fix(ai): keep chat reasoning in one lifecycle

Co-authored-by: rekram1-node <rekram1-node@users.noreply.github.com>

.changeset/calm-moves-cache.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/core": patch
----
-
-Preserve prompt cache prefixes when sessions move between locations with unchanged instructions.

.changeset/device-authorization-url.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/core": patch
----
-
-Fix OpenCode Console device authorization URLs when the server returns an origin-rooted verification path.

.changeset/idempotent-prompt-reuse.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/core": patch
----
-
-Prompt and synthetic inbox ID reuse is now idempotent: reusing an ID within the same Session succeeds and returns the first admission, ignoring the retried payload, metadata, and delivery mode. Previously reuse with a differing payload failed with a conflict. Cross-Session and cross-type reuse still fail, and control items keep their operation-specific conflict behavior.

.changeset/nested-instruction-reinjection.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/core": patch
----
-
-Nested AGENTS.md instructions are re-injected after compaction. Previously the in-memory dedup claim outlived the synthetic message that compaction dropped from model-visible history, so nested instructions were silently lost for the rest of the process lifetime. The claim now only guards in-flight loads; the synthetic message metadata in durable history is the sole lasting ledger, so any history truncation (compaction, revert) self-heals on the next read in that subtree.

.changeset/promise-plugin-adapter.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/plugin": patch
----
-
-Derive Promise plugin API request and response conversion from the canonical protocol schemas.

.changeset/quiet-forms-replace-questions.md

@@ -1,8 +0,0 @@
----
-"@opencode-ai/core": minor
-"@opencode-ai/schema": minor
-"@opencode-ai/protocol": minor
-"@opencode-ai/client": minor
----
-
-Remove the unused question request API and use session forms for question tool interactions.

.changeset/share-session-model-request.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/core": patch
----
-
-Apply shared Session model-request preparation to transient generation.

.changeset/shell-permission-scan.md

@@ -1,9 +0,0 @@
----
-"@opencode-ai/core": minor
-"@opencode-ai/schema": minor
-"@opencode-ai/protocol": minor
-"@opencode-ai/client": minor
----
-
-Add an opt-in portable shell permission scanner. Opaque commands use normal shell authorization without inferring
-external directories, while the default tree-sitter path remains unchanged.

.changeset/steer-resume-simplified.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/core": patch
----
-
-Simplify interrupt continuation: the steer-scoped resume decision now lives in SessionExecution as a post-cleanup inbox check, and the run coordinator drops its continuation state machine. Wakes arriving during cancellation cleanup now restart a normal full drain, and interrupting an idle session with continue now resumes pending steering input. Recovery-applied moves now end with the same full wake as inbox-admitted moves, retrying any stranded inbox work at the new location. Interrupting with continue now also resumes a next-in-line control item: between-turn manual compaction and moves run under any drain scope, while queued prompts remain parked.

.changeset/terminal-form-recovery.md

@@ -1,6 +0,0 @@
----
-"@opencode-ai/client": patch
-"@opencode-ai/plugin": patch
----
-
-Add form reply and cancellation operations that reconcile terminal forms in the local TUI projection.

.changeset/unify-request-envelope.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/core": patch
----
-
-Title generation and compaction summaries now build their model requests through the shared session request boundary, gaining unsupported-media filtering and image bounds while explicitly opting out of session context hooks: plugins that shape the agent conversation do not observe title or compaction requests. Title requests gain the fork-aware session prompt cache key, and compaction summaries in forked sessions reuse the fork root's prompt cache key instead of the fork's own.

.changeset/websocket-request-hooks.md

@@ -1,6 +0,0 @@
----
-"@opencode-ai/plugin": patch
-"@opencode-ai/core": patch
----
-
-Add transport-neutral Session model request hooks and provider-scoped hook registration so eligible OpenAI Responses requests can prefer WebSocket without bypassing HTTP-only middleware.

.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/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" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          bun x turbo@2.10.2 ls --affected --filter=@opencode-ai/app --output=json > affected.json
+          bun -e 'const result = await Bun.file("affected.json").json(); console.log(`app=${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
@@ -80,15 +111,35 @@ 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-ai/sdk
+            exit 0
+          fi
+          bun turbo verify:package --affected --filter=@opencode-ai/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
@@ -127,7 +178,7 @@ jobs:
 
   e2e:
     name: e2e (${{ matrix.settings.name }})
-    if: github.ref_name != 'v2' && github.head_ref != 'v2'
+    needs: affected
     strategy:
       fail-fast: false
       matrix:
@@ -138,32 +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:
@@ -171,23 +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 }}

.oxlintrc.json

@@ -41,11 +41,5 @@
     // Warn when spreading non-plain objects (Headers, class instances, etc.)
     "typescript/no-misused-spread": "warn"
   },
-  "options": {
-    "typeAware": true
-  },
-  "options": {
-    "typeAware": true
-  },
   "ignorePatterns": ["**/node_modules", "**/dist", "**/.build", "**/.sst", "**/*.d.ts", "**/sdk.gen.ts"]
 }

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", "@opencode-ai/sdk", "@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", "blume"]
+minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@brendonovich/vite-plugin-opencode", "@opencode-ai/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"

infra/app.ts

@@ -63,7 +63,8 @@ new sst.cloudflare.StaticSite("WebApp", {
   domain: "app." + domain,
   path: "packages/app",
   build: {
-    command: "bun turbo build",
+    // Preserve Sentry credentials and run source-map uploads on every deployment.
+    command: "bun run build",
     output: "./dist",
   },
 })

nix/desktop.nix

@@ -4,7 +4,7 @@
   bun,
   nodejs,
   darwin,
-  electron_41,
+  callPackage,
   makeWrapper,
   writableTmpDirAsHomeHook,
   autoPatchelfHook,
@@ -13,7 +13,7 @@
   opencode,
 }:
 let
-  electron = electron_41;
+  electron = callPackage ./electron.nix { };
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "opencode-desktop";
@@ -48,7 +48,7 @@ stdenv.mkDerivation (finalAttrs: {
     desktopName = "OpenCode";
     exec = "opencode-desktop %U";
     icon = "ai.opencode.desktop";
-    # Electron 41 derives X11 WM_CLASS from app.name.
+    # Electron derives X11 WM_CLASS from app.name.
     startupWMClass = "OpenCode";
     categories = [ "Development" ];
   });
@@ -67,12 +67,10 @@ stdenv.mkDerivation (finalAttrs: {
     # https://github.com/electron/electron/issues/31121
     # mac builds use a .app bundle which doesnt have this issue
     + lib.optionalString stdenv.isLinux ''
-      BASE_PATH=packages/desktop
-      FILES=(src/main/windows.ts)
-      for file in "''${FILES[@]}"; do
-        substituteInPlace $BASE_PATH/$file \
-          --replace-fail "process.resourcesPath" "'$out/opt/opencode-desktop/resources'"
-      done
+      substituteInPlace \
+        packages/desktop/src/main/windows/appearance.ts \
+        packages/desktop/src/main/service/desktop-cli.ts \
+        --replace-fail "process.resourcesPath" "'$out/opt/opencode-desktop/resources'"
     '';
 
   preBuild = ''

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-LvDHCOm8OAZfvb0I0L6AbdOevRoQmEJnnqrSgAiNHv8=",
-    "aarch64-linux": "sha256-O0L0iHjb4cwl9xWHIna8VFHyQzoAKzfY8oMpVNayMOg=",
-    "aarch64-darwin": "sha256-ETP8FE71NqufYDUbR7tBdsMEOVQ44wLmsZBeZiSRBRY=",
-    "x86_64-darwin": "sha256-WUcoLldDriT3QxcdlnBQhuPrxDNub0EDvvZXk/pDMpY="
+    "x86_64-linux": "sha256-EtUp4pHl9TyPtRrLGvk/X7kd2LuIxNxCpUwF5aLtzN4=",
+    "aarch64-linux": "sha256-m0j/pMZCguclR3/T9JmzCfi11YzmIvBFyR2bVhIO37Y=",
+    "aarch64-darwin": "sha256-nqefk68ZTUfNU15q1WkXaGsFzPNwOjCtMHpp6WrpNqM=",
+    "x86_64-darwin": "sha256-syD7hX62E4yCDV/wux1QKw4q/zZr24f99Y2mmzMJo6o="
   }
 }

package.json

@@ -39,24 +39,24 @@
       "packages/stats/*"
     ],
     "catalog": {
-      "@effect/opentelemetry": "4.0.0-rc.111",
-      "@effect/platform-node": "4.0.0-rc.111",
-      "@effect/platform-node-shared": "4.0.0-rc.111",
-      "@effect/sql-sqlite-bun": "4.0.0-rc.111",
+      "@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/cross-spawn": "6.0.6",
       "@octokit/rest": "22.0.0",
       "@hono/standard-validator": "0.2.0",
       "@hono/zod-validator": "0.4.2",
-      "@opentui/core": "0.5.7",
-      "@opentui/keymap": "0.5.7",
-      "@opentui/solid": "0.5.7",
+      "@opentui/core": "0.5.9",
+      "@opentui/keymap": "0.5.9",
+      "@opentui/solid": "0.5.9",
       "@tanstack/solid-virtual": "3.13.37",
-      "@shikijs/stream": "4.2.0",
+      "@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",
@@ -67,13 +67,16 @@
       "@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",
+      "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.111",
+      "effect": "4.0.0-rc.112",
       "ai": "6.0.168",
       "cross-spawn": "7.0.6",
       "hono": "4.10.7",
@@ -81,9 +84,9 @@
       "fuzzysort": "3.1.0",
       "get-east-asian-width": "1.6.0",
       "luxon": "3.6.1",
-      "marked": "18.0.7",
+      "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",
@@ -92,17 +95,17 @@
       "remeda": "2.26.0",
       "resolve.exports": "2.0.3",
       "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"
@@ -158,7 +161,8 @@
     "@effect/platform-node-shared": "catalog:",
     "@types/bun": "catalog:",
     "@types/node": "catalog:",
-    "effect": "catalog:"
+    "effect": "catalog:",
+    "solid-js": "catalog:"
   },
   "patchedDependencies": {
     "@ai-sdk/openai-compatible@2.0.41": "patches/@ai-sdk%2Fopenai-compatible@2.0.41.patch",
@@ -166,7 +170,7 @@
     "@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",
+    "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",

packages/ai/AGENTS.md

@@ -19,6 +19,14 @@ Per-type constructors live on the type, not as top-level re-exports. Use `Messag
 - 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.

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

@@ -157,9 +157,9 @@ 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",
@@ -200,8 +200,8 @@ 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)),
   },

packages/ai/src/cache-policy.ts

@@ -36,7 +36,12 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // 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", "bedrock-converse", "openrouter"])
+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" })

packages/ai/src/image-client.ts

@@ -1,7 +1,8 @@
 import { Context, Effect, Layer } from "effect"
 import { RequestExecutor } from "./route/executor.js"
+import { mergeHttpOptions, type AIError } from "./schema/index.js"
+import { sanitizeSurrogates } from "./utils/sanitize.js"
 import type { ImageOptions, ImageRequest, ImageRequestFor, ImageResponse } from "./image.js"
-import type { AIError } from "./schema/index.js"
 
 export type Execute = RequestExecutor.Interface["execute"]
 
@@ -26,7 +27,18 @@ export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer
   Effect.gen(function* () {
     const executor = yield* RequestExecutor.Service
     return Service.of({
-      generate: (request) => request.model.route.generate(request, executor.execute),
+      generate: (request) =>
+        request.model.route.generate(
+          {
+            ...sanitizeSurrogates({
+              ...request,
+              model: undefined,
+              http: mergeHttpOptions(request.model.http, request.http),
+            }),
+            model: request.model,
+          },
+          executor.execute,
+        ),
     })
   }),
 )

packages/ai/src/image.ts

@@ -1,7 +1,7 @@
 import { Effect, Schema } from "effect"
 import {
   HttpOptions,
-  InvalidRequestReason,
+  InvalidRequestError,
   AIError,
   ModelID,
   ProviderID,
@@ -158,9 +158,10 @@ export function generate(input: ImageRequest | ImageRequestInput) {
     try: () => (input instanceof ImageRequest ? input : request(input)),
     catch: (error) =>
       new AIError({
-        module: "Image",
-        method: "generate",
-        reason: new InvalidRequestReason({ message: error instanceof Error ? error.message : String(error) }),
+        reason: new InvalidRequestError({
+          message: error instanceof Error ? error.message : String(error),
+          cause: error,
+        }),
       }),
   }).pipe(Effect.flatMap((request) => ImageClient.generate(request as unknown as ImageRequestFor<ImageOptions>)))
 }

packages/ai/src/llm.ts

@@ -3,7 +3,7 @@ import { LLMClient, Service } from "./route/client.js"
 import {
   GenerationOptions,
   HttpOptions,
-  InvalidProviderOutputReason,
+  InvalidProviderOutputError,
   AIError,
   LLMEvent,
   LLMRequest,
@@ -116,9 +116,7 @@ const runGenerateObject = Effect.fn("LLM.generateObject")(function* (
   )
   if (!call || !LLMEvent.is.toolCall(call))
     return yield* new AIError({
-      module: "LLM",
-      method: "generateObject",
-      reason: new InvalidProviderOutputReason({
+      reason: new InvalidProviderOutputError({
         message: `generateObject: model did not call the forced \`${GENERATE_OBJECT_TOOL_NAME}\` tool`,
       }),
     })
@@ -126,10 +124,9 @@ const runGenerateObject = Effect.fn("LLM.generateObject")(function* (
     Effect.mapError(
       (error) =>
         new AIError({
-          module: "LLM",
-          method: "generateObject",
-          reason: new InvalidProviderOutputReason({
+          reason: new InvalidProviderOutputError({
             message: `generateObject: tool input failed schema decode: ${error.message}`,
+            cause: error,
           }),
         }),
     ),

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

@@ -1,5 +1,5 @@
 import { Buffer } from "node:buffer"
-import { Effect, Schema } from "effect"
+import { Effect, Option, Schema } from "effect"
 import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client.js"
 import { Auth } from "../route/auth.js"
@@ -69,14 +69,22 @@ export interface OptionsInput {
   // SDK Metadata:2649 {user_id?: string | null}
   readonly metadata?: { readonly user_id?: string | null }
   // SDK MessageCreateParamsContainer:2596 ContainerParams|string
-  readonly container?: string | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
+  readonly container?:
+    | string
+    | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
   readonly inference_geo?: string | null
   readonly inferenceGeo?: string | null
   readonly cache_control?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
   readonly cacheControl?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
   // SDK OutputConfig:2684 {effort, format: JSONOutputFormat}
-  readonly output_config?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
-  readonly outputConfig?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
+  readonly output_config?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
+  readonly outputConfig?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
 }
 
 export type ProviderOptionsInput = OptionsInput
@@ -157,7 +165,7 @@ type AnthropicDocumentBlock = Schema.Schema.Type<typeof AnthropicDocumentBlock>
 const AnthropicThinkingBlock = Schema.Struct({
   type: Schema.tag("thinking"),
   thinking: Schema.String,
-  signature: Schema.optional(Schema.String),
+  signature: Schema.String,
   cache_control: Schema.optional(AnthropicCacheControl),
 })
 
@@ -259,7 +267,11 @@ const AnthropicToolChoice = Schema.Union([
     type: Schema.Literals(["auto", "any", "none"]),
     disable_parallel_tool_use: Schema.optional(Schema.Boolean),
   }),
-  Schema.Struct({ type: Schema.tag("tool"), name: Schema.String, disable_parallel_tool_use: Schema.optional(Schema.Boolean) }),
+  Schema.Struct({
+    type: Schema.tag("tool"),
+    name: Schema.String,
+    disable_parallel_tool_use: Schema.optional(Schema.Boolean),
+  }),
 ])
 
 const AnthropicThinking = Schema.Union([
@@ -361,6 +373,8 @@ const AnthropicStreamBlock = Schema.Struct({
   tool_use_id: Schema.optional(Schema.String),
   content: Schema.optional(Schema.Unknown),
 })
+type AnthropicStreamBlock = Schema.Schema.Type<typeof AnthropicStreamBlock>
+const decodeAnthropicStreamBlock = Schema.decodeUnknownOption(AnthropicStreamBlock)
 
 const AnthropicStreamDelta = Schema.Struct({
   type: Schema.optional(Schema.String),
@@ -371,13 +385,15 @@ const AnthropicStreamDelta = Schema.Struct({
   stop_reason: optionalNull(Schema.String),
   stop_sequence: optionalNull(Schema.String),
 })
+type AnthropicStreamDelta = Schema.Schema.Type<typeof AnthropicStreamDelta>
+const decodeAnthropicStreamDelta = Schema.decodeUnknownOption(AnthropicStreamDelta)
 
 const AnthropicEvent = Schema.Struct({
   type: Schema.String,
   index: Schema.optional(Schema.Number),
   message: Schema.optional(Schema.Struct({ usage: Schema.optional(AnthropicUsage) })),
-  content_block: Schema.optional(AnthropicStreamBlock),
-  delta: Schema.optional(AnthropicStreamDelta),
+  content_block: Schema.optional(Schema.Unknown),
+  delta: Schema.optional(Schema.Unknown),
   usage: Schema.optional(AnthropicUsage),
   // `type` and `message` are both required per Anthropic's spec, but
   // OpenAI-compatible proxies and gateway translations occasionally drop one
@@ -390,6 +406,7 @@ const AnthropicEvent = Schema.Struct({
 type AnthropicEvent = Schema.Schema.Type<typeof AnthropicEvent>
 
 interface ParserState {
+  readonly providerMetadataKey: string
   readonly tools: ToolStream.State<number>
   readonly reasoningSignatures: Readonly<Record<number, string>>
   readonly usage?: Usage
@@ -424,18 +441,18 @@ const cacheControl = (breakpoints: Cache.Breakpoints, cache: CacheHint | undefin
   return Cache.ttlBucket(cache.ttlSeconds) === "1h" ? EPHEMERAL_1H : EPHEMERAL_5M
 }
 
-const anthropicMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ anthropic: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
 
-const signatureFromMetadata = (metadata: ProviderMetadata | undefined): string | undefined => {
-  const anthropic = metadata?.anthropic
-  if (!ProviderShared.isRecord(anthropic)) return undefined
-  return typeof anthropic.signature === "string" ? anthropic.signature : undefined
+const signatureFromMetadata = (metadata: ProviderMetadata | undefined, key: string): string | undefined => {
+  const provider = metadata?.[key]
+  if (!ProviderShared.isRecord(provider)) return undefined
+  return typeof provider.signature === "string" ? provider.signature : undefined
 }
 
-const redactedDataFromMetadata = (metadata: ProviderMetadata | undefined): string | undefined => {
-  const anthropic = metadata?.anthropic
-  if (!ProviderShared.isRecord(anthropic)) return undefined
-  return typeof anthropic.redactedData === "string" ? anthropic.redactedData : undefined
+const redactedDataFromMetadata = (metadata: ProviderMetadata | undefined, key: string): string | undefined => {
+  const provider = metadata?.[key]
+  if (!ProviderShared.isRecord(provider)) return undefined
+  return typeof provider.redactedData === "string" ? provider.redactedData : undefined
 }
 
 const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition, inputSchema: JsonSchema): AnthropicTool => ({
@@ -495,14 +512,21 @@ const serverToolResultType = (name: string): AnthropicServerToolResultType | und
   return undefined
 }
 
-const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (part: ToolResultPart) {
+const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (
+  part: ToolResultPart,
+  providerMetadataKey: string,
+) {
   const wireType = serverToolResultType(part.name)
   if (!wireType)
     return yield* invalid(`Anthropic Messages does not know how to round-trip server tool result for ${part.name}`)
   // Prefer the provider-owned replay payload; fall back to the result value for
   // histories constructed directly from provider events.
-  const payload = part.providerMetadata?.anthropic?.["result"] ?? part.result.value
-  return { type: wireType, tool_use_id: scrubToolCallID(part.id), content: payload } satisfies AnthropicServerToolResultBlock
+  const payload = part.providerMetadata?.[providerMetadataKey]?.["result"] ?? part.result.value
+  return {
+    type: wireType,
+    tool_use_id: scrubToolCallID(part.id),
+    content: payload,
+  } satisfies AnthropicServerToolResultBlock
 })
 
 const fileIdFromMetadata = (metadata: MediaPart["metadata"]): string | undefined => {
@@ -550,9 +574,7 @@ const documentContextFromMetadata = (metadata: MediaPart["metadata"]): string |
   return undefined
 }
 
-const citationsFromMetadata = (
-  metadata: MediaPart["metadata"],
-): AnthropicDocumentBlock["citations"] | undefined => {
+const citationsFromMetadata = (metadata: MediaPart["metadata"]): AnthropicDocumentBlock["citations"] | undefined => {
   if (!ProviderShared.isRecord(metadata)) return undefined
   const raw = ProviderShared.isRecord(metadata.anthropic)
     ? (metadata.anthropic.citations ?? metadata.citations)
@@ -701,6 +723,25 @@ const lowerToolResultContent = Effect.fnUntraced(function* (part: ToolResultPart
   return yield* Effect.forEach(content, lowerToolResultContentItem)
 })
 
+const requireThinkingSignature = (request: LLMRequest) => {
+  if (request.model.compatibility?.requireSignature !== undefined) return request.model.compatibility.requireSignature
+  const provider = request.model.provider.toLowerCase()
+  const model = request.model.id.toLowerCase()
+  const baseURL = (request.model.route.endpoint.baseURL ?? "").toLowerCase()
+  if (
+    provider === "kimi-for-coding" ||
+    provider === "moonshotai" ||
+    provider === "moonshotai-cn" ||
+    model.startsWith("kimi-") ||
+    baseURL.includes("api.kimi.com/coding") ||
+    baseURL.includes("api.moonshot.ai/anthropic") ||
+    baseURL.includes("api.moonshot.cn/anthropic")
+  )
+    return false
+  if (provider.includes("xiaomi") || model.includes("mimo") || baseURL.includes("xiaomimimo.com")) return false
+  return true
+}
+
 // Mid-conversation system messages became available with Opus 4.8 and version
 // 5 of the other supported Claude families. Treat later family versions as
 // compatible without assuming that every Anthropic Messages model is Claude.
@@ -720,9 +761,12 @@ const endsInServerToolUse = (message: LLMRequest["messages"][number]) => {
   return message.role === "assistant" && last?.type === "tool-call" && last.providerExecuted === true
 }
 
-const canUseNativeSystemUpdate = (messages: LLMRequest["messages"], index: number) => {
-  const previous = messages[index - 1]
-  const next = messages[index + 1]
+const canUseNativeSystemUpdate = (request: LLMRequest, index: number) => {
+  const previous = request.messages[index - 1]
+  const next = request.messages[index + 1]
+  // Vertex currently rejects/404s for a system message after local tool results,
+  // so fold it into the user tool-result turn across continuations and history.
+  if (request.model.route.id === "google-vertex-messages" && previous?.role === "tool") return false
   return (
     previous !== undefined &&
     previous.role !== "system" &&
@@ -764,12 +808,13 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
   breakpoints: Cache.Breakpoints,
 ) {
   const messages: AnthropicMessage[] = []
+  const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
 
   for (const [index, message] of request.messages.entries()) {
     if (message.role === "system") {
       if (splitsLocalToolResults(request.messages, index))
         return yield* invalid("Anthropic Messages system updates cannot split a local tool call from its tool result")
-      if (supportsNativeSystemUpdates(request) && canUseNativeSystemUpdate(request.messages, index)) {
+      if (supportsNativeSystemUpdates(request) && canUseNativeSystemUpdate(request, index)) {
         messages.push(yield* lowerNativeSystemUpdate(message, breakpoints))
         continue
       }
@@ -807,15 +852,30 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
           continue
         }
         if (part.type === "reasoning") {
-          // Mirrors Vercel's @ai-sdk/anthropic: a signature marks visible
-          // thinking; only signature-less parts carrying redactedData
-          // round-trip as opaque redacted_thinking blocks.
-          const signature = part.encrypted ?? signatureFromMetadata(part.providerMetadata)
-          const redactedData = redactedDataFromMetadata(part.providerMetadata)
+          // A signature marks visible thinking; only signature-less parts carrying
+          // redactedData round-trip as opaque redacted_thinking blocks.
+          const signature = part.encrypted ?? signatureFromMetadata(part.providerMetadata, providerMetadataKey)
+          const redactedData = redactedDataFromMetadata(part.providerMetadata, providerMetadataKey)
           if (signature === undefined && redactedData !== undefined) {
             content.push({ type: "redacted_thinking", data: redactedData })
             continue
           }
+          if (typeof signature !== "string" || signature.trim().length === 0) {
+            if (part.text.trim().length === 0) continue
+            if (!requireThinkingSignature(request)) {
+              content.push({ type: "thinking", thinking: part.text, signature: "" })
+              continue
+            }
+            // Without a signature this cannot be a valid thinking block per
+            // the SDK ThinkingBlockParam:3217 — demote to text so the
+            // conversation remains sendable.
+            content.push({
+              type: "text",
+              text: part.text,
+              cache_control: cacheControl(breakpoints, part.cache),
+            })
+            continue
+          }
           content.push({ type: "thinking", thinking: part.text, signature })
           continue
         }
@@ -824,7 +884,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
           continue
         }
         if (part.type === "tool-result" && part.providerExecuted) {
-          content.push(yield* lowerServerToolResult(part))
+          content.push(yield* lowerServerToolResult(part, providerMetadataKey))
           continue
         }
         return yield* invalid(
@@ -858,21 +918,24 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
 
 const resolveOptions = Effect.fn("AnthropicMessages.resolveOptions")(function* (request: LLMRequest) {
   const input = request.providerOptions as Record<string, unknown> | undefined
-  const rawServiceTier = (input as Record<string, unknown> | undefined)?.service_tier ?? (input as Record<string, unknown> | undefined)?.serviceTier
+  const rawServiceTier =
+    (input as Record<string, unknown> | undefined)?.service_tier ??
+    (input as Record<string, unknown> | undefined)?.serviceTier
   const service_tier =
     rawServiceTier === "auto" || rawServiceTier === "standard_only"
       ? (rawServiceTier as "auto" | "standard_only")
       : undefined
   const rawMetadata = (input as Record<string, unknown> | undefined)?.metadata
   const metadata =
-    ProviderShared.isRecord(rawMetadata) &&
-    (typeof rawMetadata.user_id === "string" || rawMetadata.user_id === null)
+    ProviderShared.isRecord(rawMetadata) && (typeof rawMetadata.user_id === "string" || rawMetadata.user_id === null)
       ? { user_id: rawMetadata.user_id as string | null }
       : undefined
   const container =
     typeof (input as Record<string, unknown> | undefined)?.container === "string" ||
     ProviderShared.isRecord((input as Record<string, unknown> | undefined)?.container)
-      ? ((input as Record<string, unknown>).container as string | { id?: string | null; skills?: ReadonlyArray<Record<string, unknown>> | null })
+      ? ((input as Record<string, unknown>).container as
+          | string
+          | { id?: string | null; skills?: ReadonlyArray<Record<string, unknown>> | null })
       : undefined
   const rawInferenceGeo =
     (input as Record<string, unknown> | undefined)?.inference_geo ??
@@ -923,8 +986,7 @@ const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function*
     input.display === "summarized" || input.display === "omitted"
       ? (input.display as "summarized" | "omitted")
       : undefined
-  if (input.type === "adaptive")
-    return { type: "adaptive" as const, ...(display === undefined ? {} : { display }) }
+  if (input.type === "adaptive") return { type: "adaptive" as const, ...(display === undefined ? {} : { display }) }
   if (input.type === "disabled") return { type: "disabled" as const }
   if (input.type !== "enabled") return undefined
   const budget =
@@ -1012,7 +1074,7 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // inclusive `inputTokens` the rest of the contract expects. Extended
 // thinking tokens are included in `output_tokens`; newer responses also
 // expose that subset through `output_tokens_details.thinking_tokens`.
-const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
+const mapUsage = (usage: AnthropicUsage | undefined, providerMetadataKey: string): Usage | undefined => {
   if (!usage) return undefined
   const nonCached = usage.input_tokens ?? undefined
   const cacheRead = usage.cache_read_input_tokens ?? undefined
@@ -1026,7 +1088,7 @@ const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
     cacheWriteInputTokens: cacheWrite,
     reasoningTokens: usage.output_tokens_details?.thinking_tokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, usage.output_tokens, undefined),
-    providerMetadata: { anthropic: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
   })
 }
 
@@ -1035,7 +1097,7 @@ const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
 // field prefers `right` when defined, falls back to `left`. `inputTokens` is
 // recomputed from the merged breakdown so the inclusive total stays
 // consistent with `nonCached + cacheRead + cacheWrite`.
-const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
+const mergeUsage = (left: Usage | undefined, right: Usage | undefined, providerMetadataKey: string) => {
   if (!left) return right
   if (!right) return left
   const nonCachedInputTokens = right.nonCachedInputTokens ?? left.nonCachedInputTokens
@@ -1053,7 +1115,9 @@ const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
     reasoningTokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, outputTokens, undefined),
     providerMetadata: {
-      anthropic: mergeJsonRecords(left.providerMetadata?.["anthropic"], right.providerMetadata?.["anthropic"]) ?? {},
+      [providerMetadataKey]:
+        mergeJsonRecords(left.providerMetadata?.[providerMetadataKey], right.providerMetadata?.[providerMetadataKey]) ??
+        {},
     },
   })
 }
@@ -1071,7 +1135,7 @@ const SERVER_TOOL_RESULT_NAMES: Record<AnthropicServerToolResultType, string> =
 
 const isServerToolResultType = (type: string): type is AnthropicServerToolResultType => type in SERVER_TOOL_RESULT_NAMES
 
-const serverToolResultEvent = (block: NonNullable<AnthropicEvent["content_block"]>): LLMEvent | undefined => {
+const serverToolResultEvent = (block: AnthropicStreamBlock, providerMetadataKey: string): LLMEvent | undefined => {
   if (!block.type || !isServerToolResultType(block.type)) return undefined
   const errorPayload =
     typeof block.content === "object" && block.content !== null && "type" in block.content
@@ -1085,7 +1149,7 @@ const serverToolResultEvent = (block: NonNullable<AnthropicEvent["content_block"
     providerExecuted: true,
     // The complete payload is irreducible provider replay state: subsequent
     // stateless requests must round-trip the typed result block verbatim.
-    providerMetadata: anthropicMetadata({ blockType: block.type, result: block.content }),
+    providerMetadata: providerMetadata(providerMetadataKey, { blockType: block.type, result: block.content }),
   })
 }
 
@@ -1094,11 +1158,14 @@ type StepResult = readonly [ParserState, ReadonlyArray<LLMEvent>]
 const NO_EVENTS: StepResult["1"] = []
 
 const onMessageStart = (state: ParserState, event: AnthropicEvent): StepResult => {
-  const usage = mapUsage(event.message?.usage)
-  return [usage ? { ...state, usage: mergeUsage(state.usage, usage) } : state, NO_EVENTS]
+  const usage = mapUsage(event.message?.usage, state.providerMetadataKey)
+  return [usage ? { ...state, usage: mergeUsage(state.usage, usage, state.providerMetadataKey) } : state, NO_EVENTS]
 }
 
-const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepResult => {
+const onContentBlockStart = (
+  state: ParserState,
+  event: AnthropicEvent & { readonly content_block: AnthropicStreamBlock },
+): StepResult => {
   const block = event.content_block
   if (!block) return [state, NO_EVENTS]
 
@@ -1144,14 +1211,16 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
   if (block.type === "thinking" && block.thinking !== undefined) {
     const events: LLMEvent[] = []
     const id = `reasoning-${event.index ?? 0}`
-    const providerMetadata =
-      block.signature === undefined ? undefined : anthropicMetadata({ signature: block.signature })
-    const lifecycle = Lifecycle.reasoningStart(state.lifecycle, events, id, providerMetadata)
+    const metadata =
+      block.signature === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { signature: block.signature })
+    const lifecycle = Lifecycle.reasoningStart(state.lifecycle, events, id, metadata)
     return [
       {
         ...state,
         lifecycle: block.thinking
-          ? Lifecycle.reasoningDelta(lifecycle, events, id, block.thinking, providerMetadata)
+          ? Lifecycle.reasoningDelta(lifecycle, events, id, block.thinking, metadata)
           : lifecycle,
         reasoningSignatures:
           event.index === undefined || block.signature === undefined
@@ -1174,14 +1243,14 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
           state.lifecycle,
           events,
           `reasoning-${event.index ?? 0}`,
-          anthropicMetadata({ redactedData: block.data }),
+          providerMetadata(state.providerMetadataKey, { redactedData: block.data }),
         ),
       },
       events,
     ]
   }
 
-  const result = serverToolResultEvent(block)
+  const result = serverToolResultEvent(block, state.providerMetadataKey)
   if (!result) return [state, NO_EVENTS]
   const events: LLMEvent[] = []
   return [{ ...state, lifecycle: Lifecycle.stepStart(state.lifecycle, events) }, [...events, result]]
@@ -1189,11 +1258,12 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
 
 const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(function* (
   state: ParserState,
-  event: AnthropicEvent,
+  event: AnthropicEvent & { readonly delta: AnthropicStreamDelta },
 ) {
   const delta = event.delta
 
   if (delta?.type === "text_delta" && delta.text) {
+    if (!state.lifecycle.text.has(`text-${event.index ?? 0}`)) return [state, NO_EVENTS] satisfies StepResult
     const events: LLMEvent[] = []
     return [
       { ...state, lifecycle: Lifecycle.textDelta(state.lifecycle, events, `text-${event.index ?? 0}`, delta.text) },
@@ -1202,6 +1272,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
   }
 
   if (delta?.type === "thinking_delta" && delta.thinking) {
+    if (!state.lifecycle.reasoning.has(`reasoning-${event.index ?? 0}`)) return [state, NO_EVENTS] satisfies StepResult
     const events: LLMEvent[] = []
     return [
       {
@@ -1214,6 +1285,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
 
   if (delta?.type === "signature_delta" && delta.signature) {
     const index = event.index ?? 0
+    if (!state.lifecycle.reasoning.has(`reasoning-${index}`)) return [state, NO_EVENTS] satisfies StepResult
     return [
       {
         ...state,
@@ -1258,7 +1330,7 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
         Lifecycle.textEnd(state.lifecycle, events, `text-${event.index}`),
         events,
         `reasoning-${event.index}`,
-        signature === undefined ? undefined : anthropicMetadata({ signature }),
+        signature === undefined ? undefined : providerMetadata(state.providerMetadataKey, { signature }),
       )
   events.push(...resultEvents)
   const reasoningSignatures = { ...state.reasoningSignatures }
@@ -1266,8 +1338,11 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
   return [{ ...state, lifecycle, tools: result.tools, reasoningSignatures }, events] satisfies StepResult
 })
 
-const onMessageDelta = (state: ParserState, event: AnthropicEvent): StepResult => {
-  const usage = mergeUsage(state.usage, mapUsage(event.usage))
+const onMessageDelta = (
+  state: ParserState,
+  event: AnthropicEvent & { readonly delta?: AnthropicStreamDelta },
+): StepResult => {
+  const usage = mergeUsage(state.usage, mapUsage(event.usage, state.providerMetadataKey), state.providerMetadataKey)
   return [
     {
       ...state,
@@ -1280,7 +1355,7 @@ const onMessageDelta = (state: ParserState, event: AnthropicEvent): StepResult =
         providerMetadata:
           event.delta?.stop_sequence === null || event.delta?.stop_sequence === undefined
             ? undefined
-            : anthropicMetadata({ stopSequence: event.delta.stop_sequence }),
+            : providerMetadata(state.providerMetadataKey, { stopSequence: event.delta.stop_sequence }),
       },
     },
     NO_EVENTS,
@@ -1312,32 +1387,80 @@ const providerErrorMessage = (event: AnthropicEvent): string => {
   return message || type || "Anthropic Messages stream error"
 }
 
-const onError = (event: AnthropicEvent) =>
-  Effect.fail(
+const onError = (event: AnthropicEvent) => {
+  const message = providerErrorMessage(event)
+  const body = ProviderShared.encodeJson(event)
+  return Effect.fail(
     new AIError({
-      module: ADAPTER,
-      method: "stream",
-      reason: classifyProviderFailure({ message: providerErrorMessage(event), code: event.error?.type }),
+      reason: classifyProviderFailure({ message, rawBody: body }),
     }),
   )
+}
+
+const isKnownStreamBlockType = (type: string) =>
+  type === "text" ||
+  type === "thinking" ||
+  type === "redacted_thinking" ||
+  type === "tool_use" ||
+  type === "server_tool_use" ||
+  isServerToolResultType(type)
+
+const isKnownStreamDeltaType = (type: string) =>
+  type === "text_delta" || type === "thinking_delta" || type === "signature_delta" || type === "input_json_delta"
+
+const invalidStreamEvent = (event: AnthropicEvent) =>
+  Effect.fail(
+    ProviderShared.eventError(
+      ADAPTER,
+      "Invalid anthropic/anthropic-messages stream event",
+      ProviderShared.encodeJson(event),
+    ),
+  )
 
 const step = (state: ParserState, event: AnthropicEvent) => {
+  if (!SSE_EVENTS.has(event.type)) return Effect.succeed<StepResult>([state, NO_EVENTS])
+  if (
+    event.type !== "content_block_start" &&
+    event.content_block !== undefined &&
+    Option.isNone(decodeAnthropicStreamBlock(event.content_block))
+  )
+    return invalidStreamEvent(event)
+  if (
+    event.type !== "content_block_delta" &&
+    event.delta !== undefined &&
+    Option.isNone(decodeAnthropicStreamDelta(event.delta))
+  )
+    return invalidStreamEvent(event)
   if (event.type === "message_start") return Effect.succeed(onMessageStart(state, event))
   if (event.type === "content_block_start") {
-    const block = event.content_block
-    if (block && (block.type === "tool_use" || block.type === "server_tool_use")) {
+    if (!ProviderShared.isRecord(event.content_block) || typeof event.content_block.type !== "string")
+      return invalidStreamEvent(event)
+    if (!isKnownStreamBlockType(event.content_block.type)) return Effect.succeed<StepResult>([state, NO_EVENTS])
+    const decoded = decodeAnthropicStreamBlock(event.content_block)
+    if (Option.isNone(decoded)) return invalidStreamEvent(event)
+    const block = decoded.value
+    if (block.type === "tool_use" || block.type === "server_tool_use") {
       if (event.index === undefined)
         return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic ${block.type} missing index`))
       if (!block.id)
-        return Effect.fail(
-          ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`),
-        )
+        return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`))
     }
-    return Effect.succeed(onContentBlockStart(state, event))
+    return Effect.succeed(onContentBlockStart(state, { ...event, content_block: block }))
+  }
+  if (event.type === "content_block_delta") {
+    if (!ProviderShared.isRecord(event.delta)) return invalidStreamEvent(event)
+    if (typeof event.delta.type === "string" && !isKnownStreamDeltaType(event.delta.type))
+      return Effect.succeed<StepResult>([state, NO_EVENTS])
+    const decoded = decodeAnthropicStreamDelta(event.delta)
+    if (Option.isNone(decoded)) return invalidStreamEvent(event)
+    return onContentBlockDelta(state, { ...event, delta: decoded.value })
   }
-  if (event.type === "content_block_delta") return onContentBlockDelta(state, event)
   if (event.type === "content_block_stop") return onContentBlockStop(state, event)
-  if (event.type === "message_delta") return Effect.succeed(onMessageDelta(state, event))
+  if (event.type === "message_delta") {
+    const decoded = decodeAnthropicStreamDelta(event.delta)
+    if (Option.isNone(decoded)) return invalidStreamEvent(event)
+    return Effect.succeed(onMessageDelta(state, { ...event, delta: decoded.value }))
+  }
   if (event.type === "message_stop") return onMessageStop(state)
   if (event.type === "error") return onError(event)
   return Effect.succeed<StepResult>([state, NO_EVENTS])
@@ -1359,7 +1482,8 @@ export const protocol = Protocol.make({
   },
   stream: {
     event: Protocol.jsonEvent(AnthropicEvent),
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
       tools: ToolStream.empty<number>(),
       reasoningSignatures: {},
       lifecycle: Lifecycle.initial(),
@@ -1373,10 +1497,9 @@ export const route = Route.make({
   provider: "anthropic",
   providerMetadataKey: "anthropic",
   protocol,
-  endpoint: Endpoint.path(
-    (input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH),
-    { baseURL: DEFAULT_BASE_URL },
-  ),
+  endpoint: Endpoint.path((input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH), {
+    baseURL: DEFAULT_BASE_URL,
+  }),
   auth: Auth.none,
   framing,
   headers: () => ({ "anthropic-version": "2023-06-01" }),

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

@@ -212,11 +212,7 @@ const BedrockEvent = Schema.Struct({
       metrics: Schema.optional(Schema.Unknown),
     }),
   ),
-  internalServerException: Schema.optional(BedrockStreamException),
-  modelStreamErrorException: Schema.optional(BedrockStreamException),
-  validationException: Schema.optional(BedrockStreamException),
-  throttlingException: Schema.optional(BedrockStreamException),
-  serviceUnavailableException: Schema.optional(BedrockStreamException),
+  exception: Schema.optional(Schema.Struct({ type: Schema.String, details: BedrockStreamException })),
 })
 type BedrockEvent = Schema.Schema.Type<typeof BedrockEvent>
 
@@ -262,19 +258,21 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
     tool: (name) => ({ tool: { name } }) as const,
   })
 
-const bedrockMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ bedrock: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
 
-const reasoningSignature = (part: ReasoningPart) => {
-  const bedrock = part.providerMetadata?.bedrock
+const reasoningSignature = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
   return (
     part.encrypted ??
-    (ProviderShared.isRecord(bedrock) && typeof bedrock.signature === "string" ? bedrock.signature : undefined)
+    (ProviderShared.isRecord(metadata) && typeof metadata.signature === "string" ? metadata.signature : undefined)
   )
 }
 
-const reasoningRedactedData = (part: ReasoningPart) => {
-  const bedrock = part.providerMetadata?.bedrock
-  return ProviderShared.isRecord(bedrock) && typeof bedrock.redactedData === "string" ? bedrock.redactedData : undefined
+const reasoningRedactedData = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
+  return ProviderShared.isRecord(metadata) && typeof metadata.redactedData === "string"
+    ? metadata.redactedData
+    : undefined
 }
 
 const lowerToolCall = (part: ToolCallPart): BedrockToolUseBlock => ({
@@ -322,6 +320,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
   breakpoints: BedrockCache.Breakpoints,
 ) {
   const messages: BedrockMessage[] = []
+  const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
 
   for (const message of request.messages) {
     if (message.role === "system") {
@@ -369,12 +368,18 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
           continue
         }
         if (part.type === "reasoning") {
-          const signature = reasoningSignature(part)
-          const redactedData = reasoningRedactedData(part)
+          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
         }
@@ -383,7 +388,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
           continue
         }
       }
-      messages.push({ role: "assistant", content })
+      if (content.length > 0) messages.push({ role: "assistant", content })
       continue
     }
 
@@ -470,7 +475,7 @@ const mapFinishReason = (reason: string): FinishReason => {
 
 // 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): Usage | undefined => {
+const mapUsage = (usage: BedrockUsageSchema | undefined, providerMetadataKey: string): Usage | undefined => {
   if (!usage) return undefined
   const inputTokens = ProviderShared.sumTokens(
     usage.inputTokens,
@@ -484,12 +489,14 @@ const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
     cacheReadInputTokens: usage.cacheReadInputTokens,
     cacheWriteInputTokens: usage.cacheWriteInputTokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, usage.outputTokens, usage.totalTokens),
-    providerMetadata: { bedrock: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
   })
 }
 
 interface ParserState {
+  readonly providerMetadataKey: string
   readonly tools: ToolStream.State<number>
+  readonly finishedTools: ReadonlySet<number>
   // Bedrock splits the finish into `messageStop` (carries `stopReason`) and
   // `metadata` (carries usage). Hold the terminal event in state so `onHalt`
   // can emit exactly one finish after both chunks have had a chance to arrive.
@@ -545,20 +552,14 @@ const step = (state: ParserState, event: BedrockEvent) =>
       const reasoning = event.contentBlockDelta.delta.reasoningContent
       const events: LLMEvent[] = []
       const redactedData = reasoning.redactedContent ?? reasoning.data
-      const providerMetadata = reasoning.signature
-        ? bedrockMetadata({ signature: reasoning.signature })
+      const metadata = reasoning.signature
+        ? providerMetadata(state.providerMetadataKey, { signature: reasoning.signature })
         : redactedData !== undefined
-          ? bedrockMetadata({ redactedData })
+          ? providerMetadata(state.providerMetadataKey, { redactedData })
           : undefined
       const lifecycle =
-        reasoning.text !== undefined || providerMetadata !== undefined
-          ? Lifecycle.reasoningDelta(
-              state.lifecycle,
-              events,
-              `reasoning-${index}`,
-              reasoning.text ?? "",
-              providerMetadata,
-            )
+        reasoning.text !== undefined || metadata !== undefined
+          ? Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text ?? "", metadata)
           : state.lifecycle
       return [
         {
@@ -574,6 +575,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
 
     if (event.contentBlockDelta?.delta?.toolUse) {
       const index = event.contentBlockDelta.contentBlockIndex
+      if (state.finishedTools.has(index)) return [state, []] as const
       const result = ToolStream.appendExisting(
         ADAPTER,
         state.tools,
@@ -600,7 +602,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
             events,
             `reasoning-${index}`,
             state.reasoningSignatures[index]
-              ? bedrockMetadata({ signature: state.reasoningSignatures[index] })
+              ? providerMetadata(state.providerMetadataKey, { signature: state.reasoningSignatures[index] })
               : undefined,
           )
       events.push(...resultEvents)
@@ -612,6 +614,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
             state.hasToolCalls,
           lifecycle,
           tools: result.tools,
+          finishedTools: resultEvents.length > 0 ? new Set([...state.finishedTools, index]) : state.finishedTools,
           reasoningSignatures: Object.fromEntries(
             Object.entries(state.reasoningSignatures).filter(([key]) => key !== String(index)),
           ),
@@ -637,7 +640,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
     }
 
     if (event.metadata) {
-      const usage = mapUsage(event.metadata.usage) ?? state.pendingFinish?.usage
+      const usage = mapUsage(event.metadata.usage, state.providerMetadataKey) ?? state.pendingFinish?.usage
       return [
         {
           ...state,
@@ -650,22 +653,14 @@ const step = (state: ParserState, event: BedrockEvent) =>
       ] as const
     }
 
-    const exception = (
-      [
-        ["internalServerException", event.internalServerException],
-        ["modelStreamErrorException", event.modelStreamErrorException],
-        ["serviceUnavailableException", event.serviceUnavailableException],
-        ["throttlingException", event.throttlingException],
-        ["validationException", event.validationException],
-      ] as const
-    ).find((entry) => entry[1] !== undefined)
-    if (exception) {
+    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({
-        module: ADAPTER,
-        method: "stream",
         reason: classifyProviderFailure({
-          message: exception[1]?.message ?? exception[1]?.originalMessage ?? "Bedrock Converse stream error",
-          code: exception[0],
+          message,
+          rawBody: body,
         }),
       })
     }
@@ -708,15 +703,17 @@ export const protocol = Protocol.make({
   },
   stream: {
     event: BedrockEvent,
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
       tools: ToolStream.empty<number>(),
+      finishedTools: new Set<number>(),
       pendingFinish: undefined,
       hasToolCalls: false,
       lifecycle: Lifecycle.initial(),
       reasoningSignatures: {},
     }),
     step,
-    onHalt,
+    onHalt: (state) => Effect.succeed(onHalt(state)),
   },
 })
 

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

@@ -1,6 +1,7 @@
 import { EventStreamCodec } from "@smithy/eventstream-codec"
 import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
-import { Effect, Stream } from "effect"
+import { Effect, Encoding, Stream } from "effect"
+import { AIError, AIErrorReason } from "../schema/index.js"
 import { Framing } from "../route/framing.js"
 import { ProviderShared } from "./shared.js"
 
@@ -49,10 +50,14 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
             `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
@@ -61,6 +66,7 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
           route,
           [code, message].filter((value): value is string => typeof value === "string").join(": ") ||
             "Bedrock Converse event-stream error",
+          body,
         )
       }
       const eventType =
@@ -70,7 +76,6 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
             ? decoded.headers[":exception-type"]?.value
             : undefined
       if (typeof eventType !== "string") continue
-      const payload = utf8.decode(decoded.body)
       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
@@ -80,9 +85,21 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
         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({ [eventType]: parsed })
+      out.push({
+        ...(messageType === "exception"
+          ? { exception: { type: eventType, details: parsed } }
+          : { [eventType]: parsed }),
+        rawBody: body,
+      })
     }
     return [cursor, out] as const
   })
@@ -95,6 +112,7 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
  */
 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.mapAccumEffect(() => initialFrameBuffer, consumeFrames(route))),
 })
 

packages/ai/src/protocols/gemini.ts

@@ -1,4 +1,4 @@
-import { Effect, Schema } from "effect"
+import { Effect, Option, Schema } from "effect"
 import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client.js"
 import { Auth } from "../route/auth.js"
@@ -6,6 +6,7 @@ import { Endpoint } from "../route/endpoint.js"
 import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
 import {
+  AIError,
   LLMEvent,
   Usage,
   type FinishReason,
@@ -17,6 +18,7 @@ import {
   type ToolCallPart,
   type ToolDefinition,
 } from "../schema/index.js"
+import { classifyProviderFailure } from "../provider-error.js"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
 import { GeminiToolSchema } from "./utils/gemini-tool-schema.js"
 import { Lifecycle } from "./utils/lifecycle.js"
@@ -125,6 +127,7 @@ const GeminiContentPart = Schema.Union([
   GeminiFunctionCallPart,
   GeminiFunctionResponsePart,
 ])
+const decodeGeminiContentPart = Schema.decodeUnknownOption(GeminiContentPart)
 
 const GeminiContent = Schema.Struct({
   role: optionalNull(Schema.Literals(["user", "model"])),
@@ -132,6 +135,11 @@ const GeminiContent = Schema.Struct({
 })
 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 })),
 })
@@ -200,7 +208,7 @@ const GeminiUsage = Schema.Struct({
 type GeminiUsage = Schema.Schema.Type<typeof GeminiUsage>
 
 const GeminiCandidate = Schema.Struct({
-  content: optionalNull(GeminiContent),
+  content: optionalNull(GeminiResponseContent),
   finishReason: optionalNull(Schema.String),
 })
 
@@ -215,6 +223,7 @@ const GeminiPromptFeedback = Schema.StructWithRest(
 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),
@@ -222,6 +231,8 @@ const GeminiEvent = Schema.Struct({
 type GeminiEvent = Schema.Schema.Type<typeof GeminiEvent>
 
 interface ParserState {
+  readonly route: string
+  readonly providerMetadataKey: string
   readonly finishReason?: string
   readonly hasToolCalls: boolean
   readonly promptFeedback?: GeminiPromptFeedback
@@ -278,22 +289,23 @@ const lowerUserPart = Effect.fn("Gemini.lowerUserPart")(function* (part: TextPar
   return { inlineData: { mimeType: media.mime, data: media.base64 } }
 })
 
-const googleMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ google: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
 
-const thoughtSignature = (providerMetadata: ProviderMetadata | undefined) => {
-  const google = providerMetadata?.google
-  return ProviderShared.isRecord(google) && typeof google.thoughtSignature === "string"
-    ? google.thoughtSignature
+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) => ({
+const lowerToolCall = (part: ToolCallPart, omitIds: boolean, metadataKey: string) => ({
   functionCall: { ...(omitIds ? {} : { id: part.id }), name: part.name, args: part.input },
-  thoughtSignature: thoughtSignature(part.providerMetadata),
+  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
@@ -335,15 +347,19 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
         if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call"]))
           return yield* ProviderShared.unsupportedContent("Gemini", "assistant", ["text", "reasoning", "tool-call"])
         if (part.type === "text") {
-          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey) })
           continue
         }
         if (part.type === "reasoning") {
-          parts.push({ text: part.text, thought: true, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({
+            text: part.text,
+            thought: true,
+            thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
+          })
           continue
         }
         if (part.type === "tool-call") {
-          const lowered = lowerToolCall(part, omitCallIds)
+          const lowered = lowerToolCall(part, omitCallIds, metadataKey)
           const signature = lowered.thoughtSignature
           parts.push({
             ...lowered,
@@ -491,7 +507,7 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
 // `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) => {
+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.
@@ -512,7 +528,7 @@ const mapUsage = (usage: GeminiUsage | undefined) => {
     cacheReadInputTokens: cached,
     reasoningTokens: thoughts,
     totalTokens: ProviderShared.totalTokens(promptTokens, outputTokens, usage.totalTokenCount ?? undefined),
-    providerMetadata: { google: usage },
+    providerMetadata: providerMetadata(metadataKey, usage),
   })
 }
 
@@ -560,10 +576,15 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
       lifecycle,
       events,
       "reasoning-0",
-      googleMetadata({ thoughtSignature: state.reasoningSignature }),
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.reasoningSignature }),
     )
   if (state.textSignature !== undefined)
-    lifecycle = Lifecycle.textEnd(lifecycle, events, "text-0", googleMetadata({ thoughtSignature: state.textSignature }))
+    lifecycle = Lifecycle.textEnd(
+      lifecycle,
+      events,
+      "text-0",
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.textSignature }),
+    )
   Lifecycle.finish(lifecycle, events, {
     reason: {
       normalized:
@@ -572,16 +593,32 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
     },
     usage: state.usage,
     providerMetadata:
-      state.promptFeedback === undefined ? undefined : googleMetadata({ promptFeedback: state.promptFeedback }),
+      state.promptFeedback === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { promptFeedback: state.promptFeedback }),
   })
   return events
 }
 
 const step = (state: ParserState, event: GeminiEvent) => {
+  if (ProviderShared.isRecord(event.error) && typeof event.error.message === "string") {
+    const body = ProviderShared.encodeJson(event)
+    return Effect.fail(
+      new AIError({
+        reason: classifyProviderFailure({
+          message: event.error.message,
+          status: typeof event.error.code === "number" ? event.error.code : undefined,
+          rawBody: body,
+        }),
+      }),
+    )
+  }
   const nextState = {
     ...state,
     promptFeedback: event.promptFeedback ?? state.promptFeedback,
-    usage: event.usageMetadata ? (mapUsage(event.usageMetadata) ?? state.usage) : state.usage,
+    usage: event.usageMetadata
+      ? (mapUsage(event.usageMetadata, state.providerMetadataKey) ?? state.usage)
+      : state.usage,
   }
   const candidate = event.candidates?.[0]
   if (!candidate?.content)
@@ -598,7 +635,21 @@ const step = (state: ParserState, event: GeminiEvent) => {
   // 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 part of candidate.content.parts ?? []) {
+  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.
@@ -611,7 +662,7 @@ const step = (state: ParserState, event: GeminiEvent) => {
           events,
           "reasoning-0",
           part.text,
-          signature ? googleMetadata({ thoughtSignature: signature }) : undefined,
+          signature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: signature }) : undefined,
         )
         continue
       }
@@ -619,14 +670,16 @@ const step = (state: ParserState, event: GeminiEvent) => {
         lifecycle,
         events,
         "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
       )
       lifecycle = Lifecycle.textDelta(
         lifecycle,
         events,
         "text-0",
         part.text,
-        textSignature ? googleMetadata({ thoughtSignature: textSignature }) : undefined,
+        textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
       )
       textSignature = undefined
       continue
@@ -646,7 +699,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
         lifecycle,
         events,
         "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
       )
       lifecycle = Lifecycle.stepStart(lifecycle, events)
       events.push(
@@ -654,8 +709,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
           id,
           name: part.functionCall.name,
           input,
-          providerMetadata:
-            part.thoughtSignature ? googleMetadata({ thoughtSignature: part.thoughtSignature }) : undefined,
+          providerMetadata: part.thoughtSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: part.thoughtSignature })
+            : undefined,
         }),
       )
       hasToolCalls = true
@@ -691,9 +747,14 @@ export const protocol = Protocol.make({
   },
   stream: {
     event: Protocol.jsonEvent(GeminiEvent),
-    initial: () => ({ hasToolCalls: false, lifecycle: Lifecycle.initial() }),
+    initial: (request) => ({
+      route: `${request.model.provider}/${request.model.route.id}`,
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
+      hasToolCalls: false,
+      lifecycle: Lifecycle.initial(),
+    }),
     step,
-    onHalt: finish,
+    onHalt: (state) => Effect.succeed(finish(state)),
   },
 })
 

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

@@ -9,15 +9,7 @@ import {
   type ImageRoute,
 } from "../image.js"
 import { Auth, type Definition as AuthDefinition } from "../route/auth.js"
-import {
-  InvalidProviderOutputReason,
-  AIError,
-  Usage,
-  mergeHttpOptions,
-  mergeJsonRecords,
-  type HttpOptions,
-  type ProviderMetadata,
-} from "../schema/index.js"
+import { AIError, Usage, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema/index.js"
 import { ProviderShared } from "./shared.js"
 import { ImageInputs } from "./utils/image-input.js"
 
@@ -124,13 +116,6 @@ const nativeOptions = (options: GoogleImageOptions | undefined) => {
   )
 }
 
-const invalidOutput = (message: string, providerMetadata?: ProviderMetadata) =>
-  new AIError({
-    module: ADAPTER,
-    method: "generate",
-    reason: new InvalidProviderOutputReason({ message, route: ADAPTER, providerMetadata }),
-  })
-
 const applyQuery = (url: string, query: Record<string, string> | undefined) => {
   if (!query) return url
   const next = new URL(url)
@@ -169,11 +154,9 @@ export const model = (input: ModelInput) => {
           HttpClientRequest.bodyText(text, "application/json"),
         ),
       )
-      const payload = yield* response.json.pipe(
-        Effect.mapError(() => invalidOutput("Failed to read the Google Images response")),
-      )
-      const decoded = yield* Schema.decodeUnknownEffect(GoogleImageResponse)(payload).pipe(
-        Effect.mapError(() => invalidOutput("Google Images returned an invalid response")),
+      const output = yield* ProviderShared.imageResponse(ADAPTER, "Google Images", response)
+      const decoded = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(GoogleImageResponse))(output.body).pipe(
+        Effect.mapError((cause) => output.invalid("Google Images returned an invalid response", cause)),
       )
       const candidates = decoded.candidates ?? []
       const candidateMetadata = candidates.map((candidate, candidateIndex) => ({
@@ -208,9 +191,10 @@ export const model = (input: ModelInput) => {
       )
       const images = yield* Effect.forEach(encoded, (item) =>
         Effect.fromResult(Encoding.decodeBase64(item.inlineData.data)).pipe(
-          Effect.mapError(() =>
-            invalidOutput(
+          Effect.mapError((cause) =>
+            output.invalid(
               `Google Images candidate ${item.candidateIndex} part ${item.partIndex} contains invalid base64 data`,
+              cause,
             ),
           ),
           Effect.map(
@@ -237,16 +221,10 @@ export const model = (input: ModelInput) => {
         const finishReasons = candidates.flatMap((candidate) =>
           candidate.finishReason === undefined ? [] : [candidate.finishReason],
         )
-        return yield* invalidOutput(
+        return yield* output.invalid(
           `Google Images returned no final images${
             finishReasons.length === 0 ? "" : ` (finish reasons: ${finishReasons.join(", ")})`
-          }; inspect reason.providerMetadata.google for prompt feedback and candidate details`,
-          {
-            google: {
-              promptFeedback: decoded.promptFeedback,
-              candidates: candidateMetadata,
-            },
-          },
+          }; inspect body for prompt feedback and candidate details`,
         )
       }
       const usage = decoded.usageMetadata
@@ -290,12 +268,11 @@ const googleImagePart = (image: ImageInput): Effect.Effect<Record<string, unknow
     return Effect.succeed({ inlineData: { mimeType: image.mediaType, data: Encoding.encodeBase64(image.data) } })
   if (image.type === "file-uri") return Effect.succeed({ fileData: { mimeType: image.mediaType, fileUri: image.uri } })
   if (image.type === "url")
-    return ImageInputs.decodeDataUrl(image.url, ADAPTER).pipe(
+    return ImageInputs.decodeDataUrl(image.url).pipe(
       Effect.flatMap((decoded) => {
         if (decoded === undefined)
           return Effect.fail(
             ImageInputs.invalid(
-              ADAPTER,
               "Google generateContent does not fetch public image URLs; use bytes, a data URL, or a Gemini file URI",
             ),
           )
@@ -305,7 +282,7 @@ const googleImagePart = (image: ImageInput): Effect.Effect<Record<string, unknow
       }),
     )
   return Effect.fail(
-    ImageInputs.invalid(ADAPTER, "Google generateContent requires Gemini file URIs rather than provider file IDs"),
+    ImageInputs.invalid("Google generateContent requires Gemini file URIs rather than provider file IDs"),
   )
 }
 

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

@@ -1,6 +1,7 @@
 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,
@@ -60,8 +61,8 @@ const driver = (options: Options, body: string): WebSocketChannelDriver => {
     observe: (_create, frame) =>
       Effect.gen(function* () {
         const event = yield* decodeEvent(frame).pipe(
-          Effect.mapError(() =>
-            ProviderShared.eventError(options.id, `Invalid ${options.name} WebSocket event`, frame),
+          Effect.mapError((cause) =>
+            ProviderShared.eventError(options.id, `Invalid ${options.name} WebSocket event`, frame, cause),
           ),
         )
         if (terminal)
@@ -73,13 +74,13 @@ const driver = (options: Options, body: string): WebSocketChannelDriver => {
         if (event.type === "error") {
           terminal = true
           yield* OpenResponses.decodeKnownErrorEvent(event).pipe(
-            Effect.mapError(() =>
-              ProviderShared.eventError(options.id, `${options.name} returned a malformed error event`, frame),
+            Effect.mapError((cause) =>
+              ProviderShared.eventError(options.id, `${options.name} returned a malformed error event`, frame, cause),
             ),
           )
           return {
             type: "provider-failure",
-            error: OpenResponses.providerFailure(options.id, event, `${options.name} stream error`),
+            error: OpenResponses.providerFailure(event, `${options.name} stream error`, frame),
           }
         }
         if (event.type === "response.failed") {
@@ -92,7 +93,7 @@ const driver = (options: Options, body: string): WebSocketChannelDriver => {
             )
           return {
             type: "provider-failure",
-            error: OpenResponses.providerFailure(options.id, event, `${options.name} response failed`),
+            error: OpenResponses.providerFailure(event, `${options.name} response failed`, frame),
           }
         }
         if (event.type === "response.created") {
@@ -175,23 +176,37 @@ export const transport = <Body>(options: Options): Transport<Body, Prepared, str
           channel,
         }
       }),
-    execute: (prepared, request, runtime, executeOptions) => {
-      if (!executeOptions?.webSocket || !prepared.channel) return http.execute(prepared.http, request, runtime)
-      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) => execution.frames)),
-          ),
-        driver: prepared.channel.driver,
-      }
-      return executeOptions.webSocket.execute(exchange)
-    },
+    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
+          },
+        }
+      }),
   }
 }
 

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

@@ -1,4 +1,4 @@
-import { AIError, TransportReason } from "../schema/index.js"
+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"
@@ -99,17 +99,17 @@ const incremental = (
 const code = (event: OpenResponses.Event) => event.code || event.error?.code || event.response?.error?.code || undefined
 
 const rejected = (
-  input: DriverInput,
   observation: Extract<ChannelObservation, { readonly type: "provider-failure" }>,
   recovery: "retry-full" | "rotate-and-retry-full",
 ): ChannelObservation => ({
   type: "rejected",
   recovery,
   error: new AIError({
-    module: input.id,
-    method: "stream",
-    reason: new TransportReason({
+    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",
@@ -137,15 +137,16 @@ export const driver = (input: DriverInput): WebSocketChannelDriver => {
     observe: (create, frame) =>
       Effect.gen(function* () {
         const event = yield* decodeEvent(frame).pipe(
-          Effect.mapError(() => ProviderShared.eventError(input.id, `Invalid ${input.name} WebSocket event`, frame)),
+          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(input, observation, "retry-full")
-          if (rejection === "websocket_connection_limit_reached")
-            return rejected(input, observation, "rotate-and-retry-full")
+          if (rejection === "previous_response_not_found") return rejected(observation, "retry-full")
+          if (rejection === "websocket_connection_limit_reached") return rejected(observation, "rotate-and-retry-full")
         }
         if (observation.type !== "completed") return observation
         const responseID = event.response?.id
@@ -154,7 +155,12 @@ export const driver = (input: DriverInput): WebSocketChannelDriver => {
           ...observation,
           checkpoint: {
             protocol: PROTOCOL,
-            value: { version: VERSION, responseID, request, output: output.slice() } satisfies CheckpointValue,
+            value: {
+              version: VERSION,
+              responseID,
+              request,
+              output: event.response?.output ? [...event.response.output] : output.slice(),
+            } satisfies CheckpointValue,
           },
         }
       }),

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

@@ -3,11 +3,16 @@ import { Tool } from "@opencode-ai/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 { HttpTransport } from "../route/transport/index.js"
 import { Protocol } from "../route/protocol.js"
 import {
   AIError,
+  AIErrorReason,
+  InvalidProviderOutputError,
   LLMEvent,
+  ProviderInternalError,
+  UnknownProviderError,
   Usage,
   type FinishReason,
   type FinishReasonDetails,
@@ -51,7 +56,12 @@ const OpenAIChatFunction = Schema.Struct({
 
 const OpenAIChatTool = Schema.Struct({
   type: Schema.tag("function"),
-  function: OpenAIChatFunction,
+  function: Schema.Struct({
+    name: Schema.String,
+    description: Schema.String,
+    parameters: JsonObject,
+    strict: Schema.optional(Schema.Boolean),
+  }),
   cache_control: Schema.optional(OpenAIChatCacheControl),
 })
 type OpenAIChatTool = Schema.Schema.Type<typeof OpenAIChatTool>
@@ -133,6 +143,7 @@ export const bodyFields = {
   store: Schema.optional(Schema.Boolean),
   prompt_cache_key: Schema.optional(Schema.String),
   reasoning_effort: Schema.optional(OpenAIOptions.OpenAIReasoningEffort),
+  tool_stream: Schema.optional(Schema.Boolean),
   max_completion_tokens: Schema.optional(Schema.Number),
   max_tokens: Schema.optional(Schema.Number),
   temperature: Schema.optional(Schema.Number),
@@ -218,17 +229,25 @@ const OpenAIChatChoice = Schema.StructWithRest(
   [Schema.Record(Schema.String, Schema.Unknown)],
 )
 
-const OpenAIChatError = Schema.Struct({
-  code: optionalNull(Schema.Union([Schema.String, Schema.Number])),
-  message: Schema.String,
-})
+const OpenAIChatError = Schema.StructWithRest(
+  Schema.Struct({
+    code: optionalNull(Schema.Union([Schema.String, Schema.Number])),
+    message: Schema.String,
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
 
-export const OpenAIChatEvent = Schema.Struct({
-  choices: optionalNull(Schema.Array(OpenAIChatChoice)),
-  usage: optionalNull(OpenAIChatUsage),
-  error: optionalNull(OpenAIChatError),
-})
+export const OpenAIChatEvent = Schema.StructWithRest(
+  Schema.Struct({
+    choices: optionalNull(Schema.Array(OpenAIChatChoice)),
+    usage: optionalNull(OpenAIChatUsage),
+    error: optionalNull(OpenAIChatError),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
 export type OpenAIChatEvent = Schema.Schema.Type<typeof OpenAIChatEvent>
+const DONE = "[DONE]" as const
+const OpenAIChatStreamEvent = Schema.Union([Schema.Literal(DONE), Protocol.jsonEvent(OpenAIChatEvent)])
 type OpenAIChatRequestMessage = LLMRequest["messages"][number]
 
 interface PendingToolDelta {
@@ -238,6 +257,7 @@ interface PendingToolDelta {
 }
 
 export interface ParserState {
+  readonly providerMetadataKey: string
   readonly tools: ToolStream.State<number>
   readonly pendingTools: Partial<Record<number, PendingToolDelta>>
   readonly toolCallEvents: ReadonlyArray<LLMEvent>
@@ -250,6 +270,7 @@ export interface ParserState {
   readonly reasoningEmitted: boolean
   readonly latestToolIndex?: number
   readonly nextToolIndex: number
+  readonly requireFinishReason: boolean
 }
 
 // =============================================================================
@@ -262,14 +283,21 @@ interface LoweringOptions {
   readonly cacheControl?: (
     cache: CacheHint | undefined,
   ) => Schema.Schema.Type<typeof OpenAIChatCacheControl> | undefined
+  readonly toolCallID?: (id: string) => string
 }
 
-const lowerTool = (tool: ToolDefinition, inputSchema: JsonSchema, options: LoweringOptions): OpenAIChatTool => ({
+const lowerTool = (
+  tool: ToolDefinition,
+  inputSchema: JsonSchema,
+  options: LoweringOptions,
+  supportsStrictMode: boolean,
+): OpenAIChatTool => ({
   type: "function",
   function: {
     name: tool.name,
     description: tool.description,
     parameters: inputSchema,
+    ...(supportsStrictMode ? { strict: false } : {}),
   },
   cache_control: options.cacheControl?.(tool.cache),
 })
@@ -282,8 +310,8 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
     tool: (name) => ({ type: "function" as const, function: { name } }),
   })
 
-const lowerToolCall = (part: ToolCallPart): OpenAIChatAssistantToolCall => ({
-  id: part.id,
+const lowerToolCall = (part: ToolCallPart, options: LoweringOptions): OpenAIChatAssistantToolCall => ({
+  id: options.toolCallID?.(part.id) ?? part.id,
   type: "function",
   function: {
     name: part.name,
@@ -301,17 +329,18 @@ const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart
 const openAICompatibleReasoningContent = (native: unknown) =>
   isRecord(native) && typeof native.reasoning_content === "string" ? native.reasoning_content : undefined
 
-const reasoningField = (part: ReasoningPart) => {
-  const field = part.providerMetadata?.openai?.reasoningField
+const reasoningField = (part: ReasoningPart, providerMetadataKey: string) => {
+  const field = part.providerMetadata?.[providerMetadataKey]?.reasoningField
   return typeof field === "string" ? field : undefined
 }
 
-const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown) => {
+const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown, providerMetadataKey: string) => {
   const observed = parts.flatMap((part) => {
-    const details = part.providerMetadata?.openai?.reasoningDetails
+    const details = part.providerMetadata?.[providerMetadataKey]?.reasoningDetails
     return Array.isArray(details) ? details : []
   })
-  if (parts.some((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))) return observed
+  if (parts.some((part) => Array.isArray(part.providerMetadata?.[providerMetadataKey]?.reasoningDetails)))
+    return observed
   if (isRecord(native) && Array.isArray(native.reasoning_details)) return native.reasoning_details
 }
 
@@ -341,8 +370,9 @@ const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (
 
 const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(function* (
   message: OpenAIChatRequestMessage,
-  configuredField?: string,
-  options: LoweringOptions = {},
+  configuredField: string | undefined,
+  requireReasoning: boolean,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
   const content: TextPart[] = []
   const reasoning: ReasoningPart[] = []
@@ -359,25 +389,31 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
       continue
     }
     if (part.type === "tool-call") {
-      toolCalls.push(lowerToolCall(part))
+      toolCalls.push(lowerToolCall(part, options))
       continue
     }
   }
   const text = reasoning.map((part) => part.text).join("")
-  const details = reasoningDetails(reasoning, message.native?.openaiCompatible)
-  const observedField = reasoning.map(reasoningField).find((value) => value !== undefined)
+  const details = reasoningDetails(reasoning, message.native?.openaiCompatible, options.providerMetadataKey)
+  const observedField = reasoning
+    .map((part) => reasoningField(part, options.providerMetadataKey))
+    .find((value) => value !== undefined)
   const nativeReasoning = openAICompatibleReasoningContent(message.native?.openaiCompatible)
-  const fullyStructured = reasoning.every((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))
+  const fullyStructured = reasoning.every((part) =>
+    Array.isArray(part.providerMetadata?.[options.providerMetadataKey]?.reasoningDetails),
+  )
   const field = (() => {
-    if (configuredField !== undefined) return configuredField
-    if (reasoning.length === 0) return undefined
+    if (configuredField !== undefined && (requireReasoning || reasoning.length > 0 || nativeReasoning !== undefined))
+      return configuredField
+    if (reasoning.length === 0) return requireReasoning ? "reasoning_content" : undefined
     if (observedField !== undefined) return observedField
     if (nativeReasoning !== undefined) return "reasoning_content"
-    if (!fullyStructured) return "reasoning_content"
+    if (!fullyStructured || requireReasoning) return "reasoning_content"
   })()
   const reasoningText = (() => {
-    if (configuredField !== undefined) return reasoning.length === 0 ? (nativeReasoning ?? "") : text
-    if (reasoning.length === 0) return nativeReasoning
+    if (configuredField !== undefined)
+      return reasoning.length === 0 ? (nativeReasoning ?? (requireReasoning ? "" : undefined)) : text
+    if (reasoning.length === 0) return nativeReasoning ?? (requireReasoning ? "" : undefined)
     return text
   })()
   const cached = message.content.findLast((part) => "cache" in part && part.cache !== undefined)
@@ -405,7 +441,7 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
     if (part.result.type !== "content") {
       messages.push({
         role: "tool",
-        tool_call_id: part.id,
+        tool_call_id: options.toolCallID?.(part.id) ?? part.id,
         content: ProviderShared.toolResultText(part),
         cache_control: options.cacheControl?.(part.cache),
       })
@@ -415,7 +451,7 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
     const text = content.filter((item) => item.type === "text").map((item) => item.text)
     messages.push({
       role: "tool",
-      tool_call_id: part.id,
+      tool_call_id: options.toolCallID?.(part.id) ?? part.id,
       content: text.join("\n"),
       cache_control: options.cacheControl?.(part.cache),
     })
@@ -431,11 +467,13 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
 
 const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
   message: OpenAIChatRequestMessage,
-  reasoningField?: string,
-  options: LoweringOptions = {},
+  reasoningField: string | undefined,
+  requireReasoning: boolean,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
   if (message.role === "user") return [yield* lowerUserMessage(message, options)]
-  if (message.role === "assistant") return [yield* lowerAssistantMessage(message, reasoningField, options)]
+  if (message.role === "assistant")
+    return [yield* lowerAssistantMessage(message, reasoningField, requireReasoning, options)]
   return (yield* lowerToolMessages(message, options)).messages
 })
 
@@ -456,12 +494,43 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
           ]
         : [{ role: "system", content: ProviderShared.joinText(request.system) }]
   const messages = [...system]
+  const modelID = request.model.id.toLowerCase()
+  const requireReasoning =
+    request.model.compatibility?.requireReasoning ??
+    (request.model.compatibility?.reasoningField !== undefined ||
+      request.model.provider === "deepseek" ||
+      request.model.route.endpoint.baseURL?.toLowerCase().includes("deepseek.com") ||
+      modelID.includes("deepseek"))
+  const reasoningField = request.model.compatibility?.reasoningField
+  const mistral = ["mistral", "devstral", "codestral", "pixtral", "mixtral"].some((family) => modelID.includes(family))
+  const lowering = {
+    ...options,
+    providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
+    toolCallID: (id: string) => {
+      if (mistral)
+        return id
+          .replace(/[^a-zA-Z0-9]/g, "")
+          .slice(0, 9)
+          .padEnd(9, "0")
+      if (modelID.includes("claude")) return id.replace(/[^a-zA-Z0-9_-]/g, "_")
+      if (request.model.provider === "openai" || request.model.provider === "azure" || modelID.startsWith("openai/"))
+        return id.slice(0, 40)
+      return id
+    },
+  }
+  const requireAssistantAfterTool = request.model.compatibility?.requireAssistantAfterTool ?? mistral
+  const bridgeTools = () => {
+    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool")
+      messages.push({ role: "assistant", content: "Done." })
+  }
   const pendingImages: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
   const flushImages = () => {
     if (pendingImages.length === 0) return
+    bridgeTools()
     messages.push({ role: "user", content: pendingImages.splice(0) })
   }
   for (const message of request.messages) {
+    if (message.role === "user") bridgeTools()
     if (message.role === "system") {
       const part = yield* ProviderShared.wrappedSystemUpdate("OpenAI Chat", message)
       if (pendingImages.length > 0) {
@@ -504,14 +573,19 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
         )
       continue
     }
+    if (
+      message.role === "assistant" &&
+      message.content.every((part) => part.type === "text" && part.text.trim() === "")
+    )
+      continue
     if (message.role === "tool") {
-      const lowered = yield* lowerToolMessages(message, options)
+      const lowered = yield* lowerToolMessages(message, lowering)
       messages.push(...lowered.messages)
       pendingImages.push(...lowered.images)
       continue
     }
     flushImages()
-    messages.push(...(yield* lowerMessage(message, request.model.compatibility?.reasoningField, options)))
+    messages.push(...(yield* lowerMessage(message, reasoningField, requireReasoning, lowering)))
   }
   flushImages()
   return messages
@@ -528,11 +602,122 @@ const hasToolHistory = (messages: ReadonlyArray<LLMRequest["messages"][number]>)
   return false
 }
 
-const lowerOptions = (request: LLMRequest) => {
+// Derive `max_tokens` vs `max_completion_tokens` from provider/baseURL when
+// explicit `compatibility.maxTokensField` is not set. Aligned with
+// models.dev provider naming: DeepSeek, Moonshot AI, Together AI, ZAI
+// (Zhipu + Coding Plan variants), Nvidia, Cerebras, Chutes, etc. still
+// require `max_tokens`.
+const detectMaxTokensField = (
+  provider: string,
+  baseURL: string | undefined,
+): "max_tokens" | "max_completion_tokens" => {
+  const p = provider.toLowerCase()
+  const url = (baseURL ?? "").toLowerCase()
+  if (
+    p === "deepseek" ||
+    url.includes("deepseek.com") ||
+    p === "moonshotai" ||
+    url.includes("api.moonshot.ai") ||
+    p === "togetherai" ||
+    url.includes("api.together.") ||
+    p === "zai" ||
+    p === "zai-coding-plan" ||
+    p === "zhipuai" ||
+    p === "zhipuai-coding-plan" ||
+    url.includes("api.z.ai") ||
+    url.includes("open.bigmodel.cn") ||
+    p === "nvidia" ||
+    url.includes("integrate.api.nvidia.com") ||
+    p === "cerebras" ||
+    url.includes("cerebras.ai") ||
+    url.includes("llm.chutes.ai") ||
+    p === "chutes" ||
+    p === "cloudflare-ai-gateway" ||
+    url.includes("gateway.ai.cloudflare.com") ||
+    p === "cloudflare-workers-ai" ||
+    url.includes("api.cloudflare.com")
+  )
+    return "max_tokens"
+  return "max_completion_tokens"
+}
+
+const detectSupportsStore = (provider: string, baseURL: string | undefined): boolean => {
+  const p = provider.toLowerCase()
+  const url = (baseURL ?? "").toLowerCase()
+  const isNvidia = p === "nvidia" || url.includes("integrate.api.nvidia.com")
+  const isMoonshot = p === "moonshotai" || p === "moonshotai-cn" || url.includes("api.moonshot.")
+  const isTogether = p === "togetherai" || p === "together" || url.includes("api.together.")
+  const isZai =
+    p === "zai" ||
+    p === "zai-coding-plan" ||
+    p === "zhipuai" ||
+    p === "zhipuai-coding-plan" ||
+    url.includes("api.z.ai") ||
+    url.includes("open.bigmodel.cn")
+  const isDeepSeek = p === "deepseek" || url.includes("deepseek.com")
+  const isCerebras = p === "cerebras" || url.includes("cerebras.ai")
+  const isXai = p === "xai" || url.includes("api.x.ai")
+  const isChutes = p === "chutes" || url.includes("chutes.ai")
+  const isCloudflareWorkersAI = p === "cloudflare-workers-ai" || url.includes("api.cloudflare.com")
+  const isCloudflareAiGateway = p === "cloudflare-ai-gateway" || url.includes("gateway.ai.cloudflare.com")
+  const isVercelAiGateway =
+    p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
+  const isAntLing = p === "ant-ling" || url.includes("api.ant-ling.com")
+  const isOpencode = p === "opencode" || url.includes("opencode.ai")
+  const isNonStandard =
+    isNvidia ||
+    isCerebras ||
+    isXai ||
+    isTogether ||
+    isChutes ||
+    isDeepSeek ||
+    isZai ||
+    isMoonshot ||
+    isOpencode ||
+    isCloudflareWorkersAI ||
+    isCloudflareAiGateway ||
+    isVercelAiGateway ||
+    isAntLing
+  return !isNonStandard
+}
+
+const detectSupportsUsageInStreaming = (): boolean => true
+
+const detectSupportsStrictMode = (provider: string, baseURL: string | undefined): boolean => {
+  const p = provider.toLowerCase()
+  const url = (baseURL ?? "").toLowerCase()
+  const isMoonshot = p === "moonshotai" || p === "moonshotai-cn" || url.includes("api.moonshot.")
+  const isTogether = p === "togetherai" || p === "together" || url.includes("api.together.")
+  const isCloudflareAiGateway = p === "cloudflare-ai-gateway" || url.includes("gateway.ai.cloudflare.com")
+  const isNvidia = p === "nvidia" || url.includes("integrate.api.nvidia.com")
+  return !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia
+}
+
+const detectZaiToolStream = (provider: string, baseURL: string | undefined, modelID: string): boolean => {
+  const p = provider.toLowerCase()
+  const url = (baseURL ?? "").toLowerCase()
+  const isZai =
+    p === "zai" ||
+    p === "zai-coding-plan" ||
+    p === "zhipuai" ||
+    p === "zhipuai-coding-plan" ||
+    url.includes("api.z.ai") ||
+    url.includes("open.bigmodel.cn")
+  if (!isZai) return false
+  const id = modelID.toLowerCase()
+  if (id === "glm-4.5" || id === "glm-4.5-air" || id === "glm-4.5-flash" || id === "glm-4.5v") return false
+  return true
+}
+
+const lowerOptions = (request: LLMRequest, supportsStore: boolean) => {
   const options = OpenAIOptions.resolve(request)
-  const cacheKey = ProviderShared.clampPromptCacheKey(request.promptCacheKey)
+  const cacheKey = ProviderShared.promptCacheKey(request)
   return {
-    ...(options.store !== undefined ? { store: options.store } : {}),
+    ...(supportsStore && options.store !== undefined ? { store: options.store } : {}),
+    // For providers that support `store`, ensure stateless `store:false` is sent
+    // even when no explicit `providerOptions.store` was supplied, mirroring the
+    // native OpenAI Chat default. Non-standard providers omit `store` entirely.
+    ...(supportsStore && options.store === undefined ? { store: false } : {}),
     ...(cacheKey ? { prompt_cache_key: cacheKey } : {}),
     ...(options.reasoningEffort ? { reasoning_effort: options.reasoningEffort } : {}),
   }
@@ -551,8 +736,19 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
     )
   const generation = request.generation
   const toolSchemaCompatibility = request.model.compatibility?.toolSchema
-  const maxTokensField = request.model.compatibility?.maxTokensField ?? "max_tokens"
+  const provider = String(request.model.provider)
+  const baseURL = request.model.route.endpoint.baseURL
+  const detectedMaxTokensField = detectMaxTokensField(provider, baseURL)
+  const maxTokensField = request.model.compatibility?.maxTokensField ?? detectedMaxTokensField
+  const supportsStore = request.model.compatibility?.supportsStore ?? detectSupportsStore(provider, baseURL)
+  const supportsUsageInStreaming =
+    request.model.compatibility?.supportsUsageInStreaming ?? detectSupportsUsageInStreaming()
+  const supportsStrictMode =
+    request.model.compatibility?.supportsStrictMode ?? detectSupportsStrictMode(provider, baseURL)
+  const zaiToolStream =
+    request.model.compatibility?.zaiToolStream ?? detectZaiToolStream(provider, baseURL, request.model.id)
   const hasHistory = hasToolHistory(request.messages)
+  const hasActiveTools = request.tools.length > 0
   return {
     model: request.model.id,
     messages: yield* lowerMessages(request, options),
@@ -566,11 +762,13 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
               tool,
               ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility),
               options,
+              supportsStrictMode,
             ),
           ),
     tool_choice: request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined,
     stream: true as const,
-    stream_options: { include_usage: true },
+    ...(supportsUsageInStreaming ? { stream_options: { include_usage: true } } : {}),
+    ...(zaiToolStream && hasActiveTools ? { tool_stream: true } : {}),
     ...(maxTokensField === "max_completion_tokens"
       ? { max_completion_tokens: generation?.maxTokens }
       : { max_tokens: generation?.maxTokens }),
@@ -580,7 +778,7 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
     presence_penalty: generation?.presencePenalty,
     seed: generation?.seed,
     stop: generation?.stop,
-    ...lowerOptions(request),
+    ...lowerOptions(request, supportsStore),
   }
 })
 
@@ -590,14 +788,36 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
 // Streaming parsers are small state machines: every event returns a new state
 // plus the common `LLMEvent`s produced by that event. Tool calls are accumulated
 // because OpenAI streams JSON arguments across multiple deltas.
-const mapFinishReason = (reason: string | null | undefined): FinishReason => {
-  if (reason === "stop") return "stop"
-  if (reason === "length") return "length"
-  if (reason === "content_filter") return "content-filter"
-  if (reason === "function_call" || reason === "tool_calls") return "tool-calls"
-  if (reason === "error") return "error"
-  return "unknown"
-}
+const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event: OpenAIChatEvent, reason: string) {
+  switch (reason) {
+    case "error":
+      return yield* new AIError({
+        reason: new UnknownProviderError({
+          message: "Provider reported an error (finish_reason: error)",
+          body: ProviderShared.encodeJson(event),
+        }),
+      })
+    case "network_error":
+      return yield* new AIError({
+        reason: new ProviderInternalError({
+          message: "Provider reported a network error (finish_reason: network_error)",
+          body: ProviderShared.encodeJson(event),
+        }),
+      })
+    case "stop":
+    case "end":
+      return "stop" as const
+    case "length":
+      return "length" as const
+    case "content_filter":
+      return "content-filter" as const
+    case "function_call":
+    case "tool_calls":
+      return "tool-calls" as const
+    default:
+      return "unknown" as const
+  }
+})
 
 // OpenAI Chat reports `prompt_tokens` (inclusive total) with a
 // cached-read and cache-write subsets, and `completion_tokens` (inclusive
@@ -607,15 +827,14 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // Providers differ on cache-hit location: OpenAI uses
 // `prompt_tokens_details.cached_tokens`, DeepSeek uses
 // `prompt_cache_hit_tokens`, and Zai uses top-level `cached_tokens`.
-const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
+const mapUsage = (usage: OpenAIChatEvent["usage"], providerMetadataKey: string): Usage | undefined => {
   if (!usage) return undefined
   const input = usage.prompt_tokens ?? undefined
   const output = usage.completion_tokens ?? undefined
-  const cached =
-    (usage.prompt_tokens_details?.cached_tokens ??
-      (usage as { prompt_cache_hit_tokens?: number | null }).prompt_cache_hit_tokens ??
-      (usage as { cached_tokens?: number | null }).cached_tokens ??
-      undefined) as number | undefined
+  const cached = (usage.prompt_tokens_details?.cached_tokens ??
+    (usage as { prompt_cache_hit_tokens?: number | null }).prompt_cache_hit_tokens ??
+    (usage as { cached_tokens?: number | null }).cached_tokens ??
+    undefined) as number | undefined
   const cacheWrite = usage.prompt_tokens_details?.cache_write_tokens ?? undefined
   const reasoning = usage.completion_tokens_details?.reasoning_tokens ?? undefined
   const nonCached = ProviderShared.subtractTokens(input, ProviderShared.sumTokens(cached, cacheWrite))
@@ -627,7 +846,7 @@ const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
     cacheWriteInputTokens: cacheWrite,
     reasoningTokens: reasoning,
     totalTokens: ProviderShared.totalTokens(input, output, usage.total_tokens ?? undefined),
-    providerMetadata: { openai: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
   })
 }
 
@@ -701,8 +920,12 @@ const conflictingReasoningTextDetails = (previous: Record<string, unknown>, curr
 const conflictingDetailValue = (previous: unknown, current: unknown) =>
   previous !== undefined && previous !== null && current !== undefined && current !== null && previous !== current
 
-const reasoningMetadata = (field: ParserState["reasoningField"], details?: ReadonlyArray<unknown>) => ({
-  openai: {
+const reasoningMetadata = (
+  providerMetadataKey: string,
+  field: ParserState["reasoningField"],
+  details?: ReadonlyArray<unknown>,
+) => ({
+  [providerMetadataKey]: {
     ...(field ? { reasoningField: field } : {}),
     ...(details ? { reasoningDetails: details } : {}),
   },
@@ -710,27 +933,32 @@ const reasoningMetadata = (field: ParserState["reasoningField"], details?: Reado
 
 const step = (state: ParserState, event: OpenAIChatEvent) =>
   Effect.gen(function* () {
-    if (event.error)
+    if (event.error) {
+      const body = ProviderShared.encodeJson(event)
       return yield* new AIError({
-        module: ADAPTER,
-        method: "stream",
         reason: classifyProviderFailure({
           message: event.error.message,
-          code: event.error.code === undefined || event.error.code === null ? undefined : String(event.error.code),
           status: typeof event.error.code === "number" ? event.error.code : undefined,
+          rawBody: body,
         }),
       })
+    }
     const events: LLMEvent[] = []
     const choice = event.choices?.[0]
     // Moonshot (and a few other OpenAI-compatible providers) attach usage to
     // `choice.usage` instead of the top-level `usage` field.
     const choiceUsage = (choice as unknown as { usage?: OpenAIChatEvent["usage"] })?.usage
-    const usage = mapUsage(event.usage) ?? (choiceUsage ? mapUsage(choiceUsage) : undefined) ?? state.usage
+    const usage =
+      mapUsage(event.usage, state.providerMetadataKey) ??
+      (choiceUsage ? mapUsage(choiceUsage, state.providerMetadataKey) : undefined) ??
+      state.usage
     const rawFinishReason = choice?.finish_reason
-    const finishReason =
-      rawFinishReason !== undefined && rawFinishReason !== null
-        ? { normalized: mapFinishReason(rawFinishReason), raw: choice?.native_finish_reason ?? rawFinishReason }
-        : state.finishReason
+    const finishReason = rawFinishReason
+      ? {
+          normalized: yield* mapFinishReason(event, rawFinishReason),
+          raw: choice?.native_finish_reason ?? rawFinishReason,
+        }
+      : state.finishReason
     const delta = choice?.delta
     const toolDeltas = delta?.tool_calls ?? []
     let tools = state.tools
@@ -749,7 +977,11 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
       toolDeltas.some((tool) => Boolean(tool.id) || Boolean(tool.function?.name) || Boolean(tool.function?.arguments))
     if (state.finishReason !== undefined) {
       if (hasLateContent)
-        return yield* ProviderShared.eventError(ADAPTER, "OpenAI Chat received content after the finish reason")
+        return yield* ProviderShared.eventError(
+          ADAPTER,
+          "OpenAI Chat received content after the finish reason",
+          ProviderShared.encodeJson(event),
+        )
       return [{ ...state, usage }, events] as const
     }
 
@@ -757,7 +989,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
     const detailDelta = Array.isArray(delta?.reasoning_details) ? delta.reasoning_details : undefined
     if (detailDelta !== undefined) appendReasoningDetails(state.reasoningDetails, detailDelta)
     const reasoningDetailsObserved = state.reasoningDetailsObserved || detailDelta !== undefined
-    const deltaMetadata = reasoningMetadata(reasoningField)
+    const deltaMetadata = reasoningMetadata(state.providerMetadataKey, reasoningField)
     const text = detailDelta?.length ? (detailText(detailDelta) ?? reasoning?.text) : reasoning?.text
     if (text !== undefined) lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", text, deltaMetadata)
     else if (
@@ -768,25 +1000,12 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
       lifecycle = Lifecycle.reasoningStart(lifecycle, events, "reasoning-0", deltaMetadata)
     const reasoningEmitted = state.reasoningEmitted || lifecycle.reasoning.has("reasoning-0")
 
-    if (delta?.content) {
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
-      )
-      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)
-    }
+    // Reasoning is one response-wide channel: it stays open alongside text and
+    // refusal output so late reasoning deltas and details join the same block,
+    // and `finishEvents` closes it once with the complete metadata.
+    if (delta?.content) lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)
 
-    if (delta?.refusal) {
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
-      )
-      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)
-    }
+    if (delta?.refusal) lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)
 
     // Compatible providers may omit indexes. Prefer durable identity, then use
     // batch position for parallel deltas or the latest call for sparse chunks.
@@ -821,14 +1040,26 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
         { id: id || undefined, name: name || undefined, text },
         "OpenAI Chat tool call delta is missing id or name",
       )
-      if (ToolStream.isError(result)) return yield* result
+      if (ToolStream.isError(result))
+        return yield* new AIError({
+          reason: AIErrorReason.make({
+            ...result.reason,
+            message: result.message,
+            cause: result.reason.cause,
+            body: ProviderShared.encodeJson(event),
+          }),
+        })
       tools = result.tools
       if (result.events.length) lifecycle = Lifecycle.stepStart(lifecycle, events)
       events.push(...result.events)
     }
 
     if (finishReason !== undefined && state.finishReason === undefined && Object.keys(pendingTools).length > 0)
-      return yield* ProviderShared.eventError(ADAPTER, "OpenAI Chat tool call delta is missing id or name")
+      return yield* ProviderShared.eventError(
+        ADAPTER,
+        "OpenAI Chat tool call delta is missing id or name",
+        ProviderShared.encodeJson(event),
+      )
 
     // Finalize accumulated tool inputs eagerly when finish_reason arrives so
     // valid calls and malformed local calls settle independently.
@@ -839,6 +1070,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
 
     return [
       {
+        providerMetadataKey: state.providerMetadataKey,
         tools: finished?.tools ?? tools,
         pendingTools,
         toolCallEvents: finished?.events ?? state.toolCallEvents,
@@ -851,16 +1083,25 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
         reasoningEmitted,
         latestToolIndex,
         nextToolIndex,
+        requireFinishReason: state.requireFinishReason,
       },
       events,
     ] as const
   })
 
-const finishEvents = (state: ParserState): ReadonlyArray<LLMEvent> => {
+const finishEvents = Effect.fn("OpenAIChat.finishEvents")(function* (state: ParserState) {
+  if (state.finishReason === undefined && state.requireFinishReason)
+    return yield* new AIError({
+      reason: new InvalidProviderOutputError({
+        message: "OpenAI Chat stream ended without finish_reason",
+        classification: "incomplete-stream",
+        route: ADAPTER,
+      }),
+    })
   const events: LLMEvent[] = []
   const toolCallEvents =
     state.finishReason === undefined && Object.keys(state.tools).length > 0
-      ? Effect.runSync(ToolStream.finishAll(ADAPTER, state.tools)).events
+      ? (yield* ToolStream.finishAll(ADAPTER, state.tools)).events
       : state.toolCallEvents
   const hasToolCalls = toolCallEvents.length > 0
   const reason = state.finishReason
@@ -869,21 +1110,29 @@ const finishEvents = (state: ParserState): ReadonlyArray<LLMEvent> => {
         normalized:
           state.finishReason.normalized === "stop" && hasToolCalls ? "tool-calls" : state.finishReason.normalized,
       }
-    : { normalized: hasToolCalls ? ("tool-calls" as const) : ("unknown" as const) }
+    : { normalized: hasToolCalls ? ("tool-calls" as const) : ("stop" as const) }
+  // Snapshot details at publish time so the emitted event never observes later
+  // mutation of the accumulated `reasoningDetails` array.
   const metadata = reasoningMetadata(
+    state.providerMetadataKey,
     state.reasoningField,
-    state.reasoningDetailsObserved ? state.reasoningDetails : undefined,
+    state.reasoningDetailsObserved ? [...state.reasoningDetails] : undefined,
   )
   const started =
     state.reasoningDetailsObserved && !state.reasoningEmitted
-      ? Lifecycle.reasoningStart(state.lifecycle, events, "reasoning-0", reasoningMetadata(state.reasoningField))
+      ? Lifecycle.reasoningStart(
+          state.lifecycle,
+          events,
+          "reasoning-0",
+          reasoningMetadata(state.providerMetadataKey, state.reasoningField),
+        )
       : state.lifecycle
   const ended = Lifecycle.reasoningEnd(started, events, "reasoning-0", metadata)
   const lifecycle = toolCallEvents.length ? Lifecycle.stepStart(ended, events) : ended
   events.push(...toolCallEvents)
   Lifecycle.finish(lifecycle, events, { reason, usage: state.usage })
   return events
-}
+})
 
 // =============================================================================
 // Protocol And OpenAI Route
@@ -901,8 +1150,9 @@ export const protocol = Protocol.make({
     from: fromRequest,
   },
   stream: {
-    event: Protocol.jsonEvent(OpenAIChatEvent),
+    event: OpenAIChatStreamEvent,
     initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
       tools: ToolStream.empty<number>(),
       pendingTools: {},
       toolCallEvents: [],
@@ -912,13 +1162,16 @@ export const protocol = Protocol.make({
       reasoningDetailsObserved: false,
       reasoningEmitted: false,
       nextToolIndex: 0,
+      requireFinishReason: request.model.compatibility?.requireFinishReason ?? true,
     }),
-    step,
+    step: (state: ParserState, event) => (event === DONE ? Effect.succeed([state, []] as const) : step(state, event)),
+    terminal: (event) => event === DONE,
     onHalt: finishEvents,
   },
 })
 
-export const httpTransport = HttpTransport.sseJson.with<OpenAIChatBody>()
+export const framing = Framing.sseWithDone
+export const httpTransport = HttpTransport.sseJson.with<OpenAIChatBody>().with({ framing })
 
 export const route = Route.make({
   id: ADAPTER,

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

@@ -1,6 +1,5 @@
 import { Route, type RouteRoutedLanguageModelInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import * as OpenAIChat from "./openai-chat.js"
 
 const ADAPTER = "openai-compatible-chat"
@@ -19,7 +18,7 @@ export const route = Route.make({
   providerMetadataKey: "openai",
   protocol: OpenAIChat.protocol,
   endpoint: Endpoint.path("/chat/completions"),
-  framing: Framing.sse,
+  framing: OpenAIChat.framing,
 })
 
 export * as OpenAICompatibleChat from "./openai-compatible-chat.js"

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

@@ -17,6 +17,7 @@ export const route = Route.make({
   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

@@ -9,14 +9,7 @@ import {
   type ImageRoute,
 } from "../image.js"
 import { Auth, type Definition as AuthDefinition } from "../route/auth.js"
-import {
-  InvalidProviderOutputReason,
-  AIError,
-  Usage,
-  mergeHttpOptions,
-  mergeJsonRecords,
-  type HttpOptions,
-} from "../schema/index.js"
+import { Usage, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema/index.js"
 import { ProviderShared } from "./shared.js"
 import { ImageInputs } from "./utils/image-input.js"
 import { OpenAIImage } from "./utils/openai-image.js"
@@ -84,13 +77,6 @@ const nativeOptions = (options: OpenAIImageOptions | undefined) => {
   }
 }
 
-const invalidOutput = (message: string) =>
-  new AIError({
-    module: ADAPTER,
-    method: "generate",
-    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
-  })
-
 const applyQuery = (url: string, query: Record<string, string> | undefined) => {
   if (!query) return url
   const next = new URL(url)
@@ -104,12 +90,12 @@ export const model = (input: ModelInput) => {
     generate: Effect.fn("OpenAIImages.generate")(function* (request: ImageRequestFor<OpenAIImageOptions>, execute) {
       const mask = request.options?.mask
       if (mask !== undefined && (request.images?.length ?? 0) === 0)
-        return yield* ImageInputs.invalid(ADAPTER, "An OpenAI image mask requires at least one input image")
+        return yield* ImageInputs.invalid("An OpenAI image mask requires at least one input image")
       const http = mergeHttpOptions(request.model.http, request.http)
       const sourceImages = request.images ?? []
       const multipartImages = yield* Effect.forEach(sourceImages, (image) => {
         if (image.type === "bytes") return Effect.succeed({ data: image.data, mediaType: image.mediaType })
-        if (image.type === "url") return ImageInputs.decodeDataUrl(image.url, ADAPTER)
+        if (image.type === "url") return ImageInputs.decodeDataUrl(image.url)
         return Effect.undefined
       })
       const multipartMask =
@@ -118,7 +104,7 @@ export const model = (input: ModelInput) => {
           : mask.type === "bytes"
             ? { data: mask.data, mediaType: mask.mediaType }
             : mask.type === "url"
-              ? yield* ImageInputs.decodeDataUrl(mask.url, ADAPTER)
+              ? yield* ImageInputs.decodeDataUrl(mask.url)
               : undefined
       const useMultipart =
         sourceImages.length > 0 &&
@@ -161,7 +147,7 @@ export const model = (input: ModelInput) => {
         return undefined
       })
       if (references.some((image) => image === undefined))
-        return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts image URLs, data URLs, bytes, and file IDs")
+        return yield* ImageInputs.invalid("OpenAI Images accepts image URLs, data URLs, bytes, and file IDs")
       const maskReference =
         mask === undefined
           ? undefined
@@ -173,7 +159,7 @@ export const model = (input: ModelInput) => {
                 ? { file_id: mask.id }
                 : undefined
       if (mask !== undefined && maskReference === undefined)
-        return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts masks as URLs, data URLs, bytes, or file IDs")
+        return yield* ImageInputs.invalid("OpenAI Images accepts masks as URLs, data URLs, bytes, or file IDs")
       const requestBody = mergeJsonRecords(
         {
           model: request.model.id,
@@ -209,11 +195,9 @@ const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
   options: OpenAIImageOptions | undefined,
   overlay: Record<string, unknown> | undefined,
 ) {
-  const payload = yield* response.json.pipe(
-    Effect.mapError(() => invalidOutput("Failed to read the OpenAI Images response")),
-  )
-  const decoded = yield* Schema.decodeUnknownEffect(OpenAIImageResponse)(payload).pipe(
-    Effect.mapError(() => invalidOutput("OpenAI Images returned an invalid response")),
+  const output = yield* ProviderShared.imageResponse(ADAPTER, "OpenAI Images", response)
+  const decoded = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(OpenAIImageResponse))(output.body).pipe(
+    Effect.mapError((cause) => output.invalid("OpenAI Images returned an invalid response", cause)),
   )
   const requestBody = mergeJsonRecords(nativeOptions(options), overlay)
   const format =
@@ -221,7 +205,7 @@ const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
   const images = yield* Effect.forEach(decoded.data, (item, index) => {
     if (item.b64_json)
       return Effect.fromResult(Encoding.decodeBase64(item.b64_json)).pipe(
-        Effect.mapError(() => invalidOutput(`OpenAI Images result ${index} contains invalid base64 data`)),
+        Effect.mapError((cause) => output.invalid(`OpenAI Images result ${index} contains invalid base64 data`, cause)),
         Effect.map(
           (data) =>
             new GeneratedImage({
@@ -241,9 +225,9 @@ const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
             item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
         }),
       )
-    return Effect.fail(invalidOutput(`OpenAI Images result ${index} has neither image data nor a URL`))
+    return Effect.fail(output.invalid(`OpenAI Images result ${index} has neither image data nor a URL`))
   })
-  if (images.length === 0) return yield* invalidOutput("OpenAI Images returned no images")
+  if (images.length === 0) return yield* output.invalid("OpenAI Images returned no images")
   return new ImageResponse({
     images,
     usage:

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

@@ -7,7 +7,7 @@ import { Protocol } from "../route/protocol.js"
 import { HttpTransport } from "../route/transport/index.js"
 import { LLMRequest, type JsonSchema, type ToolDefinition } from "../schema/index.js"
 import { OpenResponses } from "./open-responses.js"
-import { optionalArray, ProviderShared } from "./shared.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"
@@ -32,6 +32,40 @@ const OpenAIResponsesImageGenerationTool = Schema.Struct({
   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 OpenAIResponsesTools = Schema.Union([OpenResponses.Tool, OpenAIResponsesImageGenerationTool])
 
 const OpenAIResponsesToolChoice = Schema.Union([
@@ -41,6 +75,7 @@ const OpenAIResponsesToolChoice = Schema.Union([
 
 const OpenAIResponsesCoreFields = {
   ...OpenResponses.coreFields,
+  input: Schema.Array(Schema.Union([OpenResponses.InputItem, OpenAIResponsesHostedToolItem])),
   tools: optionalArray(OpenAIResponsesTools),
   tool_choice: Schema.optional(OpenAIResponsesToolChoice),
 }
@@ -51,28 +86,10 @@ const OpenAIResponsesBody = Schema.Struct({
 })
 export type OpenAIResponsesBody = Schema.Schema.Type<typeof OpenAIResponsesBody>
 
-// Replayed items are paired with stored server state by id, so a foreign or
-// synthetic token can fail request validation even when `call_id` pairing is
-// intact. Only resend ids in each item kind's own grammar; hosted tool
-// references keep generic validation because every hosted tool mints its own
-// prefix. The same allowlist approach codex uses before resending history
-// (codex-rs core/src/client.rs, `prepare_response_items_for_request`).
-const ITEM_ID_PREFIXES: Record<OpenResponses.ItemKind, ReadonlyArray<string>> = {
-  message: ["msg_"],
-  reasoning: ["rs_"],
-  "function-call": ["fc_"],
-  // Every hosted tool mints its own id prefix, so references keep generic
-  // validation only.
-  reference: [],
-}
-
 const extension = {
   id: ADAPTER,
   name: NAME,
-  acceptsItemID: (kind: OpenResponses.ItemKind, id: string) => {
-    const prefixes = ITEM_ID_PREFIXES[kind]
-    return prefixes.length === 0 || prefixes.some((prefix) => id.startsWith(prefix))
-  },
+  lowerHostedToolItem: (item: unknown) => (Schema.is(OpenAIResponsesHostedToolItem)(item) ? item : undefined),
 } satisfies OpenResponses.Extension
 
 const nativeImageToolInput = (tool: ToolDefinition) => {
@@ -105,14 +122,18 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>, tool
         : { type: "function" as const, name },
   })
 
+const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenAIResponsesBody))
+
 const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request: LLMRequest) {
   const body = yield* OpenResponses.fromRequestWithExtension(
     LLMRequest.update(request, { tools: [], toolChoice: undefined }),
     extension,
   )
   const toolSchemaCompatibility = request.model.compatibility?.toolSchema
-  return {
+  const parallelToolCalls = OpenResponses.resolveParallelToolCalls(request)
+  return yield* decodeBody({
     ...body,
+    ...(parallelToolCalls === undefined ? {} : { parallel_tool_calls: parallelToolCalls }),
     tools:
       request.tools.length === 0
         ? undefined
@@ -121,14 +142,16 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
           ),
     tool_choice:
       body.tool_choice ?? (request.toolChoice ? yield* lowerToolChoice(request.toolChoice, request.tools) : undefined),
-  } satisfies OpenAIResponsesBody
+  })
 })
 
 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(() => ProviderShared.eventError(ADAPTER, "OpenAI Responses returned invalid image base64")),
+      Effect.mapError((cause) =>
+        ProviderShared.eventError(ADAPTER, "OpenAI Responses returned invalid image base64", undefined, cause),
+      ),
     )
     const format = item.output_format ?? "png"
     return {
@@ -162,9 +185,11 @@ const HOSTED_TOOLS = {
 } as const satisfies ResponsesHostedTools.Definitions
 
 const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
-  if (event.type === "response.reasoning_text.delta" || event.type === "response.reasoning_summary.delta")
-    return event.item_id
-      ? Effect.succeed(OpenResponses.onReasoningDelta(state, event, event.item_id))
+  if (event.type === "response.reasoning_text.delta")
+    return event.item_id !== undefined
+      ? Effect.succeed(
+          OpenResponses.onReasoningDelta(state, event, OpenResponses.outputItemID(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)
@@ -205,7 +230,7 @@ export const route = Route.make({
   endpoint,
   auth,
   transport,
-  defaults: { providerOptions: { store: false } },
+  defaults: { providerOptions: { store: false, include: ["reasoning.encrypted_content"] } },
 })
 
 export * as OpenAIResponses from "./openai-responses.js"

packages/ai/src/protocols/shared.ts

@@ -2,11 +2,12 @@ import { Buffer } from "node:buffer"
 import { Tool } from "@opencode-ai/schema/tool"
 import { Effect, Schema, Stream } from "effect"
 import * as Sse from "effect/unstable/encoding/Sse"
-import { Headers, HttpClientRequest } from "effect/unstable/http"
+import { Headers, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
 import {
-  InvalidProviderOutputReason,
-  InvalidRequestReason,
+  InvalidProviderOutputError,
+  InvalidRequestError,
   AIError,
+  HttpContext,
   type ContentPart,
   type LLMRequest,
   type MediaPart,
@@ -28,10 +29,10 @@ 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 clampPromptCacheKey = (key: string | undefined): string | undefined => {
-  if (key === undefined) return undefined
-  const chars = Array.from(key)
-  if (chars.length <= OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH) return key
+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("")
 }
 
@@ -96,17 +97,15 @@ export const sumTokens = (...values: ReadonlyArray<number | undefined>): number
   return values.reduce((acc: number, value) => acc + (value ?? 0), 0)
 }
 
-export const eventError = (route: string, message: string, raw?: string) =>
+export const eventError = (route: string, message: string, body?: string, cause?: unknown) =>
   new AIError({
-    module: "ProviderShared",
-    method: "stream",
-    reason: new InvalidProviderOutputReason({ route, message, raw }),
+    reason: new InvalidProviderOutputError({ route, message, body, cause }),
   })
 
 export const parseJson = (route: string, input: string, message: string) =>
   Effect.try({
     try: () => decodeJson(input),
-    catch: () => eventError(route, message, input),
+    catch: (cause) => eventError(route, message, input, cause),
   })
 
 /**
@@ -208,27 +207,41 @@ export const errorText = (error: unknown) => {
 
 /**
  * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
- * decoder, optionally filters named events, and drops empty / `[DONE]`
- * keep-alive events so the protocol event schema sees one JSON string per
- * element. The SSE channel emits a
- * `Retry` control event on its error channel; we drop it here (we don't
- * implement client-driven retries). Decoder failures become provider output
- * errors so the public error channel stays `AIError`.
+ * 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.pipeThroughChannel(Sse.decode()),
-    Stream.catchTag("Retry", () => Stream.empty),
-    Stream.catchTag("SseError", (error) => Stream.fail(eventError("sse", error.message))),
+    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]" || (events !== undefined && event.event !== "message")),
+        (event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
     ),
     Stream.map((event) => event.data),
   )
@@ -236,13 +249,39 @@ export const sseFraming = (
 /**
  * Canonical invalid-request constructor shared by protocol lowering.
  */
-export const invalidRequest = (message: string) =>
+export const invalidRequest = (message: string, cause?: unknown) =>
   new AIError({
-    module: "ProviderShared",
-    method: "request",
-    reason: new InvalidRequestReason({ message }),
+    reason: new InvalidRequestError({ message, cause }),
   })
 
+export const imageResponse = Effect.fn("ProviderShared.imageResponse")(function* (
+  route: string,
+  name: string,
+  response: HttpClientResponse.HttpClientResponse,
+) {
+  const http = new HttpContext({ url: response.request.url, status: response.status, headers: response.headers })
+  const body = yield* response.text.pipe(
+    Effect.mapError(
+      (cause) =>
+        new AIError({
+          reason: new InvalidProviderOutputError({
+            route,
+            message: `Failed to read the ${name} response`,
+            http,
+            cause,
+          }),
+        }),
+    ),
+  )
+  return {
+    body,
+    invalid: (message: string, cause?: unknown) =>
+      new AIError({
+        reason: new InvalidProviderOutputError({ route, message, body, http, cause }),
+      }),
+  }
+})
+
 export const matchToolChoice = <Auto, None, Required, Tool>(
   route: string,
   toolChoice: NonNullable<LLMRequest["toolChoice"]>,
@@ -289,7 +328,7 @@ export const unsupportedContent = (
 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)))
+    decode(payload).pipe(Effect.mapError((error) => invalidRequest(error.message, error)))
 
 /**
  * Build an HTTP POST with a JSON body. Sets `content-type: application/json`

packages/ai/src/protocols/utils/image-input.ts

@@ -1,12 +1,10 @@
 import { Effect, Encoding } from "effect"
 import type { ImageInput } from "../../image.js"
-import { InvalidRequestReason, AIError } from "../../schema/index.js"
+import { InvalidRequestError, AIError } from "../../schema/index.js"
 
-const invalid = (module: string, message: string) =>
+const invalid = (message: string, cause?: unknown) =>
   new AIError({
-    module,
-    method: "generate",
-    reason: new InvalidRequestReason({ message }),
+    reason: new InvalidRequestError({ message, cause }),
   })
 
 export const dataUrl = (input: Extract<ImageInput, { readonly type: "bytes" }>) =>
@@ -14,13 +12,12 @@ export const dataUrl = (input: Extract<ImageInput, { readonly type: "bytes" }>)
 
 export const decodeDataUrl = (
   url: string,
-  module: string,
 ): Effect.Effect<{ readonly mediaType: string; readonly data: Uint8Array } | undefined, AIError> => {
   if (!url.startsWith("data:")) return Effect.undefined
   const match = /^data:([^;,]+);base64,(.*)$/s.exec(url)
-  if (!match) return Effect.fail(invalid(module, "Image data URLs must contain a MIME type and base64 data"))
+  if (!match) return Effect.fail(invalid("Image data URLs must contain a MIME type and base64 data"))
   return Effect.fromResult(Encoding.decodeBase64(match[2])).pipe(
-    Effect.mapError(() => invalid(module, "Image data URL contains invalid base64 data")),
+    Effect.mapError((cause) => invalid("Image data URL contains invalid base64 data", cause)),
     Effect.map((data) => ({ mediaType: match[1], data })),
   )
 }

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

@@ -62,22 +62,31 @@ export const reasoningEnd = (
   events: LLMEvent[],
   id: string,
   providerMetadata?: ProviderMetadata,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text?: string,
 ): State => {
   if (!state.reasoning.has(id)) return state
   const stepped = stepStart(state, events)
-  events.push(LLMEvent.reasoningEnd({ id, providerMetadata }))
+  events.push(LLMEvent.reasoningEnd({ id, text, providerMetadata }))
   const reasoning = new Set(stepped.reasoning)
   reasoning.delete(id)
   return { ...stepped, reasoning }
 }
 
-export const textEnd = (state: State, events: LLMEvent[], id: string, providerMetadata?: ProviderMetadata): State => {
+export const textEnd = (
+  state: State,
+  events: LLMEvent[],
+  id: string,
+  providerMetadata?: ProviderMetadata,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text?: string,
+): State => {
   if (!state.text.has(id)) return state
   const stepped = stepStart(state, events)
-  events.push(LLMEvent.textEnd({ id, providerMetadata }))
-  const text = new Set(stepped.text)
-  text.delete(id)
-  return { ...stepped, text }
+  events.push(LLMEvent.textEnd({ id, text, providerMetadata }))
+  const open = new Set(stepped.text)
+  open.delete(id)
+  return { ...stepped, text: open }
 }
 
 const closeOpenBlocks = (state: State, events: LLMEvent[]): State => {

packages/ai/src/protocols/utils/open-responses-options.ts

@@ -29,10 +29,9 @@ export type ResponseIncludable = (typeof ResponseIncludables)[number] | (string
 
 export const ServiceTiers = ["auto", "default", "flex", "priority"] as const
 export type ServiceTier = (typeof ServiceTiers)[number] | (string & {})
-export const ServiceTier = Schema.declare<ServiceTier>(
-  (value): value is ServiceTier => typeof value === "string",
-  { title: "ServiceTier" },
-)
+export const ServiceTier = Schema.declare<ServiceTier>((value): value is ServiceTier => typeof value === "string", {
+  title: "ServiceTier",
+})
 
 export const Truncations = ["auto", "disabled"] as const
 export type Truncation = (typeof Truncations)[number]
@@ -56,7 +55,6 @@ export const StreamOptions = Schema.Struct({
 })
 
 export const Options = Schema.Struct({
-  instructions: Schema.optional(Schema.String),
   store: Schema.optional(Schema.Boolean),
   metadata: Schema.optional(Schema.Record(Schema.String, Schema.String)),
   safetyIdentifier: Schema.optional(Schema.String),

packages/ai/src/protocols/utils/partial-json-options.ts

@@ -0,0 +1,64 @@
+/*
+ * Adapted from partial-json by the Promplate Dev Team:
+ * https://github.com/promplate/partial-json-parser-js/blob/main/src/options.ts
+ * Licensed under the MIT License; see partial-json.ts for the complete notice.
+ */
+
+/**
+ * allow partial strings like `"hello \u12` to be parsed as `"hello `
+ */
+export const STR = 0b000000001
+
+/**
+ * allow partial numbers like `123.` to be parsed as `123`
+ */
+export const NUM = 0b000000010
+
+/**
+ * allow partial arrays like `[1, 2,` to be parsed as `[1, 2]`
+ */
+export const ARR = 0b000000100
+
+/**
+ * allow partial objects like `{"a": 1, "b":` to be parsed as `{"a": 1}`
+ */
+export const OBJ = 0b000001000
+
+/**
+ * allow `nu` to be parsed as `null`
+ */
+export const NULL = 0b000010000
+
+/**
+ * allow `tr` to be parsed as `true`, and `fa` to be parsed as `false`
+ */
+export const BOOL = 0b000100000
+
+/**
+ * allow `Na` to be parsed as `NaN`
+ */
+export const NAN = 0b001000000
+
+/**
+ * allow `Inf` to be parsed as `Infinity`
+ */
+export const INFINITY = 0b010000000
+
+/**
+ * allow `-Inf` to be parsed as `-Infinity`
+ */
+export const _INFINITY = 0b100000000
+
+export const INF = INFINITY | _INFINITY
+export const SPECIAL = NULL | BOOL | INF | NAN
+export const ATOM = STR | NUM | SPECIAL
+export const COLLECTION = ARR | OBJ
+export const ALL = ATOM | COLLECTION
+
+/**
+ * Control what types you allow to be partially parsed.
+ * The default is to allow all types to be partially parsed, which in most cases is the best option.
+ */
+export const Allow = { STR, NUM, ARR, OBJ, NULL, BOOL, NAN, INFINITY, _INFINITY, INF, SPECIAL, ATOM, COLLECTION, ALL }
+
+export default Allow

packages/ai/src/protocols/utils/partial-json.ts

@@ -0,0 +1,282 @@
+/*
+ * Adapted from partial-json by the Promplate Dev Team:
+ * https://github.com/promplate/partial-json-parser-js
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Promplate Dev Team
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+import { Schema } from "effect"
+import { Allow } from "./partial-json-options.js"
+export * from "./partial-json-options.js"
+
+export class PartialJSON extends Error {}
+export class MalformedJSON extends Error {}
+
+const decodeJson = Schema.decodeUnknownSync(Schema.fromJsonString(Schema.Unknown))
+
+/** Parse complete or incomplete JSON, restricted by the supplied partial-value flags. */
+export function parseJSON(jsonString: string, allowPartial = Allow.ALL): unknown {
+  if (typeof jsonString !== "string") throw new TypeError(`expecting str, got ${typeof jsonString}`)
+  const input = jsonString.trim()
+  if (!input) throw new Error(`${jsonString} is empty`)
+  try {
+    return decodeJson(input)
+  } catch {}
+
+  const repaired = repairJSON(input)
+  if (repaired !== input) {
+    try {
+      return decodeJson(repaired)
+    } catch {}
+  }
+
+  try {
+    return _parseJSON(input, allowPartial)
+  } catch (error) {
+    if (repaired !== input) return _parseJSON(repaired, allowPartial)
+    throw error
+  }
+}
+
+const repairJSON = (input: string) => {
+  let repaired = ""
+  let quoted = false
+
+  for (let index = 0; index < input.length; index++) {
+    const character = input[index]
+    if (!quoted) {
+      repaired += character
+      if (character === '"') quoted = true
+      continue
+    }
+
+    if (character === '"') {
+      repaired += character
+      quoted = false
+      continue
+    }
+
+    if (character === "\\") {
+      const next = input[index + 1]
+      if (next === "u" && /^[0-9a-fA-F]{4}$/.test(input.slice(index + 2, index + 6))) {
+        repaired += input.slice(index, index + 6)
+        index += 5
+        continue
+      }
+      if (next !== undefined && '"\\/bfnrtu'.includes(next)) {
+        repaired += `\\${next}`
+        index++
+        continue
+      }
+      repaired += "\\\\"
+      continue
+    }
+
+    const code = character.charCodeAt(0)
+    repaired += code <= 0x1f ? `\\u${code.toString(16).padStart(4, "0")}` : character
+  }
+
+  return repaired
+}
+
+const _parseJSON = (jsonString: string, allow: number) => {
+  const length = jsonString.length
+  let index = 0
+
+  const markPartialJSON = (message: string): never => {
+    throw new PartialJSON(`${message} at position ${index}`)
+  }
+
+  const throwMalformedError = (message: string): never => {
+    throw new MalformedJSON(`${message} at position ${index}`)
+  }
+
+  const parseAny = (): unknown => {
+    skipBlank()
+    if (index >= length) markPartialJSON("Unexpected end of input")
+    if (jsonString[index] === '"') return parseStr()
+    if (jsonString[index] === "{") return parseObj()
+    if (jsonString[index] === "[") return parseArr()
+    if (
+      jsonString.substring(index, index + 4) === "null" ||
+      (Allow.NULL & allow && length - index < 4 && "null".startsWith(jsonString.substring(index)))
+    ) {
+      index += 4
+      return null
+    }
+    if (
+      jsonString.substring(index, index + 4) === "true" ||
+      (Allow.BOOL & allow && length - index < 4 && "true".startsWith(jsonString.substring(index)))
+    ) {
+      index += 4
+      return true
+    }
+    if (
+      jsonString.substring(index, index + 5) === "false" ||
+      (Allow.BOOL & allow && length - index < 5 && "false".startsWith(jsonString.substring(index)))
+    ) {
+      index += 5
+      return false
+    }
+    if (
+      jsonString.substring(index, index + 8) === "Infinity" ||
+      (Allow.INFINITY & allow && length - index < 8 && "Infinity".startsWith(jsonString.substring(index)))
+    ) {
+      index += 8
+      return Infinity
+    }
+    if (
+      jsonString.substring(index, index + 9) === "-Infinity" ||
+      (Allow._INFINITY & allow &&
+        1 < length - index &&
+        length - index < 9 &&
+        "-Infinity".startsWith(jsonString.substring(index)))
+    ) {
+      index += 9
+      return -Infinity
+    }
+    if (
+      jsonString.substring(index, index + 3) === "NaN" ||
+      (Allow.NAN & allow && length - index < 3 && "NaN".startsWith(jsonString.substring(index)))
+    ) {
+      index += 3
+      return NaN
+    }
+    return parseNum()
+  }
+
+  const parseStr = (): string => {
+    const start = index
+    let escape = false
+    index++
+    while (index < length && (jsonString[index] !== '"' || (escape && jsonString[index - 1] === "\\"))) {
+      escape = jsonString[index] === "\\" ? !escape : false
+      index++
+    }
+    if (jsonString.charAt(index) === '"') {
+      try {
+        return decodeJson(jsonString.substring(start, ++index - Number(escape))) as string
+      } catch (error) {
+        throwMalformedError(String(error))
+      }
+    }
+    if (Allow.STR & allow) {
+      try {
+        return decodeJson(`${jsonString.substring(start, index - Number(escape))}"`) as string
+      } catch {
+        return decodeJson(`${jsonString.substring(start, jsonString.lastIndexOf("\\"))}"`) as string
+      }
+    }
+    return markPartialJSON("Unterminated string literal")
+  }
+
+  const parseObj = (): Record<string, unknown> => {
+    index++
+    skipBlank()
+    const object: Record<string, unknown> = {}
+    try {
+      while (jsonString[index] !== "}") {
+        skipBlank()
+        if (index >= length && Allow.OBJ & allow) return object
+        const key = parseStr()
+        skipBlank()
+        index++
+        try {
+          Object.defineProperty(object, key, {
+            value: parseAny(),
+            enumerable: true,
+            configurable: true,
+            writable: true,
+          })
+        } catch (error) {
+          if (Allow.OBJ & allow) return object
+          throw error
+        }
+        skipBlank()
+        if (jsonString[index] === ",") index++
+      }
+    } catch {
+      if (Allow.OBJ & allow) return object
+      return markPartialJSON("Expected '}' at end of object")
+    }
+    index++
+    return object
+  }
+
+  const parseArr = (): unknown[] => {
+    index++
+    const array: unknown[] = []
+    try {
+      while (jsonString[index] !== "]") {
+        array.push(parseAny())
+        skipBlank()
+        if (jsonString[index] === ",") index++
+      }
+    } catch {
+      if (Allow.ARR & allow) return array
+      return markPartialJSON("Expected ']' at end of array")
+    }
+    index++
+    return array
+  }
+
+  const parseNum = (): unknown => {
+    if (index === 0) {
+      if (jsonString === "-") throwMalformedError("Not sure what '-' is")
+      try {
+        return decodeJson(jsonString)
+      } catch (error) {
+        if (Allow.NUM & allow) {
+          try {
+            return decodeJson(jsonString.substring(0, jsonString.lastIndexOf("e")))
+          } catch {}
+        }
+        throwMalformedError(String(error))
+      }
+    }
+
+    const start = index
+    if (jsonString[index] === "-") index++
+    while (jsonString[index] && !",]}".includes(jsonString[index])) index++
+    if (index === length && !(Allow.NUM & allow)) markPartialJSON("Unterminated number literal")
+
+    try {
+      return decodeJson(jsonString.substring(start, index))
+    } catch (error) {
+      if (jsonString.substring(start, index) === "-") markPartialJSON("Not sure what '-' is")
+      try {
+        return decodeJson(jsonString.substring(start, jsonString.lastIndexOf("e")))
+      } catch {
+        throwMalformedError(String(error))
+      }
+    }
+  }
+
+  const skipBlank = () => {
+    while (index < length && " \n\r\t".includes(jsonString[index])) index++
+  }
+
+  return parseAny()
+}
+
+export const parse = parseJSON

packages/ai/src/protocols/utils/responses-hosted-tools.ts

@@ -34,37 +34,35 @@ export const onDone: (
   state: OpenResponses.ParserState,
   item: Item,
   tools: Definitions,
-) => Effect.Effect<OpenResponses.StepResult, AIError> = Effect.fn("ResponsesHostedTools.onDone")(function* (
-  state,
-  item,
-  tools,
-) {
-  const tool = tools[item.type]
-  if (!tool) return [state, []] satisfies OpenResponses.StepResult
-  const providerMetadata = OpenResponses.providerMetadata(state, { itemId: item.id })
-  const events: LLMEvent[] = []
-  const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
-  events.push(
-    LLMEvent.toolCall({
-      id: item.id,
-      name: tool.name,
-      input: tool.input(item),
-      providerExecuted: true,
-      providerMetadata,
-    }),
-    LLMEvent.toolResult({
-      id: item.id,
-      name: tool.name,
-      result: tool.result
-        ? yield* tool.result(item)
-        : item.error !== undefined && item.error !== null
-          ? { type: "error", value: item.error }
-          : { type: "json", value: item },
-      providerExecuted: true,
-      providerMetadata,
-    }),
-  )
-  return [{ ...state, lifecycle }, events] satisfies OpenResponses.StepResult
-})
+) => Effect.Effect<OpenResponses.StepResult, AIError> = Effect.fn("ResponsesHostedTools.onDone")(
+  function* (state, item, tools) {
+    const tool = tools[item.type]
+    if (!tool) return [state, []] satisfies OpenResponses.StepResult
+    const providerMetadata = OpenResponses.providerMetadata(state, { itemId: item.id })
+    const events: LLMEvent[] = []
+    const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
+    events.push(
+      LLMEvent.toolCall({
+        id: item.id,
+        name: tool.name,
+        input: tool.input(item),
+        providerExecuted: true,
+        providerMetadata,
+      }),
+      LLMEvent.toolResult({
+        id: item.id,
+        name: tool.name,
+        result: tool.result
+          ? yield* tool.result(item)
+          : item.error !== undefined && item.error !== null
+            ? { type: "error", value: item.error }
+            : { type: "json", value: item },
+        providerExecuted: true,
+        providerMetadata,
+      }),
+    )
+    return [{ ...state, lifecycle }, events] satisfies OpenResponses.StepResult
+  },
+)
 
 export * as ResponsesHostedTools from "./responses-hosted-tools.js"

packages/ai/src/protocols/utils/tool-stream.ts

@@ -1,8 +1,10 @@
-import { Effect } from "effect"
-import { AIError, LLMEvent, type ProviderMetadata, type ToolCall, type ToolInputError } from "../../schema/index.js"
+import { Effect, Option } from "effect"
+import { AIError, LLMEvent, type ProviderMetadata, type ToolCall } from "../../schema/index.js"
 import { eventError, parseToolInput, type ToolAccumulator } from "../shared.js"
+import { parse } from "./partial-json.js"
 
 type StreamKey = string | number
+const parsePartialInput = Option.liftThrowable(parse)
 
 /**
  * One pending streamed tool call. Providers emit the tool identity and JSON
@@ -62,38 +64,39 @@ const inputDelta = (tool: PendingTool, text: string) =>
     id: tool.id,
     name: tool.name,
     text,
+    input: Option.getOrElse(parsePartialInput(tool.input), () => ({})),
   })
 
 const toolCall = (route: string, tool: PendingTool, inputOverride?: string) => {
   const raw = inputOverride ?? tool.input
   return parseToolInput(route, tool.name, raw).pipe(
-    Effect.map((input): ToolCall | ToolInputError =>
-      LLMEvent.toolCall({
-        id: tool.id,
-        name: tool.name,
-        input,
-        providerExecuted: tool.providerExecuted ? true : undefined,
-        providerMetadata: tool.providerMetadata,
-      }),
-    ),
     Effect.catch((error) =>
       tool.providerExecuted
         ? Effect.fail(error)
         : Effect.succeed(
-            LLMEvent.toolInputError({
-              id: tool.id,
-              name: tool.name,
-              raw,
-            }),
+            Option.getOrElse(
+              Option.map(parsePartialInput(raw), (input) => input ?? {}),
+              () => ({}),
+            ),
           ),
     ),
+    Effect.map(
+      (input): ToolCall =>
+        LLMEvent.toolCall({
+          id: tool.id,
+          name: tool.name,
+          input,
+          providerExecuted: tool.providerExecuted ? true : undefined,
+          providerMetadata: tool.providerMetadata,
+        }),
+    ),
   )
 }
 
-const finishEvents = (tool: PendingTool, event: ToolCall | ToolInputError): ReadonlyArray<LLMEvent> =>
-  event.type === "tool-input-error"
-    ? [event]
-    : [LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }), event]
+const finishEvents = (tool: PendingTool, event: ToolCall): ReadonlyArray<LLMEvent> => [
+  LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }),
+  event,
+]
 
 /** Store the updated tool and produce the optional public delta event. */
 const appendTool = <K extends StreamKey>(
@@ -176,7 +179,7 @@ export const appendExisting = <K extends StreamKey>(
 
 /**
  * Finalize one pending tool call: parse the accumulated raw JSON, remove it
- * from state, and return either a call or a non-executable local input error.
+ * from state, and recover incomplete local arguments when needed.
  * Missing keys are a no-op because some providers emit stop events for
  * non-tool content blocks.
  */

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

@@ -2,14 +2,7 @@ import { Effect, Encoding, Schema } from "effect"
 import { Headers, HttpClientRequest } from "effect/unstable/http"
 import { GeneratedImage, ImageModel, ImageResponse, type ImageRequestFor, type ImageRoute } from "../image.js"
 import { Auth, type Definition as AuthDefinition } from "../route/auth.js"
-import {
-  InvalidProviderOutputReason,
-  AIError,
-  Usage,
-  mergeHttpOptions,
-  mergeJsonRecords,
-  type HttpOptions,
-} from "../schema/index.js"
+import { Usage, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema/index.js"
 import { ProviderShared, optionalNull } from "./shared.js"
 import { ImageInputs } from "./utils/image-input.js"
 
@@ -94,13 +87,6 @@ const nativeOptions = (options: XAIImageOptions | undefined) => {
   }
 }
 
-const invalidOutput = (message: string) =>
-  new AIError({
-    module: ADAPTER,
-    method: "generate",
-    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
-  })
-
 const applyQuery = (url: string, query: Record<string, string> | undefined) => {
   if (!query) return url
   const next = new URL(url)
@@ -120,7 +106,7 @@ export const model = (input: ModelInput) => {
         return undefined
       })
       if (imageReferences.some((image) => image === undefined))
-        return yield* ImageInputs.invalid(ADAPTER, "xAI Images accepts image URLs, data URLs, bytes, and file IDs")
+        return yield* ImageInputs.invalid("xAI Images accepts image URLs, data URLs, bytes, and file IDs")
       const requestBody = mergeJsonRecords(
         {
           model: request.model.id,
@@ -149,17 +135,17 @@ export const model = (input: ModelInput) => {
           HttpClientRequest.bodyText(text, "application/json"),
         ),
       )
-      const payload = yield* response.json.pipe(
-        Effect.mapError(() => invalidOutput("Failed to read the xAI Images response")),
-      )
-      const decoded = yield* Schema.decodeUnknownEffect(XAIImageResponse)(payload).pipe(
-        Effect.mapError(() => invalidOutput("xAI Images returned an invalid response")),
+      const output = yield* ProviderShared.imageResponse(ADAPTER, "xAI Images", response)
+      const decoded = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(XAIImageResponse))(output.body).pipe(
+        Effect.mapError((cause) => output.invalid("xAI Images returned an invalid response", cause)),
       )
       const images = yield* Effect.forEach(decoded.data, (item, index) => {
         const mediaType = item.mime_type ?? "application/octet-stream"
         if (item.b64_json)
           return Effect.fromResult(Encoding.decodeBase64(item.b64_json)).pipe(
-            Effect.mapError(() => invalidOutput(`xAI Images result ${index} contains invalid base64 data`)),
+            Effect.mapError((cause) =>
+              output.invalid(`xAI Images result ${index} contains invalid base64 data`, cause),
+            ),
             Effect.map(
               (data) =>
                 new GeneratedImage({
@@ -183,9 +169,9 @@ export const model = (input: ModelInput) => {
                   : { xai: { revisedPrompt: item.revised_prompt } },
             }),
           )
-        return Effect.fail(invalidOutput(`xAI Images result ${index} has neither image data nor a URL`))
+        return Effect.fail(output.invalid(`xAI Images result ${index} has neither image data nor a URL`))
       })
-      if (images.length === 0) return yield* invalidOutput("xAI Images returned no images")
+      if (images.length === 0) return yield* output.invalid("xAI Images returned no images")
       const usage = ProviderShared.isRecord(decoded.usage) ? decoded.usage : undefined
       return new ImageResponse({
         images,

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

@@ -1,15 +1,52 @@
+import { Effect, Schema } from "effect"
 import { Protocol } from "../route/protocol.js"
+import type { LLMRequest } from "../schema/index.js"
 import { OpenResponses } from "./open-responses.js"
+import { JsonObject, optionalNull, ProviderShared } from "./shared.js"
 import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
 
 const ADAPTER = "xai-responses"
 const NAME = "xAI Responses"
 
+const XAIResponsesHostedToolItem = Schema.Union([
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("x_search_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: Schema.optional(Schema.Unknown),
+      error: Schema.optional(Schema.Unknown),
+    }),
+    [JsonObject],
+  ),
+])
+
+const XAIResponsesBody = Schema.Struct({
+  ...OpenResponses.coreFields,
+  input: Schema.Array(Schema.Union([OpenResponses.InputItem, XAIResponsesHostedToolItem])),
+  stream: Schema.Literal(true),
+})
+
 const extension = {
   id: ADAPTER,
   name: NAME,
+  lowerHostedToolItem: (item: unknown) => (Schema.is(XAIResponsesHostedToolItem)(item) ? item : undefined),
 } satisfies OpenResponses.Extension
 
+const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(XAIResponsesBody))
+const fromRequest = Effect.fn("XAIResponses.fromRequest")(function* (request: LLMRequest) {
+  return yield* decodeBody(yield* OpenResponses.fromRequestWithExtension(request, extension))
+})
+
 const HOSTED_TOOLS = {
   web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
   x_search_call: { name: "x_search", input: (item) => item.action ?? {} },
@@ -35,7 +72,10 @@ const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
 
 export const protocol = Protocol.make({
   id: ADAPTER,
-  body: OpenResponses.protocol.body,
+  body: {
+    schema: XAIResponsesBody,
+    from: fromRequest,
+  },
   stream: {
     event: OpenResponses.protocol.stream.event,
     initial: (request) => OpenResponses.initial(request, extension),

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

@@ -2,13 +2,7 @@ import { Effect, Schema } from "effect"
 import { Headers, HttpClientRequest } from "effect/unstable/http"
 import { GeneratedImage, ImageModel, ImageResponse, type ImageRequestFor, type ImageRoute } from "../image.js"
 import { Auth, type Definition as AuthDefinition } from "../route/auth.js"
-import {
-  InvalidProviderOutputReason,
-  AIError,
-  mergeHttpOptions,
-  mergeJsonRecords,
-  type HttpOptions,
-} from "../schema/index.js"
+import { mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema/index.js"
 import { ProviderShared } from "./shared.js"
 import { ImageInputs } from "./utils/image-input.js"
 
@@ -63,13 +57,6 @@ const nativeOptions = (options: ZAIImageOptions | undefined) => {
   }
 }
 
-const invalidOutput = (message: string) =>
-  new AIError({
-    module: ADAPTER,
-    method: "generate",
-    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
-  })
-
 const applyQuery = (url: string, query: Record<string, string> | undefined) => {
   if (!query) return url
   const next = new URL(url)
@@ -82,7 +69,7 @@ export const model = (input: ModelInput) => {
     id: ADAPTER,
     generate: Effect.fn("ZAIImages.generate")(function* (request: ImageRequestFor<ZAIImageOptions>, execute) {
       if ((request.images?.length ?? 0) > 0)
-        return yield* ImageInputs.invalid(ADAPTER, "Z.ai hosted image generation does not support image inputs")
+        return yield* ImageInputs.invalid("Z.ai hosted image generation does not support image inputs")
       const http = mergeHttpOptions(request.model.http, request.http)
       const requestBody = mergeJsonRecords(
         { model: request.model.id, prompt: request.prompt },
@@ -104,13 +91,11 @@ export const model = (input: ModelInput) => {
           HttpClientRequest.bodyText(text, "application/json"),
         ),
       )
-      const payload = yield* response.json.pipe(
-        Effect.mapError(() => invalidOutput("Failed to read the Z.ai Images response")),
+      const output = yield* ProviderShared.imageResponse(ADAPTER, "Z.ai Images", response)
+      const decoded = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(ZAIImageResponse))(output.body).pipe(
+        Effect.mapError((cause) => output.invalid("Z.ai Images returned an invalid response", cause)),
       )
-      const decoded = yield* Schema.decodeUnknownEffect(ZAIImageResponse)(payload).pipe(
-        Effect.mapError(() => invalidOutput("Z.ai Images returned an invalid response")),
-      )
-      if (decoded.data.length === 0) return yield* invalidOutput("Z.ai Images returned no images")
+      if (decoded.data.length === 0) return yield* output.invalid("Z.ai Images returned no images")
       return new ImageResponse({
         images: decoded.data.map(
           (item) =>

packages/ai/src/provider-error.ts

@@ -1,17 +1,16 @@
 import { Option, Schema } from "effect"
 import {
-  AuthenticationReason,
-  ContentPolicyReason,
-  InvalidRequestReason,
+  AuthenticationError,
+  ContentPolicyError,
+  InvalidRequestError,
   AIError,
   ProviderErrorEvent,
-  ProviderInternalReason,
-  QuotaExceededReason,
-  RateLimitReason,
-  UnknownProviderReason,
+  ProviderInternalError,
+  QuotaExceededError,
+  RateLimitError,
+  UnknownProviderError,
   type HttpContext,
   type HttpRateLimitDetails,
-  type ProviderMetadata,
 } from "./schema/index.js"
 
 const patterns = [
@@ -38,17 +37,19 @@ const patterns = [
   /too large for model with \d+ maximum context length/i,
   /prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i,
   /model_context_window_exceeded/i,
+  /range of input length should be/i,
   /too many tokens/i,
   /token limit exceeded/i,
+  /request_too_large/i,
 ]
 
-const payloadPatterns = [/request_too_large/i, /request entity too large/i, /payload too large/i, /request too large/i]
+const payloadPatterns = [/request entity too large/i, /payload too large/i, /request too large/i]
 
 const exclusions = [/^(throttling error|service unavailable):/i, /rate limit/i, /too many requests/i]
 
 export const isContextOverflow = (message: string) =>
   !exclusions.some((pattern) => pattern.test(message)) &&
-  (patterns.some((pattern) => pattern.test(message)) || /^400\s*(status code)?\s*\(no body\)/i.test(message))
+  (patterns.some((pattern) => pattern.test(message)) || /^4(?:00|13)\s*(status code)?\s*\(no body\)/i.test(message))
 
 export const isPayloadTooLarge = (message: string) => payloadPatterns.some((pattern) => pattern.test(message))
 
@@ -59,6 +60,7 @@ export const isContextOverflowFailure = (failure: unknown) =>
 
 const decodeJson = Schema.decodeUnknownOption(Schema.fromJsonString(Schema.Unknown))
 const QUOTA_CODES = new Set(["insufficient_quota", "usage_not_included", "billing_error"])
+const AUTH_CODES = new Set(["authentication_error", "permission_error"])
 const SERVER_CODES = new Set([
   "api_error",
   "internal_error",
@@ -74,96 +76,91 @@ const INVALID_REQUEST_CODES = new Set(["invalid_prompt", "invalid_request_error"
 const RATE_LIMIT_TEXT = /rate increased too quickly|rate[-_\s]?limit|too[_\s]?many[_\s]?requests/i
 const QUOTA_TEXT = /insufficient[-_\s]?quota|quota[-_\s]?exceeded/i
 const CONTENT_POLICY_TEXT = /content[-_\s]?policy|content_filter|safety/i
-const NETWORK_ERROR_TEXT = /network[-_\s]error/i
 
 export interface ProviderFailure {
   readonly message: string
   readonly status?: number | undefined
-  readonly code?: string | undefined
   // Raw wire payload, scanned for failure signals (codes, overflow phrases)
   // that the summary message does not carry. Not shown to users.
   readonly rawBody?: string | undefined
+  // Some SDKs supply parsed error data separately from the original response text.
+  readonly data?: unknown
+  readonly http?: HttpContext | undefined
+  readonly cause?: unknown
   readonly retryAfterMs?: number | undefined
   readonly rateLimit?: HttpRateLimitDetails | undefined
-  readonly http?: HttpContext | undefined
-  readonly providerMetadata?: ProviderMetadata | undefined
 }
 
-// Keep HTTP failures and provider-reported stream failures on one typed path so
-// session retry policy never needs provider-specific string matching.
+// Classification records affirmative evidence about a failure. Deterministic
+// failures need positive identification (a 4xx status, quota/auth/policy
+// signals); anything unrecognized stays UnknownProvider, which the session
+// retry policy treats as retry-eligible because transient failures arrive in
+// unpredictable shapes while deterministic rejections almost always carry a
+// status or known code.
 export function classifyProviderFailure(input: ProviderFailure): AIError["reason"] {
-  const body = input.http?.body ?? input.rawBody ?? ""
-  const codes = [input.code, ...providerCodes(body), ...providerCodes(input.message)]
-    .filter((code): code is string => code !== undefined)
-    .map((code) => code.toLowerCase())
+  const details = { message: input.message, body: input.rawBody, http: input.http, cause: input.cause }
+  const body = input.rawBody ?? ""
+  const codes = [...providerCodes(input.data), ...providerCodes(body), ...providerCodes(input.message)].map((code) =>
+    code.toLowerCase(),
+  )
   // Scan the raw payload too so signals missing from the summary message
   // (e.g. overflow phrases nested in a JSON error body) still classify.
   const text = [input.message, body].filter((value) => value.length > 0).join("\n")
-  const common = { message: input.message, providerMetadata: input.providerMetadata, http: input.http }
   const clientScoped = input.status === undefined || (input.status >= 400 && input.status < 500)
 
   if (
     clientScoped &&
     (codes.includes("context_length_exceeded") ||
       codes.includes("model_context_window_exceeded") ||
+      codes.includes("request_too_large") ||
       isContextOverflow(text))
   )
-    return new InvalidRequestReason({ ...common, classification: "context-overflow" })
+    return new InvalidRequestError({ ...details, classification: "context-overflow" })
   if (input.status === 413 || isPayloadTooLarge(text))
-    return new InvalidRequestReason({ ...common, classification: "payload-too-large" })
-  if (CONTENT_POLICY_TEXT.test(text)) return new ContentPolicyReason(common)
+    return new InvalidRequestError({ ...details, classification: "payload-too-large" })
+  if (CONTENT_POLICY_TEXT.test(text)) return new ContentPolicyError(details)
   if (codes.some((code) => QUOTA_CODES.has(code)) || (input.status === 429 && QUOTA_TEXT.test(text)))
-    return new QuotaExceededReason(common)
-  if (input.status === 401) return new AuthenticationReason({ ...common, kind: "invalid" })
-  if (input.status === 403) return new AuthenticationReason({ ...common, kind: "insufficient-permissions" })
-  if (codes.includes("authentication_error")) return new AuthenticationReason({ ...common, kind: "invalid" })
-  if (codes.includes("permission_error"))
-    return new AuthenticationReason({ ...common, kind: "insufficient-permissions" })
+    return new QuotaExceededError(details)
+  if (input.status === 401 || input.status === 403 || codes.some((code) => AUTH_CODES.has(code)))
+    return new AuthenticationError(details)
   if (
-    codes.some((code) => code.includes("rate_limit") || code === "too_many_requests" || code === "throttlingexception")
+    input.status === 429 ||
+    codes.some(
+      (code) => code.includes("rate_limit") || code === "too_many_requests" || code === "throttlingexception",
+    ) ||
+    RATE_LIMIT_TEXT.test(text)
   )
-    return new RateLimitReason({
-      ...common,
+    return new RateLimitError({
+      ...details,
       retryAfterMs: input.retryAfterMs,
       rateLimit: input.rateLimit,
     })
-  if (RATE_LIMIT_TEXT.test(text))
-    return new RateLimitReason({
-      ...common,
-      retryAfterMs: input.retryAfterMs,
-      rateLimit: input.rateLimit,
-    })
-  if (NETWORK_ERROR_TEXT.test(text)) return new ProviderInternalReason({ ...common, status: input.status })
-  if (codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable")))
-    return new ProviderInternalReason({
-      ...common,
-      status: input.status,
+  if (
+    input.status === 408 ||
+    input.status === 409 ||
+    (input.status !== undefined && input.status >= 500) ||
+    codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable"))
+  )
+    return new ProviderInternalError({
+      ...details,
       retryAfterMs: input.retryAfterMs,
     })
-  if (input.status === 429) {
-    return new RateLimitReason({
-      ...common,
-      retryAfterMs: input.retryAfterMs,
-      rateLimit: input.rateLimit,
-    })
-  }
-  if (input.status === 408 || input.status === 409 || (input.status !== undefined && input.status >= 500))
-    return new ProviderInternalReason({
-      ...common,
-      status: input.status,
-      retryAfterMs: input.retryAfterMs,
-    })
-  if (codes.some((code) => INVALID_REQUEST_CODES.has(code))) return new InvalidRequestReason(common)
-  if (input.status === 400 || input.status === 404 || input.status === 413 || input.status === 422)
-    return new InvalidRequestReason(common)
-  return new UnknownProviderReason({ ...common, status: input.status })
+  if (codes.some((code) => INVALID_REQUEST_CODES.has(code))) return new InvalidRequestError(details)
+  // Any remaining 4xx is a deterministic rejection of this request.
+  if (input.status !== undefined && input.status >= 400 && input.status < 500) return new InvalidRequestError(details)
+  return new UnknownProviderError(details)
 }
 
-function providerCodes(value: string) {
-  const decoded = Option.getOrUndefined(decodeJson(value))
+function providerCodes(value: unknown) {
+  const decoded = typeof value === "string" ? Option.getOrUndefined(decodeJson(value)) : value
   if (!isRecord(decoded)) return []
   const error = isRecord(decoded.error) ? decoded.error : undefined
-  return [decoded.code, error?.code, error?.type].filter((value): value is string => typeof value === "string")
+  const response = isRecord(decoded.response) ? decoded.response : undefined
+  const responseError = response && isRecord(response.error) ? response.error : undefined
+  const exception = isRecord(decoded.exception) ? decoded.exception : undefined
+  return [decoded.code, error?.code, error?.type, error?.status, responseError?.code, exception?.type].filter(
+    (value): value is string => typeof value === "string",
+  )
 }
 
 function isRecord(value: unknown): value is Record<string, unknown> {

packages/ai/src/providers/amazon-bedrock-mantle.ts

@@ -1,5 +1,5 @@
 import { Auth } from "../route/auth.js"
-import type { Route as RouteDef, RouteDefaultsInput } from "../route/client.js"
+import { Route, type RouteDefaultsInput } from "../route/client.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { OpenAIChat } from "../protocols/openai-chat.js"
 import { OpenAIResponses } from "../protocols/openai-responses.js"
@@ -23,22 +23,30 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly credentials?: Credentials
   readonly region?: string
+  readonly topP?: number
   readonly providerOptions?: OpenAIProviderOptionsInput
 }
 
-const responsesRoute = OpenAIResponses.route.with({
+const responsesRoute = Route.make({
   id: "bedrock-mantle-responses",
   provider: id,
+  providerMetadataKey: "mantle",
+  protocol: OpenAIResponses.protocol,
+  endpoint: OpenAIResponses.route.endpoint,
+  auth: OpenAIResponses.route.auth,
+  transport: OpenAIResponses.httpTransport,
+  defaults: OpenAIResponses.route.defaults,
 })
 
 const chatRoute = OpenAIChat.route.with({
   id: "bedrock-mantle-chat",
   provider: id,
+  providerMetadataKey: "mantle",
 })
 
 export const routes = [responsesRoute, chatRoute]
 
-const configuredRoute = <Body, Prepared>(route: RouteDef<Body, Prepared>, input: Config) => {
+const configuredRoute = <Body, Prepared>(route: Route<Body, Prepared>, input: Config) => {
   const region = input.region ?? input.credentials?.region ?? "us-east-1"
   const credentials = input.credentials === undefined ? undefined : { ...input.credentials, region }
   return route.with({
@@ -70,7 +78,7 @@ export const configure = (input: Config = {}) => {
 
   return {
     id,
-    model: chat,
+    model: responses,
     chat,
     responses,
     configure,
@@ -88,6 +96,7 @@ const config = (settings: Settings): Config => {
     apiKey: settings.auth === "sigv4" ? undefined : settings.apiKey,
     baseURL: settings.baseURL,
     credentials: settings.credentials,
+    generation: settings.topP === undefined ? undefined : { topP: settings.topP },
     headers: settings.headers === undefined ? undefined : { ...settings.headers },
     http: settings.body === undefined ? undefined : { body: { ...settings.body } },
     providerOptions: settings.providerOptions,
@@ -103,4 +112,4 @@ export const responsesModel: ProviderPackage.Definition<Settings, OpenAIProvider
   modelID,
   settings,
 ) => configure(config(settings)).responses(modelID)
-export const model = chatModel
+export const model = responsesModel

packages/ai/src/providers/amazon-bedrock.ts

@@ -35,6 +35,7 @@ const configuredRoute = (input: Config) => {
   return BedrockConverse.route.with({
     ...rest,
     provider: id,
+    providerMetadataKey: "bedrock",
     endpoint: { baseURL: baseURL ?? bedrockBaseURL(resolvedRegion) },
     auth: apiKey === undefined ? BedrockConverse.sigV4Auth(credentials) : Auth.bearer(apiKey),
   })

packages/ai/src/providers/amazon-bedrock/mantle.ts

@@ -1,2 +1,2 @@
-export { chatModel as model } from "../amazon-bedrock-mantle.js"
+export { responsesModel as model } from "../amazon-bedrock-mantle.js"
 export type { Settings } from "../amazon-bedrock-mantle.js"

packages/ai/src/providers/cerebras.ts

@@ -0,0 +1,58 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("cerebras")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "cerebras-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.cerebras.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.cerebras.baseURL },
+    auth: AuthOptions.bearer(input, "CEREBRAS_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning", supportsStore: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)

packages/ai/src/providers/deepinfra.ts

@@ -0,0 +1,61 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("deepinfra")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "deepinfra-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.deepinfra.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const root = baseURL?.replace(/\/+$/, "")
+  const configured = route.with({
+    ...defaults,
+    endpoint: {
+      baseURL: root === undefined ? profiles.deepinfra.baseURL : root.endsWith("/openai") ? root : `${root}/openai`,
+    },
+    auth: AuthOptions.bearer(input, "DEEPINFRA_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning_content", supportsStore: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)

packages/ai/src/providers/google-vertex-chat.ts

@@ -27,6 +27,7 @@ export interface Settings extends ProviderPackage.Settings {
 const route = OpenAICompatibleChat.route.with({
   id: "google-vertex-chat",
   provider: id,
+  providerMetadataKey: "vertex",
 })
 
 export const routes = [route]

packages/ai/src/providers/google-vertex-responses.ts

@@ -27,6 +27,7 @@ export interface Settings extends ProviderPackage.Settings {
 const route = OpenAICompatibleResponses.route.with({
   id: "google-vertex-responses",
   provider: id,
+  providerMetadataKey: "vertex",
   providerOptions: { store: false },
 })
 

packages/ai/src/providers/google-vertex.ts

@@ -68,7 +68,7 @@ const protocol = {
 const route = Route.make({
   id: "google-vertex-gemini",
   provider: id,
-  providerMetadataKey: "google",
+  providerMetadataKey: "vertex",
   protocol,
   endpoint: Endpoint.path(({ request }) => {
     const model = String(request.model.id)

packages/ai/src/providers/groq.ts

@@ -0,0 +1,115 @@
+import { Effect, Schema } from "effect"
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAIChat } from "../protocols/openai-chat.js"
+import { ProviderShared } from "../protocols/shared.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { Route, type RouteDefaultsInput } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Protocol } from "../route/protocol.js"
+import { ProviderID, type ModelID, type LLMRequest } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("groq")
+
+export type ProviderOptions = Pick<OpenAIProviderOptionsInput, "reasoningEffort"> & {
+  /** Controls visible reasoning on GPT-OSS; other models always use parsed reasoning. */
+  readonly includeReasoning?: boolean
+  readonly parallelToolCalls?: boolean
+  readonly serviceTier?: "on_demand" | "flex" | "auto" | "performance" | (string & {})
+  readonly user?: string
+}
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: ProviderOptions
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: ProviderOptions
+}
+
+const Options = Schema.Struct({
+  includeReasoning: Schema.optional(Schema.Boolean),
+  parallelToolCalls: Schema.optional(Schema.Boolean),
+  serviceTier: Schema.optional(Schema.String),
+  user: Schema.optional(Schema.String),
+})
+
+export const protocol = Protocol.make({
+  id: "groq-chat",
+  body: {
+    schema: Schema.Struct({
+      ...OpenAIChat.bodyFields,
+      reasoning_format: Schema.optional(Schema.Literal("parsed")),
+      include_reasoning: Schema.optional(Schema.Boolean),
+      parallel_tool_calls: Schema.optional(Schema.Boolean),
+      service_tier: Schema.optional(Schema.String),
+      user: Schema.optional(Schema.String),
+    }),
+    from: Effect.fn("Groq.fromRequest")(function* (request: LLMRequest) {
+      const options = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(
+        request.providerOptions ?? {},
+      )
+      const gptOSS = request.model.id.startsWith("openai/gpt-oss-")
+      return {
+        ...(yield* OpenAIChat.fromRequest(request)),
+        reasoning_format: gptOSS ? undefined : ("parsed" as const),
+        include_reasoning: gptOSS ? options.includeReasoning : undefined,
+        parallel_tool_calls: options.parallelToolCalls,
+        service_tier: options.serviceTier,
+        user: options.user,
+      }
+    }),
+  },
+  stream: OpenAIChat.protocol.stream,
+})
+
+export const route = Route.make({
+  id: "groq-chat",
+  provider: id,
+  providerMetadataKey: "openai",
+  protocol,
+  endpoint: Endpoint.path("/chat/completions", { baseURL: profiles.groq.baseURL }),
+  framing: OpenAIChat.framing,
+})
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.groq.baseURL },
+    auth: AuthOptions.bearer(input, "GROQ_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<ProviderOptions>({
+        id: modelID,
+        compatibility: {
+          maxTokensField: "max_completion_tokens",
+          reasoningField: "reasoning",
+          requireReasoning: false,
+          supportsStore: false,
+          supportsStrictMode: false,
+        },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, ProviderOptions>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
+
+export * as Groq from "./groq.js"

packages/ai/src/providers/index.ts

@@ -3,16 +3,20 @@ export * as AnthropicCompatible from "./anthropic-compatible.js"
 export * as AmazonBedrock from "./amazon-bedrock.js"
 export * as AmazonBedrockMantle from "./amazon-bedrock-mantle.js"
 export * as Azure from "./azure.js"
+export * as Cerebras from "./cerebras.js"
 export * as Cloudflare from "./cloudflare.js"
 export { CloudflareAIGateway, CloudflareWorkersAI } from "./cloudflare.js"
+export * as DeepInfra from "./deepinfra.js"
 export * as Google from "./google.js"
 export * as GoogleVertex from "./google-vertex.js"
 export * as GoogleVertexChat from "./google-vertex-chat.js"
 export * as GoogleVertexMessages from "./google-vertex-messages.js"
 export * as GoogleVertexResponses from "./google-vertex-responses.js"
+export * as Groq from "./groq.js"
 export * as OpenAI from "./openai.js"
 export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
 export * as OpenRouter from "./openrouter.js"
+export * as TogetherAI from "./togetherai.js"
 export * as XAI from "./xai.js"
 export * as ZAI from "./zai.js"

packages/ai/src/providers/openrouter.ts

@@ -1,7 +1,6 @@
 import { Effect, Schema } from "effect"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import { ProviderID, type CacheHint, type ModelID } from "../schema/index.js"
@@ -9,7 +8,7 @@ import type { ProviderPackage } from "../provider-package.js"
 import * as OpenAICompatibleProfiles from "./openai-compatible-profile.js"
 import * as OpenAIChat from "../protocols/openai-chat.js"
 import { newBreakpoints, ttlBucket } from "../protocols/utils/cache.js"
-import { isRecord, ProviderShared } from "../protocols/shared.js"
+import { isRecord } from "../protocols/shared.js"
 
 export const profile = OpenAICompatibleProfiles.profiles.openrouter
 export const id = ProviderID.make(profile.provider)
@@ -115,12 +114,10 @@ export const protocol = Protocol.make({
               reasoning_details: reasoningDetails,
             }
           })
-          const cacheKey = ProviderShared.clampPromptCacheKey(request.promptCacheKey)
           return {
             ...body,
             messages,
             ...bodyOptions(request.providerOptions),
-            ...(cacheKey ? { prompt_cache_key: cacheKey } : {}),
           } as OpenRouterBody
         }),
       ),
@@ -166,9 +163,10 @@ const bodyOptions = (input: unknown) => {
 export const route = Route.make({
   id: ADAPTER,
   provider: profile.provider,
+  providerMetadataKey: "openrouter",
   protocol,
   endpoint: Endpoint.path("/chat/completions", { baseURL: profile.baseURL }),
-  framing: Framing.sse,
+  framing: OpenAIChat.framing,
 })
 
 export const routes = [route]

packages/ai/src/providers/togetherai.ts

@@ -0,0 +1,58 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("togetherai")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "togetherai-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.togetherai.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.togetherai.baseURL },
+    auth: AuthOptions.bearer(input, ["TOGETHER_API_KEY", "TOGETHER_AI_API_KEY"]),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", supportsStore: false, supportsStrictMode: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)

packages/ai/src/providers/xai.ts

@@ -42,7 +42,7 @@ const responsesRoute = Route.make({
     name: "xAI Responses",
     rotateAfterMs: RESPONSES_WEBSOCKET_ROTATE_AFTER_MS,
   }),
-  defaults: { providerOptions: { store: false } },
+  defaults: { providerOptions: { store: false, include: ["reasoning.encrypted_content"] } },
 })
 
 const chatRoute = Route.make({

packages/ai/src/route/auth.ts

@@ -1,6 +1,6 @@
 import { Config, Effect, Redacted } from "effect"
 import { Headers } from "effect/unstable/http"
-import { AuthenticationReason, InvalidRequestReason, AIError, type HttpOptions } from "../schema/index.js"
+import { AuthenticationError, InvalidRequestError, AIError, type HttpOptions } from "../schema/index.js"
 
 export class MissingCredentialError extends Error {
   readonly _tag = "MissingCredentialError"
@@ -137,12 +137,10 @@ export function bearerHeader(name: string, source?: Secret | Credential) {
 const toAIError = (error: AuthError): AIError => {
   if (error instanceof MissingCredentialError || error instanceof Config.ConfigError) {
     return new AIError({
-      module: "Auth",
-      method: "apply",
       reason:
         error instanceof MissingCredentialError
-          ? new AuthenticationReason({ message: error.message, kind: "missing" })
-          : new InvalidRequestReason({ message: `Failed to resolve auth config: ${error.message}` }),
+          ? new AuthenticationError({ message: error.message, cause: error })
+          : new InvalidRequestError({ message: `Failed to resolve auth config: ${error.message}`, cause: error }),
     })
   }
   return error

packages/ai/src/route/client.ts

@@ -7,17 +7,19 @@ import { HttpTransport } from "./transport/index.js"
 import type { HttpMiddleware, Transport, TransportRuntime, WebSocketChannelExecutor } from "./transport/index.js"
 import type { Protocol } from "./protocol.js"
 import { applyCachePolicy } from "../cache-policy.js"
+import { sanitizeSurrogates } from "../utils/sanitize.js"
 import * as ProviderShared from "../protocols/shared.js"
 import type { ProtocolID, ProviderOptions } from "../schema/index.js"
 import {
   AIError,
+  AIErrorReason,
   GenerationOptions,
   HttpOptions,
   LLMRequest,
   LLMResponse,
   LanguageModel,
   LLMEvent,
-  InvalidProviderOutputReason,
+  InvalidProviderOutputError,
   ProviderID,
   mergeGenerationOptions,
   mergeHttpOptions,
@@ -88,6 +90,7 @@ export interface RouteDefaultsInput {
 export interface RoutePatch<Body, Prepared> extends RouteDefaultsInput {
   readonly id?: string
   readonly provider?: string | ProviderID
+  readonly providerMetadataKey?: string
   readonly auth?: Auth.Definition
   readonly transport?: Transport<Body, Prepared, unknown>
   readonly endpoint?: EndpointPatch<Body>
@@ -225,16 +228,14 @@ export interface MakeTransportInput<Body, Prepared, Frame, Event, State> {
 const streamError = (route: string, message: string, cause: Cause.Cause<unknown>) => {
   const failed = cause.reasons.find(Cause.isFailReason)?.error
   if (failed instanceof AIError) return failed
-  return ProviderShared.eventError(route, message, Cause.pretty(cause))
+  return ProviderShared.eventError(route, message, undefined, cause)
 }
 
 const incompleteStreamError = (route: string) =>
   new AIError({
-    module: "LLMClient",
-    method: "stream",
-    reason: new InvalidProviderOutputReason({
-      classification: "incomplete-stream",
+    reason: new InvalidProviderOutputError({
       message: "The provider response ended unexpectedly.",
+      classification: "incomplete-stream",
       route,
     }),
   })
@@ -263,11 +264,12 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
   const decodeEventEffect = Schema.decodeUnknownEffect(protocol.stream.event)
   const decodeEvent = (route: string) => (frame: Frame) =>
     decodeEventEffect(frame).pipe(
-      Effect.mapError(() =>
+      Effect.mapError((cause) =>
         ProviderShared.eventError(
           input.id,
           `Invalid ${route} stream event`,
           typeof frame === "string" ? frame : ProviderShared.encodeJson(frame),
+          cause,
         ),
       ),
     )
@@ -288,11 +290,16 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
       defaults: routeInput.defaults ?? {},
       body: protocol.body,
       with: (patch: RoutePatch<Body, Prepared>) => {
-        const { id, provider, auth, transport, endpoint, ...defaults } = patch
+        const { id, provider, providerMetadataKey, auth, transport, endpoint, ...defaults } = patch
         return build({
           ...routeInput,
           id: id ?? routeInput.id,
           provider: provider ?? routeInput.provider,
+          providerMetadataKey:
+            providerMetadataKey ??
+            (provider !== undefined && String(provider) !== String(routeInput.provider)
+              ? String(provider)
+              : routeInput.providerMetadataKey),
           auth: auth ?? routeInput.auth,
           endpoint: endpoint ? Endpoint.merge(routeInput.endpoint, endpoint) : routeInput.endpoint,
           transport: (transport as Transport<Body, Prepared, Frame> | undefined) ?? routeInput.transport,
@@ -317,18 +324,74 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
         return Stream.unwrap(
           routeInput.transport.execute(prepared, request, runtime, options).pipe(
             Effect.map((execution) => {
+              const terminal = protocol.stream.terminal
+              // Preserve assembled inputs; replace only serialized event fallbacks with their original wire data.
+              const frameError =
+                (frame: Frame, event: Frame | Event = frame) =>
+                (error: AIError) =>
+                  new AIError({
+                    reason: AIErrorReason.make({
+                      ...error.reason,
+                      message: error.reason.message,
+                      cause: error.reason.cause,
+                      body:
+                        error.reason.body !== undefined && error.reason.body !== ProviderShared.encodeJson(event)
+                          ? error.reason.body
+                          : (execution.body?.(frame) ??
+                            (typeof frame === "string" ? frame : ProviderShared.encodeJson(frame))),
+                    }),
+                  })
               const events = execution.frames.pipe(
-                Stream.mapEffect(decodeEvent(route)),
-                protocol.stream.terminal ? Stream.takeUntil(protocol.stream.terminal) : (stream) => stream,
-              )
-              const stream = events.pipe(
-                Stream.mapAccumEffect(
-                  () => protocol.stream.initial(request),
-                  protocol.stream.step,
-                  protocol.stream.onHalt ? { onHalt: protocol.stream.onHalt } : undefined,
+                Stream.mapEffect((frame) =>
+                  decodeEvent(route)(frame).pipe(
+                    Effect.catchCause((cause) =>
+                      Effect.fail(streamError(route, `Failed to decode ${route} event`, cause)),
+                    ),
+                    Effect.map((event) => ({ event, frame })),
+                    Effect.mapError(frameError(frame)),
+                  ),
                 ),
+                terminal ? Stream.takeUntil(({ event }) => terminal(event)) : (stream) => stream,
+              )
+              const stream = Stream.suspend(() => {
+                let state = protocol.stream.initial(request)
+                const parsed = events.pipe(
+                  Stream.mapEffect(({ event, frame }) =>
+                    protocol.stream.step(state, event).pipe(
+                      Effect.catchCause((cause) =>
+                        Effect.fail(streamError(route, `Failed to parse ${route} event`, cause)),
+                      ),
+                      Effect.map(([next, output]) => {
+                        state = next
+                        return output
+                      }),
+                      Effect.mapError(frameError(frame, event)),
+                    ),
+                  ),
+                  Stream.flatMap(Stream.fromIterable),
+                )
+                const onHalt = protocol.stream.onHalt
+                return onHalt
+                  ? parsed.pipe(
+                      Stream.concat(
+                        Stream.suspend(() => Stream.unwrap(onHalt(state).pipe(Effect.map(Stream.fromIterable)))),
+                      ),
+                    )
+                  : parsed
+              }).pipe(
                 Stream.catchCause((cause) => Stream.fail(streamError(route, `Failed to read ${route} stream`, cause))),
                 requireTerminalEvent(route),
+                Stream.mapError(
+                  (error) =>
+                    new AIError({
+                      reason: AIErrorReason.make({
+                        ...error.reason,
+                        message: error.reason.message,
+                        cause: error.reason.cause,
+                        http: error.reason.http ?? execution.http,
+                      }),
+                    }),
+                ),
               )
               return execution.complete ? stream.pipe(Stream.onEnd(execution.complete)) : stream
             }),
@@ -382,7 +445,8 @@ export function make<Body, Prepared, Frame, Event, State>(
 }
 
 const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest, options?: StreamOptions) {
-  const resolved = applyCachePolicy(resolveRequestOptions(request))
+  const original = applyCachePolicy(resolveRequestOptions(request))
+  const resolved = LLMRequest.update(original, sanitizeSurrogates({ ...LLMRequest.input(original), model: undefined }))
   const route = resolved.model.route
 
   const body = yield* route.body

packages/ai/src/route/executor.ts

@@ -7,14 +7,7 @@ import {
   HttpClientRequest,
   HttpClientResponse,
 } from "effect/unstable/http"
-import {
-  HttpContext,
-  HttpRateLimitDetails,
-  HttpRequestDetails,
-  HttpResponseDetails,
-  AIError,
-  TransportReason,
-} from "../schema/index.js"
+import { HttpContext, HttpRateLimitDetails, AIError, TransportError } from "../schema/index.js"
 import { classifyProviderFailure } from "../provider-error.js"
 
 export interface Interface {
@@ -40,17 +33,6 @@ const headerDetails = (headers: Headers.Headers) =>
 const normalizedHeaders = (headers: Headers.Headers) =>
   Object.fromEntries(Object.entries(headers).map(([key, value]) => [key.toLowerCase(), value]))
 
-const requestId = (headers: Record<string, string>) => {
-  return (
-    headers["x-request-id"] ??
-    headers["request-id"] ??
-    headers["x-amzn-requestid"] ??
-    headers["x-amz-request-id"] ??
-    headers["x-goog-request-id"] ??
-    headers["cf-ray"]
-  )
-}
-
 const retryAfterMs = (headers: Record<string, string>) => {
   const millis = Number(headers["retry-after-ms"])
   if (Number.isFinite(millis)) return Math.max(0, millis)
@@ -108,24 +90,13 @@ const rateLimitDetails = (headers: Record<string, string>, retryAfter: number |
   })
 }
 
-const requestDetails = (request: HttpClientRequest.HttpClientRequest) =>
-  new HttpRequestDetails({
-    method: request.method,
-    url: request.url,
-    headers: headerDetails(request.headers),
-  })
-
-const responseDetails = (response: HttpClientResponse.HttpClientResponse) =>
-  new HttpResponseDetails({
+export const responseHttp = (response: HttpClientResponse.HttpClientResponse) =>
+  new HttpContext({
+    url: response.request.url,
     status: response.status,
     headers: headerDetails(response.headers),
   })
 
-const responseBody = (body: string | void) => {
-  if (body === undefined) return {}
-  return { body }
-}
-
 const decodeProviderBody = Schema.decodeUnknownOption(
   Schema.fromJsonString(
     Schema.Struct({
@@ -143,81 +114,46 @@ const providerMessage = (status: number, body: string | void) => {
   )
 }
 
-const responseHttp = (input: {
-  readonly request: HttpClientRequest.HttpClientRequest
-  readonly response: HttpClientResponse.HttpClientResponse
-  readonly body: ReturnType<typeof responseBody>
-  readonly requestId?: string | undefined
-  readonly rateLimit?: HttpRateLimitDetails | undefined
-}) =>
-  new HttpContext({
-    request: requestDetails(input.request),
-    response: responseDetails(input.response),
-    ...input.body,
-    requestId: input.requestId,
-    rateLimit: input.rateLimit,
+const statusError = (response: HttpClientResponse.HttpClientResponse) =>
+  Effect.gen(function* () {
+    if (response.status < 400) return response
+    const result = yield* response.text.pipe(Effect.result)
+    return yield* httpFailure({
+      message: providerMessage(response.status, result._tag === "Success" ? result.success : undefined),
+      url: response.request.url,
+      status: response.status,
+      responseHeaders: headerDetails(response.headers),
+      responseBody: result._tag === "Success" ? result.success : undefined,
+      cause: result._tag === "Failure" ? (result.failure.cause ?? result.failure) : undefined,
+    })
   })
 
-const statusError =
-  (request: HttpClientRequest.HttpClientRequest) => (response: HttpClientResponse.HttpClientResponse) =>
-    Effect.gen(function* () {
-      if (response.status < 400) return response
-      const body = yield* response.text.pipe(Effect.catch(() => Effect.void))
-      const headers = normalizedHeaders(response.headers)
-      const retryAfter = retryAfterMs(headers)
-      const rateLimit = rateLimitDetails(headers, retryAfter)
-      const details = responseBody(body)
-      return yield* new AIError({
-        module: "RequestExecutor",
-        method: "execute",
-        reason: classifyProviderFailure({
-          status: response.status,
-          message: providerMessage(response.status, body),
-          retryAfterMs: retryAfter,
-          rateLimit,
-          http: responseHttp({
-            request,
-            response,
-            body: details,
-            requestId: requestId(headers),
-            rateLimit,
-          }),
-        }),
-      })
-    })
-
-// Classifies an HTTP failure captured outside the executor (for example by the
-// AI SDK's own fetch) onto the same reason types and HttpContext that
-// executor-driven requests produce. The originating request is not available on
-// that path, so the method is assumed (language model calls are always POST),
-// request headers are empty.
-export const classifyHttpFailure = (input: {
+/** Preserve HTTP diagnostics for executor and externally captured failures alike. */
+export const httpFailure = (input: {
   readonly message: string
-  readonly url: string
+  readonly url?: string | undefined
   readonly status?: number | undefined
-  readonly code?: string | undefined
+  readonly data?: unknown
   readonly responseHeaders?: Record<string, string> | undefined
   readonly responseBody?: string | undefined
+  readonly cause?: unknown
 }) => {
   const headers = normalizedHeaders(Headers.fromInput(input.responseHeaders))
   const retryAfter = retryAfterMs(headers)
   const rateLimit = rateLimitDetails(headers, retryAfter)
-  const details = responseBody(input.responseBody)
-  return classifyProviderFailure({
-    message: input.message,
-    status: input.status,
-    code: input.code,
-    retryAfterMs: retryAfter,
-    rateLimit,
-    http: new HttpContext({
-      request: new HttpRequestDetails({ method: "POST", url: input.url, headers: {} }),
-      response:
-        input.status === undefined
-          ? undefined
-          : new HttpResponseDetails({ status: input.status, headers: headerDetails(Headers.fromInput(headers)) }),
-      ...details,
-      requestId: requestId(headers),
+  return new AIError({
+    reason: classifyProviderFailure({
+      message: input.message,
+      status: input.status,
+      data: input.data,
+      rawBody: input.responseBody,
+      retryAfterMs: retryAfter,
       rateLimit,
+      cause: input.cause,
+      http:
+        input.status === undefined || input.url === undefined
+          ? undefined
+          : new HttpContext({ url: input.url, status: input.status, headers }),
     }),
   })
 }
@@ -244,25 +180,25 @@ const httpError = (input: {
   readonly error: unknown
   readonly request: HttpClientRequest.HttpClientRequest
   readonly operation: HttpOperation
+  readonly http?: HttpContext
 }) => {
   const request = HttpClientError.isHttpClientError(input.error) ? input.error.request : input.request
   const transportError = (failure: { readonly message: string; readonly code?: string | undefined }) =>
     new AIError({
-      module: "RequestExecutor",
-      method: input.operation,
-      reason: new TransportReason({
+      reason: new TransportError({
         message: failure.message,
+        cause: source,
+        http: input.http,
         transport: "http",
         operation: input.operation,
         code: failure.code,
         url: request.url,
-        http: new HttpContext({ request: requestDetails(request) }),
       }),
     })
 
   const source =
     HttpClientError.isHttpClientError(input.error) && "cause" in input.error.reason
-      ? input.error.reason.cause
+      ? (input.error.reason.cause ?? input.error)
       : input.error
   const native = nativeTransportFailure(source)
   const code = native?.code
@@ -286,6 +222,13 @@ const httpError = (input: {
   })
 }
 
+export const responseStream = (response: HttpClientResponse.HttpClientResponse): Stream.Stream<Uint8Array, AIError> =>
+  response.stream.pipe(
+    Stream.mapError((error) =>
+      httpError({ error, request: response.request, operation: "read", http: responseHttp(response) }),
+    ),
+  )
+
 export const stream = (
   executor: Interface,
   request: HttpClientRequest.HttpClientRequest,
@@ -294,9 +237,7 @@ export const stream = (
   Stream.unwrap(
     Effect.gen(function* () {
       const response = yield* executor.execute(request, middleware)
-      return response.stream.pipe(
-        Stream.mapError((error) => httpError({ error, request: response.request, operation: "read" })),
-      )
+      return responseStream(response)
     }),
   )
 
@@ -309,7 +250,7 @@ export const layer: Layer.Layer<Service, never, HttpClient.HttpClient> = Layer.e
         if (!middleware)
           return yield* http.execute(request).pipe(
             Effect.mapError((error) => httpError({ error, request, operation: "request" })),
-            Effect.flatMap(statusError(request)),
+            Effect.flatMap(statusError),
           )
 
         const response = yield* middleware(request, (input) =>
@@ -317,7 +258,7 @@ export const layer: Layer.Layer<Service, never, HttpClient.HttpClient> = Layer.e
             .execute(input)
             .pipe(Effect.mapError((cause) => (cause instanceof Error ? cause : new Error(String(cause))))),
         ).pipe(Effect.mapError((error) => httpError({ error, request, operation: "request" })))
-        return yield* statusError(response.request)(response)
+        return yield* statusError(response)
       })
     return Service.of({
       execute: executeOnce,

packages/ai/src/route/framing.ts

@@ -8,8 +8,8 @@ import type { AIError } from "../schema/index.js"
  * `Framing` is the byte-stream-shaped seam between transport and protocol:
  *
  * - SSE (`Framing.sse`) — UTF-8 decode the body, run the SSE channel decoder,
- *   drop empty / `[DONE]` keep-alives. Each emitted frame is the JSON `data:`
- *   payload of one event.
+ *   and emit the `data:` payload of each non-empty event. The default drops
+ *   `[DONE]`; protocols that use it as a terminal select `sseWithDone`.
  * - AWS event stream — length-prefixed binary frames with CRC checksums.
  *   Each emitted frame is one parsed binary event record.
  *
@@ -19,11 +19,19 @@ import type { AIError } from "../schema/index.js"
 export interface Definition<Frame> {
   readonly id: string
   readonly frame: (bytes: Stream.Stream<Uint8Array, AIError>) => Stream.Stream<Frame, AIError>
+  /** Original wire representation when framing transforms the provider payload. */
+  readonly body?: (frame: Frame) => string | undefined
 }
 
 /** Server-Sent Events framing. Used by every JSON-streaming HTTP provider. */
 export const sse: Definition<string> = { id: "sse", frame: ProviderShared.sseFraming }
 
+/** Server-Sent Events framing that retains the conventional `[DONE]` sentinel. */
+export const sseWithDone: Definition<string> = {
+  id: "sse",
+  frame: (bytes) => ProviderShared.sseFraming(bytes, undefined, true),
+}
+
 /** SSE framing restricted to protocol-recognized event names. */
 export const sseEvents = (events: ReadonlySet<string>): Definition<string> => ({
   id: "sse",

packages/ai/src/route/protocol.ts

@@ -59,8 +59,8 @@ export interface ProtocolStream<Frame, Event, State> {
   readonly step: (state: State, event: Event) => Effect.Effect<readonly [State, ReadonlyArray<LLMEvent>], AIError>
   /** Optional request-completion signal for transports that do not end naturally. */
   readonly terminal?: (event: Event) => boolean
-  /** Optional flush emitted when the framed stream ends. */
-  readonly onHalt?: (state: State) => ReadonlyArray<LLMEvent>
+  /** Optional effectful flush emitted when the framed stream ends. */
+  readonly onHalt?: (state: State) => Effect.Effect<ReadonlyArray<LLMEvent>, AIError>
 }
 
 /**

packages/ai/src/route/transport/http.ts

@@ -88,8 +88,13 @@ export const httpJson = <Body, Frame>(input: HttpJsonInput<Body, Frame>): HttpJs
       }
     }),
   execute: (prepared, _request, runtime) =>
-    Effect.succeed({
-      frames: prepared.framing.frame(RequestExecutor.stream(runtime.http, prepared.request, prepared.middleware)),
+    Effect.gen(function* () {
+      const response = yield* runtime.http.execute(prepared.request, prepared.middleware)
+      return {
+        frames: prepared.framing.frame(RequestExecutor.responseStream(response)),
+        http: RequestExecutor.responseHttp(response),
+        body: prepared.framing.body,
+      }
     }),
 })
 

packages/ai/src/route/transport/index.ts

@@ -3,7 +3,7 @@ import { Endpoint } from "../endpoint.js"
 import { Auth } from "../auth.js"
 import type { HttpMiddleware, Interface as RequestExecutorInterface } from "../executor.js"
 import type { WebSocketChannelExecutor } from "./websocket-channel.js"
-import type { AIError, LLMRequest } from "../../schema/index.js"
+import type { AIError, HttpContext, LLMRequest } from "../../schema/index.js"
 
 export interface TransportRuntime {
   readonly http: RequestExecutorInterface
@@ -11,6 +11,8 @@ export interface TransportRuntime {
 
 export interface TransportExecution<Frame> {
   readonly frames: Stream.Stream<Frame, AIError>
+  readonly http?: HttpContext
+  body?(frame: Frame): string | undefined
   /** Optional successful-consumption acknowledgement. HTTP leaves this absent. */
   readonly complete?: Effect.Effect<void>
 }

packages/ai/src/route/transport/websocket-channel.ts

@@ -1,6 +1,6 @@
 import type { Effect, Scope, Stream } from "effect"
 import type { Headers } from "effect/unstable/http"
-import type { AIError } from "../../schema/index.js"
+import type { AIError, HttpContext } from "../../schema/index.js"
 
 export interface WebSocketChannelExecutor {
   readonly execute: (
@@ -10,6 +10,7 @@ export interface WebSocketChannelExecutor {
 
 export interface WebSocketChannelExecution {
   readonly frames: Stream.Stream<string, AIError>
+  readonly http?: HttpContext
   /** Commits staged state after the decoded Route stream ends successfully. */
   readonly complete: Effect.Effect<void>
 }

packages/ai/src/route/transport/websocket.ts

@@ -1,7 +1,13 @@
 import { Cause, Effect, Queue, Stream } from "effect"
 import { Headers } from "effect/unstable/http"
 import { Socket } from "effect/unstable/socket"
-import { AIError, TransportReason, type TransportOperation } from "../../schema/index.js"
+import {
+  AIError,
+  AIErrorReason,
+  TransportError,
+  type HttpContext,
+  type TransportOperation,
+} from "../../schema/index.js"
 import * as HttpTransport from "./http.js"
 import type { Transport } from "./index.js"
 import type {
@@ -17,6 +23,7 @@ export interface WebSocketRequest {
 }
 
 export interface WebSocketConnection {
+  readonly http?: HttpContext
   readonly sendText: (message: string) => Effect.Effect<void, AIError>
   readonly messages: Stream.Stream<string | Uint8Array, AIError>
   readonly close: Effect.Effect<void, never>
@@ -33,21 +40,22 @@ type WebSocketConstructorWithHeaders = (
 
 const MAX_FRAME_BYTES = 16 * 1024 * 1024
 const transportError = (
-  method: string,
   message: string,
   input: {
     readonly operation: TransportOperation
     readonly url?: string
     readonly code?: string
-    readonly phase?: TransportReason["phase"]
-    readonly delivery?: TransportReason["delivery"]
+    readonly phase?: TransportError["phase"]
+    readonly delivery?: TransportError["delivery"]
+    readonly body?: string
+    readonly cause?: unknown
   },
 ) =>
   new AIError({
-    module: "WebSocketConnector",
-    method,
-    reason: new TransportReason({
+    reason: new TransportError({
       message,
+      body: input.body,
+      cause: input.cause,
       transport: "websocket",
       operation: input.operation,
       url: input.url,
@@ -59,22 +67,16 @@ const transportError = (
 
 const annotateTransportError = (
   error: AIError,
-  input: { readonly phase: TransportReason["phase"]; readonly delivery: TransportReason["delivery"] },
+  input: { readonly phase: TransportError["phase"]; readonly delivery: TransportError["delivery"] },
 ) =>
   error.reason._tag === "Transport"
     ? new AIError({
-        module: error.module,
-        method: error.method,
-        reason: new TransportReason({
+        reason: new TransportError({
+          ...error.reason,
           message: error.reason.message,
-          transport: error.reason.transport,
-          operation: error.reason.operation,
-          code: error.reason.code,
-          url: error.reason.url,
-          http: error.reason.http,
+          cause: error.reason.cause,
           phase: input.phase,
           delivery: input.delivery,
-          recovery: error.reason.recovery,
         }),
       })
     : error
@@ -95,7 +97,7 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
   if (ws.readyState === globalThis.WebSocket.OPEN) return Effect.void
   if (ws.readyState === globalThis.WebSocket.CLOSING || ws.readyState === globalThis.WebSocket.CLOSED) {
     return Effect.fail(
-      transportError("open", `WebSocket closed before opening (state ${ws.readyState})`, {
+      transportError(`WebSocket closed before opening (state ${ws.readyState})`, {
         url: input.url,
         operation: "request",
         code: "closed",
@@ -124,7 +126,8 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
       cleanup()
       resume(
         Effect.fail(
-          transportError("open", `Failed to open WebSocket: ${eventMessage(event)}`, {
+          transportError(`Failed to open WebSocket: ${eventMessage(event)}`, {
+            cause: "error" in event ? (event.error ?? event) : event,
             url: input.url,
             operation: "request",
             phase: "connect",
@@ -137,7 +140,9 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
       cleanup()
       resume(
         Effect.fail(
-          transportError("open", `WebSocket closed before opening with code ${event.code}`, {
+          transportError(`WebSocket closed before opening with code ${event.code}`, {
+            body: event.reason,
+            cause: event,
             url: input.url,
             operation: "request",
             code: String(event.code),
@@ -169,7 +174,8 @@ export const toWebSocketUrl = (value: string) =>
       throw new Error(`Unsupported WebSocket URL protocol ${url.protocol}`)
     },
     catch: (error) =>
-      transportError("prepare", error instanceof Error ? error.message : "Invalid WebSocket URL", {
+      transportError(error instanceof Error ? error.message : "Invalid WebSocket URL", {
+        cause: error,
         url: value,
         operation: "request",
         code: "invalid-url",
@@ -189,7 +195,8 @@ export const open = (input: WebSocketRequest) =>
           headers: input.headers,
         }),
       catch: (error) =>
-        transportError("open", error instanceof Error ? error.message : "Failed to construct WebSocket", {
+        transportError(error instanceof Error ? error.message : "Failed to construct WebSocket", {
+          cause: error,
           url: input.url,
           operation: "request",
           phase: "connect",
@@ -214,7 +221,8 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", "WebSocket message exceeds the 16 MiB limit", {
+          transportError("WebSocket message exceeds the 16 MiB limit", {
+            body: typeof message === "string" ? message : new TextDecoder().decode(message),
             url: input.url,
             operation: "read",
             code: "message-too-large",
@@ -231,7 +239,8 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", "WebSocket inbound queue overflow", {
+          transportError("WebSocket inbound queue overflow", {
+            body: typeof message === "string" ? message : new TextDecoder().decode(message),
             url: input.url,
             operation: "read",
             code: "queue-overflow",
@@ -248,7 +257,8 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", "Unsupported WebSocket message payload", {
+          transportError("Unsupported WebSocket message payload", {
+            cause: event,
             url: input.url,
             operation: "read",
             code: "message",
@@ -261,7 +271,8 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", `WebSocket error: ${eventMessage(event)}`, {
+          transportError(`WebSocket error: ${eventMessage(event)}`, {
+            cause: "error" in event ? (event.error ?? event) : event,
             url: input.url,
             operation: "read",
             code: "message",
@@ -274,7 +285,9 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", `WebSocket closed with code ${event.code}`, {
+          transportError(`WebSocket closed with code ${event.code}`, {
+            body: event.reason,
+            cause: event,
             url: input.url,
             operation: "read",
             code: String(event.code),
@@ -298,7 +311,7 @@ export const fromWebSocket = (
         Effect.suspend(() => {
           if (ws.readyState !== globalThis.WebSocket.OPEN)
             return Effect.fail(
-              transportError("sendText", `WebSocket is not open (state ${ws.readyState})`, {
+              transportError(`WebSocket is not open (state ${ws.readyState})`, {
                 url: input.url,
                 operation: "write",
                 phase: "send",
@@ -308,7 +321,8 @@ export const fromWebSocket = (
           return Effect.try({
             try: () => ws.send(message),
             catch: (error) =>
-              transportError("sendText", error instanceof Error ? error.message : "Failed to send WebSocket message", {
+              transportError(error instanceof Error ? error.message : "Failed to send WebSocket message", {
+                cause: error,
                 url: input.url,
                 operation: "write",
                 phase: "send",
@@ -349,10 +363,23 @@ export const makeDirect = (connector: WebSocketConnector): WebSocketChannelExecu
         (connection) => connection.close,
       )
       const create = yield* exchange.driver.create(undefined)
-      yield* connection.sendText(create.message)
+      yield* connection.sendText(create.message).pipe(
+        Effect.mapError(
+          (error) =>
+            new AIError({
+              reason: AIErrorReason.make({
+                ...error.reason,
+                message: error.reason.message,
+                cause: error.reason.cause,
+                http: error.reason.http ?? connection.http,
+              }),
+            }),
+        ),
+      )
       const decoder = new TextDecoder()
       let observed = false
       return {
+        http: connection.http,
         frames: connection.messages.pipe(
           Stream.map((message) => {
             observed = true
@@ -364,9 +391,49 @@ export const makeDirect = (connector: WebSocketConnector): WebSocketChannelExecu
               delivery: observed ? "accepted" : "ambiguous",
             }),
           ),
-          Stream.mapEffect((frame) => exchange.driver.observe(create, frame)),
+          Stream.mapEffect((frame) =>
+            exchange.driver.observe(create, frame).pipe(
+              Effect.mapError(
+                (error) =>
+                  new AIError({
+                    reason: AIErrorReason.make({
+                      ...error.reason,
+                      message: error.reason.message,
+                      cause: error.reason.cause,
+                      body: frame,
+                    }),
+                  }),
+              ),
+              Effect.map((observation) =>
+                "error" in observation
+                  ? {
+                      ...observation,
+                      error: new AIError({
+                        reason: AIErrorReason.make({
+                          ...observation.error.reason,
+                          message: observation.error.reason.message,
+                          cause: observation.error.reason.cause,
+                          body: frame,
+                        }),
+                      }),
+                    }
+                  : observation,
+              ),
+            ),
+          ),
           Stream.takeUntil(observationTerminal),
           Stream.mapEffect(observationFrame),
+          Stream.mapError(
+            (error) =>
+              new AIError({
+                reason: AIErrorReason.make({
+                  ...error.reason,
+                  message: error.reason.message,
+                  cause: error.reason.cause,
+                  http: error.reason.http ?? connection.http,
+                }),
+              }),
+          ),
         ),
         complete: Effect.void,
       }
@@ -417,7 +484,7 @@ export const json = <Body, Message>(input: JsonInput<Body, Message>): JsonTransp
     const webSocket = options?.webSocket
     if (!webSocket) {
       return Effect.fail(
-        transportError("json", "WebSocket JSON transport requires StreamOptions.webSocket", {
+        transportError("WebSocket JSON transport requires StreamOptions.webSocket", {
           url: prepared.url,
           operation: "request",
           code: "unavailable",
@@ -435,7 +502,7 @@ export const json = <Body, Message>(input: JsonInput<Body, Message>): JsonTransp
       connect: { url: prepared.url, headers: prepared.headers },
       fallback: () =>
         Stream.fail(
-          transportError("fallback", "WebSocket JSON transport does not provide HTTP fallback", {
+          transportError("WebSocket JSON transport does not provide HTTP fallback", {
             url: prepared.url,
             operation: "request",
             code: "websocket",

packages/ai/src/schema/errors.ts

@@ -1,19 +1,13 @@
 import { Schema } from "effect"
 import { Tool } from "@opencode-ai/schema/tool"
 import { ModelID, ProviderID, RouteID } from "./ids.js"
-import { ProviderMetadata } from "./messages.js"
 
 export const ProviderFailureClassification = Schema.Literals(["context-overflow", "payload-too-large"])
 export type ProviderFailureClassification = typeof ProviderFailureClassification.Type
 
-export class HttpRequestDetails extends Schema.Class<HttpRequestDetails>("AI.HttpRequestDetails")({
-  method: Schema.String,
+export class HttpContext extends Schema.Class<HttpContext>("AI.HttpContext")({
   url: Schema.String,
-  headers: Schema.Record(Schema.String, Schema.String),
-}) {}
-
-export class HttpResponseDetails extends Schema.Class<HttpResponseDetails>("AI.HttpResponseDetails")({
-  status: Schema.Number,
+  status: Schema.Int.check(Schema.isBetween({ minimum: 100, maximum: 599 })),
   headers: Schema.Record(Schema.String, Schema.String),
 }) {}
 
@@ -24,74 +18,58 @@ export class HttpRateLimitDetails extends Schema.Class<HttpRateLimitDetails>("AI
   reset: Schema.optional(Schema.Record(Schema.String, Schema.String)),
 }) {}
 
-export class HttpContext extends Schema.Class<HttpContext>("AI.HttpContext")({
-  request: HttpRequestDetails,
-  response: Schema.optional(HttpResponseDetails),
-  body: Schema.optional(Schema.String),
-  bodyTruncated: Schema.optional(Schema.Boolean),
-  requestId: Schema.optional(Schema.String),
-  rateLimit: Schema.optional(HttpRateLimitDetails),
-}) {}
-
-export class InvalidRequestReason extends Schema.Class<InvalidRequestReason>("AI.Error.InvalidRequest")({
-  _tag: Schema.tag("InvalidRequest"),
+const ReasonFields = {
   message: Schema.String,
-  parameter: Schema.optional(Schema.String),
-  classification: Schema.optional(ProviderFailureClassification),
-  providerMetadata: Schema.optional(ProviderMetadata),
+  // Preserve the complete original response or triggering event before decoding narrows it.
+  body: Schema.optional(Schema.String),
   http: Schema.optional(HttpContext),
-}) {}
+  cause: Schema.optional(Schema.Defect({ includeStack: true })),
+}
 
-export class NoRouteReason extends Schema.Class<NoRouteReason>("AI.Error.NoRoute")({
-  _tag: Schema.tag("NoRoute"),
+export class InvalidRequestError extends Schema.TaggedError<InvalidRequestError>("AI.Error.InvalidRequest")(
+  "InvalidRequest",
+  {
+    ...ReasonFields,
+    parameter: Schema.optional(Schema.String),
+    classification: Schema.optional(ProviderFailureClassification),
+  },
+) {}
+
+export class NoRouteError extends Schema.TaggedError<NoRouteError>("AI.Error.NoRoute")("NoRoute", {
+  ...ReasonFields,
   route: RouteID,
   provider: ProviderID,
   model: ModelID,
-}) {
-  get message() {
-    return `No AI route for ${this.provider}/${this.model} using ${this.route}`
-  }
-}
-
-export class AuthenticationReason extends Schema.Class<AuthenticationReason>("AI.Error.Authentication")({
-  _tag: Schema.tag("Authentication"),
-  message: Schema.String,
-  kind: Schema.Literals(["missing", "invalid", "expired", "insufficient-permissions", "unknown"]),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
 }) {}
 
-export class RateLimitReason extends Schema.Class<RateLimitReason>("AI.Error.RateLimit")({
-  _tag: Schema.tag("RateLimit"),
-  message: Schema.String,
+export class AuthenticationError extends Schema.TaggedError<AuthenticationError>("AI.Error.Authentication")(
+  "Authentication",
+  ReasonFields,
+) {}
+
+export class RateLimitError extends Schema.TaggedError<RateLimitError>("AI.Error.RateLimit")("RateLimit", {
+  ...ReasonFields,
   retryAfterMs: Schema.optional(Schema.Number),
   rateLimit: Schema.optional(HttpRateLimitDetails),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
 }) {}
 
-export class QuotaExceededReason extends Schema.Class<QuotaExceededReason>("AI.Error.QuotaExceeded")({
-  _tag: Schema.tag("QuotaExceeded"),
-  message: Schema.String,
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class QuotaExceededError extends Schema.TaggedError<QuotaExceededError>("AI.Error.QuotaExceeded")(
+  "QuotaExceeded",
+  ReasonFields,
+) {}
 
-export class ContentPolicyReason extends Schema.Class<ContentPolicyReason>("AI.Error.ContentPolicy")({
-  _tag: Schema.tag("ContentPolicy"),
-  message: Schema.String,
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class ContentPolicyError extends Schema.TaggedError<ContentPolicyError>("AI.Error.ContentPolicy")(
+  "ContentPolicy",
+  ReasonFields,
+) {}
 
-export class ProviderInternalReason extends Schema.Class<ProviderInternalReason>("AI.Error.ProviderInternal")({
-  _tag: Schema.tag("ProviderInternal"),
-  message: Schema.String,
-  status: Schema.optional(Schema.Number),
-  retryAfterMs: Schema.optional(Schema.Number),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class ProviderInternalError extends Schema.TaggedError<ProviderInternalError>("AI.Error.ProviderInternal")(
+  "ProviderInternal",
+  {
+    ...ReasonFields,
+    retryAfterMs: Schema.optional(Schema.Number),
+  },
+) {}
 
 export const TransportType = Schema.Literals(["http", "websocket"])
 export type TransportType = typeof TransportType.Type
@@ -99,14 +77,12 @@ export type TransportType = typeof TransportType.Type
 export const TransportOperation = Schema.Literals(["request", "read", "write"])
 export type TransportOperation = typeof TransportOperation.Type
 
-export class TransportReason extends Schema.Class<TransportReason>("AI.Error.Transport")({
-  _tag: Schema.tag("Transport"),
-  message: Schema.String,
+export class TransportError extends Schema.TaggedError<TransportError>("AI.Error.Transport")("Transport", {
+  ...ReasonFields,
   transport: TransportType,
   operation: TransportOperation,
   code: Schema.optional(Schema.String),
   url: Schema.optional(Schema.String),
-  http: Schema.optional(HttpContext),
   phase: Schema.optional(
     Schema.Literals(["prepare", "queue", "connect", "send", "receive", "decode", "complete", "fallback", "close"]),
   ),
@@ -116,51 +92,40 @@ export class TransportReason extends Schema.Class<TransportReason>("AI.Error.Tra
   ),
 }) {}
 
-export class InvalidProviderOutputReason extends Schema.Class<InvalidProviderOutputReason>(
+export class InvalidProviderOutputError extends Schema.TaggedError<InvalidProviderOutputError>(
   "AI.Error.InvalidProviderOutput",
-)({
-  _tag: Schema.tag("InvalidProviderOutput"),
-  message: Schema.String,
+)("InvalidProviderOutput", {
+  ...ReasonFields,
   classification: Schema.optional(Schema.Literals(["incomplete-stream"])),
   route: Schema.optional(Schema.String),
-  raw: Schema.optional(Schema.String),
-  providerMetadata: Schema.optional(ProviderMetadata),
 }) {}
 
-export class UnknownProviderReason extends Schema.Class<UnknownProviderReason>("AI.Error.UnknownProvider")({
-  _tag: Schema.tag("UnknownProvider"),
-  message: Schema.String,
-  status: Schema.optional(Schema.Number),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class UnknownProviderError extends Schema.TaggedError<UnknownProviderError>("AI.Error.UnknownProvider")(
+  "UnknownProvider",
+  ReasonFields,
+) {}
 
 export const AIErrorReason = Schema.Union([
-  InvalidRequestReason,
-  NoRouteReason,
-  AuthenticationReason,
-  RateLimitReason,
-  QuotaExceededReason,
-  ContentPolicyReason,
-  ProviderInternalReason,
-  TransportReason,
-  InvalidProviderOutputReason,
-  UnknownProviderReason,
+  InvalidRequestError,
+  NoRouteError,
+  AuthenticationError,
+  RateLimitError,
+  QuotaExceededError,
+  ContentPolicyError,
+  ProviderInternalError,
+  TransportError,
+  InvalidProviderOutputError,
+  UnknownProviderError,
 ]).pipe(Schema.toTaggedUnion("_tag"))
 export type AIErrorReason = Schema.Schema.Type<typeof AIErrorReason>
 
 export class AIError extends Schema.TaggedError<AIError>()("AI.Error", {
-  module: Schema.String,
-  method: Schema.String,
   reason: AIErrorReason,
-  // Raw provider payload as a string, so classified failures never lose the
-  // original error detail even when the pretty message is a summary.
-  body: Schema.optional(Schema.String),
 }) {
   override readonly cause = this.reason
 
-  override get message() {
-    return `${this.module}.${this.method}: ${this.reason.message}`
+  override get message(): string {
+    return this.reason.message
   }
 }
 

packages/ai/src/schema/events.ts

@@ -112,6 +112,8 @@ export type TextDelta = Schema.Schema.Type<typeof TextDelta>
 export const TextEnd = Schema.Struct({
   type: Schema.tag("text-end"),
   id: ContentBlockID,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text: Schema.optional(Schema.String),
   providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.TextEnd" })
 export type TextEnd = Schema.Schema.Type<typeof TextEnd>
@@ -134,6 +136,8 @@ export type ReasoningDelta = Schema.Schema.Type<typeof ReasoningDelta>
 export const ReasoningEnd = Schema.Struct({
   type: Schema.tag("reasoning-end"),
   id: ContentBlockID,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text: Schema.optional(Schema.String),
   providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.ReasoningEnd" })
 export type ReasoningEnd = Schema.Schema.Type<typeof ReasoningEnd>
@@ -152,6 +156,8 @@ export const ToolInputDelta = Schema.Struct({
   id: ToolCallID,
   name: Schema.String,
   text: Schema.String,
+  /** Best-effort parse of all input fragments received through this delta. */
+  input: Schema.optional(Schema.Unknown),
 }).annotate({ identifier: "LLM.Event.ToolInputDelta" })
 export type ToolInputDelta = Schema.Schema.Type<typeof ToolInputDelta>
 
@@ -326,17 +332,32 @@ export const LLMEvent = Object.assign(llmEventTagged, {
 })
 export type LLMEvent = Schema.Schema.Type<typeof llmEventTagged>
 
+/** Joins deltas per fragment, letting an authoritative end value replace that fragment's accumulated deltas. */
+const joinFragments = <Delta extends { id: string; text: string }, End extends { id: string; text?: string }>(
+  events: ReadonlyArray<LLMEvent>,
+  isDelta: (event: LLMEvent) => event is Extract<LLMEvent, Delta>,
+  isEnd: (event: LLMEvent) => event is Extract<LLMEvent, End>,
+) => {
+  const order: string[] = []
+  const parts = new Map<string, string>()
+  for (const event of events) {
+    if (isDelta(event)) {
+      if (!parts.has(event.id)) order.push(event.id)
+      parts.set(event.id, (parts.get(event.id) ?? "") + event.text)
+    }
+    if (isEnd(event) && event.text !== undefined) {
+      if (!parts.has(event.id)) order.push(event.id)
+      parts.set(event.id, event.text)
+    }
+  }
+  return order.map((id) => parts.get(id)).join("")
+}
+
 const responseText = (events: ReadonlyArray<LLMEvent>) =>
-  events
-    .filter(LLMEvent.is.textDelta)
-    .map((event) => event.text)
-    .join("")
+  joinFragments(events, LLMEvent.is.textDelta, LLMEvent.is.textEnd)
 
 const responseReasoning = (events: ReadonlyArray<LLMEvent>) =>
-  events
-    .filter(LLMEvent.is.reasoningDelta)
-    .map((event) => event.text)
-    .join("")
+  joinFragments(events, LLMEvent.is.reasoningDelta, LLMEvent.is.reasoningEnd)
 
 const responseUsage = (events: ReadonlyArray<LLMEvent>) =>
   events.reduce<Usage | undefined>(
@@ -443,10 +464,11 @@ const reduceTextDelta = (state: ResponseState, event: TextDelta): ResponseState
 const reduceTextEnd = (state: ResponseState, event: TextEnd): ResponseState => {
   const current = state.textParts[event.id]
   if (!current) return state
+  const text = event.text ?? current.text
   const providerMetadata = event.providerMetadata ?? current.providerMetadata
   return {
-    ...replaceContent(state, current.contentIndex, textContent(current.text, providerMetadata)),
-    textParts: { ...state.textParts, [event.id]: { ...current, providerMetadata } },
+    ...replaceContent(state, current.contentIndex, textContent(text, providerMetadata)),
+    textParts: { ...state.textParts, [event.id]: { ...current, text, providerMetadata } },
   }
 }
 
@@ -476,10 +498,11 @@ const reduceReasoningDelta = (state: ResponseState, event: ReasoningDelta): Resp
 const reduceReasoningEnd = (state: ResponseState, event: ReasoningEnd): ResponseState => {
   const current = state.reasoningParts[event.id]
   if (!current) return state
+  const text = event.text ?? current.text
   const providerMetadata = event.providerMetadata ?? current.providerMetadata
   return {
-    ...replaceContent(state, current.contentIndex, reasoningContent(current.text, providerMetadata)),
-    reasoningParts: { ...state.reasoningParts, [event.id]: { ...current, providerMetadata } },
+    ...replaceContent(state, current.contentIndex, reasoningContent(text, providerMetadata)),
+    reasoningParts: { ...state.reasoningParts, [event.id]: { ...current, text, providerMetadata } },
   }
 }
 
@@ -577,12 +600,12 @@ export class LLMResponse extends Schema.Class<LLMResponse>("LLM.Response")({
   usage: Schema.optional(Usage),
   finishReason: FinishReasonDetails,
 }) {
-  /** Concatenated assistant text assembled from streamed `text-delta` events. */
+  /** Concatenated assistant text; each fragment's `text-end` value replaces its accumulated deltas when present. */
   get text() {
     return responseText(this.events)
   }
 
-  /** Concatenated reasoning text assembled from streamed `reasoning-delta` events. */
+  /** Concatenated reasoning text; each fragment's `reasoning-end` value replaces its accumulated deltas when present. */
   get reasoning() {
     return responseReasoning(this.events)
   }

packages/ai/src/schema/options.ts

@@ -153,8 +153,16 @@ export class LanguageModelCompatibility extends Schema.Class<LanguageModelCompat
 )({
   toolSchema: Schema.optional(LanguageModelToolSchemaCompatibility),
   reasoningField: Schema.optional(Schema.String),
+  /** Require every assistant message to include its reasoning field, even when empty. */
+  requireReasoning: Schema.optional(Schema.Boolean),
   maxTokensField: Schema.optional(LanguageModelMaxTokensFieldCompatibility),
   requireFinishReason: Schema.optional(Schema.Boolean),
+  requireAssistantAfterTool: Schema.optional(Schema.Boolean),
+  supportsStore: Schema.optional(Schema.Boolean),
+  supportsUsageInStreaming: Schema.optional(Schema.Boolean),
+  supportsStrictMode: Schema.optional(Schema.Boolean),
+  zaiToolStream: Schema.optional(Schema.Boolean),
+  requireSignature: Schema.optional(Schema.Boolean),
 }) {}
 
 export namespace LanguageModelCompatibility {

packages/ai/src/utils/sanitize.ts

@@ -0,0 +1,12 @@
+import { isRecord } from "./record.js"
+
+export const sanitizeSurrogates = <T>(value: T): T => {
+  if (typeof value === "string") return value.toWellFormed() as T
+  if (Array.isArray(value)) return value.map(sanitizeSurrogates) as T
+  if (value instanceof Uint8Array || value instanceof Error) return value
+  if (isRecord(value))
+    return Object.fromEntries(
+      Object.entries(value).map(([key, entry]) => [key.toWellFormed(), sanitizeSurrogates(entry)]),
+    ) as T
+  return value
+}

packages/ai/test/adapter.test.ts

@@ -1,4 +1,4 @@
-import { describe, expect } from "bun:test"
+import { describe, expect, test } from "bun:test"
 import { Effect, Schema, Stream } from "effect"
 import { LLM, LLMRequest, LLMResponse } from "../src/index.js"
 import { Route, Endpoint, LLMClient, Protocol, type FramingDef } from "../src/route.js"
@@ -148,15 +148,13 @@ describe("llm route", () => {
     }),
   )
 
-  it.effect("builds models from configured routes", () =>
-    Effect.gen(function* () {
-      const configured = fake.with({ provider: "fake-provider", endpoint: { baseURL: "https://fake.local" } })
+  test("builds models from configured routes", () => {
+    const configured = fake.with({ provider: "fake-provider", endpoint: { baseURL: "https://fake.local" } })
 
-      expect(configured.model({ id: "fake-model" })).toMatchObject({
-        provider: "fake-provider",
-      })
-    }),
-  )
+    expect(configured.model({ id: "fake-model" })).toMatchObject({
+      provider: "fake-provider",
+    })
+  })
 
   it.effect("does not register duplicate route ids globally", () =>
     Effect.gen(function* () {

packages/ai/test/cache-policy.test.ts

@@ -3,7 +3,7 @@ import { Effect } from "effect"
 import { CacheHint, LLM, Message } from "../src/index.js"
 import { Auth } from "../src/route.js"
 import { compileRequest } from "../src/route/client.js"
-import { AmazonBedrock } from "../src/providers.js"
+import { AmazonBedrock, GoogleVertexMessages } from "../src/providers.js"
 import * as AnthropicMessages from "../src/protocols/anthropic-messages.js"
 import * as Gemini from "../src/protocols/gemini.js"
 import * as OpenAIChat from "../src/protocols/openai-chat.js"
@@ -86,6 +86,27 @@ describe("applyCachePolicy", () => {
     }),
   )
 
+  it.effect("'auto' emits Anthropic cache markers on Vertex", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: GoogleVertexMessages.configure({ accessToken: "test", location: "global", project: "test" }).model(
+            "claude-opus-4-8",
+          ),
+          system: "You are concise.",
+          tools: [{ name: "lookup", description: "Look up a value", inputSchema: { type: "object", properties: {} } }],
+          prompt: "hi",
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        tools: [{ name: "lookup", cache_control: { type: "ephemeral" } }],
+        system: [{ type: "text", text: "You are concise.", cache_control: { type: "ephemeral" } }],
+        messages: [{ role: "user", content: [{ type: "text", text: "hi", cache_control: { type: "ephemeral" } }] }],
+      })
+    }),
+  )
+
   it.effect("'auto' is a no-op on OpenAI (implicit caching protocol)", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(

packages/ai/test/compile.test.ts

@@ -1,7 +1,7 @@
 import { describe, expect, test } from "bun:test"
 import { Effect, Ref, Schema } from "effect"
 import { HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
-import { LLM, mergeProviderOptions } from "../src/index.js"
+import { LLM, Message, ToolCallPart, mergeProviderOptions } from "../src/index.js"
 import { AnthropicMessages, OpenAIChat } from "../src/protocols.js"
 import { Auth, LLMClient } from "../src/route.js"
 import { compileRequest } from "../src/route/client.js"
@@ -66,7 +66,7 @@ describe("request option precedence", () => {
       expect(prepared.body).toMatchObject({
         model: "gpt-4o-mini",
         stream: true,
-        max_tokens: 30,
+        max_completion_tokens: 30,
         temperature: 0.5,
         top_p: 0.9,
         frequency_penalty: 0.25,
@@ -247,6 +247,73 @@ describe("request option precedence", () => {
     }),
   )
 
+  it.effect("sanitizes outbound JSON without an HTTP overlay", () =>
+    LLMClient.generate(
+      LLM.request({
+        model: OpenAIChat.route
+          .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
+          .model({ id: "gpt-4o-mini" }),
+        prompt: "hello \uD800 \u{1F600}",
+      }),
+    ).pipe(
+      Effect.provide(
+        dynamicResponse((input) =>
+          Effect.gen(function* () {
+            expect(decodeJson(input.text)).toMatchObject({
+              messages: [{ role: "user", content: "hello \uFFFD \u{1F600}" }],
+            })
+            return input.respond(sseEvents(deltaChunk({}, "stop")), {
+              headers: { "content-type": "text/event-stream" },
+            })
+          }),
+        ),
+      ),
+    ),
+  )
+
+  it.effect("sanitizes unpaired surrogates throughout outbound JSON", () =>
+    LLMClient.generate(
+      LLM.request({
+        model: OpenAIChat.route
+          .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
+          .model({ id: "gpt-4o-mini" }),
+        system: "system \uD800 \u{1F600}",
+        messages: [
+          Message.user("user \uDC00"),
+          Message.assistant([
+            Message.text("assistant \uD800"),
+            ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "input \uDC00" } }),
+          ]),
+          Message.tool({ id: "call_1", name: "lookup", result: { output: "result \uD800" } }),
+        ],
+        http: { body: { metadata: { "key\uD800": ["overlay \uDC00", "valid \u{1F600}"] } } },
+      }),
+    ).pipe(
+      Effect.provide(
+        dynamicResponse((input) =>
+          Effect.gen(function* () {
+            expect(decodeJson(input.text)).toMatchObject({
+              messages: [
+                { role: "system", content: "system \uFFFD \u{1F600}" },
+                { role: "user", content: "user \uFFFD" },
+                {
+                  role: "assistant",
+                  content: "assistant \uFFFD",
+                  tool_calls: [{ function: { arguments: '{"query":"input \uFFFD"}' } }],
+                },
+                { role: "tool", content: '{"output":"result \uFFFD"}' },
+              ],
+              metadata: { "key\uFFFD": ["overlay \uFFFD", "valid \u{1F600}"] },
+            })
+            return input.respond(sseEvents(deltaChunk({}, "stop")), {
+              headers: { "content-type": "text/event-stream" },
+            })
+          }),
+        ),
+      ),
+    ),
+  )
+
   it.effect("applies raw body overlays after protocol lowering", () =>
     LLMClient.generate(
       LLM.request({

packages/ai/test/executor.test.ts

@@ -1,10 +1,10 @@
 import { describe, expect } from "bun:test"
-import { Deferred, Effect, Fiber, Layer, Ref, Stream } from "effect"
-import { Headers, HttpClient, HttpClientError, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
-import { LLM, AIError } from "../src/index.js"
+import { Deferred, Effect, Fiber, Ref, Stream } from "effect"
+import { Headers, HttpClientError, HttpClientRequest } from "effect/unstable/http"
+import { LLM, AIError, HttpContext, InvalidProviderOutputError, TransportError } from "../src/index.js"
 import { LLMClient, RequestExecutor, WebSocketTransport, type WebSocketChannelExecutor } from "../src/route.js"
-import * as OpenAIChat from "../src/protocols/openai-chat.js"
-import * as OpenAI from "../src/providers/openai.js"
+import { route } from "../src/protocols/openai-chat.js"
+import { configure } from "../src/providers/openai.js"
 import { dynamicResponse, fixedResponse, systemError } from "./lib/http.js"
 import { deltaChunk } from "./lib/openai-chunks.js"
 import { sseEvents, sseRaw } from "./lib/sse.js"
@@ -18,66 +18,86 @@ const secretRequest = HttpClientRequest.post("https://provider.test/v1/chat?api_
   HttpClientRequest.setHeaders(Headers.fromInput({ authorization: "Bearer header-secret-456" })),
 )
 
-const responsesLayer = (responses: ReadonlyArray<Response>) =>
-  RequestExecutor.layer.pipe(
-    Layer.provide(
-      Layer.unwrap(
-        Effect.gen(function* () {
-          const cursor = yield* Ref.make(0)
-          return Layer.succeed(
-            HttpClient.HttpClient,
-            HttpClient.make((request) =>
-              Effect.gen(function* () {
-                const index = yield* Ref.getAndUpdate(cursor, (value) => value + 1)
-                return HttpClientResponse.fromWeb(request, responses[index] ?? responses[responses.length - 1])
-              }),
-            ),
-          )
-        }),
-      ),
-    ),
-  )
-
-const countedResponsesLayer = (attempts: Ref.Ref<number>, responses: ReadonlyArray<Response>) =>
-  RequestExecutor.layer.pipe(
-    Layer.provide(
-      Layer.unwrap(
-        Effect.gen(function* () {
-          const cursor = yield* Ref.make(0)
-          return Layer.succeed(
-            HttpClient.HttpClient,
-            HttpClient.make((request) =>
-              Effect.gen(function* () {
-                yield* Ref.update(attempts, (value) => value + 1)
-                const index = yield* Ref.getAndUpdate(cursor, (value) => value + 1)
-                return HttpClientResponse.fromWeb(request, responses[index] ?? responses[responses.length - 1])
-              }),
-            ),
-          )
-        }),
-      ),
-    ),
-  )
-
 const expectAIError = (error: unknown) => {
   expect(error).toBeInstanceOf(AIError)
   if (!(error instanceof AIError)) throw new Error("expected AIError")
+  expect(error.reason).toBeInstanceOf(Error)
+  expect(error.cause).toBe(error.reason)
   return error
 }
 
-const errorHttp = (error: AIError) => ("http" in error.reason ? error.reason.http : undefined)
 const largeProviderMessage = `Upstream request failed: ${"validation failed; ".repeat(1_000)}`
 
 describe("RequestExecutor", () => {
+  it.effect("preserves externally captured HTTP errors without inventing response context", () =>
+    Effect.sync(() => {
+      const cause = new Error("upstream request failed")
+      const body = '{"error":{"message":"Rate limited","trace":"original"}}'
+      const error = RequestExecutor.httpFailure({
+        message: "Rate limited",
+        url: request.url,
+        status: 429,
+        responseHeaders: { "Retry-After": "2", "X-Request-ID": "req_external" },
+        responseBody: body,
+        cause,
+      })
+
+      expect(error.message).toBe("Rate limited")
+      expect(error.reason).toMatchObject({ _tag: "RateLimit", retryAfterMs: 2000 })
+      expect(error.reason.body).toBe(body)
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason.http).toEqual(
+        new HttpContext({
+          url: request.url,
+          status: 429,
+          headers: { "retry-after": "2", "x-request-id": "req_external" },
+        }),
+      )
+      expect(RequestExecutor.httpFailure({ message: "No response", url: request.url }).reason.http).toBeUndefined()
+      expect(RequestExecutor.httpFailure({ message: "No URL", status: 500 }).reason.http).toBeUndefined()
+    }),
+  )
+
+  it.effect("retains the original body-read failure on an HTTP status error", () =>
+    Effect.gen(function* () {
+      const cause = new Error("response body disconnected")
+      const error = yield* Effect.gen(function* () {
+        const executor = yield* RequestExecutor.Service
+        return yield* executor.execute(request).pipe(Effect.flip)
+      }).pipe(
+        Effect.provide(
+          fixedResponse(
+            new ReadableStream({
+              start(controller) {
+                controller.error(cause)
+              },
+            }),
+            {
+              status: 503,
+              headers: { "x-request-id": "req_failed_body" },
+            },
+          ),
+        ),
+      )
+
+      expect(error.reason._tag).toBe("ProviderInternal")
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason.body).toBeUndefined()
+      expect(error.reason.http).toMatchObject({ status: 503, headers: { "x-request-id": "req_failed_body" } })
+    }),
+  )
+
   it.effect("parses response body failures at the executor seam", () =>
     Effect.gen(function* () {
       const executor = yield* RequestExecutor.Service
       const error = yield* RequestExecutor.stream(executor, secretRequest).pipe(Stream.runDrain, Effect.flip)
 
       expectAIError(error)
+      expect(error.message).toBe("ECONNRESET: disconnected query-secret-123 header-secret-456")
+      expect(error.reason.http).toMatchObject({ status: 200, url: secretRequest.url })
+      expect(error.reason.cause).toMatchObject({ code: "ECONNRESET" })
       expect(error.reason).toMatchObject({
         _tag: "Transport",
-        message: "ECONNRESET: disconnected query-secret-123 header-secret-456",
         transport: "http",
         operation: "read",
         code: "ECONNRESET",
@@ -85,15 +105,14 @@ describe("RequestExecutor", () => {
       })
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response(
-            new ReadableStream({
-              start(controller) {
-                controller.error(systemError("ECONNRESET", "disconnected query-secret-123 header-secret-456"))
-              },
-            }),
-          ),
-        ]),
+        fixedResponse(
+          new ReadableStream({
+            start(controller) {
+              controller.error(systemError("ECONNRESET", "disconnected query-secret-123 header-secret-456"))
+            },
+          }),
+          {},
+        ),
       ),
     ),
   )
@@ -104,23 +123,23 @@ describe("RequestExecutor", () => {
       const error = yield* RequestExecutor.stream(executor, secretRequest).pipe(Stream.runDrain, Effect.flip)
 
       expectAIError(error)
+      expect(error.message).toBe("ECONNRESET: socket closed")
+      expect(error.reason.cause).toBeInstanceOf(TypeError)
       expect(error.reason).toMatchObject({
         _tag: "Transport",
-        message: "ECONNRESET: socket closed",
         operation: "read",
         code: "ECONNRESET",
       })
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response(
-            new ReadableStream({
-              pull(controller) {
-                controller.error(new TypeError("fetch failed", { cause: systemError("ECONNRESET", "socket closed") }))
-              },
-            }),
-          ),
-        ]),
+        fixedResponse(
+          new ReadableStream({
+            pull(controller) {
+              controller.error(new TypeError("fetch failed", { cause: systemError("ECONNRESET", "socket closed") }))
+            },
+          }),
+          {},
+        ),
       ),
     ),
   )
@@ -133,8 +152,10 @@ describe("RequestExecutor", () => {
         .pipe(Effect.flip)
 
       expectAIError(error)
-      expect(error.reason.message).toBe("plugin rejected request")
-    }).pipe(Effect.provide(responsesLayer([]))),
+      expect(error.message).toBe("plugin rejected request")
+      expect(error.reason.cause).toBeInstanceOf(Error)
+      expect(error.reason.http).toBeUndefined()
+    }).pipe(Effect.provide(dynamicResponse(() => Effect.die(new Error("unexpected HTTP request"))))),
   )
 
   it.effect("reports the request sent by middleware", () =>
@@ -152,16 +173,11 @@ describe("RequestExecutor", () => {
         .pipe(Effect.flip)
 
       expectAIError(error)
+      expect(error.message).toBe("ECONNRESET: proxy disconnected proxy-secret")
+      expect(error.reason.http).toBeUndefined()
       expect(error.reason).toMatchObject({
         _tag: "Transport",
-        message: "ECONNRESET: proxy disconnected proxy-secret",
         url: "https://proxy.test/v1/chat?api_key=proxy-secret",
-        http: {
-          request: {
-            url: "https://proxy.test/v1/chat?api_key=proxy-secret",
-            headers: { authorization: "Bearer proxy-secret" },
-          },
-        },
       })
     }).pipe(
       Effect.provide(
@@ -188,11 +204,9 @@ describe("RequestExecutor", () => {
       expect(error.reason).toMatchObject({ _tag: "InvalidRequest", classification: "context-overflow" })
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response('{"error":{"code":"context_length_exceeded","message":"prompt too long"}}', {
-            status: 400,
-          }),
-        ]),
+        fixedResponse('{"error":{"code":"context_length_exceeded","message":"prompt too long"}}', {
+          status: 400,
+        }),
       ),
     ),
   )
@@ -206,9 +220,29 @@ describe("RequestExecutor", () => {
       expect(error.reason).toMatchObject({
         _tag: "InvalidRequest",
         classification: "payload-too-large",
-        http: { response: { status: 413 } },
       })
-    }).pipe(Effect.provide(responsesLayer([new Response("request too large", { status: 413 })]))),
+      expect(error.reason.http?.status).toBe(413)
+    }).pipe(Effect.provide(fixedResponse("request too large", { status: 413 }))),
+  )
+
+  it.effect("classifies Anthropic request_too_large as context overflow", () =>
+    Effect.gen(function* () {
+      const executor = yield* RequestExecutor.Service
+      const error = yield* executor.execute(request).pipe(Effect.flip)
+
+      expectAIError(error)
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidRequest",
+        classification: "context-overflow",
+      })
+      expect(error.reason.http?.status).toBe(413)
+    }).pipe(
+      Effect.provide(
+        fixedResponse('{"error":{"type":"request_too_large","message":"Request exceeds the maximum size"}}', {
+          status: 413,
+        }),
+      ),
+    ),
   )
 
   it.effect("does not classify ordinary invalid requests as context overflow", () =>
@@ -219,8 +253,8 @@ describe("RequestExecutor", () => {
       expectAIError(error)
       expect(error.reason).toMatchObject({ _tag: "InvalidRequest" })
       expect("classification" in error.reason ? error.reason.classification : undefined).toBeUndefined()
-      expect(error.reason.message).toBe("Provider request failed with HTTP 400")
-    }).pipe(Effect.provide(responsesLayer([new Response("invalid parameter", { status: 400 })]))),
+      expect(error.message).toBe("Provider request failed with HTTP 400")
+    }).pipe(Effect.provide(fixedResponse("invalid parameter", { status: 400 }))),
   )
 
   it.effect("preserves structured provider messages from large error bodies", () =>
@@ -229,20 +263,18 @@ describe("RequestExecutor", () => {
       const error = yield* executor.execute(request).pipe(Effect.flip)
 
       expectAIError(error)
-      expect(error.reason).toMatchObject({ _tag: "InvalidRequest", message: largeProviderMessage })
-      expect(errorHttp(error)?.body).toContain(largeProviderMessage)
-      expect(errorHttp(error)?.bodyTruncated).toBeUndefined()
+      expect(error.reason).toMatchObject({ _tag: "InvalidRequest" })
+      expect(error.message).toBe(largeProviderMessage)
+      expect(error.reason.body).toContain(largeProviderMessage)
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response(
-            JSON.stringify({
-              model: "gpt-5.6-sol",
-              error: { type: "invalid_request", message: largeProviderMessage },
-            }),
-            { status: 400 },
-          ),
-        ]),
+        fixedResponse(
+          JSON.stringify({
+            model: "test-model",
+            error: { type: "invalid_request", message: largeProviderMessage },
+          }),
+          { status: 400 },
+        ),
       ),
     ),
   )
@@ -255,9 +287,9 @@ describe("RequestExecutor", () => {
       expectAIError(error)
       expect(error.reason).toMatchObject({
         _tag: "InvalidRequest",
-        message: "Provider request failed with HTTP 400",
       })
-    }).pipe(Effect.provide(responsesLayer([new Response('{"error":{"message":"  "}}', { status: 400 })]))),
+      expect(error.message).toBe("Provider request failed with HTTP 400")
+    }).pipe(Effect.provide(fixedResponse('{"error":{"message":"  "}}', { status: 400 }))),
   )
 
   it.effect("classifies provider rate limits hidden behind HTTP 400", () =>
@@ -269,7 +301,7 @@ describe("RequestExecutor", () => {
 
           expectAIError(error)
           expect(error.reason).toMatchObject({ _tag: "RateLimit" })
-        }).pipe(Effect.provide(responsesLayer([new Response(body, { status: 400 })])))
+        }).pipe(Effect.provide(fixedResponse(body, { status: 400 })))
 
       yield* classify("Request rate increased too quickly")
       yield* classify('{"type":"error","error":{"type":"too_many_requests"}}')
@@ -286,7 +318,7 @@ describe("RequestExecutor", () => {
 
           expectAIError(error)
           expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
-        }).pipe(Effect.provide(responsesLayer([new Response(body, { status: 400 })])))
+        }).pipe(Effect.provide(fixedResponse(body, { status: 400 })))
 
       yield* classify('{"code":"resource_exhausted"}')
       yield* classify('{"code":"service_unavailable"}')
@@ -299,38 +331,27 @@ describe("RequestExecutor", () => {
       const error = yield* executor.execute(request).pipe(Effect.flip)
 
       expectAIError(error)
-      expect(error).toMatchObject({
-        reason: {
-          _tag: "RateLimit",
-          retryAfterMs: 0,
-          rateLimit: { retryAfterMs: 0 },
-          http: {
-            requestId: "req_123",
-            request: {
-              method: "POST",
-              url: "https://provider.test/v1/chat?api_key=secret&key=secret&debug=1",
-              headers: { authorization: "Bearer secret", "x-safe": "visible" },
-            },
-            response: {
-              status: 429,
-              headers: {
-                "retry-after-ms": "0",
-                "x-request-id": "req_123",
-                "x-api-key": "secret",
-              },
-            },
+      expect(error.reason).toMatchObject({
+        _tag: "RateLimit",
+        retryAfterMs: 0,
+        rateLimit: { retryAfterMs: 0 },
+        http: {
+          url: "https://provider.test/v1/chat?api_key=secret&key=secret&debug=1",
+          status: 429,
+          headers: {
+            "retry-after-ms": "0",
+            "x-request-id": "req_123",
+            "x-api-key": "secret",
           },
         },
       })
-      expect(errorHttp(error)?.body).toBe("rate limited")
+      expect(error.reason.body).toBe("rate limited")
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response("rate limited", {
-            status: 429,
-            headers: { "retry-after-ms": "0", "x-request-id": "req_123", "x-api-key": "secret" },
-          }),
-        ]),
+        fixedResponse("rate limited", {
+          status: 429,
+          headers: { "retry-after-ms": "0", "x-request-id": "req_123", "x-api-key": "secret" },
+        }),
       ),
     ),
   )
@@ -341,10 +362,9 @@ describe("RequestExecutor", () => {
       const error = yield* executor.execute(request).pipe(Effect.flip)
 
       expectAIError(error)
-      expect(errorHttp(error)?.request.headers["x-safe"]).toBe("visible")
-      expect(errorHttp(error)?.response?.headers["x-safe"]).toBe("response-secret")
+      expect(error.reason.http?.headers["x-safe"]).toBe("response-secret")
     }).pipe(
-      Effect.provide(responsesLayer([new Response("bad", { status: 400, headers: { "x-safe": "response-secret" } })])),
+      Effect.provide(fixedResponse("bad", { status: 400, headers: { "x-safe": "response-secret" } })),
       Effect.provideService(Headers.CurrentRedactedNames, ["x-safe"]),
     ),
   )
@@ -364,20 +384,18 @@ describe("RequestExecutor", () => {
       })
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response("rate limited", {
-            status: 429,
-            headers: {
-              "retry-after-ms": "0",
-              "x-ratelimit-limit-requests": "500",
-              "x-ratelimit-limit-tokens": "30000",
-              "x-ratelimit-remaining-requests": "499",
-              "x-ratelimit-remaining-tokens": "29900",
-              "x-ratelimit-reset-requests": "1s",
-              "x-ratelimit-reset-tokens": "10s",
-            },
-          }),
-        ]),
+        fixedResponse("rate limited", {
+          status: 429,
+          headers: {
+            "retry-after-ms": "0",
+            "x-ratelimit-limit-requests": "500",
+            "x-ratelimit-limit-tokens": "30000",
+            "x-ratelimit-remaining-requests": "499",
+            "x-ratelimit-remaining-tokens": "29900",
+            "x-ratelimit-reset-requests": "1s",
+            "x-ratelimit-reset-tokens": "10s",
+          },
+        }),
       ),
     ),
   )
@@ -388,8 +406,8 @@ describe("RequestExecutor", () => {
       const error = yield* executor.execute(request).pipe(Effect.flip)
 
       expectAIError(error)
-      expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
-      expect(errorHttp(error)?.rateLimit).toEqual({
+      expect(error.reason).toMatchObject({ _tag: "RateLimit" })
+      expect(error.reason._tag === "RateLimit" ? error.reason.rateLimit : undefined).toEqual({
         retryAfterMs: 0,
         limit: { requests: "100", "input-tokens": "10000" },
         remaining: { requests: "12", "input-tokens": "9000" },
@@ -397,20 +415,18 @@ describe("RequestExecutor", () => {
       })
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response("overloaded", {
-            status: 529,
-            headers: {
-              "retry-after-ms": "0",
-              "anthropic-ratelimit-requests-limit": "100",
-              "anthropic-ratelimit-requests-remaining": "12",
-              "anthropic-ratelimit-requests-reset": "2026-05-06T12:00:00Z",
-              "anthropic-ratelimit-input-tokens-limit": "10000",
-              "anthropic-ratelimit-input-tokens-remaining": "9000",
-              "anthropic-ratelimit-input-tokens-reset": "2026-05-06T12:00:10Z",
-            },
-          }),
-        ]),
+        fixedResponse("rate limited", {
+          status: 429,
+          headers: {
+            "retry-after-ms": "0",
+            "anthropic-ratelimit-requests-limit": "100",
+            "anthropic-ratelimit-requests-remaining": "12",
+            "anthropic-ratelimit-requests-reset": "2026-05-06T12:00:00Z",
+            "anthropic-ratelimit-input-tokens-limit": "10000",
+            "anthropic-ratelimit-input-tokens-remaining": "9000",
+            "anthropic-ratelimit-input-tokens-reset": "2026-05-06T12:00:10Z",
+          },
+        }),
       ),
     ),
   )
@@ -423,15 +439,20 @@ describe("RequestExecutor", () => {
         return yield* executor.execute(request).pipe(Effect.flip)
       }).pipe(
         Effect.provide(
-          countedResponsesLayer(attempts, [
-            new Response("busy", { status: 503, headers: { "retry-after-ms": "0" } }),
-            new Response("ok", { status: 200 }),
-          ]),
+          dynamicResponse((input) =>
+            Effect.gen(function* () {
+              const attempt = yield* Ref.getAndUpdate(attempts, (value) => value + 1)
+              return attempt === 0
+                ? input.respond("busy", { status: 503, headers: { "retry-after-ms": "0" } })
+                : input.respond("ok", { status: 200 })
+            }),
+          ),
         ),
       )
 
       expectAIError(error)
-      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", status: 503 })
+      expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
+      expect(error.reason.http?.status).toBe(503)
       expect(yield* Ref.get(attempts)).toBe(1)
     }),
   )
@@ -444,15 +465,14 @@ describe("RequestExecutor", () => {
           const error = yield* executor.execute(request).pipe(Effect.flip)
 
           expectAIError(error)
-          expect(error.reason).toMatchObject({ _tag: "ProviderInternal", status })
+          expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
+          expect(error.reason.http?.status).toBe(status)
         }).pipe(
           Effect.provide(
-            responsesLayer([
-              new Response("provider failure", {
-                status,
-                headers: { "retry-after-ms": "0" },
-              }),
-            ]),
+            fixedResponse("provider failure", {
+              status,
+              headers: { "retry-after-ms": "0" },
+            }),
           ),
         )
 
@@ -463,21 +483,28 @@ describe("RequestExecutor", () => {
 
   it.effect("preserves large authentication error bodies", () =>
     Effect.gen(function* () {
-      const executor = yield* RequestExecutor.Service
-      const error = yield* executor.execute(request).pipe(Effect.flip)
+      const attempts = yield* Ref.make(0)
+      const error = yield* Effect.gen(function* () {
+        const executor = yield* RequestExecutor.Service
+        return yield* executor.execute(request).pipe(Effect.flip)
+      }).pipe(
+        Effect.provide(
+          dynamicResponse((input) =>
+            Effect.gen(function* () {
+              const attempt = yield* Ref.getAndUpdate(attempts, (value) => value + 1)
+              return attempt === 0
+                ? input.respond("x".repeat(20_000), { status: 401 })
+                : input.respond("should not retry", { status: 200 })
+            }),
+          ),
+        ),
+      )
 
       expectAIError(error)
       expect(error.reason).toMatchObject({ _tag: "Authentication" })
-      expect(errorHttp(error)?.bodyTruncated).toBeUndefined()
-      expect(errorHttp(error)?.body).toHaveLength(20_000)
-    }).pipe(
-      Effect.provide(
-        responsesLayer([
-          new Response("x".repeat(20_000), { status: 401 }),
-          new Response("should not retry", { status: 200 }),
-        ]),
-      ),
-    ),
+      expect(error.reason.body).toHaveLength(20_000)
+      expect(yield* Ref.get(attempts)).toBe(1)
+    }),
   )
 
   it.effect("preserves response body fields", () =>
@@ -486,16 +513,12 @@ describe("RequestExecutor", () => {
       const error = yield* executor.execute(request).pipe(Effect.flip)
 
       expectAIError(error)
-      expect(errorHttp(error)?.body).toBe(
-        '{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}',
-      )
+      expect(error.reason.body).toBe('{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}')
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response('{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}', {
-            status: 400,
-          }),
-        ]),
+        fixedResponse('{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}', {
+          status: 400,
+        }),
       ),
     ),
   )
@@ -506,12 +529,10 @@ describe("RequestExecutor", () => {
       const error = yield* executor.execute(secretRequest).pipe(Effect.flip)
 
       expectAIError(error)
-      expect(errorHttp(error)?.body).toBe("provider echoed query-secret-123 and authorization header-secret-456")
+      expect(error.reason.body).toBe("provider echoed query-secret-123 and authorization header-secret-456")
     }).pipe(
       Effect.provide(
-        responsesLayer([
-          new Response("provider echoed query-secret-123 and authorization header-secret-456", { status: 400 }),
-        ]),
+        fixedResponse("provider echoed query-secret-123 and authorization header-secret-456", { status: 400 }),
       ),
     ),
   )
@@ -519,9 +540,7 @@ describe("RequestExecutor", () => {
   it.effect("does not re-execute after a successful response reaches stream parsing", () =>
     Effect.gen(function* () {
       const attempts = yield* Ref.make(0)
-      const model = OpenAIChat.route
-        .with({ endpoint: { baseURL: "https://api.openai.test/v1" } })
-        .model({ id: "gpt-4o-mini" })
+      const model = route.with({ endpoint: { baseURL: "https://api.openai.test/v1" } }).model({ id: "gpt-4o-mini" })
       const error = yield* LLMClient.generate(LLM.request({ model, prompt: "Say hello." })).pipe(
         Effect.provide(
           dynamicResponse((input) =>
@@ -543,19 +562,170 @@ describe("RequestExecutor", () => {
 
       expectAIError(error)
       expect(error.reason).toMatchObject({ _tag: "InvalidProviderOutput" })
+      expect(error.reason.body).toBe("not-json")
+      expect(error.reason.cause).toBeDefined()
+      expect(error.reason.http).toMatchObject({ status: 200, headers: { "content-type": "text/event-stream" } })
       expect(yield* Ref.get(attempts)).toBe(1)
     }),
   )
 })
 
 describe("WebSocket channel execution", () => {
-  const model = OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses("gpt-4.1-mini")
+  const model = configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses("gpt-4.1-mini")
   const request = LLM.request({ model, prompt: "Say hello." })
   const frames = [
     JSON.stringify({ type: "response.output_text.delta", item_id: "msg_1", delta: "Hi" }),
     JSON.stringify({ type: "response.completed", response: { id: "resp_1" } }),
   ]
 
+  it.effect("preserves close reasons and native event causes without fabricated HTTP metadata", () =>
+    Effect.gen(function* () {
+      class TestSocket extends EventTarget {
+        readyState = globalThis.WebSocket.OPEN
+        send() {}
+        close() {}
+      }
+      const socket = new TestSocket()
+      const connection = yield* WebSocketTransport.fromWebSocket(
+        // oxlint-disable-next-line typescript-eslint/no-unsafe-type-assertion
+        socket as unknown as globalThis.WebSocket,
+        { url: "wss://provider.test/responses", headers: Headers.empty },
+      )
+      const event = new CloseEvent("close", { code: 1011, reason: "upstream trace: req_close" })
+      socket.dispatchEvent(event)
+      const error = yield* connection.messages.pipe(Stream.runDrain, Effect.flip)
+
+      expect(error.reason).toMatchObject({ _tag: "Transport", code: "1011", phase: "close" })
+      expect(error.message).toBe("WebSocket closed with code 1011")
+      expect(error.reason.body).toBe(event.reason)
+      expect(error.reason.cause).toBe(event)
+      expect(error.reason.http).toBeUndefined()
+      yield* connection.close
+    }),
+  )
+
+  it.effect("preserves opening event errors and native send exceptions", () =>
+    Effect.gen(function* () {
+      const cause = new Error("native send failed")
+      class TestSocket extends EventTarget {
+        readyState = globalThis.WebSocket.CONNECTING
+        send() {
+          throw cause
+        }
+        close() {}
+      }
+      const socket = new TestSocket()
+      const open = WebSocketTransport.fromWebSocket(
+        // oxlint-disable-next-line typescript-eslint/no-unsafe-type-assertion
+        socket as unknown as globalThis.WebSocket,
+        { url: "wss://provider.test/responses", headers: Headers.empty },
+      )
+      const fiber = yield* open.pipe(Effect.flip, Effect.forkChild({ startImmediately: true }))
+      const event = new ErrorEvent("error", { message: "handshake rejected", error: cause })
+      socket.dispatchEvent(event)
+      const error = yield* Fiber.join(fiber)
+      expect(error.reason.cause).toBe(cause)
+      expect(error.message).toContain("handshake rejected")
+      expect(error.reason.http).toBeUndefined()
+
+      socket.readyState = globalThis.WebSocket.OPEN
+      const connection = yield* open
+      const sent = yield* connection.sendText("create").pipe(Effect.flip)
+      expect(sent.reason.cause).toBe(cause)
+      expect(sent.message).toBe(cause.message)
+      yield* connection.close
+    }),
+  )
+
+  it.effect("preserves raw driver failures and known upgrade metadata", () =>
+    Effect.gen(function* () {
+      const cause = new Error("driver validation failed")
+      const frame = '{ "error": "failed", "trace": "original" }'
+      const http = new HttpContext({
+        url: "https://provider.test/responses",
+        status: 101,
+        headers: { upgrade: "websocket" },
+      })
+      const executor = WebSocketTransport.makeDirect({
+        open: () =>
+          Effect.succeed({
+            http,
+            sendText: () => Effect.void,
+            messages: Stream.make(frame),
+            close: Effect.void,
+          }),
+      })
+      const execution = yield* executor.execute({
+        id: "exchange_error",
+        connect: { url: "wss://provider.test/responses", headers: Headers.empty },
+        fallback: () => Stream.empty,
+        driver: {
+          create: () => Effect.succeed({ message: "create", mode: "full" }),
+          observe: () =>
+            Effect.succeed({
+              type: "provider-failure",
+              error: new AIError({
+                reason: new InvalidProviderOutputError({
+                  message: "Driver failed",
+                  cause,
+                  body: "narrowed",
+                }),
+              }),
+            }),
+        },
+      })
+      const error = yield* execution.frames.pipe(Stream.runDrain, Effect.flip)
+
+      expect(error.message).toBe("Driver failed")
+      expect(error.reason.body).toBe(frame)
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason.http).toBe(http)
+      expect(execution.http).toBe(http)
+    }),
+  )
+
+  it.effect("retains diagnostic fields when annotating transport delivery", () =>
+    Effect.gen(function* () {
+      const cause = new Error("connection closed")
+      const executor = WebSocketTransport.makeDirect({
+        open: () =>
+          Effect.succeed({
+            sendText: () => Effect.void,
+            messages: Stream.fail(
+              new AIError({
+                reason: new TransportError({
+                  message: "Socket closed",
+                  transport: "websocket",
+                  operation: "read",
+                  phase: "close",
+                  recovery: "retry-full",
+                  body: "server close detail",
+                  cause,
+                }),
+              }),
+            ),
+            close: Effect.void,
+          }),
+      })
+      const execution = yield* executor.execute({
+        id: "exchange_closed",
+        connect: { url: "wss://provider.test/responses", headers: Headers.empty },
+        fallback: () => Stream.empty,
+        driver: {
+          create: () => Effect.succeed({ message: "create", mode: "full" }),
+          observe: (_create, frame) => Effect.succeed({ type: "frame", frame }),
+        },
+      })
+      const error = yield* execution.frames.pipe(Stream.runDrain, Effect.flip)
+
+      expect(error.message).toBe("Socket closed")
+      expect(error.reason.body).toBe("server close detail")
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason).toMatchObject({ phase: "close", delivery: "ambiguous", recovery: "retry-full" })
+      expect(error.reason.http).toBeUndefined()
+    }),
+  )
+
   it.effect("runs a channel driver through the direct executor", () =>
     Effect.gen(function* () {
       const sent = yield* Ref.make("")

packages/ai/test/fixtures/recordings/azure/chat-streams-text.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/chat-streams-text",
     "recordedAt": "2026-08-23T17:21:53.198Z"
   },

packages/ai/test/fixtures/recordings/azure/responses-calls-a-tool.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/responses-calls-a-tool",
     "recordedAt": "2026-08-23T17:21:55.170Z"
   },

packages/ai/test/fixtures/recordings/azure/responses-continues-after-a-tool-result.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/responses-continues-after-a-tool-result",
     "recordedAt": "2026-08-23T17:21:56.397Z"
   },

packages/ai/test/fixtures/recordings/azure/responses-streams-text.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/responses-streams-text",
     "recordedAt": "2026-08-23T17:21:54.158Z"
   },

packages/ai/test/fixtures/recordings/cerebras-chat/cerebras-gpt-oss-120b-text.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "cerebras",
+    "route": "cerebras-chat",
+    "transport": "http",
+    "model": "gpt-oss-120b",
+    "tags": ["prefix:cerebras-chat", "provider:cerebras", "text", "golden"],
+    "name": "cerebras-chat/cerebras-gpt-oss-120b-text",
+    "recordedAt": "2026-08-25T23:55:27.619Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.cerebras.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-oss-120b\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":256}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"role\":\"assistant\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\"The\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" user says: \\\"Reply exactly with: Hello!\\\" So we must output exactly\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" \\\"Hello!\\\" with no extra characters, no formatting\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\". Ensure no extra spaces or new\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\"lines? Probably just \\\"Hello!\\\".\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" Usually we output exactly that.\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"content\":\"Hello!\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\",\"usage\":{\"total_tokens\":142,\"completion_tokens\":58,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0,\"reasoning_tokens\":46},\"prompt_tokens\":84,\"prompt_tokens_details\":{\"cached_tokens\":0}},\"time_info\":{\"created\":1787702127.645281,\"queue_time\":0.003817115,\"prompt_time\":0.001587193,\"completion_time\":0.029805929,\"total_time\":0.036823272705078125}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/cerebras-chat/cerebras-gpt-oss-120b-tool-call.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "cerebras",
+    "route": "cerebras-chat",
+    "transport": "http",
+    "model": "gpt-oss-120b",
+    "tags": ["prefix:cerebras-chat", "provider:cerebras", "tool", "tool-call", "golden"],
+    "name": "cerebras-chat/cerebras-gpt-oss-120b-tool-call",
+    "recordedAt": "2026-08-25T23:55:28.454Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.cerebras.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-oss-120b\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":512}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"role\":\"assistant\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\"We\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\" need to call the function get_weather with city \\\"\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\"Paris\\\".\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"name\":\"get_weather\",\"arguments\":\"\"},\"type\":\"function\",\"id\":\"3d860cefe\",\"index\":0}]},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},\"type\":\"function\",\"index\":0}]},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\",\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\",\"usage\":{\"total_tokens\":174,\"completion_tokens\":37,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0,\"reasoning_tokens\":13},\"prompt_tokens\":137,\"prompt_tokens_details\":{\"cached_tokens\":0}},\"time_info\":{\"created\":1787702127.8019717,\"queue_time\":0.31196235,\"prompt_time\":0.005234764,\"completion_time\":0.020198402,\"total_time\":0.702225923538208}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/cloudflare-workers-ai/cloudflare-workers-ai-gpt-oss-20b-tools-tool-call.json

@@ -18,7 +18,7 @@
         "headers": {
           "content-type": "application/json"
         },
-        "body": "{\"model\":\"@cf/openai/gpt-oss-20b\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":120,\"temperature\":0}"
+        "body": "{\"model\": \"@cf/openai/gpt-oss-20b\", \"messages\": [{\"role\": \"system\", \"content\": \"Call tools exactly as requested.\"}, {\"role\": \"user\", \"content\": \"Call get_weather with city exactly Paris.\"}], \"tools\": [{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get current weather for a city.\", \"parameters\": {\"type\": \"object\", \"properties\": {\"city\": {\"type\": \"string\"}}, \"required\": [\"city\"], \"additionalProperties\": false}, \"strict\": false}}], \"tool_choice\": {\"type\": \"function\", \"function\": {\"name\": \"get_weather\"}}, \"stream\": true, \"stream_options\": {\"include_usage\": true}, \"max_tokens\": 120, \"temperature\": 0}"
       },
       "response": {
         "status": 200,

packages/ai/test/fixtures/recordings/deepinfra-chat/deepinfra-llama-3-3-70b-text.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "text", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-text",
+    "recordedAt": "2026-08-26T00:34:03.019Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":40,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Hello\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"!\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":3.46e-6,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/deepinfra-chat/deepinfra-llama-3-3-70b-tool-call.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-call", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-call",
+    "recordedAt": "2026-08-26T00:34:04.173Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_SMfBjXa8eCmHLyjfeyARxe3a\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":0.000027480000000000005,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/deepinfra-chat/deepinfra-llama-3-3-70b-tool-loop.json

@@ -0,0 +1,50 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-loop", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-loop",
+    "recordedAt": "2026-08-26T00:34:05.656Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":0.000028780000000000002,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}],\"reasoning_content\":\"\"},{\"role\":\"tool\",\"tool_call_id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Paris\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" is\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" sunny\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\".\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":0.0000292,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/google-vertex/calls-a-tool.json

@@ -1,11 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
     "name": "google-vertex/calls-a-tool",
     "recordedAt": "2026-08-23T17:21:51.036Z"
   },

packages/ai/test/fixtures/recordings/google-vertex/continues-after-a-tool-result.json

@@ -1,11 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
     "name": "google-vertex/continues-after-a-tool-result",
     "recordedAt": "2026-08-23T17:21:51.853Z"
   },

packages/ai/test/fixtures/recordings/google-vertex/streams-text.json

@@ -1,11 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
     "name": "google-vertex/streams-text",
     "recordedAt": "2026-08-23T17:21:50.112Z"
   },