the same delta under
+              // Anthropic's shorter `data` spelling.
+              data: Schema.optional(Schema.String),
+            }),
+          ),
+        }),
+      ),
+    }),
+  ),
+  contentBlockStop: Schema.optional(Schema.Struct({ contentBlockIndex: Schema.Number })),
+  messageStop: Schema.optional(
+    Schema.Struct({
+      stopReason: Schema.String,
+      additionalModelResponseFields: Schema.optional(Schema.Unknown),
+    }),
+  ),
+  metadata: Schema.optional(
+    Schema.Struct({
+      usage: Schema.optional(BedrockUsageSchema),
+      metrics: Schema.optional(Schema.Unknown),
+    }),
+  ),
+  exception: Schema.optional(Schema.Struct({ type: Schema.String, details: BedrockStreamException })),
+})
+type BedrockEvent = Schema.Schema.Type<typeof BedrockEvent>
+
+// =============================================================================
+// Request Lowering
+// =============================================================================
+const lowerToolSpec = (tool: ToolDefinition, inputSchema: JsonSchema): BedrockToolSpec => ({
+  toolSpec: {
+    name: tool.name,
+    ...(tool.description.trim().length > 0 ? { description: tool.description } : {}),
+    inputSchema: { json: inputSchema },
+  },
+})
+
+const lowerTools = (
+  compatibility: LanguageModelToolSchemaCompatibility | undefined,
+  breakpoints: BedrockCache.Breakpoints,
+  tools: ReadonlyArray<ToolDefinition>,
+): BedrockTool[] => {
+  const result: BedrockTool[] = []
+  for (const tool of tools) {
+    result.push(lowerToolSpec(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, compatibility)))
+    const cachePoint = BedrockCache.block(breakpoints, tool.cache)
+    if (cachePoint) result.push(cachePoint)
+  }
+  return result
+}
+
+const textWithCache = (
+  breakpoints: BedrockCache.Breakpoints,
+  text: string,
+  cache: CacheHint | undefined,
+): Array<BedrockTextBlock | BedrockCache.CachePointBlock> => {
+  const cachePoint = BedrockCache.block(breakpoints, cache)
+  return cachePoint ? [{ text }, cachePoint] : [{ text }]
+}
+
+const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
+  ProviderShared.matchToolChoice("Bedrock Converse", toolChoice, {
+    auto: () => ({ auto: {} }) as const,
+    none: () => undefined,
+    required: () => ({ any: {} }) as const,
+    tool: (name) => ({ tool: { name } }) as const,
+  })
+
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
+
+const reasoningSignature = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
+  if (part.encrypted !== undefined) return part.encrypted
+  if (ProviderShared.isRecord(metadata) && typeof metadata.signature === "string") return metadata.signature
+}
+
+const reasoningRedactedData = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
+  if (ProviderShared.isRecord(metadata) && typeof metadata.redactedData === "string") return metadata.redactedData
+}
+
+const removeEmptyToolInputKeys = (input: unknown): unknown => {
+  if (Array.isArray(input)) return input.map(removeEmptyToolInputKeys)
+  if (!ProviderShared.isRecord(input)) return input
+  return Object.fromEntries(
+    Object.entries(input).flatMap(([key, value]) => (key === "" ? [] : [[key, removeEmptyToolInputKeys(value)]])),
+  )
+}
+
+const lowerToolCall = (part: ToolCallPart, normalizeID: (id: string) => string): BedrockToolUseBlock => ({
+  toolUse: {
+    toolUseId: normalizeID(part.id),
+    // Models can emit names that Converse rejects when replayed in history.
+    name: part.name.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64) || "_",
+    input: removeEmptyToolInputKeys(part.input),
+  },
+})
+
+const lowerToolResultContent = Effect.fn("BedrockConverse.lowerToolResultContent")(function* (
+  part: ToolResultPart,
+  documentNames: Set<string>,
+) {
+  if (part.result.type === "text" || part.result.type === "error")
+    return [{ text: ProviderShared.toolResultText(part) }]
+  if (part.result.type === "json") return [{ json: part.result.value }]
+
+  const content: Array<Schema.Schema.Type<typeof BedrockToolResultContentItem>> = []
+  for (const item of part.result.value) {
+    if (item.type === "text") {
+      content.push({ text: item.text })
+      continue
+    }
+    const media = yield* BedrockMedia.lower(ProviderShared.toolFileMedia(item), documentNames)
+    content.push(...media)
+  }
+  return content
+})
+
+const lowerToolResult = Effect.fn("BedrockConverse.lowerToolResult")(function* (
+  part: ToolResultPart,
+  documentNames: Set<string>,
+  normalizeID: (id: string) => string,
+) {
+  return {
+    toolResult: {
+      toolUseId: normalizeID(part.id),
+      content: yield* lowerToolResultContent(part, documentNames),
+      status: part.result.type === "error" ? "error" : "success",
+    },
+  } satisfies BedrockToolResultBlock
+})
+
+const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
+  request: LLMRequest,
+  breakpoints: BedrockCache.Breakpoints,
+) {
+  const messages: BedrockMessage[] = []
+  const documentNames = new Set<string>()
+  // Mistral can reject replay IDs even when they satisfy Converse's broader ID syntax.
+  const normalizeID = request.model.id.includes("mistral.") ? MistralToolID.normalizer(request) : (id: string) => id
+  const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
+
+  for (const message of request.messages) {
+    if (message.role === "system") {
+      const part = yield* ProviderShared.wrappedSystemUpdate("Bedrock Converse", message)
+      const content = textWithCache(breakpoints, part.text, part.cache)
+      const previous = messages.at(-1)
+      if (previous?.role === "user")
+        messages[messages.length - 1] = { role: "user", content: [...previous.content, ...content] }
+      else messages.push({ role: "user", content })
+      continue
+    }
+
+    if (message.role === "user") {
+      const content: BedrockUserBlock[] = []
+      for (const part of message.content) {
+        if (!ProviderShared.supportsContent(part, ["text", "media"]))
+          return yield* ProviderShared.unsupportedContent("Bedrock Converse", "user", ["text", "media"])
+        if (part.type === "text") {
+          content.push(...textWithCache(breakpoints, part.text, part.cache))
+          continue
+        }
+        if (part.type === "media") {
+          content.push(...(yield* BedrockMedia.lower(part, documentNames)))
+          continue
+        }
+      }
+      const previous = messages.at(-1)
+      if (previous?.role === "user")
+        messages[messages.length - 1] = { role: "user", content: [...previous.content, ...content] }
+      else messages.push({ role: "user", content })
+      continue
+    }
+
+    if (message.role === "assistant") {
+      const content: BedrockAssistantBlock[] = []
+      for (const part of message.content) {
+        if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call"]))
+          return yield* ProviderShared.unsupportedContent("Bedrock Converse", "assistant", [
+            "text",
+            "reasoning",
+            "tool-call",
+          ])
+        if (part.type === "text") {
+          content.push(...textWithCache(breakpoints, part.text, part.cache))
+          continue
+        }
+        if (part.type === "reasoning") {
+          const signature = reasoningSignature(part, providerMetadataKey)
+          const redactedData = reasoningRedactedData(part, providerMetadataKey)
+          if (signature === undefined && redactedData !== undefined) {
+            content.push({ reasoningContent: { redactedContent: redactedData } })
+            continue
+          }
+          if (signature === undefined || signature.trim().length === 0) {
+            // Interrupted streams and model switches can leave unsigned reasoning.
+            // Preserve readable history as text rather than replay invalid reasoningContent.
+            if (part.text.trim().length > 0) content.push(...textWithCache(breakpoints, part.text, part.cache))
+            continue
+          }
+          content.push({ reasoningContent: { reasoningText: { text: part.text, signature } } })
+          continue
+        }
+        if (part.type === "tool-call") {
+          content.push(lowerToolCall(part, normalizeID))
+          continue
+        }
+      }
+      if (content.length > 0) messages.push({ role: "assistant", content })
+      continue
+    }
+
+    const content: BedrockUserBlock[] = []
+    for (const part of message.content) {
+      if (!ProviderShared.supportsContent(part, ["tool-result"]))
+        return yield* ProviderShared.unsupportedContent("Bedrock Converse", "tool", ["tool-result"])
+      content.push(yield* lowerToolResult(part, documentNames, normalizeID))
+      const cachePoint = BedrockCache.block(breakpoints, part.cache)
+      if (cachePoint) content.push(cachePoint)
+    }
+    const previous = messages.at(-1)
+    if (previous?.role === "user")
+      messages[messages.length - 1] = { role: "user", content: [...previous.content, ...content] }
+    else messages.push({ role: "user", content })
+  }
+
+  return messages
+})
+
+// System prompts share the cache-point convention: emit the text block, then
+// optionally a positional `cachePoint` marker.
+const lowerSystem = (breakpoints: BedrockCache.Breakpoints, system: ReadonlyArray<LLMRequest["system"][number]>) => {
+  const content = system
+    .filter((part) => part.text.length > 0)
+    .flatMap((part) => textWithCache(breakpoints, part.text, part.cache))
+  return content.length === 0 ? undefined : content
+}
+
+const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request: LLMRequest) {
+  const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
+  const flattened = ProviderShared.flattenToolRequest(request)
+  const generation = request.generation
+  // Bedrock-Claude shares Anthropic's 4-breakpoint cap. Spend the budget in
+  // tools → system → messages order to favour the highest-impact prefixes.
+  const breakpoints = BedrockCache.breakpoints(request.model.id)
+  const toolConfig = (() => {
+    if (flattened.tools.length === 0) return undefined
+    return {
+      tools: lowerTools(request.model.compatibility?.toolSchema, breakpoints, flattened.tools),
+      // Converse has no native "none". Keep definitions stable for prompt
+      // caching and omit only the unsupported choice.
+      toolChoice,
+    }
+  })()
+  const system = lowerSystem(breakpoints, request.system)
+  const messages = yield* lowerMessages(flattened.request, breakpoints)
+  if (breakpoints.dropped > 0) {
+    yield* Effect.logWarning(
+      `Bedrock Converse: dropped ${breakpoints.dropped} cache breakpoint(s); the API allows at most ${BedrockCache.BEDROCK_BREAKPOINT_CAP} per request.`,
+    )
+  }
+  const inferenceConfig = (() => {
+    if (
+      generation?.maxTokens === undefined &&
+      generation?.temperature === undefined &&
+      generation?.topP === undefined &&
+      (generation?.stop === undefined || generation.stop.length === 0)
+    )
+      return undefined
+    return {
+      maxTokens: generation?.maxTokens,
+      temperature: generation?.temperature,
+      topP: generation?.topP,
+      stopSequences: generation?.stop,
+    }
+  })()
+  return {
+    modelId: request.model.id,
+    messages,
+    system,
+    inferenceConfig,
+    toolConfig,
+    // Converse's base inferenceConfig has no topK; Anthropic/Nova accept it
+    // as a model-specific field, so it goes through additionalModelRequestFields.
+    additionalModelRequestFields: generation?.topK === undefined ? undefined : { top_k: generation.topK },
+  }
+})
+
+// =============================================================================
+// Stream Parsing
+// =============================================================================
+const mapFinishReason = (reason: string): FinishReason => {
+  if (reason === "end_turn" || reason === "stop_sequence") return "stop"
+  if (reason === "max_tokens" || reason === "model_context_window_exceeded") return "length"
+  if (reason === "tool_use") return "tool-calls"
+  if (reason === "content_filtered" || reason === "guardrail_intervened") return "content-filter"
+  return "unknown"
+}
+
+// AWS reports inputTokens separately from cache reads and writes.
+// Bedrock does not break reasoning out of outputTokens for current models.
+const mapUsage = (usage: BedrockUsageSchema | undefined, providerMetadataKey: string): Usage | undefined => {
+  if (!usage) return undefined
+  const inputTokens = ProviderShared.sumTokens(
+    usage.inputTokens,
+    usage.cacheReadInputTokens,
+    usage.cacheWriteInputTokens,
+  )
+  return new Usage({
+    inputTokens,
+    outputTokens: usage.outputTokens,
+    nonCachedInputTokens: usage.inputTokens,
+    cacheReadInputTokens: usage.cacheReadInputTokens,
+    cacheWriteInputTokens: usage.cacheWriteInputTokens,
+    totalTokens: ProviderShared.totalTokens(inputTokens, usage.outputTokens, usage.totalTokens),
+    providerMetadata: { [providerMetadataKey]: usage },
+  })
+}
+
+interface ParserState {
+  readonly providerMetadataKey: string
+  readonly tools: ToolStream.State<number>
+  // Bedrock splits the finish into `messageStop` (carries `stopReason`) and
+  // `metadata` (carries usage). Hold both in state so `onHalt` can emit exactly
+  // one finish after both chunks have had a chance to arrive.
+  readonly finishReason: FinishReasonDetails | undefined
+  readonly usage: Usage | undefined
+  readonly hasToolCalls: boolean
+  readonly lifecycle: Lifecycle.State
+  readonly reasoningSignatures: Readonly<Record<number, string>>
+  readonly reasoningRedactedContent: Readonly<Record<number, ReadonlyArray<Uint8Array>>>
+}
+
+const encodeRedactedContent = (chunks: ReadonlyArray<Uint8Array>) => {
+  const bytes = new Uint8Array(chunks.reduce((total, chunk) => total + chunk.length, 0))
+  chunks.reduce((offset, chunk) => {
+    bytes.set(chunk, offset)
+    return offset + chunk.length
+  }, 0)
+  return Encoding.encodeBase64(bytes)
+}
+
+const step = (state: ParserState, event: BedrockEvent) =>
+  Effect.gen(function* () {
+    if (event.contentBlockStart?.start?.toolUse) {
+      const index = event.contentBlockStart.contentBlockIndex
+      const events: LLMEvent[] = []
+      const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
+      return [
+        {
+          ...state,
+          lifecycle,
+          tools: ToolStream.start(state.tools, index, {
+            id: event.contentBlockStart.start.toolUse.toolUseId,
+            name: event.contentBlockStart.start.toolUse.name,
+          }),
+        },
+        [
+          ...events,
+          LLMEvent.toolInputStart({
+            id: event.contentBlockStart.start.toolUse.toolUseId,
+            name: event.contentBlockStart.start.toolUse.name,
+          }),
+        ],
+      ] as const
+    }
+
+    if (event.contentBlockDelta?.delta?.text) {
+      const events: LLMEvent[] = []
+      return [
+        {
+          ...state,
+          lifecycle: Lifecycle.textDelta(
+            state.lifecycle,
+            events,
+            `text-${event.contentBlockDelta.contentBlockIndex}`,
+            event.contentBlockDelta.delta.text,
+          ),
+        },
+        events,
+      ] as const
+    }
+
+    if (event.contentBlockDelta?.delta?.reasoningContent) {
+      const index = event.contentBlockDelta.contentBlockIndex
+      const reasoning = event.contentBlockDelta.delta.reasoningContent
+      const events: LLMEvent[] = []
+      const redactedChunks = yield* (() => {
+        if (reasoning.redactedContent === undefined) return Effect.succeed(undefined)
+        return Effect.fromResult(Encoding.decodeBase64(reasoning.redactedContent)).pipe(
+          Effect.map((chunk) => [...(state.reasoningRedactedContent[index] ?? []), chunk]),
+          Effect.mapError((cause) =>
+            ProviderShared.eventError(
+              ADAPTER,
+              "Bedrock Converse reasoningContent.redactedContent contains invalid base64 data",
+              undefined,
+              cause,
+            ),
+          ),
+        )
+      })()
+      const redactedData = redactedChunks === undefined ? reasoning.data : encodeRedactedContent(redactedChunks)
+      const metadata = (() => {
+        if (reasoning.signature) return providerMetadata(state.providerMetadataKey, { signature: reasoning.signature })
+        if (redactedData !== undefined) return providerMetadata(state.providerMetadataKey, { redactedData })
+      })()
+      const lifecycle = (() => {
+        if (reasoning.text === undefined && metadata === undefined) return state.lifecycle
+        return Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text ?? "", metadata)
+      })()
+      const reasoningRedactedContent = (() => {
+        if (redactedChunks !== undefined) return { ...state.reasoningRedactedContent, [index]: redactedChunks }
+        if (reasoning.data === undefined) return state.reasoningRedactedContent
+        return Object.fromEntries(
+          Object.entries(state.reasoningRedactedContent).filter(([key]) => key !== String(index)),
+        )
+      })()
+      const reasoningSignatures = (() => {
+        if (!reasoning.signature) return state.reasoningSignatures
+        return { ...state.reasoningSignatures, [index]: reasoning.signature }
+      })()
+      return [
+        {
+          ...state,
+          lifecycle,
+          reasoningSignatures,
+          reasoningRedactedContent,
+        },
+        events,
+      ] as const
+    }
+
+    if (event.contentBlockDelta?.delta?.toolUse) {
+      // A delta for a block that is not open, whether it already stopped or never
+      // started, has nothing to attach to and is dropped.
+      const result = ToolStream.append(
+        state.tools,
+        event.contentBlockDelta.contentBlockIndex,
+        event.contentBlockDelta.delta.toolUse.input,
+      )
+      if (!result) return [state, []] as const
+      const events: LLMEvent[] = []
+      const lifecycle = result.events.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
+      events.push(...result.events)
+      return [{ ...state, lifecycle, tools: result.tools }, events] as const
+    }
+
+    if (event.contentBlockStop) {
+      const index = event.contentBlockStop.contentBlockIndex
+      const result = yield* ToolStream.finish(ADAPTER, state.tools, index)
+      const events: LLMEvent[] = []
+      const resultEvents = result.events ?? []
+      const lifecycle = (() => {
+        if (resultEvents.length) return Lifecycle.stepStart(state.lifecycle, events)
+        const metadata = (() => {
+          const signature = state.reasoningSignatures[index]
+          if (signature) return providerMetadata(state.providerMetadataKey, { signature })
+          const redactedContent = state.reasoningRedactedContent[index]
+          if (redactedContent)
+            return providerMetadata(state.providerMetadataKey, {
+              redactedData: encodeRedactedContent(redactedContent),
+            })
+        })()
+        return Lifecycle.reasoningEnd(
+          Lifecycle.textEnd(state.lifecycle, events, `text-${index}`),
+          events,
+          `reasoning-${index}`,
+          metadata,
+        )
+      })()
+      events.push(...resultEvents)
+      return [
+        {
+          ...state,
+          hasToolCalls:
+            resultEvents.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
+            state.hasToolCalls,
+          lifecycle,
+          tools: result.tools,
+          reasoningSignatures: Object.fromEntries(
+            Object.entries(state.reasoningSignatures).filter(([key]) => key !== String(index)),
+          ),
+          reasoningRedactedContent: Object.fromEntries(
+            Object.entries(state.reasoningRedactedContent).filter(([key]) => key !== String(index)),
+          ),
+        },
+        events,
+      ] as const
+    }
+
+    if (event.messageStop) {
+      if (
+        event.messageStop.stopReason === "malformed_model_output" ||
+        event.messageStop.stopReason === "malformed_tool_use"
+      )
+        return yield* ProviderShared.eventError(
+          ADAPTER,
+          `Bedrock Converse stopped with ${event.messageStop.stopReason}`,
+          ProviderShared.encodeJson(event),
+        )
+      return [
+        {
+          ...state,
+          finishReason: {
+            normalized: mapFinishReason(event.messageStop.stopReason),
+            raw: event.messageStop.stopReason,
+          },
+        },
+        [],
+      ] as const
+    }
+
+    if (event.metadata) {
+      const usage = mapUsage(event.metadata.usage, state.providerMetadataKey) ?? state.usage
+      return [
+        {
+          ...state,
+          usage,
+        },
+        [],
+      ] as const
+    }
+
+    if (event.exception) {
+      const message =
+        event.exception.details.message ?? event.exception.details.originalMessage ?? "Bedrock Converse stream error"
+      const body = ProviderShared.encodeJson(event)
+      return yield* new AIError({
+        reason: classifyProviderFailure({
+          message,
+          rawBody: body,
+        }),
+      })
+    }
+
+    return [state, []] as const
+  })
+
+const framing = BedrockEventStream.framing(ADAPTER)
+
+const onHalt = (state: ParserState): ReadonlyArray<LLMEvent> => {
+  if (!state.finishReason) return []
+  const normalized = (() => {
+    if (state.finishReason.normalized === "stop" && state.hasToolCalls) return "tool-calls"
+    return state.finishReason.normalized
+  })()
+  const events: LLMEvent[] = []
+  Lifecycle.finish(state.lifecycle, events, {
+    reason: {
+      ...state.finishReason,
+      normalized,
+    },
+    usage: state.usage,
+  })
+  return events
+}
+
+// =============================================================================
+// Protocol And Bedrock Route
+// =============================================================================
+/**
+ * The Bedrock Converse protocol — request body construction, body schema, and
+ * the streaming-event state machine.
+ */
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: {
+    schema: BedrockConverseBody,
+    from: fromRequest,
+  },
+  stream: {
+    event: BedrockEvent,
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
+      tools: ToolStream.empty<number>(),
+      finishReason: undefined,
+      usage: undefined,
+      hasToolCalls: false,
+      lifecycle: Lifecycle.initial(),
+      reasoningSignatures: {},
+      reasoningRedactedContent: {},
+    }),
+    step,
+    onHalt: (state) => Effect.succeed(onHalt(state)),
+  },
+})
+
+export const route = Route.make({
+  id: ADAPTER,
+  provider: "bedrock",
+  providerMetadataKey: "bedrock",
+  protocol,
+  // Bedrock's URL embeds the region in the route endpoint host and the
+  // validated modelId in the path. We read the validated body so the URL
+  // matches the body that gets signed.
+  endpoint: Endpoint.path<BedrockConverseBody>(
+    ({ body }) => `/model/${encodeURIComponent(body.modelId)}/converse-stream`,
+  ),
+  auth: BedrockAuth.auth,
+  framing,
+})
+
+export const sigV4Auth = BedrockAuth.sigV4
+
+export * as BedrockConverse from "./bedrock-converse.js"

packages/ai/src/protocols/bedrock-event-stream.ts

@@ -0,0 +1,142 @@
+import { EventStreamCodec } from "@smithy/eventstream-codec"
+import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
+import { Effect, Encoding, Stream } from "effect"
+import { AIError, AIErrorReason, InvalidProviderOutputError } from "../schema/index.js"
+import { Framing } from "../route/framing.js"
+import { ProviderShared } from "./shared.js"
+
+// Bedrock streams responses using the AWS event stream binary protocol — each
+// frame is `[length:4][headers-length:4][prelude-crc:4][headers][payload][crc:4]`.
+// We use `@smithy/eventstream-codec` to validate framing and CRCs, then
+// reconstruct the JSON wrapping by `:event-type` so the chunk schema can match.
+const eventCodec = new EventStreamCodec(toUtf8, fromUtf8)
+const utf8 = new TextDecoder()
+
+// Cursor-tracking buffer state. Bytes accumulate in `buffer`; `offset` is the
+// read position. Reading by `subarray` is zero-copy. We only allocate a fresh
+// buffer when a new network chunk arrives and we need to append.
+interface FrameBufferState {
+  readonly buffer: Uint8Array
+  readonly offset: number
+}
+
+const initialFrameBuffer: FrameBufferState = { buffer: new Uint8Array(0), offset: 0 }
+
+type FrameInput = { readonly _tag: "Chunk"; readonly bytes: Uint8Array } | { readonly _tag: "End" }
+
+const endOfStream: FrameInput = { _tag: "End" }
+
+const appendChunk = (state: FrameBufferState, chunk: Uint8Array): FrameBufferState => {
+  const remaining = state.buffer.length - state.offset
+  // Compact: drop the consumed prefix and append the new chunk in one alloc.
+  // This bounds buffer growth to at most one network chunk past the live
+  // window, regardless of stream length.
+  const next = new Uint8Array(remaining + chunk.length)
+  next.set(state.buffer.subarray(state.offset), 0)
+  next.set(chunk, remaining)
+  return { buffer: next, offset: 0 }
+}
+
+const consumeFrames = (route: string) => (state: FrameBufferState, input: FrameInput) =>
+  Effect.gen(function* () {
+    if (input._tag === "End") {
+      const remaining = state.buffer.subarray(state.offset)
+      if (remaining.length > 0)
+        return yield* new AIError({
+          reason: new InvalidProviderOutputError({
+            route,
+            classification: "incomplete-stream",
+            message: `Incomplete Bedrock Converse event-stream frame: ${remaining.length} buffered bytes remain at end of stream`,
+            body: Encoding.encodeBase64(remaining),
+          }),
+        })
+      return [state, []] as const
+    }
+
+    let cursor = appendChunk(state, input.bytes)
+    const out: object[] = []
+    while (cursor.buffer.length - cursor.offset >= 4) {
+      const view = cursor.buffer.subarray(cursor.offset)
+      const totalLength = new DataView(view.buffer, view.byteOffset, view.byteLength).getUint32(0, false)
+      if (view.length < totalLength) break
+
+      const decoded = yield* Effect.try({
+        try: () => eventCodec.decode(view.subarray(0, totalLength)),
+        catch: (error) =>
+          ProviderShared.eventError(
+            route,
+            `Failed to decode Bedrock Converse event-stream frame: ${
+              error instanceof Error ? error.message : String(error)
+            }`,
+            Encoding.encodeBase64(view.subarray(0, totalLength)),
+            error,
+          ),
+      })
+      cursor = { buffer: cursor.buffer, offset: cursor.offset + totalLength }
+
+      const payload = utf8.decode(decoded.body)
+      const body = ProviderShared.encodeJson({ headers: decoded.headers, body: payload })
+      const messageType = decoded.headers[":message-type"]?.value
+      if (messageType === "error") {
+        const code = decoded.headers[":error-code"]?.value
+        const message = decoded.headers[":error-message"]?.value
+        return yield* ProviderShared.eventError(
+          route,
+          [code, message].filter((value): value is string => typeof value === "string").join(": ") ||
+            "Bedrock Converse event-stream error",
+          body,
+        )
+      }
+      const eventType =
+        messageType === "event"
+          ? decoded.headers[":event-type"]?.value
+          : messageType === "exception"
+            ? decoded.headers[":exception-type"]?.value
+            : undefined
+      if (typeof eventType !== "string") continue
+      if (!payload) continue
+      // The AWS event stream pads short payloads with a `p` field. Drop it
+      // before handing the object to the chunk schema. JSON decode goes
+      // through the shared Schema-driven codec to satisfy the package rule
+      // against ad-hoc `JSON.parse` calls.
+      const parsed = (yield* ProviderShared.parseJson(
+        route,
+        payload,
+        "Failed to parse Bedrock Converse event-stream payload",
+      ).pipe(
+        Effect.mapError(
+          (error) =>
+            new AIError({
+              reason: AIErrorReason.make({ ...error.reason, message: error.message, cause: error.reason.cause, body }),
+            }),
+        ),
+      )) as Record<string, unknown>
+      delete parsed.p
+      out.push({
+        ...(messageType === "exception"
+          ? { exception: { type: eventType, details: parsed } }
+          : { [eventType]: parsed }),
+        rawBody: body,
+      })
+    }
+    return [cursor, out] as const
+  })
+
+/**
+ * AWS event-stream framing for Bedrock Converse. Each frame is decoded by
+ * `@smithy/eventstream-codec` (length + header + payload + CRC) and rewrapped
+ * under its `:event-type` header so the chunk schema can match the JSON
+ * payload directly.
+ */
+export const framing = (route: string): Framing.Definition<object> => ({
+  id: "aws-event-stream",
+  body: (frame) => ("rawBody" in frame && typeof frame.rawBody === "string" ? frame.rawBody : undefined),
+  frame: (bytes) =>
+    bytes.pipe(
+      Stream.map((bytes): FrameInput => ({ _tag: "Chunk", bytes })),
+      Stream.concat(Stream.succeed(endOfStream)),
+      Stream.mapAccumEffect(() => initialFrameBuffer, consumeFrames(route)),
+    ),
+})
+
+export * as BedrockEventStream from "./bedrock-event-stream.js"

packages/ai/src/protocols/gemini.ts

@@ -0,0 +1,887 @@
+import { Effect, Option, Schema } from "effect"
+import { Tool } from "@opencode/schema/tool"
+import { Route } from "../route/client.js"
+import { Auth } from "../route/auth.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Framing } from "../route/framing.js"
+import { Protocol } from "../route/protocol.js"
+import {
+  AIError,
+  LLMEvent,
+  Usage,
+  type FinishReason,
+  type JsonSchema,
+  type LLMRequest,
+  type MediaPart,
+  type ProviderMetadata,
+  type TextPart,
+  type ToolCallPart,
+  type ToolDefinition,
+} from "../schema/index.js"
+import { classifyProviderFailure } from "../provider-error.js"
+import { Media } from "../media.js"
+import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
+import { GeminiToolSchema } from "./utils/gemini-tool-schema.js"
+import { Lifecycle } from "./utils/lifecycle.js"
+import { ToolSchemaProjection } from "./utils/tool-schema.js"
+
+const ADAPTER = "gemini"
+// Google documents this sentinel for replaying Gemini 3 function calls after their original signature was lost.
+const SKIP_THOUGHT_SIGNATURE_VALIDATOR = "skip_thought_signature_validator"
+export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
+
+// Gemini 3 rejects replayed function calls without a thought signature. Google's SDKs avoid that in normal chats by
+// retaining complete model responses, but OpenCode reconstructs durable history and may encounter an unsigned call
+// from an older or external session. Model IDs are open-ended, so unknown Gemini aliases inherit current behavior.
+const requiresThoughtSignatureFallback = (modelID: string) => {
+  if (!/(^|\/)gemini-/i.test(modelID)) return false
+  if (/(^|\/)gemini-(?:1|2)(?:[.-]|$)/i.test(modelID)) return false
+  if (/(^|\/)gemini-pro(?:-vision)?$/i.test(modelID)) return false
+  return !/(^|\/)gemini-robotics-er-1\.5(?:[.-]|$)/i.test(modelID)
+}
+
+// Gemini 3 accepts media nested inside function responses; matched Gemini 2.5 variants reject it,
+// so their tool-result attachments lower as a separate user turn instead.
+const routesLegacyToolMedia = (modelID: string) => /gemini-2[.-]5(?:[.-]|$)/i.test(modelID)
+
+// Blacklist: Gemini 1.x/2.x ignore or reject explicit function call ids.
+// Every other model id (Gemini 3+, gemma, anything unrecognized) gets them.
+const omitsFunctionCallIds = (modelID: string) => {
+  const match = /^gemini(?:-live)?-(\d+)/i.exec(modelID)
+  return match !== null && Number(match[1]) < 3
+}
+
+export interface OptionsInput {
+  readonly [key: string]: unknown
+  readonly cachedContent?: string
+  readonly safetySettings?: ReadonlyArray<{
+    readonly category:
+      | "HARM_CATEGORY_UNSPECIFIED"
+      | "HARM_CATEGORY_HATE_SPEECH"
+      | "HARM_CATEGORY_DANGEROUS_CONTENT"
+      | "HARM_CATEGORY_HARASSMENT"
+      | "HARM_CATEGORY_SEXUALLY_EXPLICIT"
+      | "HARM_CATEGORY_CIVIC_INTEGRITY"
+      | (string & {})
+    readonly threshold:
+      | "HARM_BLOCK_THRESHOLD_UNSPECIFIED"
+      | "BLOCK_LOW_AND_ABOVE"
+      | "BLOCK_MEDIUM_AND_ABOVE"
+      | "BLOCK_ONLY_HIGH"
+      | "BLOCK_NONE"
+      | "OFF"
+      | (string & {})
+  }>
+  readonly serviceTier?: "standard" | "flex" | "priority" | (string & {})
+  readonly thinkingConfig?: {
+    readonly thinkingBudget?: number
+    readonly includeThoughts?: boolean
+    readonly thinkingLevel?: "minimal" | "low" | "medium" | "high" | (string & {})
+  }
+}
+
+export type ProviderOptionsInput = OptionsInput
+
+// =============================================================================
+// Request Body Schema
+// =============================================================================
+// Gemini is known to send explicit `null` for optional streaming fields
+// (usage counts, flags, whole subtrees), so every response-side optional uses
+// `optionalNull` instead of bare `Schema.optional`. The same part/content
+// schemas lower the outbound request body; encoding drops `undefined` keys,
+// so the shared schemas stay safe there.
+const GeminiTextPart = Schema.Struct({
+  text: Schema.String,
+  thought: optionalNull(Schema.Boolean),
+  thoughtSignature: optionalNull(Schema.String),
+})
+
+const GeminiInlineDataPart = Schema.Struct({
+  inlineData: Schema.Struct({
+    mimeType: Schema.String,
+    data: Schema.String,
+  }),
+  thoughtSignature: optionalNull(Schema.String),
+})
+type GeminiInlineDataPart = Schema.Schema.Type<typeof GeminiInlineDataPart>
+
+/** Gemini Files API reference; the only remote input Gemini accepts. */
+const GeminiFileDataPart = Schema.Struct({
+  fileData: Schema.Struct({
+    mimeType: Schema.String,
+    fileUri: Schema.String,
+  }),
+})
+
+const GeminiFunctionCallPart = Schema.Struct({
+  functionCall: Schema.Struct({
+    id: optionalNull(Schema.String),
+    name: Schema.String,
+    args: Schema.optional(Schema.Unknown),
+  }),
+  thoughtSignature: optionalNull(Schema.String),
+})
+
+const GeminiFunctionResponsePart = Schema.Struct({
+  functionResponse: Schema.Struct({
+    id: Schema.optional(Schema.String),
+    name: Schema.String,
+    response: Schema.Unknown,
+    parts: Schema.optional(Schema.Array(GeminiInlineDataPart)),
+  }),
+})
+
+const GeminiContentPart = Schema.Union([
+  GeminiTextPart,
+  GeminiInlineDataPart,
+  GeminiFileDataPart,
+  GeminiFunctionCallPart,
+  GeminiFunctionResponsePart,
+])
+const decodeGeminiContentPart = Schema.decodeUnknownOption(GeminiContentPart)
+
+const GeminiContent = Schema.Struct({
+  role: optionalNull(Schema.Literals(["user", "model"])),
+  parts: optionalNull(Schema.Array(GeminiContentPart)),
+})
+type GeminiContent = Schema.Schema.Type<typeof GeminiContent>
+
+const GeminiResponseContent = Schema.Struct({
+  role: optionalNull(Schema.Literals(["user", "model"])),
+  parts: optionalNull(Schema.Array(Schema.Unknown)),
+})
+
+const GeminiSystemInstruction = Schema.Struct({
+  parts: Schema.Array(Schema.Struct({ text: Schema.String })),
+})
+
+const GeminiFunctionDeclaration = Schema.Struct({
+  name: Schema.String,
+  description: Schema.String,
+  parameters: Schema.optional(JsonObject),
+})
+
+const GeminiTool = Schema.Struct({
+  functionDeclarations: Schema.Array(GeminiFunctionDeclaration),
+})
+
+const GeminiToolConfig = Schema.Struct({
+  functionCallingConfig: Schema.Struct({
+    mode: Schema.Literals(["AUTO", "NONE", "ANY"]),
+    allowedFunctionNames: optionalArray(Schema.String),
+  }),
+})
+
+const GeminiThinkingConfig = Schema.Struct({
+  thinkingBudget: Schema.optional(Schema.Number),
+  includeThoughts: Schema.optional(Schema.Boolean),
+  thinkingLevel: Schema.optional(Schema.String),
+})
+
+const GeminiSafetySetting = Schema.Struct({
+  category: Schema.String,
+  threshold: Schema.String,
+})
+
+const GeminiGenerationConfig = Schema.Struct({
+  maxOutputTokens: Schema.optional(Schema.Number),
+  temperature: Schema.optional(Schema.Number),
+  topP: Schema.optional(Schema.Number),
+  topK: Schema.optional(Schema.Number),
+  frequencyPenalty: Schema.optional(Schema.Number),
+  presencePenalty: Schema.optional(Schema.Number),
+  seed: Schema.optional(Schema.Number),
+  stopSequences: optionalArray(Schema.String),
+  thinkingConfig: Schema.optional(GeminiThinkingConfig),
+})
+
+const GeminiBodyFields = {
+  cachedContent: Schema.optional(Schema.String),
+  contents: Schema.Array(GeminiContent),
+  labels: Schema.optional(Schema.Record(Schema.String, Schema.String)),
+  safetySettings: optionalArray(GeminiSafetySetting),
+  serviceTier: Schema.optional(Schema.String),
+  systemInstruction: Schema.optional(GeminiSystemInstruction),
+  tools: optionalArray(GeminiTool),
+  toolConfig: Schema.optional(GeminiToolConfig),
+  generationConfig: Schema.optional(GeminiGenerationConfig),
+}
+const GeminiBody = Schema.Struct(GeminiBodyFields)
+export type GeminiBody = Schema.Schema.Type<typeof GeminiBody>
+
+const GeminiUsage = Schema.Struct({
+  cachedContentTokenCount: optionalNull(Schema.Number),
+  thoughtsTokenCount: optionalNull(Schema.Number),
+  promptTokenCount: optionalNull(Schema.Number),
+  candidatesTokenCount: optionalNull(Schema.Number),
+  totalTokenCount: optionalNull(Schema.Number),
+})
+type GeminiUsage = Schema.Schema.Type<typeof GeminiUsage>
+
+const GeminiCandidate = Schema.Struct({
+  content: optionalNull(GeminiResponseContent),
+  finishReason: optionalNull(Schema.String),
+})
+
+const GeminiPromptFeedback = Schema.StructWithRest(
+  Schema.Struct({
+    blockReason: optionalNull(Schema.String),
+    blockReasonMessage: optionalNull(Schema.String),
+    safetyRatings: optionalNull(Schema.Unknown),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type GeminiPromptFeedback = Schema.Schema.Type<typeof GeminiPromptFeedback>
+
+const GeminiEvent = Schema.Struct({
+  error: Schema.optional(Schema.Unknown),
+  candidates: optionalNull(Schema.Array(GeminiCandidate)),
+  promptFeedback: optionalNull(GeminiPromptFeedback),
+  usageMetadata: optionalNull(GeminiUsage),
+})
+type GeminiEvent = Schema.Schema.Type<typeof GeminiEvent>
+
+interface ParserState {
+  readonly route: string
+  readonly providerMetadataKey: string
+  readonly finishReason?: string
+  readonly hasToolCalls: boolean
+  readonly promptFeedback?: GeminiPromptFeedback
+  readonly usage?: Usage
+  readonly lifecycle: Lifecycle.State
+  readonly reasoningSignature?: string
+  readonly textSignature?: string
+  readonly reasoningId?: string
+  readonly textId?: string
+  readonly nextReasoningId: number
+  readonly nextTextId: number
+  readonly seenCallIds?: ReadonlySet<string>
+}
+
+// =============================================================================
+// Tool Schema Conversion
+// =============================================================================
+// Tool-schema conversion has two distinct concerns:
+//
+// 1. Sanitize — fix common authoring mistakes Gemini rejects: integer/number
+//    enums (must be strings), `required` entries that don't match a property,
+//    untyped arrays (`items` must be present), and `properties`/`required`
+//    keys on non-object scalars. Mirrors OpenCode's historical Gemini rules.
+//
+// 2. Project — lossy mapping from JSON Schema to Gemini's schema dialect:
+//    drop empty root parameter schemas while preserving nested empty objects,
+//    expand type arrays into `anyOf`, derive `nullable: true` from null members,
+//    coerce `const` to `[const]` enum, recurse properties/items, and propagate
+//    only an allowlisted set of keys (description, required, format, type,
+//    nullable, enum, properties, items, allOf, anyOf, oneOf, minLength).
+//    Anything outside the allowlist (e.g. `additionalProperties`, `$ref`) is
+//    silently dropped.
+//
+// Sanitize runs first, then project. The implementation lives in
+// `utils/gemini-tool-schema` so this protocol keeps the same shape as the other
+// provider protocols.
+
+// =============================================================================
+// Request Lowering
+// =============================================================================
+const lowerTool = (tool: ToolDefinition, inputSchema: JsonSchema) => ({
+  name: tool.name,
+  description: tool.description,
+  parameters: GeminiToolSchema.convert(inputSchema),
+})
+
+const lowerToolConfig = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
+  ProviderShared.matchToolChoice("Gemini", toolChoice, {
+    auto: () => ({ functionCallingConfig: { mode: "AUTO" as const } }),
+    none: () => ({ functionCallingConfig: { mode: "NONE" as const } }),
+    required: () => ({ functionCallingConfig: { mode: "ANY" as const } }),
+    tool: (name) => ({ functionCallingConfig: { mode: "ANY" as const, allowedFunctionNames: [name] } }),
+  })
+
+// Gemini does not fetch public URLs; inline payloads and Gemini Files references are the accepted inputs.
+const lowerContentPart = Effect.fn("Gemini.lowerContentPart")(function* (part: TextPart | MediaPart) {
+  if (part.type === "text") return { text: part.text }
+  const source = part.media.source
+  if (source.type === "ref" && source.provider === "google")
+    return { fileData: { mimeType: part.media.mediaType, fileUri: source.id } }
+  const media = yield* ProviderShared.requireInlineMedia("Gemini", part.media)
+  return { inlineData: { mimeType: media.mime, data: media.base64 } }
+})
+
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
+
+const thoughtSignature = (metadata: ProviderMetadata | undefined, key: string) => {
+  const value = metadata?.[key]
+  return ProviderShared.isRecord(value) && typeof value.thoughtSignature === "string"
+    ? value.thoughtSignature
+    : undefined
+}
+
+const lowerToolCall = (part: ToolCallPart, omitIds: boolean, metadataKey: string) => ({
+  functionCall: { ...(omitIds ? {} : { id: part.id }), name: part.name, args: part.input },
+  thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
+})
+
+const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMRequest) {
+  const contents: GeminiContent[] = []
+  const metadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
+  const omitCallIds = omitsFunctionCallIds(request.model.id)
+  const legacyToolMedia = routesLegacyToolMedia(request.model.id)
+  let pendingMedia: GeminiInlineDataPart[] | undefined
+  const flushMedia = () => {
+    if (!pendingMedia) return
+    contents.push({ role: "user", parts: [{ text: "Attached media from tool result:" }, ...pendingMedia] })
+    pendingMedia = undefined
+  }
+
+  for (const message of request.messages) {
+    if (message.role !== "tool") flushMedia()
+    if (message.role === "system") {
+      const part = yield* ProviderShared.wrappedSystemUpdate("Gemini", message)
+      const previous = contents.at(-1)
+      // Gemini rejects a continuation whose function-response turn carries extra
+      // parts, so an update after a tool result starts its own user turn.
+      if (previous?.role === "user" && !(previous.parts ?? []).some((item) => "functionResponse" in item))
+        contents[contents.length - 1] = { role: "user", parts: [...(previous.parts ?? []), { text: part.text }] }
+      else contents.push({ role: "user", parts: [{ text: part.text }] })
+      continue
+    }
+
+    if (message.role === "user") {
+      const parts: Array<Schema.Schema.Type<typeof GeminiContentPart>> = []
+      for (const part of message.content) {
+        if (!ProviderShared.supportsContent(part, ["text", "media"]))
+          return yield* ProviderShared.unsupportedContent("Gemini", "user", ["text", "media"])
+        parts.push(yield* lowerContentPart(part))
+      }
+      contents.push({ role: "user", parts })
+      continue
+    }
+
+    if (message.role === "assistant") {
+      const parts: Array<Schema.Schema.Type<typeof GeminiContentPart>> = []
+      // Parallel Gemini 3 calls may carry one signature on the first call; unsigned sibling calls are valid.
+      let hasSignedToolCall = false
+      for (const part of message.content) {
+        if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call", "media"]))
+          return yield* ProviderShared.unsupportedContent("Gemini", "assistant", [
+            "text",
+            "reasoning",
+            "tool-call",
+            "media",
+          ])
+        if (part.type === "text") {
+          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey) })
+          continue
+        }
+        // Generated images replay as model-role inline data so multi-turn image editing keeps the prior output.
+        if (part.type === "media") {
+          const lowered = yield* lowerContentPart(part)
+          parts.push({ ...lowered, thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey) })
+          continue
+        }
+        if (part.type === "reasoning") {
+          parts.push({
+            text: part.text,
+            thought: true,
+            thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
+          })
+          continue
+        }
+        if (part.type === "tool-call") {
+          const lowered = lowerToolCall(part, omitCallIds, metadataKey)
+          const signature = lowered.thoughtSignature
+          parts.push({
+            ...lowered,
+            thoughtSignature:
+              signature ??
+              (requiresThoughtSignatureFallback(request.model.id) && !hasSignedToolCall
+                ? SKIP_THOUGHT_SIGNATURE_VALIDATOR
+                : undefined),
+          })
+          if (signature !== undefined) hasSignedToolCall = true
+          continue
+        }
+      }
+      contents.push({ role: "model", parts })
+      continue
+    }
+
+    const parts: Array<Schema.Schema.Type<typeof GeminiContentPart>> = []
+    for (const part of message.content) {
+      if (!ProviderShared.supportsContent(part, ["tool-result"]))
+        return yield* ProviderShared.unsupportedContent("Gemini", "tool", ["tool-result"])
+      if (part.result.type !== "content") {
+        parts.push({
+          functionResponse: {
+            ...(omitCallIds ? {} : { id: part.id }),
+            name: part.name,
+            response: {
+              name: part.name,
+              content: ProviderShared.toolResultText(part),
+            },
+          },
+        })
+        continue
+      }
+      const content: ReadonlyArray<Tool.Content> = part.result.value
+      const text = content.filter((item) => item.type === "text").map((item) => item.text)
+      const media: GeminiInlineDataPart[] = []
+      for (const item of content) {
+        if (item.type === "text") continue
+        const value = yield* ProviderShared.requireInlineMedia("Gemini", ProviderShared.toolFileMedia(item).media)
+        media.push({ inlineData: { mimeType: value.mime, data: value.base64 } })
+      }
+      if (legacyToolMedia && media.length > 0) (pendingMedia ??= []).push(...media)
+      parts.push({
+        functionResponse: {
+          ...(omitCallIds ? {} : { id: part.id }),
+          name: part.name,
+          response: {
+            name: part.name,
+            content: text.join("\n"),
+          },
+          parts: legacyToolMedia || media.length === 0 ? undefined : media,
+        },
+      })
+    }
+    // Gemini requires every response to a parallel call batch in one user turn,
+    // so consecutive tool results join the open function-response turn.
+    const previous = contents.at(-1)
+    if (previous?.role === "user" && (previous.parts ?? []).some((item) => "functionResponse" in item))
+      contents[contents.length - 1] = { role: "user", parts: [...(previous.parts ?? []), ...parts] }
+    else contents.push({ role: "user", parts })
+  }
+
+  flushMedia()
+  return contents
+})
+
+const resolveOptions = (request: LLMRequest) => {
+  const input = request.providerOptions
+  const value = input?.thinkingConfig
+  const thinkingConfig = {
+    thinkingBudget:
+      ProviderShared.isRecord(value) && typeof value.thinkingBudget === "number" ? value.thinkingBudget : undefined,
+    includeThoughts:
+      ProviderShared.isRecord(value) && typeof value.includeThoughts === "boolean"
+        ? value.includeThoughts
+        : ProviderShared.isRecord(value)
+          ? true
+          : undefined,
+    thinkingLevel:
+      ProviderShared.isRecord(value) && typeof value.thinkingLevel === "string" ? value.thinkingLevel : undefined,
+  }
+  return {
+    cachedContent: typeof input?.cachedContent === "string" ? input.cachedContent : undefined,
+    safetySettings: mapSafetySettings(input?.safetySettings),
+    serviceTier: typeof input?.serviceTier === "string" ? input.serviceTier : undefined,
+    thinkingConfig: Object.values(thinkingConfig).some((item) => item !== undefined) ? thinkingConfig : undefined,
+  }
+}
+
+function mapSafetySettings(value: unknown) {
+  if (!Array.isArray(value)) return undefined
+  const settings = value.flatMap((item) =>
+    ProviderShared.isRecord(item) && typeof item.category === "string" && typeof item.threshold === "string"
+      ? [{ category: item.category, threshold: item.threshold }]
+      : [],
+  )
+  return settings
+}
+
+const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMRequest) {
+  const flattened = ProviderShared.flattenToolRequest(request)
+  const hasTools = flattened.tools.length > 0
+  const generation = request.generation
+  const options = resolveOptions(request)
+  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
+  const generationConfig = {
+    maxOutputTokens: generation?.maxTokens,
+    temperature: generation?.temperature,
+    topP: generation?.topP,
+    topK: generation?.topK,
+    frequencyPenalty: generation?.frequencyPenalty,
+    presencePenalty: generation?.presencePenalty,
+    seed: generation?.seed,
+    stopSequences: generation?.stop,
+    thinkingConfig: options.thinkingConfig,
+  }
+
+  return {
+    cachedContent: options.cachedContent,
+    contents: yield* lowerMessages(flattened.request),
+    safetySettings: options.safetySettings,
+    serviceTier: options.serviceTier,
+    systemInstruction:
+      request.system.length === 0 ? undefined : { parts: [{ text: ProviderShared.joinText(request.system) }] },
+    tools: hasTools
+      ? [
+          {
+            functionDeclarations: flattened.tools.map((tool) =>
+              lowerTool(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility)),
+            ),
+          },
+        ]
+      : undefined,
+    toolConfig: hasTools && request.toolChoice ? yield* lowerToolConfig(request.toolChoice) : undefined,
+    generationConfig: Object.values(generationConfig).some((value) => value !== undefined)
+      ? generationConfig
+      : undefined,
+  }
+})
+
+// =============================================================================
+// Stream Parsing
+// =============================================================================
+// Gemini reports `promptTokenCount` (inclusive total) with a
+// `cachedContentTokenCount` subset. `candidatesTokenCount` is *exclusive*
+// of `thoughtsTokenCount` — visible-only, not a total — so we sum the two
+// to produce the inclusive `outputTokens` the rest of the contract expects.
+const mapUsage = (usage: GeminiUsage | undefined, metadataKey: string) => {
+  if (!usage) return undefined
+  // Explicit provider nulls decode as `null`; normalize to `undefined` so the
+  // token arithmetic below treats them like absent counts.
+  const promptTokens = usage.promptTokenCount ?? undefined
+  const cached = usage.cachedContentTokenCount ?? undefined
+  const thoughts = usage.thoughtsTokenCount ?? undefined
+  const visible = usage.candidatesTokenCount ?? undefined
+  const nonCached = ProviderShared.subtractTokens(promptTokens, cached)
+  // `candidatesTokenCount` is visible-only; sum with thoughts to produce the
+  // inclusive `outputTokens` the contract expects. Only compute the total
+  // when the visible component is reported — otherwise we'd fabricate an
+  // inclusive number from a partial breakdown.
+  const outputTokens = visible !== undefined ? visible + (thoughts ?? 0) : undefined
+  return new Usage({
+    inputTokens: promptTokens,
+    outputTokens,
+    nonCachedInputTokens: nonCached,
+    cacheReadInputTokens: cached,
+    reasoningTokens: thoughts,
+    totalTokens: ProviderShared.totalTokens(promptTokens, outputTokens, usage.totalTokenCount ?? undefined),
+    providerMetadata: providerMetadata(metadataKey, usage),
+  })
+}
+
+const mapFinishReason = (finishReason: string | undefined, hasToolCalls: boolean): FinishReason => {
+  if (finishReason === undefined) return hasToolCalls ? "tool-calls" : "unknown"
+  if (finishReason === "STOP") return hasToolCalls ? "tool-calls" : "stop"
+  if (finishReason === "MAX_TOKENS") return "length"
+  if (
+    finishReason === "IMAGE_SAFETY" ||
+    finishReason === "RECITATION" ||
+    finishReason === "SAFETY" ||
+    finishReason === "BLOCKLIST" ||
+    finishReason === "PROHIBITED_CONTENT" ||
+    finishReason === "SPII" ||
+    finishReason === "MODEL_ARMOR" ||
+    finishReason === "IMAGE_PROHIBITED_CONTENT" ||
+    finishReason === "IMAGE_RECITATION" ||
+    finishReason === "LANGUAGE"
+  )
+    return "content-filter"
+  if (
+    finishReason === "MALFORMED_FUNCTION_CALL" ||
+    finishReason === "UNEXPECTED_TOOL_CALL" ||
+    finishReason === "NO_IMAGE" ||
+    finishReason === "TOO_MANY_TOOL_CALLS" ||
+    finishReason === "MISSING_THOUGHT_SIGNATURE" ||
+    finishReason === "MALFORMED_RESPONSE"
+  )
+    return "error"
+  return "unknown"
+}
+
+const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
+  // `?? undefined` normalizes an explicit `null` blockReason back to absent so
+  // the "nothing to finish" check below keeps its meaning.
+  const promptBlockReason =
+    state.finishReason === undefined ? (state.promptFeedback?.blockReason ?? undefined) : undefined
+  const finishReason = state.finishReason ?? promptBlockReason
+  if (finishReason === undefined && state.usage === undefined) return []
+
+  const events: LLMEvent[] = []
+  let lifecycle = state.lifecycle
+  if (state.reasoningId !== undefined)
+    lifecycle = Lifecycle.reasoningEnd(
+      lifecycle,
+      events,
+      state.reasoningId,
+      state.reasoningSignature === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { thoughtSignature: state.reasoningSignature }),
+    )
+  if (state.textId !== undefined)
+    lifecycle = Lifecycle.textEnd(
+      lifecycle,
+      events,
+      state.textId,
+      state.textSignature === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { thoughtSignature: state.textSignature }),
+    )
+  Lifecycle.finish(lifecycle, events, {
+    reason: {
+      normalized:
+        promptBlockReason === undefined ? mapFinishReason(finishReason, state.hasToolCalls) : "content-filter",
+      raw: finishReason,
+    },
+    usage: state.usage,
+    providerMetadata:
+      state.promptFeedback === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { promptFeedback: state.promptFeedback }),
+  })
+  return events
+}
+
+const step = (state: ParserState, event: GeminiEvent) => {
+  if (ProviderShared.isRecord(event.error)) {
+    const body = ProviderShared.encodeJson(event)
+    return Effect.fail(
+      new AIError({
+        reason: classifyProviderFailure({
+          message:
+            typeof event.error.message === "string" && event.error.message.length > 0
+              ? event.error.message
+              : typeof event.error.status === "string" && event.error.status.length > 0
+                ? event.error.status
+                : "Gemini provider error",
+          status: typeof event.error.code === "number" ? event.error.code : undefined,
+          rawBody: body,
+        }),
+      }),
+    )
+  }
+  if ("error" in event)
+    return Effect.fail(
+      ProviderShared.eventError(state.route, `Invalid ${state.route} stream event`, ProviderShared.encodeJson(event)),
+    )
+  const nextState = {
+    ...state,
+    promptFeedback: event.promptFeedback ?? state.promptFeedback,
+    usage: event.usageMetadata
+      ? (mapUsage(event.usageMetadata, state.providerMetadataKey) ?? state.usage)
+      : state.usage,
+  }
+  const candidate = event.candidates?.[0]
+  if (candidate?.finishReason && mapFinishReason(candidate.finishReason, state.hasToolCalls) === "error")
+    return Effect.fail(
+      ProviderShared.eventError(
+        state.route,
+        `Gemini stopped with ${candidate.finishReason}`,
+        ProviderShared.encodeJson(event),
+      ),
+    )
+  if (!candidate?.content)
+    return Effect.succeed([
+      { ...nextState, finishReason: candidate?.finishReason ?? nextState.finishReason },
+      [],
+    ] as const)
+
+  const events: LLMEvent[] = []
+  let hasToolCalls = nextState.hasToolCalls
+  let lifecycle = nextState.lifecycle
+  let reasoningSignature = nextState.reasoningSignature
+  let textSignature = nextState.textSignature
+  let reasoningId = nextState.reasoningId
+  let textId = nextState.textId
+  let nextReasoningId = nextState.nextReasoningId
+  let nextTextId = nextState.nextTextId
+  // Supplier ids must be tracked across chunks of the same response, not just within one event's parts.
+  const seenCallIds = new Set(nextState.seenCallIds)
+
+  for (const input of candidate.content.parts ?? []) {
+    if (
+      ProviderShared.isRecord(input) &&
+      !("text" in input) &&
+      !("inlineData" in input) &&
+      !("functionCall" in input) &&
+      !("functionResponse" in input)
+    )
+      continue
+    const decoded = decodeGeminiContentPart(input)
+    if (Option.isNone(decoded))
+      return Effect.fail(
+        ProviderShared.eventError(ADAPTER, `Invalid ${state.route} stream event`, ProviderShared.encodeJson(event)),
+      )
+    const part = decoded.value
+    const signature = "thoughtSignature" in part && part.thoughtSignature ? part.thoughtSignature : undefined
+    // Gemini attaches replay signatures to thought parts, visible text, or function calls;
+    // each block kind must retain the signature attached to its own parts.
+    if (signature !== undefined && "thought" in part && part.thought) reasoningSignature = signature
+    else if (signature !== undefined && "text" in part) textSignature = signature
+    // Image-capable Gemini models return generated images as inline data parts; surface them as first-class output.
+    if ("inlineData" in part) {
+      lifecycle = Lifecycle.stepStart(lifecycle, events)
+      events.push(
+        LLMEvent.media({
+          media: Media.base64(part.inlineData.data, part.inlineData.mimeType),
+          providerMetadata: signature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: signature })
+            : undefined,
+        }),
+      )
+      continue
+    }
+    if ("text" in part && part.text.length > 0) {
+      if (part.thought) {
+        if (textId !== undefined) {
+          lifecycle = Lifecycle.textEnd(
+            lifecycle,
+            events,
+            textId,
+            textSignature
+              ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature })
+              : undefined,
+          )
+          textId = undefined
+          textSignature = undefined
+        }
+        if (reasoningId === undefined) {
+          reasoningId = `reasoning-${nextReasoningId}`
+          nextReasoningId += 1
+        }
+        lifecycle = Lifecycle.reasoningDelta(
+          lifecycle,
+          events,
+          reasoningId,
+          part.text,
+          signature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: signature }) : undefined,
+        )
+        continue
+      }
+      if (reasoningId !== undefined) {
+        lifecycle = Lifecycle.reasoningEnd(
+          lifecycle,
+          events,
+          reasoningId,
+          reasoningSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+            : undefined,
+        )
+        reasoningId = undefined
+        reasoningSignature = undefined
+      }
+      if (textId === undefined) {
+        textId = `text-${nextTextId}`
+        nextTextId += 1
+      }
+      lifecycle = Lifecycle.textDelta(
+        lifecycle,
+        events,
+        textId,
+        part.text,
+        textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
+      )
+      textSignature = undefined
+      continue
+    }
+
+    if ("functionCall" in part) {
+      const input = part.functionCall.args === undefined ? {} : part.functionCall.args
+      // Gemini 2.0+ supplies a unique function call ID on the part; when omitted (e.g. Gemini 1.5),
+      // generate a globally unique ID rather than a per-request counter to prevent cross-request collisions in downstream registries.
+      // A repeated supplier id would replay as two identical calls, so only the first occurrence keeps it.
+      // A `null` supplier id normalizes to absent so the generated-id fallback applies.
+      const supplied = part.functionCall.id ?? undefined
+      const duplicate = supplied !== undefined && seenCallIds.has(supplied)
+      if (supplied !== undefined) seenCallIds.add(supplied)
+      const id = supplied !== undefined && !duplicate ? supplied : `tool_${crypto.randomUUID().replaceAll("-", "")}`
+      if (reasoningId !== undefined) {
+        lifecycle = Lifecycle.reasoningEnd(
+          lifecycle,
+          events,
+          reasoningId,
+          reasoningSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+            : undefined,
+        )
+        reasoningId = undefined
+        reasoningSignature = undefined
+      }
+      if (textId !== undefined) {
+        lifecycle = Lifecycle.textEnd(
+          lifecycle,
+          events,
+          textId,
+          textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
+        )
+        textId = undefined
+        textSignature = undefined
+      }
+      lifecycle = Lifecycle.stepStart(lifecycle, events)
+      events.push(
+        LLMEvent.toolCall({
+          id,
+          name: part.functionCall.name,
+          input,
+          providerMetadata: part.thoughtSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: part.thoughtSignature })
+            : undefined,
+        }),
+      )
+      hasToolCalls = true
+    }
+  }
+
+  return Effect.succeed([
+    {
+      ...nextState,
+      hasToolCalls,
+      lifecycle,
+      reasoningSignature,
+      textSignature,
+      reasoningId,
+      textId,
+      nextReasoningId,
+      nextTextId,
+      seenCallIds,
+      finishReason: candidate.finishReason ?? nextState.finishReason,
+    },
+    events,
+  ] as const)
+}
+
+// =============================================================================
+// Protocol And Gemini Route
+// =============================================================================
+/**
+ * The Gemini protocol — request body construction, body schema, and the
+ * streaming-event state machine shared by Google AI Studio and Vertex Gemini.
+ */
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: {
+    schema: GeminiBody,
+    from: fromRequest,
+  },
+  stream: {
+    event: Protocol.jsonEvent(GeminiEvent),
+    initial: (request) => ({
+      route: `${request.model.provider}/${request.model.route.id}`,
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
+      hasToolCalls: false,
+      lifecycle: Lifecycle.initial(),
+      nextReasoningId: 0,
+      nextTextId: 0,
+    }),
+    step,
+    onHalt: (state) => Effect.succeed(finish(state)),
+  },
+})
+
+export const route = Route.make({
+  id: ADAPTER,
+  provider: "google",
+  providerMetadataKey: "google",
+  protocol,
+  // Gemini's path embeds the model id and pins SSE framing at the URL level.
+  endpoint: Endpoint.path(({ request }) => `/models/${request.model.id}:streamGenerateContent?alt=sse`, {
+    baseURL: DEFAULT_BASE_URL,
+  }),
+  auth: Auth.none,
+  framing: Framing.sse,
+})
+
+export * as Gemini from "./gemini.js"

packages/ai/src/protocols/google-images.ts

@@ -0,0 +1,309 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { ProviderID, mergeJsonRecords, type AIError } from "../schema/index.js"
+import { ProviderShared } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const ADAPTER = "google-images"
+const NAME = "Google Images"
+const PROVIDER = ProviderID.make("google")
+export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type GoogleImageString<Known extends string> = Known | (string & {})
+
+/** Provider-native options. Common fields (`aspectRatio`, `seed`, `images`) live on the request. */
+export type GoogleImageOptions = {
+  readonly imageSize?: GoogleImageString<"1K" | "2K" | "4K">
+  readonly thinkingLevel?: GoogleImageString<"MINIMAL" | "LOW" | "MEDIUM" | "HIGH">
+  readonly includeThoughts?: boolean
+} & Record<string, unknown>
+
+export type Request = ImageRequestFor<GoogleImageOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Response schema
+// ---------------------------------------------------------------------------
+
+const GoogleUsage = Schema.StructWithRest(
+  Schema.Struct({
+    cachedContentTokenCount: Schema.optional(Schema.Number),
+    thoughtsTokenCount: Schema.optional(Schema.Number),
+    promptTokenCount: Schema.optional(Schema.Number),
+    candidatesTokenCount: Schema.optional(Schema.Number),
+    totalTokenCount: Schema.optional(Schema.Number),
+    promptTokensDetails: Schema.optional(Schema.Unknown),
+    candidatesTokensDetails: Schema.optional(Schema.Unknown),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const GoogleImageResponse = Schema.Struct({
+  candidates: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        index: Schema.optional(Schema.Number),
+        content: Schema.optional(
+          Schema.Struct({
+            parts: Schema.Array(
+              Schema.Struct({
+                text: Schema.optional(Schema.String),
+                thought: Schema.optional(Schema.Boolean),
+                thoughtSignature: Schema.optional(Schema.String),
+                inlineData: Schema.optional(
+                  Schema.Struct({
+                    mimeType: Schema.String,
+                    data: Schema.String,
+                  }),
+                ),
+              }),
+            ),
+          }),
+        ),
+        finishReason: Schema.optional(Schema.String),
+        finishMessage: Schema.optional(Schema.String),
+        safetyRatings: Schema.optional(Schema.Unknown),
+        citationMetadata: Schema.optional(Schema.Unknown),
+        groundingMetadata: Schema.optional(Schema.Unknown),
+      }),
+    ),
+  ),
+  usageMetadata: Schema.optional(GoogleUsage),
+  modelVersion: Schema.optional(Schema.String),
+  responseId: Schema.optional(Schema.String),
+  promptFeedback: Schema.optional(Schema.Unknown),
+})
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const generationConfig = (request: Request) => {
+  const { imageSize, thinkingLevel, includeThoughts, ...native } = request.providerOptions ?? {}
+  const imageConfig = { aspectRatio: request.aspectRatio, imageSize }
+  const thinkingConfig = { thinkingLevel, includeThoughts }
+  return (
+    mergeJsonRecords(
+      {
+        responseModalities: ["IMAGE"],
+        imageConfig: Object.values(imageConfig).some((value) => value !== undefined) ? imageConfig : undefined,
+        seed: request.seed,
+        thinkingConfig: Object.values(thinkingConfig).some((value) => value !== undefined) ? thinkingConfig : undefined,
+      },
+      native,
+    ) ?? { responseModalities: ["IMAGE"] }
+  )
+}
+
+// Gemini does not fetch public URLs; inline payloads or Gemini Files references are the only accepted inputs.
+const imagePart = (asset: Media.Asset): Effect.Effect<Record<string, unknown>, AIError> => {
+  const inline = asset.inline()
+  if (inline) return Effect.succeed({ inlineData: { mimeType: inline.mime, data: inline.base64 } })
+  const id = MediaInput.refID(asset, PROVIDER)
+  if (id) return Effect.succeed({ fileData: { mimeType: asset.mediaType, fileUri: id } })
+  if (asset.source.type === "ref")
+    return Effect.fail(
+      ProviderShared.invalidRequest(
+        "Google generateContent requires Gemini file references rather than other providers' file IDs",
+      ),
+    )
+  return Effect.fail(
+    ProviderShared.invalidRequest(
+      "Google generateContent does not fetch public image URLs; use bytes, a data URL, or a Gemini file reference",
+    ),
+  )
+}
+
+const fromRequest = Effect.fn("GoogleImages.fromRequest")(function* (request: Request) {
+  if (request.n !== undefined && request.n > 1)
+    return yield* ProviderShared.unsupportedOperation({
+      operation: "image.n",
+      provider: PROVIDER,
+      route: ADAPTER,
+      message: `${NAME} generates one image per request; call it once per image instead of n=${request.n}`,
+    })
+  const parts = yield* Effect.forEach(request.images ?? [], imagePart)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        contents: [{ role: "user", parts: [{ text: request.prompt }, ...parts] }],
+        generationConfig: generationConfig(request),
+      },
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeResponse = Effect.fn("GoogleImages.decodeResponse")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+) {
+  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, GoogleImageResponse)(response)
+  const decoded = output.value
+  const candidates = decoded.candidates ?? []
+  const candidateMetadata = candidates.map((candidate, candidateIndex) => ({
+    index: candidate.index ?? candidateIndex,
+    finishReason: candidate.finishReason,
+    finishMessage: candidate.finishMessage,
+    safetyRatings: candidate.safetyRatings,
+    citationMetadata: candidate.citationMetadata,
+    groundingMetadata: candidate.groundingMetadata,
+    parts: (candidate.content?.parts ?? []).map((part) =>
+      part.inlineData === undefined
+        ? { type: "text", text: part.text, thought: part.thought, thoughtSignature: part.thoughtSignature }
+        : {
+            type: "inlineData",
+            mediaType: part.inlineData.mimeType,
+            thought: part.thought,
+            thoughtSignature: part.thoughtSignature,
+          },
+    ),
+  }))
+  // Thought parts are drafts; only non-thought inline data is a final image.
+  const encoded = candidates.flatMap((candidate, candidateIndex) =>
+    (candidate.content?.parts ?? []).flatMap((part, partIndex) =>
+      part.inlineData === undefined || part.thought === true
+        ? []
+        : [
+            {
+              candidate,
+              candidateIndex,
+              partIndex,
+              inlineData: part.inlineData,
+              thoughtSignature: part.thoughtSignature,
+            },
+          ],
+    ),
+  )
+  const images = yield* Effect.forEach(encoded, (item) =>
+    MediaInput.decodedAsset(
+      output.invalid,
+      `${NAME} candidate ${item.candidateIndex} part ${item.partIndex}`,
+      item.inlineData.data,
+      item.inlineData.mimeType,
+      {
+        providerMetadata: {
+          google: {
+            candidateIndex: item.candidate.index ?? item.candidateIndex,
+            partIndex: item.partIndex,
+            finishReason: item.candidate.finishReason,
+            safetyRatings: item.candidate.safetyRatings,
+            citationMetadata: item.candidate.citationMetadata,
+            groundingMetadata: item.candidate.groundingMetadata,
+            thoughtSignature: item.thoughtSignature,
+          },
+        },
+      },
+    ),
+  )
+  if (images.length === 0) {
+    const finishReasons = candidates.flatMap((candidate) =>
+      candidate.finishReason === undefined ? [] : [candidate.finishReason],
+    )
+    return yield* output.invalid(
+      `${NAME} returned no final images${
+        finishReasons.length === 0 ? "" : ` (finish reasons: ${finishReasons.join(", ")})`
+      }; inspect body for prompt feedback and candidate details`,
+    )
+  }
+  // Candidates that stopped for a safety or policy reason are partial results, not a silent drop.
+  const notices = [
+    ...(decoded.promptFeedback === undefined
+      ? []
+      : [
+          {
+            type: "filtered" as const,
+            message: `${NAME} reported prompt feedback`,
+            providerMetadata: { google: { promptFeedback: decoded.promptFeedback } },
+          },
+        ]),
+    ...candidates.flatMap((candidate, index) =>
+      candidate.finishReason === undefined || candidate.finishReason === "STOP"
+        ? []
+        : [
+            {
+              type: "filtered" as const,
+              message: `${NAME} candidate ${candidate.index ?? index} finished with ${candidate.finishReason}${
+                candidate.finishMessage === undefined ? "" : `: ${candidate.finishMessage}`
+              }`,
+              providerMetadata: {
+                google: {
+                  candidateIndex: candidate.index ?? index,
+                  finishReason: candidate.finishReason,
+                  finishMessage: candidate.finishMessage,
+                  safetyRatings: candidate.safetyRatings,
+                },
+              },
+            },
+          ],
+    ),
+  ]
+  const usage = decoded.usageMetadata
+  const outputTokens =
+    usage?.candidatesTokenCount === undefined ? undefined : usage.candidatesTokenCount + (usage.thoughtsTokenCount ?? 0)
+  return new ImageResponse({
+    images,
+    notices: notices.length === 0 ? undefined : notices,
+    usage:
+      usage === undefined
+        ? undefined
+        : {
+            type: "tokens",
+            input: usage.promptTokenCount,
+            output: outputTokens,
+            total: ProviderShared.totalTokens(usage.promptTokenCount, outputTokens, usage.totalTokenCount),
+            details: {
+              reasoningTokens: usage.thoughtsTokenCount,
+              cacheReadInputTokens: usage.cachedContentTokenCount,
+              google: usage,
+            },
+          },
+    providerMetadata: {
+      google: {
+        modelVersion: decoded.modelVersion,
+        responseId: decoded.responseId,
+        promptFeedback: decoded.promptFeedback,
+        candidates: candidateMetadata,
+      },
+    },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.inline<Request, ImageResponse>({
+  id: ADAPTER,
+  name: NAME,
+  unsupported: ["mask", "size", "format"],
+  body: { from: fromRequest },
+  response: { decode: decodeResponse },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  ImageModel.fromRoute<GoogleImageOptions>(
+    {
+      id: ADAPTER,
+      provider: PROVIDER,
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      path: ({ request }) => `/models/${request.model.id}:generateContent`,
+    },
+    input,
+  )
+
+export const GoogleImages = {
+  protocol,
+  model,
+} as const

packages/ai/src/protocols/index.ts

@@ -0,0 +1,12 @@
+export * as AnthropicMessages from "./anthropic-messages.js"
+export * as BedrockConverse from "./bedrock-converse.js"
+export * as Gemini from "./gemini.js"
+export * as MistralChat from "./mistral-chat.js"
+export * as OpenAIChat from "./openai-chat.js"
+export * as OpenAIImages from "./openai-images.js"
+export * as OpenAICompatibleChat from "./openai-compatible-chat.js"
+export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
+export * as OpenAIResponses from "./openai-responses.js"
+export * as OpenResponses from "./open-responses.js"
+export * as OpenResponsesChannel from "./open-responses-channel.js"
+export * as XAIResponses from "./xai-responses.js"

packages/ai/src/protocols/meta-images.ts

@@ -0,0 +1,165 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { ProviderID, mergeJsonRecords, type AIError } from "../schema/index.js"
+import { JsonObject, ProviderShared, optionalNull } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const ADAPTER = "meta-images"
+const NAME = "Meta Images"
+const PROVIDER = ProviderID.make("meta")
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+type OpenString<Known extends string> = Known | (string & {})
+
+/** Provider-native options. Common fields (`n`, `size`, `format`, `images`) live on the request. */
+export type ImageOptions = {
+  readonly responseFormat?: OpenString<"b64_json" | "url">
+  readonly reasoningStrength?: OpenString<"low" | "high">
+  readonly toolEnablement?: {
+    readonly enable_image_search?: boolean
+    readonly enable_web_search?: boolean
+    readonly enable_shell?: boolean
+  }
+  readonly [key: string]: unknown
+}
+
+export type Request = ImageRequestFor<ImageOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Request body and response schemas
+// ---------------------------------------------------------------------------
+
+const Body = Schema.StructWithRest(
+  Schema.Struct({
+    model: Schema.String,
+    prompt: Schema.String,
+    images: Schema.optional(Schema.Array(JsonObject)),
+    n: Schema.optional(Schema.Number),
+    /** Aspect ratio hint, not an exact output resolution. */
+    size: Schema.optional(Schema.String),
+    output_format: Schema.optional(Schema.String),
+    response_format: Schema.optional(Schema.String),
+    reasoning_strength: Schema.optional(Schema.String),
+    tool_enablement: Schema.optional(Schema.Record(Schema.String, Schema.Boolean)),
+  }),
+  [JsonObject],
+)
+
+const Response = Schema.Struct({
+  data: Schema.Array(Schema.Struct({ b64_json: optionalNull(Schema.String), url: optionalNull(Schema.String) })),
+  output_format: Schema.optional(Schema.String),
+  usage: Schema.optional(
+    Schema.Struct({
+      input_tokens: Schema.optional(Schema.Number),
+      output_tokens: Schema.optional(Schema.Number),
+      total_tokens: Schema.optional(Schema.Number),
+    }),
+  ),
+})
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const isEdit = (request: Request) => (request.images?.length ?? 0) > 0
+
+const reference = (asset: Media.Asset): Effect.Effect<Record<string, unknown>, AIError> => {
+  const inline = asset.inline()
+  if (inline) return Effect.succeed({ image_url: inline.dataUrl })
+  const url = ProviderShared.mediaUrl(asset)
+  if (url) return Effect.succeed({ image_url: url })
+  return Effect.fail(ProviderShared.invalidRequest(`${NAME} accepts image bytes and URLs`))
+}
+
+const fromRequest = Effect.fn("MetaImages.fromRequest")(function* (request: Request) {
+  const images = yield* Effect.forEach(request.images ?? [], reference)
+  const { responseFormat, reasoningStrength, toolEnablement, ...native } = request.providerOptions ?? {}
+  const payload = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Body))(
+    mergeJsonRecords(
+      {
+        model: request.model.id,
+        prompt: request.prompt,
+        images: images.length === 0 ? undefined : images,
+        n: request.n,
+        size: request.size,
+        output_format: request.format,
+        response_format: responseFormat,
+        reasoning_strength: reasoningStrength,
+        tool_enablement: toolEnablement,
+      },
+      native,
+      request.http?.body,
+    ),
+  )
+  return MediaProtocol.json(payload)
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeResponse = Effect.fn("MetaImages.decodeResponse")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.DecodeContext<Request>,
+) {
+  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, Response)(response)
+  const decoded = output.value
+  const requested = context.body.type === "json" ? context.body.value.output_format : undefined
+  const format = decoded.output_format ?? (typeof requested === "string" ? requested : "webp")
+  const mediaType = `image/${format}`
+  const images = yield* Effect.forEach(decoded.data, (item, index) => {
+    if (item.b64_json)
+      return MediaInput.decodedAsset(output.invalid, `${NAME} result ${index}`, item.b64_json, mediaType, {
+        info: { format },
+      })
+    if (item.url) return Effect.succeed(Media.url(item.url, { mediaType, info: { format } }))
+    return Effect.fail(output.invalid(`${NAME} result ${index} has neither image data nor a URL`))
+  })
+  if (images.length === 0) return yield* output.invalid(`${NAME} returned no images`)
+  return new ImageResponse({
+    images,
+    usage:
+      decoded.usage === undefined
+        ? undefined
+        : {
+            type: "tokens",
+            input: decoded.usage.input_tokens,
+            output: decoded.usage.output_tokens,
+            total: decoded.usage.total_tokens,
+            details: { meta: decoded.usage },
+          },
+    providerMetadata: { meta: { outputFormat: format } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.inline<Request, ImageResponse>({
+  id: ADAPTER,
+  name: NAME,
+  unsupported: ["mask", "aspectRatio", "seed"],
+  body: { from: fromRequest },
+  response: { decode: decodeResponse },
+})
+
+export const model = (input: MediaRoute.ModelInput & { readonly baseURL: string }) =>
+  ImageModel.fromRoute<ImageOptions>(
+    {
+      id: ADAPTER,
+      provider: PROVIDER,
+      protocol,
+      path: ({ request }) => `/images/${isEdit(request) ? "edits" : "generations"}`,
+    },
+    input,
+  )
+
+export * as MetaImages from "./meta-images.js"

packages/ai/src/protocols/meta-messages.ts

@@ -0,0 +1,52 @@
+import { Effect, Schema } from "effect"
+import { Protocol } from "../route/protocol.js"
+import type { LLMRequest } from "../schema/index.js"
+import { AnthropicMessages } from "./anthropic-messages.js"
+import { MetaResponses } from "./meta-responses.js"
+import { JsonObject, optionalArray, ProviderShared } from "./shared.js"
+
+const WebSearch = Schema.Struct({
+  type: Schema.Literal("web_search"),
+  name: Schema.Literal("web_search"),
+  user_location: MetaResponses.WebSearch.fields.user_location,
+})
+const Body = Schema.Struct({
+  ...AnthropicMessages.AnthropicMessagesBody.fields,
+  tools: optionalArray(
+    Schema.Union([
+      Schema.Struct({ name: Schema.String, description: Schema.String, input_schema: JsonObject }),
+      WebSearch,
+    ]),
+  ),
+})
+
+const fromRequest = Effect.fn("MetaMessages.fromRequest")(function* (request: LLMRequest) {
+  const projected = ProviderShared.flattenToolRequest(request)
+  const body = yield* AnthropicMessages.protocol.body.from(projected.request)
+  return {
+    ...body,
+    tools:
+      body.tools === undefined
+        ? undefined
+        : yield* Effect.forEach(body.tools, (tool, index) =>
+            Effect.gen(function* () {
+              const native = projected.tools[index]?.native
+              if (native === undefined) return tool
+              const search = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(MetaResponses.WebSearch))(
+                native.meta,
+              )
+              if (search.search_context_size !== undefined)
+                return yield* ProviderShared.invalidRequest("Meta Messages does not support searchContextSize")
+              return { type: "web_search" as const, name: "web_search" as const, user_location: search.user_location }
+            }),
+          ),
+  }
+})
+
+export const protocol = Protocol.make({
+  id: "meta-messages",
+  body: { schema: Body, from: fromRequest },
+  stream: AnthropicMessages.protocol.stream,
+})
+
+export * as MetaMessages from "./meta-messages.js"

packages/ai/src/protocols/meta-responses.ts

@@ -0,0 +1,238 @@
+import { Effect, Encoding, Schema } from "effect"
+import { Protocol } from "../route/protocol.js"
+import { HttpTransport } from "../route/transport/index.js"
+import { LLMEvent, LLMRequest, Message, ToolResultPart } from "../schema/index.js"
+import { OpenResponses } from "./open-responses.js"
+import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
+import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
+import { ToolSchemaProjection } from "./utils/tool-schema.js"
+import { MetaImage } from "./utils/meta-image.js"
+
+const ADAPTER = "meta-responses"
+const NAME = "Meta Responses"
+
+export const WebSearch = Schema.Struct({
+  type: Schema.Literal("web_search"),
+  search_context_size: Schema.optional(Schema.String),
+  user_location: Schema.optional(
+    Schema.Struct({
+      type: Schema.Literal("approximate"),
+      city: Schema.optional(Schema.String),
+      region: Schema.optional(Schema.String),
+      country: Schema.optional(Schema.String),
+      timezone: Schema.optional(Schema.String),
+    }),
+  ),
+})
+
+export const ImageGeneration = Schema.Struct({
+  type: Schema.Literal("image_generation"),
+  size: Schema.optional(Schema.String),
+  output_format: Schema.optional(Schema.String),
+  reasoning_strength: Schema.optional(Schema.String),
+  enable_image_search: Schema.optional(Schema.Boolean),
+  enable_web_search: Schema.optional(Schema.Boolean),
+  enable_shell: Schema.optional(Schema.Boolean),
+})
+
+const NativeTool = Schema.Union([WebSearch, ImageGeneration])
+const ImageItem = Schema.Struct({
+  type: Schema.Literal("image_generation_call"),
+  id: Schema.String,
+  status: Schema.optional(Schema.String),
+  result: optionalNull(Schema.String),
+  output_format: Schema.optional(Schema.String),
+  error: Schema.optional(Schema.Unknown),
+})
+
+const Body = Schema.Struct({
+  ...OpenResponses.coreFields,
+  input: Schema.Array(Schema.Union([OpenResponses.InputItem, ImageItem])),
+  tools: optionalArray(Schema.Union([OpenResponses.Tool, NativeTool])),
+  stream: Schema.Literal(true),
+})
+
+const MessageAnnotations = Schema.Struct({
+  content: Schema.Array(Schema.Struct({ annotations: optionalArray(JsonObject) })),
+})
+
+interface ParserState extends OpenResponses.ParserState {
+  readonly completedItems: ReadonlySet<string>
+}
+
+const adapter = {
+  id: ADAPTER,
+  name: NAME,
+  restoreHostedToolItem: (item: unknown) => (Schema.is(ImageItem)(item) ? item : undefined),
+} satisfies OpenResponses.ProviderAdapter
+
+const fromRequest = Effect.fn("MetaResponses.fromRequest")(function* (request: LLMRequest) {
+  const key = request.model.route.providerMetadataKey ?? String(request.model.provider)
+  const projected = ProviderShared.flattenToolRequest(
+    LLMRequest.update(request, {
+      messages: request.messages.map((message) =>
+        Message.make({
+          ...message,
+          content: message.content.map((part) => {
+            if (
+              part.type !== "tool-result" ||
+              !part.providerExecuted ||
+              part.name !== "image_generation" ||
+              part.result.type !== "content" ||
+              part.providerMetadata?.[key]?.itemId !== part.id
+            )
+              return part
+            // Meta's signed image ID carries edit state; replay the handle, not the image bytes as a user message.
+            return ToolResultPart.make({
+              ...part,
+              result: {
+                type: "json",
+                value: { type: "image_generation_call", id: part.id, status: "completed", result: null },
+              },
+            })
+          }),
+        }),
+      ),
+    }),
+  )
+  return yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Body))({
+    ...(yield* OpenResponses.lowerConversation(projected.request, adapter)),
+    ...OpenResponses.lowerGeneration(request),
+    tools:
+      projected.tools.length === 0
+        ? undefined
+        : yield* Effect.forEach(projected.tools, (tool) =>
+            Effect.gen(function* () {
+              if (tool.native === undefined)
+                return yield* OpenResponses.lowerTool(
+                  NAME,
+                  tool,
+                  ToolSchemaProjection.modelCompatibility(tool.inputSchema, request.model.compatibility?.toolSchema),
+                )
+              return yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(NativeTool))(tool.native.meta)
+            }),
+          ),
+    tool_choice:
+      OpenResponses.allowedToolChoice(request) ??
+      (request.toolChoice ? yield* OpenResponses.lowerToolChoice(NAME, request.toolChoice) : undefined),
+  })
+})
+
+const HOSTED_TOOLS = {
+  web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
+  image_generation_call: {
+    name: "image_generation",
+    input: () => ({}),
+    result: Effect.fn("MetaResponses.imageResult")(function* (raw: ResponsesHostedTools.Item) {
+      const item = yield* Schema.decodeUnknownEffect(ImageItem)(raw).pipe(
+        Effect.mapError((cause) =>
+          ProviderShared.eventError(
+            ADAPTER,
+            "Meta returned an invalid image item",
+            ProviderShared.encodeJson(raw),
+            cause,
+          ),
+        ),
+      )
+      if (item.error !== undefined && item.error !== null) return { type: "error" as const, value: item.error }
+      if (!item.result)
+        return yield* ProviderShared.eventError(
+          ADAPTER,
+          "Meta returned an image without data",
+          ProviderShared.encodeJson(raw),
+        )
+      const data = yield* Effect.fromResult(Encoding.decodeBase64(item.result)).pipe(
+        Effect.mapError((cause) =>
+          ProviderShared.eventError(
+            ADAPTER,
+            "Meta returned invalid image base64",
+            ProviderShared.encodeJson(raw),
+            cause,
+          ),
+        ),
+      )
+      const mime = MetaImage.mediaType(data, item.output_format)
+      return {
+        type: "content" as const,
+        value: [{ type: "file" as const, uri: `data:${mime};base64,${item.result}`, mime }],
+      }
+    }),
+  },
+} satisfies ResponsesHostedTools.Definitions
+
+const onEvent = Effect.fn("MetaResponses.onEvent")(function* (
+  state: OpenResponses.ParserState,
+  input: OpenResponses.Event,
+) {
+  const event = OpenResponses.normalize(state, input)
+  if (event.type === "response.output_item.done" && event.item && ResponsesHostedTools.isItem(event.item, HOSTED_TOOLS))
+    return yield* ResponsesHostedTools.onDone(state, event.item, HOSTED_TOOLS)
+  const result = yield* OpenResponses.step(state, event)
+  if (event.type !== "response.output_item.done" || event.item?.type !== "message") return result
+  const message = yield* Schema.decodeUnknownEffect(MessageAnnotations)(event.item).pipe(
+    Effect.mapError((cause) =>
+      ProviderShared.eventError(
+        ADAPTER,
+        "Meta returned invalid message annotations",
+        ProviderShared.encodeJson(event),
+        cause,
+      ),
+    ),
+  )
+  const annotations = message.content.flatMap((part) => part.annotations ?? [])
+  if (annotations.length === 0) return result
+  return [
+    result[0],
+    result[1].map((item) =>
+      LLMEvent.is.textEnd(item)
+        ? LLMEvent.textEnd({
+            ...item,
+            providerMetadata: {
+              ...item.providerMetadata,
+              [state.providerMetadataKey]: { ...item.providerMetadata?.[state.providerMetadataKey], annotations },
+            },
+          })
+        : item,
+    ),
+  ] satisfies OpenResponses.StepResult
+})
+
+const step = Effect.fn("MetaResponses.step")(function* (state: ParserState, input: OpenResponses.Event) {
+  const completedItems = new Set(state.completedItems)
+  const event = OpenResponses.normalize(state, input)
+  if (event.type === "response.output_item.done" && event.item && completedItems.has(event.item.id))
+    return [state, []] as const
+  const events: LLMEvent[] = []
+  let current: OpenResponses.ParserState = state
+  // Muse Image delivers its image and optional summary only in response.completed.
+  // Recover terminal-only items in order, without duplicating Spark's streamed items.
+  if (event.type === "response.completed") {
+    for (const [index, item] of (event.response?.output ?? []).entries()) {
+      const done = OpenResponses.normalize(current, { type: "response.output_item.done", item, output_index: index })
+      // Spark changes reasoning IDs in the terminal snapshot; output indices still identify the streamed items.
+      if (!done.item || completedItems.has(done.item.id) || completedItems.has(state.outputItems[index] ?? "")) continue
+      const result = yield* onEvent(current, done)
+      current = result[0]
+      events.push(...result[1])
+      completedItems.add(done.item.id)
+    }
+  }
+  const result = yield* onEvent(current, event)
+  if (event.type === "response.output_item.done" && event.item) completedItems.add(event.item.id)
+  return [{ ...result[0], completedItems }, [...events, ...result[1]]] as const
+})
+
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: { schema: Body, from: fromRequest },
+  stream: {
+    event: OpenResponses.protocol.stream.event,
+    initial: (request): ParserState => ({ ...OpenResponses.initial(request, adapter), completedItems: new Set() }),
+    step,
+    terminal: OpenResponses.terminal,
+  },
+})
+
+export const httpTransport = HttpTransport.sseJson.with<Schema.Schema.Type<typeof Body>>()
+
+export * as MetaResponses from "./meta-responses.js"

packages/ai/src/protocols/mistral-chat.ts

@@ -0,0 +1,767 @@
+import { Effect, Schema } from "effect"
+import { Auth } from "../route/auth.js"
+import { Route } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Framing } from "../route/framing.js"
+import { Protocol } from "../route/protocol.js"
+import { HttpTransport } from "../route/transport/index.js"
+import {
+  AIError,
+  InvalidProviderOutputError,
+  LLMEvent,
+  ProviderInternalError,
+  UnknownProviderError,
+  Usage,
+  type FinishReasonDetails,
+  type LLMRequest,
+  type MediaPart,
+  type ToolCallPart,
+  type ToolDefinition,
+} from "../schema/index.js"
+import { classifyProviderFailure } from "../provider-error.js"
+import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
+import { Lifecycle } from "./utils/lifecycle.js"
+import { MistralToolID } from "./utils/mistral-tool-id.js"
+import { ToolStream } from "./utils/tool-stream.js"
+
+const ADAPTER = "mistral-chat"
+const DONE = "[DONE]" as const
+export const DEFAULT_BASE_URL = "https://api.mistral.ai/v1"
+export const PATH = "/chat/completions"
+
+const MistralTextContent = Schema.Struct({
+  type: Schema.Literal("text"),
+  text: Schema.String,
+})
+
+const MistralThinkingUnit = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.optional(Schema.String),
+    text: Schema.optional(Schema.String),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralThinkingUnit = Schema.Schema.Type<typeof MistralThinkingUnit>
+
+const MistralThinkingContent = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.Literal("thinking"),
+    thinking: Schema.Array(MistralThinkingUnit),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralThinkingContent = Schema.Schema.Type<typeof MistralThinkingContent>
+const isMistralThinkingContent = Schema.is(MistralThinkingContent)
+
+const MistralUserContent = Schema.Union([
+  MistralTextContent,
+  Schema.Struct({ type: Schema.Literal("image_url"), image_url: Schema.String }),
+  Schema.Struct({ type: Schema.Literal("document_url"), document_url: Schema.String }),
+])
+type MistralUserContent = Schema.Schema.Type<typeof MistralUserContent>
+
+const MistralAssistantToolCall = Schema.Struct({
+  id: Schema.String,
+  type: Schema.Literal("function"),
+  function: Schema.Struct({ name: Schema.String, arguments: Schema.String }),
+})
+type MistralAssistantToolCall = Schema.Schema.Type<typeof MistralAssistantToolCall>
+
+const MistralMessage = Schema.Union([
+  Schema.Struct({ role: Schema.Literal("system"), content: Schema.String }),
+  Schema.Struct({
+    role: Schema.Literal("user"),
+    content: Schema.Union([Schema.String, Schema.Array(MistralUserContent)]),
+  }),
+  Schema.Struct({
+    role: Schema.Literal("assistant"),
+    content: Schema.Union([Schema.String, Schema.Array(Schema.Union([MistralTextContent, MistralThinkingContent]))]),
+    tool_calls: optionalArray(MistralAssistantToolCall),
+    prefix: Schema.optional(Schema.Literal(true)),
+  }),
+  Schema.Struct({
+    role: Schema.Literal("tool"),
+    tool_call_id: Schema.String,
+    name: Schema.String,
+    content: Schema.Union([Schema.String, Schema.Array(MistralUserContent)]),
+  }),
+]).pipe(Schema.toTaggedUnion("role"))
+type MistralMessage = Schema.Schema.Type<typeof MistralMessage>
+
+const MistralTool = Schema.Struct({
+  type: Schema.Literal("function"),
+  function: Schema.Struct({
+    name: Schema.String,
+    description: Schema.String,
+    parameters: JsonObject,
+    strict: Schema.Literal(false),
+  }),
+})
+type MistralTool = Schema.Schema.Type<typeof MistralTool>
+
+const MistralOptions = Schema.Struct({
+  safePrompt: Schema.optional(Schema.Boolean),
+  documentImageLimit: Schema.optional(Schema.Number),
+  documentPageLimit: Schema.optional(Schema.Number),
+  parallelToolCalls: Schema.optional(Schema.Boolean),
+  reasoningEffort: Schema.optional(Schema.String),
+  promptMode: Schema.optional(Schema.Literal("reasoning")),
+  promptCacheKey: Schema.optional(Schema.String),
+})
+
+export type ReasoningEffort = "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | (string & {})
+
+export type ProviderOptionsInput = {
+  readonly safePrompt?: boolean
+  readonly documentImageLimit?: number
+  readonly documentPageLimit?: number
+  readonly parallelToolCalls?: boolean
+  readonly reasoningEffort?: ReasoningEffort
+  readonly promptMode?: "reasoning"
+  readonly promptCacheKey?: string
+  readonly [key: string]: unknown
+}
+
+const MistralBody = Schema.Struct({
+  model: Schema.String,
+  messages: Schema.Array(MistralMessage),
+  tools: optionalArray(MistralTool),
+  tool_choice: Schema.optional(
+    Schema.Union([
+      Schema.Literals(["auto", "none", "any"]),
+      Schema.Struct({ type: Schema.Literal("function"), function: Schema.Struct({ name: Schema.String }) }),
+    ]),
+  ),
+  stream: Schema.Literal(true),
+  max_tokens: Schema.optional(Schema.Number),
+  random_seed: Schema.optional(Schema.Number),
+  temperature: Schema.optional(Schema.Number),
+  top_p: Schema.optional(Schema.Number),
+  frequency_penalty: Schema.optional(Schema.Number),
+  presence_penalty: Schema.optional(Schema.Number),
+  stop: optionalArray(Schema.String),
+  prompt_cache_key: Schema.optional(Schema.String),
+  safe_prompt: Schema.optional(Schema.Boolean),
+  document_image_limit: Schema.optional(Schema.Number),
+  document_page_limit: Schema.optional(Schema.Number),
+  parallel_tool_calls: Schema.optional(Schema.Boolean),
+  reasoning_effort: Schema.optional(Schema.String),
+  prompt_mode: Schema.optional(Schema.Literal("reasoning")),
+})
+export type MistralBody = Schema.Schema.Type<typeof MistralBody>
+
+const MistralUsageDetails = Schema.StructWithRest(Schema.Struct({ cached_tokens: optionalNull(Schema.Number) }), [
+  Schema.Record(Schema.String, Schema.Unknown),
+])
+
+const MistralUsage = Schema.StructWithRest(
+  Schema.Struct({
+    prompt_tokens: optionalNull(Schema.Number),
+    completion_tokens: optionalNull(Schema.Number),
+    total_tokens: optionalNull(Schema.Number),
+    num_cached_tokens: optionalNull(Schema.Number),
+    prompt_token_details: optionalNull(MistralUsageDetails),
+    prompt_tokens_details: optionalNull(MistralUsageDetails),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const MistralOutputContent = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.String,
+    text: optionalNull(Schema.String),
+    thinking: optionalNull(Schema.Unknown),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralOutputContent = Schema.Schema.Type<typeof MistralOutputContent>
+
+const MistralToolDelta = Schema.Struct({
+  index: optionalNull(Schema.Number),
+  id: optionalNull(Schema.String),
+  function: optionalNull(
+    Schema.Struct({
+      name: optionalNull(Schema.String),
+      arguments: optionalNull(Schema.Union([Schema.String, JsonObject])),
+    }),
+  ),
+})
+type MistralToolDelta = Schema.Schema.Type<typeof MistralToolDelta>
+
+const MistralChoice = Schema.StructWithRest(
+  Schema.Struct({
+    delta: optionalNull(
+      Schema.StructWithRest(
+        Schema.Struct({
+          content: optionalNull(Schema.Union([Schema.String, Schema.Array(MistralOutputContent)])),
+          tool_calls: optionalNull(Schema.Array(MistralToolDelta)),
+        }),
+        [Schema.Record(Schema.String, Schema.Unknown)],
+      ),
+    ),
+    finish_reason: optionalNull(Schema.String),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const MistralError = Schema.StructWithRest(
+  Schema.Struct({
+    message: Schema.String,
+    code: optionalNull(Schema.Union([Schema.String, Schema.Number])),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const MistralEvent = Schema.StructWithRest(
+  Schema.Struct({
+    choices: optionalNull(Schema.Array(MistralChoice)),
+    usage: optionalNull(MistralUsage),
+    error: optionalNull(MistralError),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralEvent = Schema.Schema.Type<typeof MistralEvent>
+const MistralStreamEvent = Schema.Union([Schema.Literal(DONE), Protocol.jsonEvent(MistralEvent)])
+
+const lowerMedia = Effect.fn("MistralChat.lowerMedia")(function* (part: MediaPart) {
+  const mime = part.media.mediaType.toLowerCase()
+  const url =
+    ProviderShared.mediaUrl(part.media) ??
+    (yield* ProviderShared.requireInlineMedia("Mistral Chat", part.media)).dataUrl
+  if (mime.startsWith("image/")) return { type: "image_url" as const, image_url: url }
+  if (mime === "application/pdf") return { type: "document_url" as const, document_url: url }
+  return yield* ProviderShared.invalidRequest(`Mistral Chat does not support media type ${part.media.mediaType}`)
+})
+
+const lowerUser = Effect.fn("MistralChat.lowerUser")(function* (message: LLMRequest["messages"][number]) {
+  const content: MistralUserContent[] = []
+  for (const part of message.content) {
+    if (part.type === "text") {
+      content.push({ type: "text", text: part.text })
+      continue
+    }
+    if (part.type === "media") {
+      content.push(yield* lowerMedia(part))
+      continue
+    }
+    return yield* ProviderShared.unsupportedContent("Mistral Chat", "user", ["text", "media"])
+  }
+  if (content.every((part) => part.type === "text"))
+    return { role: "user" as const, content: content.map((part) => part.text).join("") }
+  return { role: "user" as const, content }
+})
+
+const lowerToolCall = (part: ToolCallPart, normalizeID: (id: string) => string): MistralAssistantToolCall => ({
+  id: normalizeID(part.id),
+  type: "function",
+  function: { name: part.name, arguments: ProviderShared.encodeJson(part.input) },
+})
+
+const lowerAssistant = Effect.fn("MistralChat.lowerAssistant")(function* (
+  message: LLMRequest["messages"][number],
+  normalizeID: (id: string) => string,
+  prefix: boolean,
+) {
+  const structured = message.content.some(
+    (part) => part.type === "reasoning" && isMistralThinkingContent(part.providerMetadata?.mistral?.thinking),
+  )
+  const content: Array<Schema.Schema.Type<typeof MistralTextContent> | MistralThinkingContent> = []
+  const text: string[] = []
+  const toolCalls: MistralAssistantToolCall[] = []
+  for (const part of message.content) {
+    if (part.type === "text") {
+      if (structured) content.push({ type: "text", text: part.text })
+      else text.push(part.text)
+      continue
+    }
+    if (part.type === "reasoning") {
+      const native = part.providerMetadata?.mistral?.thinking
+      if (structured && isMistralThinkingContent(native)) content.push(native)
+      else if (structured) content.push({ type: "text", text: part.text })
+      else text.push(part.text)
+      continue
+    }
+    if (part.type === "tool-call") {
+      toolCalls.push(lowerToolCall(part, normalizeID))
+      continue
+    }
+    return yield* ProviderShared.unsupportedContent("Mistral Chat", "assistant", ["text", "reasoning", "tool-call"])
+  }
+  return {
+    role: "assistant" as const,
+    content: structured ? content : text.join(""),
+    ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),
+    ...(prefix ? { prefix: true as const } : {}),
+  }
+})
+
+const lowerToolResults = Effect.fn("MistralChat.lowerToolResults")(function* (
+  message: LLMRequest["messages"][number],
+  normalizeID: (id: string) => string,
+) {
+  const output: MistralMessage[] = []
+  for (const part of message.content) {
+    if (part.type !== "tool-result")
+      return yield* ProviderShared.unsupportedContent("Mistral Chat", "tool", ["tool-result"])
+    if (part.result.type !== "content") {
+      output.push({
+        role: "tool",
+        tool_call_id: normalizeID(part.id),
+        name: part.name,
+        content: ProviderShared.toolResultText(part),
+      })
+      continue
+    }
+    const content: MistralUserContent[] = []
+    for (const item of part.result.value) {
+      if (item.type === "text") {
+        content.push({ type: "text", text: item.text })
+        continue
+      }
+      content.push(yield* lowerMedia(ProviderShared.toolFileMedia(item)))
+    }
+    output.push({
+      role: "tool",
+      tool_call_id: normalizeID(part.id),
+      name: part.name,
+      content: content.some((item) => item.type !== "text")
+        ? content
+        : content.map((item) => (item.type === "text" ? item.text : "")).join(""),
+    })
+  }
+  return output
+})
+
+const lowerMessages = Effect.fn("MistralChat.lowerMessages")(function* (request: LLMRequest) {
+  const normalizeID = MistralToolID.normalizer(request)
+  const messages: MistralMessage[] =
+    request.system.length === 0 ? [] : [{ role: "system", content: ProviderShared.joinText(request.system) }]
+  for (const message of request.messages) {
+    if (message.role === "system") {
+      const update = yield* ProviderShared.wrappedSystemUpdate("Mistral Chat", message)
+      messages.push({
+        role: "user",
+        content: update.text,
+      })
+      continue
+    }
+    if (message.role === "user") {
+      messages.push(yield* lowerUser(message))
+      continue
+    }
+    if (message.role === "assistant") {
+      const hasToolCalls = message.content.some((part) => part.type === "tool-call")
+      const hasNativeThinking = message.content.some(
+        (part) => part.type === "reasoning" && isMistralThinkingContent(part.providerMetadata?.mistral?.thinking),
+      )
+      const text = message.content
+        .flatMap((part) => (part.type === "text" || part.type === "reasoning" ? [part.text] : []))
+        .join("")
+      if (!hasToolCalls && !hasNativeThinking && text.trim() === "") continue
+      messages.push(yield* lowerAssistant(message, normalizeID, !hasToolCalls && message === request.messages.at(-1)))
+      continue
+    }
+    messages.push(...(yield* lowerToolResults(message, normalizeID)))
+  }
+  return messages
+})
+
+const lowerTool = (tool: ToolDefinition): MistralTool => ({
+  type: "function",
+  function: { name: tool.name, description: tool.description, parameters: tool.inputSchema, strict: false },
+})
+
+export const fromRequest = Effect.fn("MistralChat.fromRequest")(function* (request: LLMRequest) {
+  const options = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(MistralOptions))(
+    request.providerOptions ?? {},
+  )
+  const selected = request.toolChoice?.type === "tool" ? request.toolChoice.name : undefined
+  if (request.toolChoice?.type === "tool" && !selected)
+    return yield* ProviderShared.invalidRequest("Mistral Chat tool choice requires a tool name")
+  if (options.reasoningEffort !== undefined && options.promptMode !== undefined)
+    return yield* ProviderShared.invalidRequest(
+      "Mistral Chat reasoningEffort and promptMode provider options are mutually exclusive",
+    )
+  const toolChoice = request.toolChoice
+    ? yield* ProviderShared.matchToolChoice("Mistral Chat", request.toolChoice, {
+        auto: () => "auto" as const,
+        none: () => "none" as const,
+        required: () => "any" as const,
+        tool: (name) => ({ type: "function" as const, function: { name } }),
+      })
+    : undefined
+  const flattened = ProviderShared.flattenToolRequest(request)
+  return {
+    model: request.model.id,
+    messages: yield* lowerMessages(flattened.request),
+    tools: flattened.tools.length > 0 ? flattened.tools.map(lowerTool) : undefined,
+    tool_choice: toolChoice,
+    stream: true as const,
+    max_tokens: request.generation?.maxTokens,
+    random_seed: request.generation?.seed,
+    temperature: request.generation?.temperature,
+    top_p: request.generation?.topP,
+    frequency_penalty: request.generation?.frequencyPenalty,
+    presence_penalty: request.generation?.presencePenalty,
+    stop: request.generation?.stop,
+    prompt_cache_key: request.cache === "none" ? undefined : (options.promptCacheKey ?? request.promptCacheKey),
+    safe_prompt: options.safePrompt,
+    document_image_limit: options.documentImageLimit,
+    document_page_limit: options.documentPageLimit,
+    parallel_tool_calls:
+      options.parallelToolCalls ?? (request.toolChoice?.disableParallelToolUse === true ? false : undefined),
+    reasoning_effort: options.reasoningEffort,
+    prompt_mode: options.promptMode,
+  }
+})
+
+type ToolKey = string | number
+interface PendingTool {
+  readonly id: string
+  readonly name?: string
+  readonly input: string
+}
+
+interface ActiveContent {
+  readonly type: "text" | "reasoning"
+  readonly id: string
+  readonly thinking?: MistralThinkingContent
+}
+
+export interface ParserState {
+  readonly tools: ToolStream.State<ToolKey>
+  readonly pendingTools: Partial<Record<ToolKey, PendingTool>>
+  readonly toolIDs: ReadonlyMap<string, string>
+  readonly usedToolIDs: ReadonlySet<string>
+  readonly completedTools: ReadonlyArray<LLMEvent>
+  readonly latestToolKey?: ToolKey
+  readonly generatedTools: number
+  readonly lifecycle: Lifecycle.State
+  readonly active?: ActiveContent
+  readonly nextContent: number
+  readonly usage?: Usage
+  readonly finishReason?: FinishReasonDetails
+}
+
+const mapUsage = (usage: MistralEvent["usage"]): Usage | undefined => {
+  if (!usage) return undefined
+  const input = usage.prompt_tokens ?? undefined
+  const reported =
+    usage.num_cached_tokens ??
+    usage.prompt_tokens_details?.cached_tokens ??
+    usage.prompt_token_details?.cached_tokens ??
+    undefined
+  const cached = input === undefined || reported === undefined ? undefined : Math.max(0, Math.min(input, reported))
+  const output = usage.completion_tokens ?? undefined
+  return new Usage({
+    inputTokens: input,
+    outputTokens: output,
+    nonCachedInputTokens: ProviderShared.subtractTokens(input, cached),
+    cacheReadInputTokens: cached,
+    totalTokens: ProviderShared.totalTokens(input, output, usage.total_tokens ?? undefined),
+    providerMetadata: { mistral: usage },
+  })
+}
+
+const mapFinishReason = (reason: string) => {
+  switch (reason) {
+    case "stop":
+      return "stop" as const
+    case "length":
+    case "model_length":
+      return "length" as const
+    case "tool_calls":
+      return "tool-calls" as const
+    case "content_filter":
+      return "content-filter" as const
+    case "error":
+    case "network_error":
+      return "error" as const
+    default:
+      return "unknown" as const
+  }
+}
+
+const thinkingUnits = (value: unknown): ReadonlyArray<MistralThinkingUnit> => {
+  if (typeof value === "string") return [{ type: "text", text: value }]
+  if (!Array.isArray(value)) return []
+  return value.filter(Schema.is(MistralThinkingUnit))
+}
+
+const thinkingText = (thinking: ReadonlyArray<MistralThinkingUnit>) =>
+  thinking.flatMap((unit) => (typeof unit.text === "string" ? [unit.text] : [])).join("")
+
+const thinkingMetadata = (thinking: MistralThinkingContent) => ({ mistral: { thinking } })
+
+const closeActive = (state: ParserState, events: LLMEvent[]) => {
+  if (!state.active) return state
+  const lifecycle =
+    state.active.type === "text"
+      ? Lifecycle.textEnd(state.lifecycle, events, state.active.id)
+      : Lifecycle.reasoningEnd(
+          state.lifecycle,
+          events,
+          state.active.id,
+          thinkingMetadata(state.active.thinking ?? { type: "thinking", thinking: [] }),
+          thinkingText(state.active.thinking?.thinking ?? []),
+        )
+  return { ...state, lifecycle, active: undefined }
+}
+
+const appendText = (state: ParserState, events: LLMEvent[], text: string) => {
+  if (text.length === 0) return state
+  const current = state.active?.type === "text" ? state : closeActive(state, events)
+  const active = current.active ?? { type: "text" as const, id: `text-${current.nextContent}` }
+  return {
+    ...current,
+    lifecycle: Lifecycle.textDelta(current.lifecycle, events, active.id, text),
+    active,
+    nextContent: current.active ? current.nextContent : current.nextContent + 1,
+  }
+}
+
+const appendThinking = (state: ParserState, events: LLMEvent[], part: MistralOutputContent) => {
+  const current = state.active?.type === "reasoning" ? state : closeActive(state, events)
+  const units = thinkingUnits(part.thinking)
+  const active = current.active ?? { type: "reasoning" as const, id: `reasoning-${current.nextContent}` }
+  const thinking = {
+    ...active.thinking,
+    ...part,
+    type: "thinking" as const,
+    thinking: [...(active.thinking?.thinking ?? []), ...units],
+  }
+  const text = thinkingText(units)
+  return {
+    ...current,
+    lifecycle:
+      text.length > 0
+        ? Lifecycle.reasoningDelta(current.lifecycle, events, active.id, text, thinkingMetadata(thinking))
+        : Lifecycle.reasoningStart(current.lifecycle, events, active.id, thinkingMetadata(thinking)),
+    active: { ...active, thinking },
+    nextContent: current.active ? current.nextContent : current.nextContent + 1,
+  }
+}
+
+const appendContent = (
+  state: ParserState,
+  events: LLMEvent[],
+  content: string | ReadonlyArray<MistralOutputContent>,
+) => {
+  if (typeof content === "string") return appendText(state, events, content)
+  return content.reduce((current, part) => {
+    if (part.type === "text") return appendText(current, events, part.text ?? "")
+    if (part.type === "thinking") return appendThinking(current, events, part)
+    return closeActive(current, events)
+  }, state)
+}
+
+const normalizeStreamToolID = (state: ParserState, source: string) => {
+  if (MistralToolID.valid.test(source))
+    return { id: source, state: { ...state, usedToolIDs: new Set([...state.usedToolIDs, source]) } }
+  const previous = state.toolIDs.get(source)
+  if (previous) return { id: previous, state }
+  let attempt = 0
+  let id = MistralToolID.hash(source)
+  while (state.usedToolIDs.has(id)) id = MistralToolID.hash(`${source}:${++attempt}`)
+  return {
+    id,
+    state: {
+      ...state,
+      toolIDs: new Map([...state.toolIDs, [source, id]]),
+      usedToolIDs: new Set([...state.usedToolIDs, id]),
+    },
+  }
+}
+
+const toolText = (tool: MistralToolDelta) => {
+  const value = tool.function?.arguments
+  if (typeof value === "string") return value
+  return value === null || value === undefined ? "" : ProviderShared.encodeJson(value)
+}
+
+const appendTools = Effect.fn("MistralChat.appendTools")(function* (
+  initial: ParserState,
+  events: LLMEvent[],
+  deltas: ReadonlyArray<MistralToolDelta>,
+) {
+  if (deltas.length === 0) return initial
+  let state = closeActive(initial, events)
+  for (const [position, delta] of deltas.entries()) {
+    const wireID = delta.id?.trim() || undefined
+    const providedID = wireID === "null" ? undefined : wireID
+    const key =
+      delta.index ??
+      (providedID
+        ? `id:${providedID}`
+        : deltas.length > 1
+          ? `position:${position}`
+          : (state.latestToolKey ?? `missing:${state.generatedTools}`))
+    const existing = state.tools[key]
+    const pending = state.pendingTools[key]
+    const source = providedID ?? `generated:${String(key)}`
+    const normalized =
+      existing || pending ? { id: existing?.id ?? pending?.id ?? "", state } : normalizeStreamToolID(state, source)
+    state = normalized.state
+    const name = existing?.name ?? pending?.name ?? (delta.function?.name?.trim() || undefined)
+    const text = `${pending?.input ?? ""}${toolText(delta)}`
+    if (!name) {
+      state = {
+        ...state,
+        pendingTools: { ...state.pendingTools, [key]: { id: normalized.id, input: text } },
+        latestToolKey: key,
+        generatedTools: state.generatedTools + (!providedID && !pending ? 1 : 0),
+      }
+      continue
+    }
+    const result = ToolStream.appendOrStart(
+      ADAPTER,
+      state.tools,
+      key,
+      { id: normalized.id, name, text },
+      "Mistral Chat tool call delta is missing a name",
+    )
+    if (ToolStream.isError(result)) return yield* result
+    if (result.events.length > 0) state = { ...state, lifecycle: Lifecycle.stepStart(state.lifecycle, events) }
+    events.push(...result.events)
+    const pendingTools = { ...state.pendingTools }
+    delete pendingTools[key]
+    state = {
+      ...state,
+      tools: result.tools,
+      pendingTools,
+      latestToolKey: key,
+      generatedTools: state.generatedTools + (!providedID && !existing && !pending ? 1 : 0),
+    }
+  }
+  return state
+})
+
+const hasLateContent = (event: MistralEvent) => {
+  const delta = event.choices?.[0]?.delta
+  if (typeof delta?.content === "string" && delta.content.length > 0) return true
+  if (Array.isArray(delta?.content) && delta.content.length > 0) return true
+  return (delta?.tool_calls ?? []).some(
+    (tool) => Boolean(tool.id) || Boolean(tool.function?.name) || tool.function?.arguments !== undefined,
+  )
+}
+
+const step = Effect.fn("MistralChat.step")(function* (state: ParserState, event: MistralEvent) {
+  if (event.error) {
+    const body = ProviderShared.encodeJson(event)
+    return yield* new AIError({
+      reason: classifyProviderFailure({
+        message: event.error.message,
+        status: typeof event.error.code === "number" ? event.error.code : undefined,
+        rawBody: body,
+      }),
+    })
+  }
+  const events: LLMEvent[] = []
+  const usage = mapUsage(event.usage) ?? state.usage
+  if (state.finishReason) {
+    if (hasLateContent(event))
+      return yield* ProviderShared.eventError(
+        ADAPTER,
+        "Mistral Chat received content after the finish reason",
+        ProviderShared.encodeJson(event),
+      )
+    return [{ ...state, usage }, events] as const
+  }
+  const choice = event.choices?.[0]
+  const withContent = choice?.delta?.content == null ? state : appendContent(state, events, choice.delta.content)
+  const withTools = yield* appendTools(withContent, events, choice?.delta?.tool_calls ?? [])
+  if (!choice?.finish_reason) return [{ ...withTools, usage }, events] as const
+
+  const finishReason = {
+    normalized: mapFinishReason(choice.finish_reason),
+    raw: choice.finish_reason,
+  }
+  if (finishReason.normalized === "error") {
+    const details = {
+      message: `Mistral Chat stopped with ${finishReason.raw}`,
+      body: ProviderShared.encodeJson(event),
+    }
+    return yield* new AIError({
+      reason:
+        finishReason.raw === "network_error" ? new ProviderInternalError(details) : new UnknownProviderError(details),
+    })
+  }
+  const incomplete = finishReason.normalized === "length" || finishReason.normalized === "content-filter"
+  if (!incomplete && Object.keys(withTools.pendingTools).length > 0)
+    return yield* ProviderShared.eventError(
+      ADAPTER,
+      "Mistral Chat tool call delta is missing a name",
+      ProviderShared.encodeJson(event),
+    )
+  const finished =
+    !incomplete && Object.keys(withTools.tools).length > 0
+      ? yield* ToolStream.finishAll(ADAPTER, withTools.tools)
+      : undefined
+  return [
+    {
+      ...withTools,
+      tools: finished?.tools ?? withTools.tools,
+      completedTools: finished?.events ?? withTools.completedTools,
+      usage,
+      finishReason,
+    },
+    events,
+  ] as const
+})
+
+const finishEvents = Effect.fn("MistralChat.finishEvents")(function* (state: ParserState) {
+  if (!state.finishReason)
+    return yield* new AIError({
+      reason: new InvalidProviderOutputError({
+        message: "Mistral Chat stream ended without finish_reason",
+        classification: "incomplete-stream",
+        route: ADAPTER,
+      }),
+    })
+  const events: LLMEvent[] = []
+  const closed = closeActive(state, events)
+  const lifecycle = closed.completedTools.length > 0 ? Lifecycle.stepStart(closed.lifecycle, events) : closed.lifecycle
+  events.push(...closed.completedTools)
+  const reason =
+    state.finishReason.normalized === "stop" && closed.completedTools.some(LLMEvent.is.toolCall)
+      ? { ...state.finishReason, normalized: "tool-calls" as const }
+      : state.finishReason
+  Lifecycle.finish(lifecycle, events, { reason, usage: closed.usage })
+  return events
+})
+
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: { schema: MistralBody, from: fromRequest },
+  stream: {
+    event: MistralStreamEvent,
+    initial: (): ParserState => ({
+      tools: ToolStream.empty<ToolKey>(),
+      pendingTools: {},
+      toolIDs: new Map(),
+      usedToolIDs: new Set(),
+      completedTools: [],
+      generatedTools: 0,
+      lifecycle: Lifecycle.initial(),
+      nextContent: 0,
+    }),
+    step: (state: ParserState, event) => (event === DONE ? Effect.succeed([state, []] as const) : step(state, event)),
+    terminal: (event) => event === DONE,
+    onHalt: finishEvents,
+  },
+})
+
+export const framing = Framing.sseWithDone
+export const httpTransport = HttpTransport.sseJson.with<MistralBody>().with({ framing })
+
+export const route = Route.make({
+  id: ADAPTER,
+  provider: "mistral",
+  providerMetadataKey: "mistral",
+  protocol,
+  endpoint: Endpoint.path(PATH, { baseURL: DEFAULT_BASE_URL }),
+  auth: Auth.none,
+  transport: httpTransport,
+})
+
+export * as MistralChat from "./mistral-chat.js"

packages/ai/src/protocols/open-responses-channel.ts

@@ -0,0 +1,218 @@
+import { Effect, Schema, Stream } from "effect"
+import { Headers } from "effect/unstable/http"
+import { Framing } from "../route/framing.js"
+import type { HttpContext } from "../schema/index.js"
+import {
+  HttpTransport,
+  WebSocketTransport,
+  type Transport,
+  type WebSocketChannelDriver,
+  type WebSocketChannelExchange,
+} from "../route/transport/index.js"
+import * as ProviderShared from "./shared.js"
+import { OpenResponses } from "./open-responses.js"
+import { OpenResponsesContinuation } from "./open-responses-continuation.js"
+
+const WebSocketResponseCreate = Schema.StructWithRest(Schema.Struct({ type: Schema.tag("response.create") }), [
+  Schema.Record(Schema.String, Schema.Unknown),
+])
+const decodeMessage = ProviderShared.validateWith(Schema.decodeUnknownEffect(WebSocketResponseCreate))
+const encodeMessage = Schema.encodeSync(Schema.fromJsonString(WebSocketResponseCreate))
+
+export interface Options {
+  readonly id: string
+  readonly name: string
+  readonly rotateAfterMs?: number
+  readonly enabled?: (url: string) => boolean
+  readonly url?: (url: string) => string
+  readonly headers?: (headers: Headers.Headers) => Headers.Headers
+  readonly continuation?: OpenResponsesContinuation.Shape
+}
+
+export interface Prepared {
+  readonly http: HttpTransport.HttpPrepared<string>
+  readonly channel?: {
+    readonly url: string
+    readonly headers: Headers.Headers
+    readonly rotateAfterMs?: number
+    readonly driver: WebSocketChannelDriver
+  }
+}
+
+const message = (body: unknown) =>
+  Effect.gen(function* () {
+    if (!ProviderShared.isRecord(body))
+      return yield* ProviderShared.invalidRequest("Open Responses WebSocket body must be a JSON object")
+    const { stream: _stream, stream_options: _streamOptions, background: _background, ...request } = body
+    const decoded = yield* decodeMessage({ ...request, type: "response.create" })
+    return { request: decoded, message: encodeMessage(decoded) }
+  })
+
+const driver = (options: Options, body: string): WebSocketChannelDriver => {
+  let responseID: string | undefined
+  let terminal = false
+  return {
+    create: () =>
+      Effect.sync(() => {
+        responseID = undefined
+        terminal = false
+        return { message: body, mode: "full" }
+      }),
+    observe: (_create, frame) =>
+      Effect.gen(function* () {
+        const event = yield* OpenResponses.decodeChannelEvent(frame).pipe(
+          Effect.mapError((cause) =>
+            ProviderShared.eventError(options.id, `Invalid ${options.name} WebSocket event`, frame, cause),
+          ),
+        )
+        if (terminal)
+          return yield* ProviderShared.eventError(
+            options.id,
+            `${options.name} emitted ${event.type} after a terminal event`,
+            frame,
+          )
+        if (event.type === "error") {
+          terminal = true
+          yield* OpenResponses.decodeKnownErrorEvent(event).pipe(
+            Effect.mapError((cause) =>
+              ProviderShared.eventError(options.id, `${options.name} returned a malformed error event`, frame, cause),
+            ),
+          )
+          return {
+            type: "provider-failure",
+            error: OpenResponses.providerFailure(event, `${options.name} stream error`, frame),
+          }
+        }
+        if (event.type === "response.failed") {
+          terminal = true
+          if (responseID && event.response?.id && event.response.id !== responseID)
+            return yield* ProviderShared.eventError(
+              options.id,
+              `${options.name} response ID changed during execution`,
+              frame,
+            )
+          return {
+            type: "provider-failure",
+            error: OpenResponses.providerFailure(event, `${options.name} response failed`, frame),
+          }
+        }
+        if (event.type === "response.created") {
+          const created = event.response?.id
+          if (responseID)
+            return yield* ProviderShared.eventError(
+              options.id,
+              `${options.name} emitted duplicate response.created`,
+              frame,
+            )
+          if (!created)
+            return yield* ProviderShared.eventError(
+              options.id,
+              `${options.name} response.created is missing response.id`,
+              frame,
+            )
+          responseID = created
+          return { type: "frame", frame }
+        }
+        // Keepalives and provider notifications carry no response state and may precede response.created.
+        if (!event.type.startsWith("response.")) return { type: "frame", frame }
+        if (!responseID)
+          return yield* ProviderShared.eventError(
+            options.id,
+            `${options.name} emitted ${event.type} before response.created`,
+            frame,
+          )
+        if (event.response?.id && event.response.id !== responseID)
+          return yield* ProviderShared.eventError(
+            options.id,
+            `${options.name} response ID changed during execution`,
+            frame,
+          )
+        if (event.type === "response.completed") {
+          terminal = true
+          return { type: "completed", frame }
+        }
+        if (event.type === "response.incomplete") {
+          terminal = true
+          return { type: "incomplete", frame }
+        }
+        return { type: "frame", frame }
+      }),
+  }
+}
+
+export const transport = <Body>(options: Options): Transport<Body, Prepared, string> => {
+  const http = HttpTransport.sseJson.with<Body>()
+  return {
+    id: http.id,
+    prepare: (input) =>
+      Effect.gen(function* () {
+        const parts = yield* HttpTransport.jsonRequestParts(input)
+        const headers = Headers.remove(options.headers?.(parts.headers) ?? parts.headers, "content-length")
+        const channel =
+          input.webSocket && (options.enabled?.(parts.url) ?? true)
+            ? yield* Effect.gen(function* () {
+                const create = yield* message(parts.jsonBody)
+                const base = driver(options, create.message)
+                return {
+                  url: yield* WebSocketTransport.toWebSocketUrl(options.url?.(parts.url) ?? parts.url),
+                  headers,
+                  rotateAfterMs: options.rotateAfterMs,
+                  driver: OpenResponsesContinuation.driver({
+                    id: options.id,
+                    name: options.name,
+                    request: create.request,
+                    message: create.message,
+                    base,
+                    continuation: options.continuation,
+                  }),
+                }
+              })
+            : undefined
+        if (input.webSocket && !channel)
+          yield* Effect.logWarning(`${options.name} does not offer WebSocket for this endpoint; using HTTP`, {
+            url: parts.url,
+          })
+        return {
+          http: {
+            request: ProviderShared.jsonPost({ url: parts.url, body: parts.bodyText, headers: parts.headers }),
+            framing: Framing.sse,
+            middleware: input.middleware,
+          },
+          channel,
+        }
+      }),
+    execute: (prepared, request, runtime, executeOptions) =>
+      Effect.gen(function* () {
+        if (!executeOptions?.webSocket || !prepared.channel) return yield* http.execute(prepared.http, request, runtime)
+        let fallbackHttp: HttpContext | undefined
+        const exchange: WebSocketChannelExchange = {
+          id: request.id ?? "request",
+          connect: {
+            url: prepared.channel.url,
+            headers: prepared.channel.headers,
+            rotateAfterMs: prepared.channel.rotateAfterMs,
+          },
+          fallback: () =>
+            Stream.unwrap(
+              http.execute(prepared.http, request, runtime).pipe(
+                Effect.map((execution) => {
+                  fallbackHttp = execution.http
+                  return execution.frames
+                }),
+              ),
+            ),
+          driver: prepared.channel.driver,
+        }
+        const execution = yield* executeOptions.webSocket.execute(exchange)
+        return {
+          frames: execution.frames,
+          complete: execution.complete,
+          get http() {
+            return fallbackHttp ?? execution.http
+          },
+        }
+      }),
+  }
+}
+
+export const OpenResponsesChannel = { transport } as const

packages/ai/src/protocols/open-responses-continuation.ts

@@ -0,0 +1,208 @@
+import { AIError, TransportError } from "../schema/index.js"
+import type { ChannelCheckpoint, ChannelObservation, WebSocketChannelDriver } from "../route/transport/index.js"
+import { Effect, Option, Schema } from "effect"
+import * as ProviderShared from "./shared.js"
+import { OpenResponses } from "./open-responses.js"
+
+const PROTOCOL = "open-responses.websocket.v1"
+const VERSION = 1
+
+interface CheckpointValue {
+  readonly version: typeof VERSION
+  readonly responseID: string
+  readonly request: Readonly<Record<string, unknown>>
+  readonly output: ReadonlyArray<unknown>
+}
+
+/**
+ * Fields to send next to `previous_response_id` on an incremental step, or undefined to send the step in full.
+ * Whether omitted fields carry over from the continued response is provider behavior the route must know.
+ */
+export type Shape = (request: Readonly<Record<string, unknown>>) => Readonly<Record<string, unknown>> | undefined
+
+export interface DriverInput {
+  readonly id: string
+  readonly name: string
+  readonly request: Readonly<Record<string, unknown>>
+  readonly message: string
+  readonly base: WebSocketChannelDriver
+  readonly continuation?: Shape
+}
+
+const checkpointValue = (checkpoint: ChannelCheckpoint | undefined): CheckpointValue | undefined => {
+  if (checkpoint?.protocol !== PROTOCOL || !ProviderShared.isRecord(checkpoint.value)) return undefined
+  if (checkpoint.value.version !== VERSION) return undefined
+  if (typeof checkpoint.value.responseID !== "string" || checkpoint.value.responseID.trim().length === 0)
+    return undefined
+  if (!ProviderShared.isRecord(checkpoint.value.request) || !Array.isArray(checkpoint.value.output)) return undefined
+  return {
+    version: VERSION,
+    responseID: checkpoint.value.responseID,
+    request: checkpoint.value.request,
+    output: checkpoint.value.output,
+  }
+}
+
+const canonical = (value: unknown): string => {
+  if (value === undefined) return "undefined"
+  if (Array.isArray(value)) return `[${value.map(canonical).join(",")}]`
+  if (!ProviderShared.isRecord(value)) return ProviderShared.encodeJson(value)
+  return `{${Object.keys(value)
+    .filter((key) => value[key] !== undefined)
+    .sort()
+    .map((key) => `${ProviderShared.encodeJson(key)}:${canonical(value[key])}`)
+    .join(",")}}`
+}
+
+const json = (value: unknown) => {
+  if (typeof value !== "string") return value
+  return Option.getOrElse(Schema.decodeUnknownOption(ProviderShared.Json)(value), () => value)
+}
+
+const comparable = (value: unknown) => {
+  if (!ProviderShared.isRecord(value)) return value
+  if (value.type === "message" && value.role === "assistant")
+    return {
+      role: "assistant",
+      // Annotations and logprobs describe the response, not the text replayed in model input.
+      content: Array.isArray(value.content)
+        ? value.content.map((part) =>
+            ProviderShared.isRecord(part) && part.type === "output_text" ? { type: part.type, text: part.text } : part,
+          )
+        : value.content,
+      ...(value.phase === undefined ? {} : { phase: value.phase }),
+    }
+  if (value.type === "function_call")
+    return {
+      type: value.type,
+      call_id: value.call_id,
+      name: value.name,
+      arguments: json(value.arguments),
+    }
+  if (value.type === "reasoning")
+    return {
+      type: value.type,
+      summary: value.summary,
+      encrypted_content: value.encrypted_content,
+    }
+  return value
+}
+
+const invariant = (request: Readonly<Record<string, unknown>>) => {
+  const { type: _type, input: _input, previous_response_id: _previousResponseID, ...rest } = request
+  return rest
+}
+
+const incremental = (
+  request: Readonly<Record<string, unknown>>,
+  checkpoint: CheckpointValue,
+): ReadonlyArray<unknown> | undefined => {
+  const input = request.input
+  const previousInput = checkpoint.request.input
+  if (!Array.isArray(input) || !Array.isArray(previousInput)) return undefined
+  if (canonical(invariant(request)) !== canonical(invariant(checkpoint.request))) return undefined
+  const baseline = [...previousInput, ...checkpoint.output]
+  if (input.length <= baseline.length) return undefined
+  if (!baseline.every((item, index) => canonical(comparable(item)) === canonical(comparable(input[index]))))
+    return undefined
+  return input.slice(baseline.length)
+}
+
+const code = (event: OpenResponses.Event) => event.code || event.error?.code || event.response?.error?.code || undefined
+
+const rejected = (
+  observation: Extract<ChannelObservation, { readonly type: "provider-failure" }>,
+  recovery: "retry-full" | "rotate-and-retry-full",
+): ChannelObservation => ({
+  type: "rejected",
+  recovery,
+  error: new AIError({
+    reason: new TransportError({
+      message: observation.error.message,
+      body: observation.error.reason.body,
+      http: observation.error.reason.http,
+      cause: observation.error.reason.cause,
+      transport: "websocket",
+      operation: "read",
+      phase: "receive",
+      delivery: "rejected",
+      recovery,
+    }),
+  }),
+})
+
+export const driver = (input: DriverInput): WebSocketChannelDriver => {
+  const { previous_response_id: _previousResponseID, ...request } = input.request
+  const shape = input.continuation ?? ((fields: Readonly<Record<string, unknown>>) => fields)
+  let output: OpenResponses.StreamItem[] = []
+  return {
+    create: (checkpoint) =>
+      Effect.sync(() => {
+        output = []
+        const previous = checkpointValue(checkpoint)
+        // Ask the route first: diffing the whole history is wasted when it declines the continuation.
+        const fields = previous ? shape(request) : undefined
+        const delta = previous && fields ? incremental(request, previous) : undefined
+        if (!previous || !fields || !delta)
+          return { message: ProviderShared.encodeJson(request), mode: "full" as const }
+        return {
+          message: ProviderShared.encodeJson({ ...fields, input: delta, previous_response_id: previous.responseID }),
+          mode: "incremental" as const,
+        }
+      }),
+    observe: (create, frame) =>
+      Effect.gen(function* () {
+        const event = yield* OpenResponses.decodeChannelEvent(frame).pipe(
+          Effect.mapError((cause) =>
+            ProviderShared.eventError(input.id, `Invalid ${input.name} WebSocket event`, frame, cause),
+          ),
+        )
+        const observation = yield* input.base.observe(create, frame)
+        if (event.type === "response.output_item.done" && event.item) output.push(event.item)
+        if (observation.type === "provider-failure") {
+          const rejection = code(event)
+          if (rejection === "previous_response_not_found") return rejected(observation, "retry-full")
+          if (rejection === "websocket_connection_limit_reached") return rejected(observation, "rotate-and-retry-full")
+          // Only the continuation distinguishes an incremental send from a full one, so an unclassified
+          // invalid request there is retried full; Codex reports a stale previous_response_id that way, with
+          // no code. Classified failures such as context overflow keep their runner-owned recovery.
+          if (
+            create.mode === "incremental" &&
+            observation.error.reason._tag === "InvalidRequest" &&
+            observation.error.reason.classification === undefined
+          )
+            return rejected(observation, "retry-full")
+        }
+        if (observation.type !== "completed") return observation
+        // A trigger installs a different context window. Clear the append baseline, retaining the socket.
+        if (
+          Array.isArray(request.input) &&
+          request.input.some((item) => ProviderShared.isRecord(item) && item.type === "compaction_trigger")
+        )
+          return observation
+        const responseID = event.response?.id
+        if (!responseID || responseID.trim().length === 0) return observation
+        return {
+          ...observation,
+          checkpoint: {
+            protocol: PROTOCOL,
+            value: {
+              version: VERSION,
+              responseID,
+              request,
+              // Completion can re-encrypt reasoning. Callers replay the item already emitted by output_item.done.
+              output: event.response?.output?.length
+                ? event.response.output.map((item) =>
+                    item.type === "reasoning" && item.id !== undefined
+                      ? (output.find((done) => done.type === item.type && done.id === item.id) ?? item)
+                      : item,
+                  )
+                : output.slice(),
+            } satisfies CheckpointValue,
+          },
+        }
+      }),
+  }
+}
+
+export * as OpenResponsesContinuation from "./open-responses-continuation.js"

packages/ai/src/protocols/openai-compatible-chat.ts

@@ -0,0 +1,24 @@
+import { Route, type RouteRoutedLanguageModelInput } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import * as OpenAIChat from "./openai-chat.js"
+
+const ADAPTER = "openai-compatible-chat"
+
+export type OpenAICompatibleChatLanguageModelInput = RouteRoutedLanguageModelInput
+
+/**
+ * Route for non-OpenAI providers that expose an OpenAI Chat-compatible
+ * `/chat/completions` endpoint. Reuses `OpenAIChat.protocol` end-to-end and
+ * overrides only the route id so providers can be resolved per-family without
+ * colliding with native OpenAI. Provider helpers configure the route endpoint
+ * before model selection.
+ */
+export const route = Route.make({
+  id: ADAPTER,
+  providerMetadataKey: "openai",
+  protocol: OpenAIChat.protocol,
+  endpoint: Endpoint.path("/chat/completions"),
+  framing: OpenAIChat.framing,
+})
+
+export * as OpenAICompatibleChat from "./openai-compatible-chat.js"

packages/ai/src/protocols/openai-compatible-responses.ts

@@ -0,0 +1,23 @@
+import { Route, type RouteRoutedLanguageModelInput } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { OpenResponses } from "./open-responses.js"
+
+const ADAPTER = "openai-compatible-responses"
+
+export type OpenAICompatibleResponsesLanguageModelInput = RouteRoutedLanguageModelInput
+
+/**
+ * Deployment adapter for providers that expose an Open Responses-compatible
+ * `/responses` endpoint. Provider helpers configure identity, endpoint, and
+ * auth while the semantic protocol remains provider-neutral.
+ */
+export const route = Route.make({
+  id: ADAPTER,
+  providerMetadataKey: "openresponses",
+  protocol: OpenResponses.protocol,
+  endpoint: Endpoint.path(OpenResponses.PATH),
+  transport: OpenResponses.httpTransport,
+  defaults: { providerOptions: { store: false, include: ["reasoning.encrypted_content"] } },
+})
+
+export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"

packages/ai/src/protocols/openai-images.ts

@@ -0,0 +1,207 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { ProviderID, mergeJsonRecords, type AIError } from "../schema/index.js"
+import { ProviderShared } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const ADAPTER = "openai-images"
+const NAME = "OpenAI Images"
+const PROVIDER = ProviderID.make("openai")
+export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
+export const PATH = "/images/generations"
+export const EDIT_PATH = "/images/edits"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type OpenAIImageString<Known extends string> = Known | (string & {})
+
+/** Provider-native options. Common fields (`n`, `size`, `format`, `images`, `mask`) live on the request. */
+export type OpenAIImageOptions = {
+  readonly quality?: OpenAIImageString<"auto" | "low" | "medium" | "high" | "standard" | "hd">
+  readonly background?: OpenAIImageString<"auto" | "opaque" | "transparent">
+  readonly moderation?: OpenAIImageString<"auto" | "low">
+  readonly outputCompression?: number
+} & Record<string, unknown>
+
+export type Request = ImageRequestFor<OpenAIImageOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Response schema
+// ---------------------------------------------------------------------------
+
+const OpenAIImageResponse = Schema.Struct({
+  data: Schema.Array(
+    Schema.Struct({
+      b64_json: Schema.optional(Schema.String),
+      url: Schema.optional(Schema.String),
+      revised_prompt: Schema.optional(Schema.String),
+    }),
+  ),
+  output_format: Schema.optional(Schema.String),
+  usage: Schema.optional(
+    Schema.Struct({
+      input_tokens: Schema.optional(Schema.Number),
+      output_tokens: Schema.optional(Schema.Number),
+      total_tokens: Schema.optional(Schema.Number),
+      input_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+      output_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+    }),
+  ),
+})
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+/** Multipart field names the route owns; `http.body` overlays cannot smuggle replacements for them. */
+const RESERVED_FORM_FIELDS = new Set(["model", "prompt", "image", "image[]", "images", "mask"])
+
+const nativeOptions = (options: OpenAIImageOptions | undefined) => {
+  if (!options) return undefined
+  const { outputCompression, ...native } = options
+  return { output_compression: outputCompression, ...native }
+}
+
+const isEdit = (request: Request) => (request.images?.length ?? 0) > 0
+
+const isInline = (asset: Media.Asset) => asset.inline() !== undefined
+
+const blob = (data: Uint8Array, mediaType: string) => {
+  const buffer = new ArrayBuffer(data.byteLength)
+  new Uint8Array(buffer).set(data)
+  return new Blob([buffer], { type: mediaType })
+}
+
+const reference = (asset: Media.Asset): Effect.Effect<Record<string, unknown>, AIError> => {
+  const inline = asset.inline()
+  if (inline) return Effect.succeed({ image_url: inline.dataUrl })
+  const url = ProviderShared.mediaUrl(asset)
+  if (url) return Effect.succeed({ image_url: url })
+  const id = MediaInput.refID(asset, PROVIDER)
+  if (id) return Effect.succeed({ file_id: id })
+  return Effect.fail(
+    ProviderShared.invalidRequest("OpenAI Images accepts image URLs, data URLs, bytes, and OpenAI file IDs"),
+  )
+}
+
+const fromRequest = Effect.fn("OpenAIImages.fromRequest")(function* (request: Request) {
+  const images = request.images ?? []
+  const mask = request.mask
+  if (mask !== undefined && images.length === 0)
+    return yield* ProviderShared.invalidRequest("An OpenAI image mask requires at least one input image")
+  const fields = mergeJsonRecords(
+    { n: request.n, size: request.size, output_format: request.format },
+    nativeOptions(request.providerOptions),
+    request.http?.body,
+  )
+
+  // Owned bytes go through multipart edits; remote URLs and file IDs use the JSON edits body instead.
+  if (images.length > 0 && images.every(isInline) && (mask === undefined || isInline(mask))) {
+    const form = new FormData()
+    form.append("model", request.model.id)
+    form.append("prompt", request.prompt)
+    Object.entries(fields ?? {}).forEach(([key, value]) => {
+      if (RESERVED_FORM_FIELDS.has(key)) return
+      form.append(key, typeof value === "string" ? value : ProviderShared.encodeJson(value))
+    })
+    const uploads = yield* Effect.forEach(images, (image) => MediaInput.inlineBytes(ADAPTER, image))
+    uploads.forEach((data, index) => form.append("image[]", blob(data, images[index].mediaType), `image-${index}`))
+    if (mask !== undefined)
+      form.append("mask", blob(yield* MediaInput.inlineBytes(ADAPTER, mask), mask.mediaType), "mask")
+    return MediaProtocol.multipart(form)
+  }
+
+  const references = yield* Effect.forEach(images, reference)
+  const maskReference = mask === undefined ? undefined : yield* reference(mask)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        model: request.model.id,
+        prompt: request.prompt,
+        images: references.length === 0 ? undefined : references,
+        mask: maskReference,
+      },
+      fields,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const requestedFormat = (body: MediaProtocol.Body) => {
+  const value = body.type === "json" ? body.value.output_format : body.value.get("output_format")
+  return typeof value === "string" ? value : undefined
+}
+
+const decodeResponse = Effect.fn("OpenAIImages.decodeResponse")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.DecodeContext<Request>,
+) {
+  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, OpenAIImageResponse)(response)
+  const decoded = output.value
+  const format = decoded.output_format ?? requestedFormat(context.body) ?? "png"
+  const mediaType = `image/${format}`
+  const images = yield* Effect.forEach(decoded.data, (item, index) => {
+    const providerMetadata =
+      item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } }
+    if (item.b64_json)
+      return MediaInput.decodedAsset(output.invalid, `${NAME} result ${index}`, item.b64_json, mediaType, {
+        info: { format },
+        providerMetadata,
+      })
+    if (item.url) return Effect.succeed(Media.url(item.url, { mediaType, info: { format }, providerMetadata }))
+    return Effect.fail(output.invalid(`${NAME} result ${index} has neither image data nor a URL`))
+  })
+  if (images.length === 0) return yield* output.invalid(`${NAME} returned no images`)
+  return new ImageResponse({
+    images,
+    usage:
+      decoded.usage === undefined
+        ? undefined
+        : {
+            type: "tokens",
+            input: decoded.usage.input_tokens,
+            output: decoded.usage.output_tokens,
+            total: decoded.usage.total_tokens,
+            details: { openai: decoded.usage },
+          },
+    providerMetadata: { openai: { outputFormat: format } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.inline<Request, ImageResponse>({
+  id: ADAPTER,
+  name: NAME,
+  unsupported: ["aspectRatio", "seed"],
+  body: { from: fromRequest },
+  response: { decode: decodeResponse },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  ImageModel.fromRoute<OpenAIImageOptions>(
+    {
+      id: ADAPTER,
+      provider: PROVIDER,
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      path: ({ request }) => (isEdit(request) ? EDIT_PATH : PATH),
+    },
+    input,
+  )
+
+export const OpenAIImages = {
+  protocol,
+  model,
+} as const

packages/ai/src/protocols/openai-responses.ts

@@ -0,0 +1,345 @@
+import { Effect, Encoding, Schema } from "effect"
+import { Headers } from "effect/unstable/http"
+import { Route } from "../route/client.js"
+import { Auth } from "../route/auth.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Protocol } from "../route/protocol.js"
+import { HttpTransport } from "../route/transport/index.js"
+import { LLMRequest, mergeJsonRecords, type JsonSchema, type ToolDefinition, type ToolEntry } from "../schema/index.js"
+import { resolveEffortUpdates } from "../effort-updates.js"
+import { OpenResponses } from "./open-responses.js"
+import { OpenResponsesOptions } from "./utils/open-responses-options.js"
+import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
+import { OpenAIImage } from "./utils/openai-image.js"
+import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
+import { ToolSchemaProjection } from "./utils/tool-schema.js"
+import { OpenResponsesChannel } from "./open-responses-channel.js"
+import { ResponsesCompaction } from "./utils/responses-compaction.js"
+import { ResponsesCheckpoint } from "./utils/responses-checkpoint.js"
+
+const ADAPTER = "openai-responses"
+const NAME = "OpenAI Responses"
+const WEBSOCKET_PROTOCOL_HEADER = "responses_websockets=2026-02-06"
+const WEBSOCKET_ROTATE_AFTER_MS = 55 * 60 * 1000
+export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
+export const PATH = OpenResponses.PATH
+
+export const ContextManagement = Schema.Array(
+  Schema.Struct({
+    type: Schema.Literal("compaction"),
+    compactThreshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+  }),
+)
+export type ContextManagement = typeof ContextManagement.Type
+
+const OpenAIResponsesImageGenerationTool = Schema.Struct({
+  type: Schema.tag("image_generation"),
+  action: Schema.optional(Schema.Literals(["auto", "generate", "edit"])),
+  background: Schema.optional(Schema.Literals(["auto", "opaque", "transparent"])),
+  input_fidelity: Schema.optional(Schema.Literals(["low", "high"])),
+  output_compression: Schema.optional(Schema.Int.check(Schema.isBetween({ minimum: 0, maximum: 100 }))),
+  output_format: Schema.optional(Schema.Literals(["png", "jpeg", "webp"])),
+  partial_images: Schema.optional(Schema.Int.check(Schema.isGreaterThanOrEqualTo(0))),
+  quality: Schema.optional(Schema.Literals(["auto", "low", "medium", "high"])),
+  size: Schema.optional(OpenAIImage.Size),
+})
+
+const OpenAIResponsesHostedToolItem = Schema.Union([
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("computer_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      call_id: Schema.optional(Schema.String),
+      action: optionalNull(JsonObject),
+      pending_safety_checks: Schema.optional(Schema.Array(JsonObject)),
+    }),
+    [JsonObject],
+  ),
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("web_search_preview_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      action: optionalNull(JsonObject),
+    }),
+    [JsonObject],
+  ),
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("image_generation_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      result: optionalNull(Schema.String),
+      output_format: Schema.optional(Schema.Literals(["png", "jpeg", "webp"])),
+      revised_prompt: optionalNull(Schema.String),
+    }),
+    [JsonObject],
+  ),
+])
+
+const OpenAIResponsesNamespace = Schema.Struct({
+  type: Schema.tag("namespace"),
+  name: Schema.String,
+  description: Schema.String,
+  tools: Schema.Array(OpenResponses.Tool),
+})
+
+const OpenAIResponsesTools = Schema.Union([
+  OpenResponses.Tool,
+  OpenAIResponsesNamespace,
+  OpenAIResponsesImageGenerationTool,
+])
+
+const OpenAIResponsesToolChoice = Schema.Union([
+  OpenResponses.ToolChoice,
+  Schema.Struct({ type: Schema.tag("image_generation") }),
+])
+
+const OpenAIResponsesInputItem = Schema.Union([
+  OpenResponses.InputItem,
+  OpenAIResponsesHostedToolItem,
+  OpenResponses.ConfigurationUpdate,
+])
+
+const OpenAIResponsesCoreFields = {
+  ...OpenResponses.coreFields,
+  input: Schema.Array(OpenAIResponsesInputItem),
+  tools: optionalArray(OpenAIResponsesTools),
+  tool_choice: Schema.optional(OpenAIResponsesToolChoice),
+  context_management: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        type: Schema.Literal("compaction"),
+        compact_threshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+      }),
+    ),
+  ),
+}
+
+const OpenAIResponsesBody = Schema.Struct({
+  ...OpenAIResponsesCoreFields,
+  stream: Schema.Literal(true),
+})
+export type OpenAIResponsesBody = Schema.Schema.Type<typeof OpenAIResponsesBody>
+
+/** Request control, never conversation content. */
+export const CompactionTrigger = Schema.Struct({ type: Schema.Literal("compaction_trigger") })
+const CheckpointBody = Schema.Struct({
+  ...OpenAIResponsesBody.fields,
+  input: Schema.Array(Schema.Union([OpenAIResponsesInputItem, CompactionTrigger])),
+  store: Schema.Literal(false),
+  prompt_cache_retention: optionalNull(Schema.String),
+  prompt_cache_options: optionalNull(
+    Schema.Struct({ mode: Schema.optional(Schema.String), ttl: Schema.optional(Schema.String) }),
+  ),
+})
+
+const adapter = {
+  id: ADAPTER,
+  name: NAME,
+  restoreHostedToolItem: (item: unknown) => (Schema.is(OpenAIResponsesHostedToolItem)(item) ? item : undefined),
+} satisfies OpenResponses.ProviderAdapter
+
+// Only GPT-6 Astra accepts `configuration_update`, and never alongside automatic `context_management` compaction.
+const supportsEffortUpdates = (request: LLMRequest) => {
+  if (request.providerOptions?.contextManagement !== undefined) return false
+  const override = request.model.compatibility?.supportsEffortUpdates
+  if (override !== undefined) return override
+  return /(?:^|\/)gpt-6-astra$/i.test(request.model.id)
+}
+
+const nativeImageToolInput = (tool: ToolDefinition) => {
+  const native = tool.native?.openai
+  return ProviderShared.isRecord(native) && native.type === "image_generation" ? native : undefined
+}
+
+const nativeImageTool = (tool: ToolDefinition) => {
+  const native = nativeImageToolInput(tool)
+  return Schema.is(OpenAIResponsesImageGenerationTool)(native) ? native : undefined
+}
+
+const lowerTool = Effect.fn("OpenAIResponses.lowerTool")(function* (tool: ToolDefinition, inputSchema: JsonSchema) {
+  const native = nativeImageToolInput(tool)
+  if (native !== undefined) {
+    if (Schema.is(OpenAIResponsesImageGenerationTool)(native)) return native
+    return yield* ProviderShared.invalidRequest("OpenAI Responses image generation tool options are invalid")
+  }
+  return yield* OpenResponses.lowerTool(NAME, tool, inputSchema)
+})
+
+// Native namespaces hold only function tools, so deeper levels flatten into
+// the leaf names the same way non-native protocols flatten the whole tree.
+const lowerToolEntry = Effect.fn("OpenAIResponses.lowerToolEntry")(function* (
+  tool: ToolEntry,
+  compatibility: Parameters<typeof ToolSchemaProjection.modelCompatibility>[1],
+) {
+  if (tool.type === "tool")
+    return yield* lowerTool(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, compatibility))
+  // OpenAI requires a namespace description; fall back to a generic one so a
+  // missing description never blocks the request.
+  return {
+    type: "namespace" as const,
+    name: tool.name,
+    description: tool.description ?? `Tools in the ${tool.name} namespace.`,
+    tools: yield* Effect.forEach(ProviderShared.flattenTools(tool.tools), (leaf) =>
+      OpenResponses.lowerTool(NAME, leaf, ToolSchemaProjection.modelCompatibility(leaf.inputSchema, compatibility)),
+    ),
+  }
+})
+
+const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>, tools: ReadonlyArray<ToolEntry>) =>
+  ProviderShared.matchToolChoice(NAME, toolChoice, {
+    auto: () => "auto" as const,
+    none: () => "none" as const,
+    required: () => "required" as const,
+    tool: (name) =>
+      tools.some((tool) => tool.type === "tool" && tool.name === name && nativeImageTool(tool) !== undefined)
+        ? ({ type: "image_generation" } as const)
+        : { type: "function" as const, name },
+  })
+
+const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenAIResponsesBody))
+
+const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request: LLMRequest) {
+  const management = yield* ProviderShared.validateWith(
+    Schema.decodeUnknownEffect(Schema.UndefinedOr(ContextManagement)),
+  )(request.providerOptions?.contextManagement)
+  const options = OpenResponsesOptions.resolve(request)
+  const updates = resolveEffortUpdates(request, options.reasoningEffort)
+  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
+  return yield* decodeBody({
+    ...(yield* OpenResponses.lowerConversation(updates.request, adapter)),
+    ...OpenResponses.lowerGeneration(request, { ...options, reasoningEffort: updates.effort }),
+    context_management: management?.map((edit) => ({ type: edit.type, compact_threshold: edit.compactThreshold })),
+    tools:
+      request.tools.length === 0
+        ? undefined
+        : yield* Effect.forEach(request.tools, (tool) => lowerToolEntry(tool, toolSchemaCompatibility)),
+    tool_choice:
+      request.tools.length === 0
+        ? undefined
+        : (OpenResponses.allowedToolChoice(request) ??
+          (request.toolChoice ? yield* lowerToolChoice(request.toolChoice, request.tools) : undefined)),
+  })
+})
+
+const checkpointBody = {
+  schema: CheckpointBody,
+  from: Effect.fn("OpenAIResponses.checkpointBody")(function* (request: LLMRequest) {
+    const native = yield* fromRequest(LLMRequest.update(request, { toolChoice: undefined }))
+    const overlay = request.http?.body
+    // Complete history is required for stateless replay and SSE recovery. Raw input overrides bypass that contract.
+    if (
+      overlay?.input !== undefined ||
+      overlay?.previous_response_id !== undefined ||
+      overlay?.conversation !== undefined
+    )
+      return yield* ProviderShared.invalidRequest(
+        "Trigger compaction requires complete canonical history, not an input or continuation override",
+      )
+    return yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(CheckpointBody))({
+      ...mergeJsonRecords(native, overlay),
+      input: [...native.input, { type: "compaction_trigger" }],
+      stream: true,
+      store: false,
+      parallel_tool_calls: true,
+      tool_choice: undefined,
+      context_management: undefined,
+      text: undefined,
+      max_output_tokens: undefined,
+      max_tool_calls: undefined,
+    })
+  }),
+}
+
+const hostedToolResult = Effect.fn("OpenAIResponses.hostedToolResult")(function* (item: ResponsesHostedTools.Item) {
+  const isError = item.error !== undefined && item.error !== null
+  if (item.type === "image_generation_call" && item.result) {
+    yield* Effect.fromResult(Encoding.decodeBase64(item.result)).pipe(
+      Effect.mapError((cause) =>
+        ProviderShared.eventError(ADAPTER, "OpenAI Responses returned invalid image base64", undefined, cause),
+      ),
+    )
+    const format = item.output_format ?? "png"
+    return {
+      type: "content" as const,
+      value: [
+        {
+          type: "file" as const,
+          uri: `data:image/${format};base64,${item.result}`,
+          mime: `image/${format}`,
+        },
+      ],
+    }
+  }
+  return isError ? { type: "error" as const, value: item.error } : { type: "json" as const, value: item }
+})
+
+const HOSTED_TOOLS = {
+  web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
+  web_search_preview_call: { name: "web_search_preview", input: (item) => item.action ?? {} },
+  file_search_call: { name: "file_search", input: (item) => ({ queries: item.queries ?? [] }) },
+  code_interpreter_call: {
+    name: "code_interpreter",
+    input: (item) => ({ code: item.code, container_id: item.container_id }),
+  },
+  computer_call: { name: "computer_use", input: (item) => item.action ?? {} },
+  image_generation_call: { name: "image_generation", input: () => ({}), result: hostedToolResult },
+  mcp_call: {
+    name: "mcp",
+    input: (item) => ({ server_label: item.server_label, name: item.name, arguments: item.arguments }),
+  },
+} as const satisfies ResponsesHostedTools.Definitions
+
+const step = (state: OpenResponses.ParserState, input: OpenResponses.Event) => {
+  const event = OpenResponses.normalize(state, input)
+  if (event.type === "response.reasoning_text.delta")
+    return event.item_id !== undefined
+      ? Effect.succeed(OpenResponses.onReasoningDelta(state, event, event.item_id))
+      : ProviderShared.eventError(ADAPTER, `${event.type} is missing item_id`)
+  if (event.type === "response.output_item.done" && event.item && ResponsesHostedTools.isItem(event.item, HOSTED_TOOLS))
+    return ResponsesHostedTools.onDone(state, event.item, HOSTED_TOOLS)
+  return OpenResponses.step(state, event)
+}
+
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: {
+    schema: OpenAIResponsesBody,
+    from: fromRequest,
+  },
+  stream: {
+    event: OpenResponses.protocol.stream.event,
+    initial: (request) => OpenResponses.initial(request, adapter),
+    step,
+    terminal: OpenResponses.terminal,
+  },
+  supportsEffortUpdates,
+})
+
+const endpoint = Endpoint.path<OpenAIResponsesBody>(PATH, { baseURL: DEFAULT_BASE_URL })
+const auth = Auth.none
+
+export const httpTransport = HttpTransport.sseJson.with<OpenAIResponsesBody>()
+export const channelTransport = OpenResponsesChannel.transport<OpenAIResponsesBody>
+export const transport = channelTransport({
+  id: ADAPTER,
+  name: NAME,
+  rotateAfterMs: WEBSOCKET_ROTATE_AFTER_MS,
+  headers: (headers) => Headers.set(headers, "openai-beta", headers["openai-beta"] ?? WEBSOCKET_PROTOCOL_HEADER),
+})
+
+export const route = Route.make({
+  compact: { endpoint: ResponsesCompaction.make(adapter), trigger: ResponsesCheckpoint.make(checkpointBody) },
+  id: ADAPTER,
+  provider: "openai",
+  providerMetadataKey: "openai",
+  protocol,
+  endpoint,
+  auth,
+  transport,
+  defaults: { providerOptions: { store: false, include: ["reasoning.encrypted_content"] } },
+})
+
+export * as OpenAIResponses from "./openai-responses.js"

packages/ai/src/protocols/shared.ts

@@ -0,0 +1,387 @@
+import { Tool } from "@opencode/schema/tool"
+import { Effect, Schema, Stream } from "effect"
+import * as Sse from "effect/unstable/encoding/Sse"
+import { Headers, HttpClientRequest } from "effect/unstable/http"
+import { Media } from "../media.js"
+import {
+  InvalidProviderOutputError,
+  InvalidRequestError,
+  UnsupportedOperationError,
+  AIError,
+  LLMRequest,
+  Message,
+  ToolDefinition,
+  type ContentPart,
+  type MediaPart,
+  type ProviderID,
+  type TextPart,
+  type ToolEntry,
+  type ToolResultPart,
+} from "../schema/index.js"
+import { isRecord } from "../utils/record.js"
+export { isRecord }
+
+export const Json = Schema.fromJsonString(Schema.Unknown)
+export const decodeJson = Schema.decodeUnknownSync(Json)
+export const encodeJson = Schema.encodeSync(Json)
+const isJson = Schema.is(Schema.Json)
+export const JsonObject = Schema.Record(Schema.String, Schema.Unknown)
+export const optionalArray = <const S extends Schema.Top>(schema: S) => Schema.optional(Schema.Array(schema))
+export const optionalNull = <const S extends Schema.Top>(schema: S) => Schema.optional(Schema.NullOr(schema))
+
+export const OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH = 64
+
+// OpenAI limits `prompt_cache_key` to 64 chars; DeepSeek and Zai inherit the same
+// limit via their OpenAI-compatible APIs. Clamp with unicode-aware slicing.
+export const promptCacheKey = (request: LLMRequest): string | undefined => {
+  if (request.cache === "none" || request.promptCacheKey === undefined) return undefined
+  const chars = Array.from(request.promptCacheKey)
+  if (chars.length <= OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH) return request.promptCacheKey
+  return chars.slice(0, OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH).join("")
+}
+
+/**
+ * Streaming tool-call accumulator. Adapters that build a tool call across
+ * multiple `tool-input-delta` chunks store the partial JSON input string here
+ * and finalize it with `parseToolInput` once the call completes.
+ */
+export interface ToolAccumulator {
+  readonly id: string
+  readonly name: string
+  readonly namespace?: string
+  readonly input: string
+}
+
+/**
+ * `Usage.totalTokens` policy shared by every route. Honors a provider-
+ * supplied total; otherwise falls back to `inputTokens + outputTokens` only
+ * when at least one is defined. Returns `undefined` when neither input nor
+ * output is known so routes don't publish a misleading `0`.
+ *
+ * Under the inclusive `AI.Usage` contract, `inputTokens` includes cached input
+ * and `outputTokens` includes reasoning. Protocol mappers normalize those
+ * inclusive values before calling this helper. The provider-supplied total is
+ * the source of truth when present; otherwise their sum is the canonical total.
+ */
+export const totalTokens = (
+  inputTokens: number | undefined,
+  outputTokens: number | undefined,
+  total: number | undefined,
+) => {
+  if (total !== undefined) return total
+  if (inputTokens === undefined && outputTokens === undefined) return undefined
+  return (inputTokens ?? 0) + (outputTokens ?? 0)
+}
+
+/**
+ * Subtract `subtrahend` from `total`, clamping to zero if the provider
+ * reports a non-sensical breakdown (e.g. `cached_tokens > prompt_tokens`).
+ * Used by protocol mappers when deriving a non-overlapping breakdown field
+ * from a provider's inclusive total — `nonCachedInputTokens` from
+ * `inputTokens - cacheReadInputTokens - cacheWriteInputTokens`.
+ *
+ * If `total` is `undefined`, returns `undefined` (we don't fabricate
+ * counts). If `subtrahend` is `undefined`, returns `total` unchanged. The
+ * provider-native breakdown stays available on `Usage.providerMetadata` for debugging.
+ */
+export const subtractTokens = (total: number | undefined, subtrahend: number | undefined): number | undefined => {
+  if (total === undefined) return undefined
+  if (subtrahend === undefined) return total
+  return Math.max(0, total - subtrahend)
+}
+
+/**
+ * Sum a list of optional token counts, returning `undefined` only when
+ * every value is `undefined` (so we don't fabricate a `0`). Used by
+ * protocol mappers to derive the inclusive `inputTokens` total from a
+ * provider that natively reports a non-overlapping breakdown
+ * (e.g. Anthropic, whose `input_tokens` is already non-cached only).
+ */
+export const sumTokens = (...values: ReadonlyArray<number | undefined>): number | undefined => {
+  if (values.every((value) => value === undefined)) return undefined
+  return values.reduce((acc: number, value) => acc + (value ?? 0), 0)
+}
+
+export const eventError = (route: string, message: string, body?: string, cause?: unknown) =>
+  new AIError({
+    reason: new InvalidProviderOutputError({ route, message, body, cause }),
+  })
+
+export const parseJson = (route: string, input: string, message: string) =>
+  Effect.try({
+    try: () => decodeJson(input),
+    catch: (cause) => eventError(route, message, input, cause),
+  })
+
+/**
+ * Join the `text` field of a list of parts with newlines. Used by routes
+ * that flatten system / message content arrays into a single provider string
+ * (OpenAI Chat `system` content, OpenAI Responses `system` content, Gemini
+ * `systemInstruction.parts[].text`).
+ */
+export const joinText = (parts: ReadonlyArray<{ readonly text: string }>) => parts.map((part) => part.text).join("\n")
+
+const escapeSystemUpdateText = (text: string) =>
+  text.replaceAll("&", "&amp;").replaceAll("<", "&lt;").replaceAll(">", "&gt;")
+
+/**
+ * Stable fallback representation for chronological `Message.system(...)`
+ * updates on routes that do not support that privileged role natively. The
+ * wrapper remains visibly lower-authority user text, preserves the original
+ * temporal position, and XML-escapes content so it cannot close the wrapper.
+ */
+export const wrapSystemUpdate = (parts: ReadonlyArray<{ readonly text: string }>) =>
+  `<system-update>\n${escapeSystemUpdateText(joinText(parts))}\n</system-update>`
+
+/**
+ * Chronological system updates deliberately accept text only. Do not insert
+ * raw retrieved, tool, or web content into privileged updates: keep untrusted
+ * data in ordinary user/tool messages instead.
+ */
+export const systemUpdateText = Effect.fn("ProviderShared.systemUpdateText")(function* (
+  route: string,
+  message: LLMRequest["messages"][number],
+) {
+  const content: TextPart[] = []
+  for (const part of message.content) {
+    if (!supportsContent(part, ["text"])) return yield* unsupportedContent(route, "system", ["text"])
+    content.push(part)
+  }
+  return content
+})
+
+/** Lower an unsupported privileged update into visible, in-order user text. */
+export const wrappedSystemUpdate = Effect.fn("ProviderShared.wrappedSystemUpdate")(function* (
+  route: string,
+  message: LLMRequest["messages"][number],
+) {
+  const content = yield* systemUpdateText(route, message)
+  return { type: "text" as const, text: wrapSystemUpdate(content), cache: content.at(-1)?.cache }
+})
+
+/**
+ * Parse the streamed JSON input of a tool call. Treats an empty string as
+ * `"{}"` — providers occasionally finish a tool call without ever emitting
+ * input deltas (e.g. zero-arg tools). The error message is uniform across
+ * routes: `Invalid JSON input for <route> tool call <name>`.
+ */
+export const parseToolInput = (route: string, name: string, raw: string) =>
+  parseJson(route, raw || "{}", `Invalid JSON input for ${route} tool call ${name}`)
+
+/** Inline view or a typed `InvalidRequest` for routes that cannot fetch URLs or dereference provider refs. */
+export const requireInlineMedia = (route: string, asset: Media.Asset): Effect.Effect<Media.Inline, AIError> => {
+  const inline = asset.inline()
+  return inline ? Effect.succeed(inline) : Effect.fail(inlineRequired(route, asset))
+}
+
+export const inlineRequired = (route: string, asset: Media.Asset) =>
+  invalidRequest(
+    `${route} requires inline media (bytes or base64); ${asset.source.type} sources must be materialized first`,
+  )
+
+/** The remote URL of a `url` asset, for protocols that accept `http(s)` references natively. */
+export const mediaUrl = (asset: Media.Asset) => (asset.source.type === "url" ? asset.source.url : undefined)
+
+/**
+ * Lift a tool-result file into a `MediaPart`. Tool files carry either a data URL, an `http(s)` URL, or raw base64 in
+ * `uri`; the declared `mime` wins over any data-URL prefix so tool authors control the type the model sees.
+ */
+export const toolFileMedia = (item: Tool.FileContent): MediaPart => {
+  const parsed = Media.parseDataUrl(item.uri)
+  const asset = parsed
+    ? Media.from({ ...parsed.source, mediaType: item.mime })
+    : /^https?:\/\//.test(item.uri)
+      ? Media.url(item.uri, { mediaType: item.mime })
+      : Media.base64(item.uri, item.mime)
+  return Message.media(asset, { filename: item.name })
+}
+
+export const trimBaseUrl = (value: string) => value.replace(/\/+$/, "")
+
+export const toolResultText = (part: ToolResultPart) => {
+  if (part.result.type === "text") return String(part.result.value)
+  if (part.result.type === "error") {
+    const value = part.result.value
+    const prototype =
+      typeof value === "object" && value !== null && !Array.isArray(value) && Object.getPrototypeOf(value)
+    const structured = Array.isArray(value) || prototype === Object.prototype || prototype === null
+    return structured && isJson(value) ? encodeJson(value) : String(value)
+  }
+  return encodeJson(part.result.value)
+}
+
+export const errorText = (error: unknown) => {
+  if (error instanceof Error) return error.message
+  if (typeof error === "string") return error
+  if (typeof error === "number" || typeof error === "boolean" || typeof error === "bigint") return String(error)
+  if (error === null) return "null"
+  if (error === undefined) return "undefined"
+  return "Unknown stream error"
+}
+
+/**
+ * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
+ * decoder, optionally filters named events, and drops empty events. `[DONE]`
+ * is dropped by default or retained for protocols that use it as their stream
+ * boundary. Retry control events are ignored without interrupting the stream.
+ * Decoder failures become provider output errors so the public error channel
+ * stays `AIError`.
+ */
+export const sseFraming = (
+  bytes: Stream.Stream<Uint8Array, AIError>,
+  events?: ReadonlySet<string>,
+  includeDone = false,
+): Stream.Stream<string, AIError> =>
+  bytes.pipe(
+    Stream.decodeText(),
+    Stream.mapAccumEffect(
+      () => {
+        const output: Sse.Event[] = []
+        return {
+          output,
+          parser: Sse.makeParser((event) => {
+            if (event._tag === "Event") output.push(event)
+          }),
+        }
+      },
+      (state, chunk) =>
+        Effect.gen(function* () {
+          const error = state.parser.feed(chunk)
+          if (error) return yield* eventError("sse", error.message, chunk, error)
+          return [state, state.output.splice(0)] as const
+        }),
+    ),
+    Stream.filter(
+      (event) =>
+        (events === undefined || events.has(event.event)) &&
+        event.data.length > 0 &&
+        (event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
+    ),
+    Stream.map((event) => event.data),
+  )
+
+/**
+ * Canonical invalid-request constructor shared by protocol lowering.
+ */
+export const invalidRequest = (message: string, cause?: unknown) =>
+  new AIError({
+    reason: new InvalidRequestError({ message, cause }),
+  })
+
+/**
+ * Canonical constructor for operations the selected route does not implement.
+ * Prefer this over `invalidRequest` when the failure is a missing route
+ * capability rather than a malformed caller input, so consumers can branch on
+ * `reason._tag` plus `reason.operation` instead of matching message text.
+ */
+export const unsupportedOperation = (input: {
+  readonly operation: string
+  readonly message: string
+  readonly provider?: ProviderID
+  readonly route?: string
+  readonly cause?: unknown
+}) =>
+  new AIError({
+    reason: new UnsupportedOperationError({
+      operation: input.operation,
+      message: input.message,
+      provider: input.provider,
+      route: input.route,
+      cause: input.cause,
+    }),
+  })
+
+/**
+ * Lower namespaces to flat definitions for protocols without a native
+ * namespace construct. Leaf names join their namespace path with `_` because
+ * `.` is not broadly accepted in provider tool names.
+ */
+export const flattenTools = (tools: ReadonlyArray<ToolEntry>, path: ReadonlyArray<string> = []) => {
+  const flat = tools.flatMap((tool): ReadonlyArray<ToolDefinition> => {
+    if (tool.type === "namespace") return flattenTools(tool.tools, [...path, tool.name])
+    if (path.length === 0) return [tool]
+    return [new ToolDefinition({ ...tool, name: [...path, tool.name].join("_") })]
+  })
+  return [...new Map(flat.map((tool) => [tool.name, tool])).values()]
+}
+
+export const flattenToolRequest = (request: LLMRequest) => {
+  const messages = request.messages.map((message) => {
+    const content = message.content.map((part) => {
+      if ((part.type !== "tool-call" && part.type !== "tool-result") || part.namespace === undefined) return part
+      return { ...part, name: `${part.namespace}_${part.name}`, namespace: undefined }
+    })
+    return content.every((part, index) => part === message.content[index])
+      ? message
+      : new Message({ ...message, content })
+  })
+  return {
+    tools: flattenTools(request.tools),
+    request: messages.every((message, index) => message === request.messages[index])
+      ? request
+      : LLMRequest.update(request, { messages }),
+  }
+}
+
+export const matchToolChoice = <Auto, None, Required, Tool>(
+  route: string,
+  toolChoice: NonNullable<LLMRequest["toolChoice"]>,
+  cases: {
+    readonly auto: () => Auto
+    readonly none: () => None
+    readonly required: () => Required
+    readonly tool: (name: string) => Tool
+  },
+) =>
+  Effect.gen(function* () {
+    if (toolChoice.type === "auto") return cases.auto()
+    if (toolChoice.type === "none") return cases.none()
+    if (toolChoice.type === "required") return cases.required()
+    if (!toolChoice.name) return yield* invalidRequest(`${route} tool choice requires a tool name`)
+    return cases.tool(toolChoice.name)
+  })
+
+type ContentType = ContentPart["type"]
+
+const formatContentTypes = (types: ReadonlyArray<ContentType>) => {
+  if (types.length <= 1) return types[0] ?? ""
+  if (types.length === 2) return `${types[0]} and ${types[1]}`
+  return `${types.slice(0, -1).join(", ")}, and ${types.at(-1)}`
+}
+
+export const supportsContent = <const Type extends ContentType>(
+  part: ContentPart,
+  types: ReadonlyArray<Type>,
+): part is Extract<ContentPart, { readonly type: Type }> => (types as ReadonlyArray<ContentType>).includes(part.type)
+
+export const unsupportedContent = (
+  route: string,
+  role: LLMRequest["messages"][number]["role"],
+  types: ReadonlyArray<ContentType>,
+) => invalidRequest(`${route} ${role} messages only support ${formatContentTypes(types)} content for now`)
+
+/**
+ * Build a `validate` step from a Schema decoder. Replaces the per-route
+ * lambda body `(payload) => decode(payload).pipe(Effect.mapError((e) =>
+ * invalid(e.message)))`. Any decode error is translated into
+ * `AIError` carrying the original parse-error message.
+ */
+export const validateWith =
+  <A, I, E extends { readonly message: string }>(decode: (input: I) => Effect.Effect<A, E>) =>
+  (payload: I) =>
+    decode(payload).pipe(Effect.mapError((error) => invalidRequest(error.message, error)))
+
+/**
+ * Build an HTTP POST with a JSON body. Sets `content-type: application/json`
+ * automatically after caller-supplied headers so routes cannot accidentally
+ * send JSON with a stale content type. The body is passed pre-encoded so
+ * routes can choose between
+ * `Schema.encodeSync(payload)` and `ProviderShared.encodeJson(payload)`.
+ */
+export const jsonPost = (input: { readonly url: string; readonly body: string; readonly headers?: Headers.Input }) =>
+  HttpClientRequest.post(input.url).pipe(
+    HttpClientRequest.setHeaders(Headers.set(Headers.fromInput(input.headers), "content-type", "application/json")),
+    HttpClientRequest.bodyText(input.body, "application/json"),
+  )
+
+export * as ProviderShared from "./shared.js"

packages/ai/src/protocols/utils/bedrock-auth.ts

@@ -0,0 +1,146 @@
+import { AwsV4Signer } from "aws4fetch"
+import { Effect } from "effect"
+import { Headers } from "effect/unstable/http"
+import { Auth, type AuthInput } from "../../route/auth.js"
+import { AIError, AuthenticationError } from "../../schema/index.js"
+import { ProviderShared } from "../shared.js"
+
+/**
+ * AWS credentials for SigV4 signing. Bedrock also supports Bearer API key auth,
+ * which provider facades configure as route auth instead of SigV4.
+ */
+export interface Credentials {
+  readonly region: string
+  readonly accessKeyId: string
+  readonly secretAccessKey: string
+  readonly sessionToken?: string
+}
+
+/** Static credentials or an effect resolved before every request. */
+export type CredentialSource = Credentials | Effect.Effect<Credentials, AIError>
+
+export interface DefaultChainOptions {
+  readonly region: string
+  /** Shared config profile passed to the AWS default chain. */
+  readonly profile?: string
+}
+
+/**
+ * Resolve credentials through the AWS default provider chain: environment
+ * variables, shared config and SSO caches, web identity tokens, process
+ * credentials, and container or instance metadata. A fresh chain runs on every
+ * request so credentials rotated on disk without an expiration (for example
+ * shared-config keys rewritten by a corporate SSO tool) are always re-read;
+ * the SDK's own memoization would otherwise pin them for the process lifetime.
+ */
+export const defaultChain = (options: DefaultChainOptions): Effect.Effect<Credentials, AIError> =>
+  Effect.tryPromise({
+    try: async () => {
+      const { fromNodeProviderChain } = await import("@aws-sdk/credential-providers")
+      const identity = await fromNodeProviderChain(options.profile === undefined ? {} : { profile: options.profile })()
+      return {
+        region: options.region,
+        accessKeyId: identity.accessKeyId,
+        secretAccessKey: identity.secretAccessKey,
+        ...(identity.sessionToken === undefined ? {} : { sessionToken: identity.sessionToken }),
+      }
+    },
+    catch: (error) =>
+      new AIError({
+        reason: new AuthenticationError({
+          message: `AWS default credential chain failed: ${ProviderShared.errorText(error)}`,
+          cause: error,
+        }),
+      }),
+  })
+
+const signRequest = (input: {
+  readonly url: string
+  readonly body: string
+  readonly headers: Headers.Headers
+  readonly credentials: Credentials
+  readonly service: string
+  readonly name: string
+}) =>
+  Effect.tryPromise({
+    try: async () => {
+      const signed = await new AwsV4Signer({
+        url: input.url,
+        method: "POST",
+        headers: Object.entries(input.headers),
+        body: input.body,
+        region: input.credentials.region,
+        accessKeyId: input.credentials.accessKeyId,
+        secretAccessKey: input.credentials.secretAccessKey,
+        sessionToken: input.credentials.sessionToken,
+        service: input.service,
+      }).sign()
+      return Object.fromEntries(signed.headers.entries())
+    },
+    catch: (error) =>
+      ProviderShared.invalidRequest(
+        `${input.name} SigV4 signing failed: ${error instanceof Error ? error.message : String(error)}`,
+      ),
+  })
+
+/** Sign the exact JSON bytes with SigV4 using credentials configured on the route. */
+export const sigV4 = (
+  source: CredentialSource | undefined,
+  options: { readonly service?: string; readonly name?: string } = {},
+) =>
+  Auth.custom((input: AuthInput) => {
+    return Effect.gen(function* () {
+      if (!source) {
+        return yield* ProviderShared.invalidRequest(
+          `${options.name ?? "Bedrock Converse"} requires either route bearer auth or AWS credentials configured on the route`,
+        )
+      }
+      const credentials = Effect.isEffect(source) ? yield* source : source
+      const headersForSigning = Headers.set(input.headers, "content-type", "application/json")
+      const signed = yield* signRequest({
+        url: input.url,
+        body: input.body,
+        headers: headersForSigning,
+        credentials,
+        service: options.service ?? "bedrock",
+        name: options.name ?? "Bedrock Converse",
+      })
+      return Headers.setAll(headersForSigning, signed)
+    })
+  })
+
+/** Bedrock route auth defaults to SigV4 and expects credentials from route configuration. */
+export const auth = sigV4(undefined)
+
+export const resolveRegion = (input: {
+  readonly region?: string
+  readonly credentials?: { readonly region: string }
+}) =>
+  input.region ?? input.credentials?.region ?? process.env.AWS_REGION ?? process.env.AWS_DEFAULT_REGION ?? "us-east-1"
+
+export interface ResolveAuthInput {
+  readonly apiKey?: string
+  readonly credentials?: Credentials
+  readonly profile?: string
+}
+
+export interface ResolveAuthOptions {
+  readonly service?: string
+  readonly name?: string
+  /** `sigv4` ignores an ambient `AWS_BEARER_TOKEN_BEDROCK`; `bearer` is validated by the caller. */
+  readonly mode?: "bearer" | "sigv4"
+}
+
+/**
+ * Bearer wins over SigV4 and explicit static credentials win over the default
+ * chain, matching the AWS SDK's own precedence for `AWS_BEARER_TOKEN_BEDROCK`.
+ * The region is applied to the SigV4 scope so it always matches the endpoint host.
+ */
+export const resolveAuth = (input: ResolveAuthInput, region: string, options: ResolveAuthOptions = {}) => {
+  const apiKey = options.mode === "sigv4" ? undefined : (input.apiKey ?? process.env.AWS_BEARER_TOKEN_BEDROCK)
+  if (apiKey !== undefined) return Auth.bearer(apiKey)
+  if (input.credentials !== undefined) return sigV4({ ...input.credentials, region }, options)
+  return sigV4(defaultChain({ region, profile: input.profile }), options)
+}
+
+export * as BedrockAuth from "./bedrock-auth.js"

packages/ai/src/protocols/utils/bedrock-cache.ts

@@ -0,0 +1,58 @@
+import { Schema } from "effect"
+import type { CacheHint } from "../../schema/index.js"
+import { newBreakpoints, ttlBucket } from "./cache.js"
+
+// Bedrock cache markers are positional: emit a `cachePoint` block immediately
+// after the content the caller wants treated as a cacheable prefix. Bedrock
+// accepts optional `ttl: "5m" | "1h"` on cachePoint.
+export const CachePointBlock = Schema.Struct({
+  cachePoint: Schema.Struct({
+    type: Schema.tag("default"),
+    ttl: Schema.optional(Schema.Literals(["5m", "1h"])),
+  }),
+})
+export type CachePointBlock = Schema.Schema.Type<typeof CachePointBlock>
+
+const LEGACY_CLAUDE = ["anthropic.claude-instant", "anthropic.claude-v1", "anthropic.claude-v2", "anthropic.claude-3-"]
+
+// These legacy Claude releases support explicit caching, but only for five minutes.
+const CLAUDE_5M = [
+  "anthropic.claude-3-5-sonnet-20241022-v2:0",
+  "anthropic.claude-3-5-haiku-20241022-v1:0",
+  "anthropic.claude-3-7-sonnet-20250219-v1:0",
+  "anthropic.claude-sonnet-4-20250514-v1:0",
+  "anthropic.claude-opus-4-20250514-v1:0",
+  "anthropic.claude-opus-4-1-20250805-v1:0",
+]
+
+// Callers share the four-breakpoint budget across system, messages, and tools.
+export const BEDROCK_BREAKPOINT_CAP = 4
+
+export const breakpoints = (modelID: string) => {
+  // Substring matching also handles regional prefixes and model-bearing ARNs.
+  const short = CLAUDE_5M.some((id) => modelID.includes(id))
+  return {
+    ...newBreakpoints(BEDROCK_BREAKPOINT_CAP),
+    // Assume modern Claude releases retain caching support; older generations need an explicit exception.
+    // Other model families use implicit caching where available.
+    supported: modelID.includes("anthropic.claude-") && (short || !LEGACY_CLAUDE.some((id) => modelID.includes(id))),
+    ttl1h: !short,
+  }
+}
+export type Breakpoints = ReturnType<typeof breakpoints>
+
+const DEFAULT_5M: CachePointBlock = { cachePoint: { type: "default" } }
+const DEFAULT_1H: CachePointBlock = { cachePoint: { type: "default", ttl: "1h" } }
+
+export const block = (breakpoints: Breakpoints, cache: CacheHint | undefined): CachePointBlock | undefined => {
+  if (!breakpoints.supported) return undefined
+  if (cache?.type !== "ephemeral" && cache?.type !== "persistent") return undefined
+  if (breakpoints.remaining <= 0) {
+    breakpoints.dropped += 1
+    return undefined
+  }
+  breakpoints.remaining -= 1
+  return breakpoints.ttl1h && ttlBucket(cache.ttlSeconds) === "1h" ? DEFAULT_1H : DEFAULT_5M
+}
+
+export * as BedrockCache from "./bedrock-cache.js"

packages/ai/src/protocols/utils/bedrock-media.ts

@@ -0,0 +1,120 @@
+import { Effect, Encoding, Schema } from "effect"
+import type { MediaPart } from "../../schema/index.js"
+import { ProviderShared } from "../shared.js"
+
+// Bedrock Converse accepts image `format` as the file extension and
+// `source.bytes` as base64 in the JSON wire format.
+export const ImageFormat = Schema.Literals(["png", "jpeg", "gif", "webp"])
+export type ImageFormat = Schema.Schema.Type<typeof ImageFormat>
+
+export const ImageBlock = Schema.Struct({
+  image: Schema.Struct({
+    format: ImageFormat,
+    source: Schema.Struct({ bytes: Schema.String }),
+  }),
+})
+export type ImageBlock = Schema.Schema.Type<typeof ImageBlock>
+
+// Bedrock document blocks require a user-facing name so the model can refer to
+// the uploaded document.
+export const DocumentFormat = Schema.Literals(["pdf", "csv", "doc", "docx", "xls", "xlsx", "html", "txt", "md"])
+export type DocumentFormat = Schema.Schema.Type<typeof DocumentFormat>
+
+export const DocumentBlock = Schema.Struct({
+  document: Schema.Struct({
+    format: DocumentFormat,
+    name: Schema.String,
+    source: Schema.Struct({ bytes: Schema.String }),
+  }),
+})
+export type DocumentBlock = Schema.Schema.Type<typeof DocumentBlock>
+
+const IMAGE_FORMATS = {
+  "image/png": "png",
+  "image/jpeg": "jpeg",
+  "image/jpg": "jpeg",
+  "image/gif": "gif",
+  "image/webp": "webp",
+} as const satisfies Record<string, ImageFormat>
+
+const DOCUMENT_FORMATS = {
+  "application/pdf": "pdf",
+  "text/csv": "csv",
+  "application/msword": "doc",
+  "application/vnd.openxmlformats-officedocument.wordprocessingml.document": "docx",
+  "application/vnd.ms-excel": "xls",
+  "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet": "xlsx",
+  "text/html": "html",
+  "text/plain": "txt",
+  "text/markdown": "md",
+} as const satisfies Record<string, DocumentFormat>
+
+const documentBlock = (name: string, format: DocumentFormat, bytes: string): DocumentBlock => ({
+  document: {
+    format,
+    name,
+    source: { bytes },
+  },
+})
+
+function documentName(filename: string | undefined, names: Set<string>) {
+  const base =
+    (filename ?? "")
+      .replace(/\.[^.]*$/, "")
+      .replace(/[^a-zA-Z0-9 ()[\]-]/g, " ")
+      .replace(/\s+/g, " ")
+      .trim()
+      .slice(0, 200)
+      .trim() || "document"
+  let name = base
+  // Converse requires labels to be unique across the entire request, including tool results.
+  for (let index = 2; names.has(name); index++) {
+    const suffix = ` ${index}`
+    name = `${base.slice(0, 200 - suffix.length).trimEnd()}${suffix}`
+  }
+  names.add(name)
+  return name
+}
+
+const mediaBase64 = Effect.fn("BedrockMedia.mediaBase64")(function* (part: MediaPart) {
+  const media = yield* ProviderShared.requireInlineMedia("Bedrock Converse", part.media)
+  const bytes = yield* Effect.fromResult(Encoding.decodeBase64(media.base64)).pipe(
+    Effect.mapError((cause) =>
+      ProviderShared.invalidRequest("Bedrock Converse media data must be valid base64", cause),
+    ),
+  )
+  return Encoding.encodeBase64(bytes)
+})
+
+// Route by MIME. Known image/document formats lower into a typed block; anything
+// else fails with a clear error instead of silently degrading to a malformed
+// document block. Image MIME types not in `IMAGE_FORMATS` (e.g. `image/svg+xml`)
+// get an image-specific error so the caller knows it's a format-support issue,
+// not a kind-detection issue.
+export const lower = Effect.fn("BedrockMedia.lower")(function* (part: MediaPart, documentNames: Set<string>) {
+  const mime = part.media.mediaType.toLowerCase()
+  const imageFormat = IMAGE_FORMATS[mime as keyof typeof IMAGE_FORMATS]
+  if (imageFormat) {
+    return [{ image: { format: imageFormat, source: { bytes: yield* mediaBase64(part) } } } satisfies ImageBlock]
+  }
+  if (mime.startsWith("image/"))
+    return yield* ProviderShared.invalidRequest(
+      `Bedrock Converse does not support image media type ${part.media.mediaType}`,
+    )
+  const documentFormat = DOCUMENT_FORMATS[mime as keyof typeof DOCUMENT_FORMATS]
+  if (documentFormat) {
+    const name = documentName(part.filename, documentNames)
+    const block = documentBlock(name, documentFormat, yield* mediaBase64(part))
+    return part.filename !== undefined && part.filename !== name
+      ? [
+          {
+            text: `Attached file ${ProviderShared.encodeJson(part.filename)} has document label ${ProviderShared.encodeJson(name)}.`,
+          },
+          block,
+        ]
+      : [block]
+  }
+  return yield* ProviderShared.invalidRequest(`Bedrock Converse does not support media type ${part.media.mediaType}`)
+})
+
+export * as BedrockMedia from "./bedrock-media.js"

packages/ai/src/protocols/utils/cache.ts

@@ -0,0 +1,13 @@
+// Shared counter and TTL mapping for provider cache-marker lowering.
+
+export interface Breakpoints {
+  remaining: number
+  dropped: number
+}
+
+export const newBreakpoints = (cap: number): Breakpoints => ({ remaining: cap, dropped: 0 })
+
+// Requests of at least one hour use the explicit `"1h"` bucket; shorter
+// requests omit the wire TTL and use the provider default.
+export const ttlBucket = (ttlSeconds: number | undefined): "1h" | undefined =>
+  ttlSeconds !== undefined && ttlSeconds >= 3600 ? "1h" : undefined