test: run only affected unit suites

test: skip unaffected app e2e tests (#45030 )
Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>
2026-08-25 19:16:15 +00:00 · 2026-08-25 16:40:08 +00:00 · 2026-08-26 00:33:14 +08:00 · 2026-08-25 12:30:20 -04:00 · 2026-08-25 11:27:40 -05:00 · 2026-08-25 16:26:05 +00:00
814 changed files with 66986 additions and 26726 deletions
@@ -0,0 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Fix OpenCode Console device authorization URLs when the server returns an origin-rooted verification path.
@@ -0,0 +1,6 @@
+---
+"@opencode-ai/client": patch
+"@opencode-ai/plugin": patch
+---
+
+Add form reply and cancellation operations that reconcile terminal forms in the local TUI projection.
@@ -27,7 +27,6 @@ jobs:
        working-directory: packages/www
        run: bun run build
        env:
-          BLUME_ENV: ${{ github.ref_name == 'v2' && 'production' || 'dev' }}
          CLOUDFLARE_ENV: ${{ github.ref_name == 'v2' && 'production' || 'dev' }}

      - name: Deploy
@@ -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!

@@ -91,7 +91,7 @@ jobs:

      - uses: ./.github/actions/setup-bun
        with:
-          bun-version: canary # Bun 1.4 until its stable release is published
+          bun-version: 1.4.0

      - name: Setup git committer
        id: committer
@@ -113,7 +113,7 @@ jobs:
        id: build
        run: ./packages/cli/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
        env:
-          BUN_COMPILE_RELEASE: canary
+          BUN_COMPILE_RELEASE: bun-v1.4.0
          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
          GH_REPO: ${{ needs.version.outputs.repo }}
@@ -195,9 +195,33 @@ jobs:
          path: packages/cli/dist/cli-*
          if-no-files-found: error

-  build-node-cli:
+  build-node-app-archive:
    needs: version
-    if: github.repository == 'anomalyco/opencode' && false # Temporarily disabled
+    runs-on: blacksmith-4vcpu-ubuntu-2404
+    timeout-minutes: 30
+    if: github.repository == 'anomalyco/opencode'
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+
+      - name: Build app archive
+        run: bun packages/cli/script/build-node.ts --app-archive-only --app-archive=.cache/app-archive.bin --skip-install
+        env:
+          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
+          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
+
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: opencode-node-app-archive
+          path: packages/cli/.cache/app-archive.bin
+          if-no-files-found: error
+
+  build-node-cli:
+    needs:
+      - version
+      - build-node-app-archive
+    if: github.repository == 'anomalyco/opencode'
    strategy:
      fail-fast: false
      matrix:
@@ -210,6 +234,7 @@ jobs:
            host: macos-26
          - target: windows-arm64
            host: blacksmith-4vcpu-windows-2025
+            bun_install_flags: --cpu=*
          - target: windows-x64
            host: blacksmith-4vcpu-windows-2025
    runs-on: ${{ matrix.settings.host }}
@@ -221,14 +246,19 @@ jobs:

      - uses: ./.github/actions/setup-bun
        with:
-          install-flags: --os=* --cpu=*
+          install-flags: ${{ matrix.settings.bun_install_flags }}

      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
        with:
          node-version: "26.4.0"

+      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        with:
+          name: opencode-node-app-archive
+          path: packages/cli/.cache
+
      - name: Build
-        run: bun packages/cli/script/build-node.ts --target=${{ matrix.settings.target }} --skip-install --outdir=dist/node
+        run: bun packages/cli/script/build-node.ts --target=${{ matrix.settings.target }} --skip-install --outdir=dist/node --app-archive=.cache/app-archive.bin
        env:
          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
@@ -543,6 +573,7 @@ jobs:
      - version
      - build-cli
      - sign-cli-macos
+      - build-node-app-archive
      - build-node-cli
      - sign-cli-windows
      - build-electron
@@ -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.pull_request.base.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
@@ -80,9 +111,27 @@ 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.pull_request.base.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
+        working-directory: packages/sdk
+        run: bun run verify:package

      - name: Verify compiled service lifecycle
        if: always()
@@ -122,7 +171,8 @@ jobs:

  e2e:
    name: e2e (${{ matrix.settings.name }})
-    if: github.ref_name != 'v2' && github.head_ref != 'v2'
+    needs: affected
+    if: needs.affected.outputs.app == 'true' && github.ref_name != 'v2' && github.head_ref != 'v2'
    strategy:
      fail-fast: false
      matrix:
@@ -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", "@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"]

 [test]
 root = "./do-not-run-tests-from-root"
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-PatsUdaitHvSUpS5gkC5J2rsUNB5vwJKqHdlOFaKk70=",
-    "aarch64-linux": "sha256-gTRQMAADH/SpQ8yh+YS2IcnmQxnJcU4FUah2YfrKeP8=",
-    "aarch64-darwin": "sha256-QTqlwmugYh+iu5Sh/Hxv01NXH/OhzcQ8ObVUbA9A8AM=",
-    "x86_64-darwin": "sha256-0DPAbNCVw2nUMWkIGEhB6saMdxRRwAJi7wAoWCQc7xQ="
+    "x86_64-linux": "sha256-3Jx1Q7hl+Y0Log/k2vd5y6dzBpzFKWlhShPESxn1Rm4=",
+    "aarch64-linux": "sha256-EiiI6g01oBIrExCMAUgT3w82P0fvu4FAJhI32C+ze0I=",
+    "aarch64-darwin": "sha256-s+w49HRp1+ewtiTaU65tPWjUiO1NQw3kzfemMEEQZb0=",
+    "x86_64-darwin": "sha256-/Ee5V7pnL/qm3c4ZHeWEjH7FhGVXArXryOugbG5vsz8="
  }
 }
@@ -39,20 +39,20 @@
      "packages/stats/*"
    ],
    "catalog": {
-      "@effect/opentelemetry": "4.0.0-rc.110",
-      "@effect/platform-node": "4.0.0-rc.110",
-      "@effect/platform-node-shared": "4.0.0-rc.110",
-      "@effect/sql-sqlite-bun": "4.0.0-rc.110",
+      "@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",
      "@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": "0.5.6",
-      "@opentui/keymap": "0.5.6",
-      "@opentui/solid": "0.5.6",
-      "@tanstack/solid-virtual": "3.13.32",
+      "@opentui/core": "0.5.8",
+      "@opentui/keymap": "0.5.8",
+      "@opentui/solid": "0.5.8",
+      "@tanstack/solid-virtual": "3.13.37",
      "@shikijs/stream": "4.2.0",
      "@standard-schema/spec": "1.1.0",
      "ulid": "3.0.1",
@@ -71,9 +71,9 @@
      "@tailwindcss/vite": "4.1.11",
      "diff": "8.0.2",
      "dompurify": "3.3.1",
-      "drizzle-kit": "1.0.0-rc.2",
-      "drizzle-orm": "1.0.0-rc.2",
-      "effect": "4.0.0-rc.110",
+      "drizzle-kit": "1.0.0-rc.5-ab785fc",
+      "drizzle-orm": "1.0.0-rc.5-169397b",
+      "effect": "4.0.0-rc.111",
      "ai": "6.0.168",
      "cross-spawn": "7.0.6",
      "hono": "4.10.7",
@@ -166,7 +166,6 @@
    "@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",
-    "drizzle-orm@1.0.0-rc.2": "patches/drizzle-orm@1.0.0-rc.2.patch",
    "solid-js@1.9.10": "patches/solid-js@1.9.10.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",
@@ -174,7 +173,7 @@
    "@ai-sdk/google@3.0.73": "patches/@ai-sdk%2Fgoogle@3.0.73.patch",
    "@pierre/trees@1.0.0-beta.4": "patches/@pierre%2Ftrees@1.0.0-beta.4.patch",
    "@modelcontextprotocol/sdk@1.29.0": "patches/@modelcontextprotocol%2Fsdk@1.29.0.patch",
-    "@tanstack/virtual-core@3.17.3": "patches/@tanstack%2Fvirtual-core@3.17.3.patch",
+    "@tanstack/virtual-core@3.17.8": "patches/@tanstack%2Fvirtual-core@3.17.8.patch",
    "@ff-labs/fff-bun@0.10.5": "patches/@ff-labs%2Ffff-bun@0.10.5.patch"
  }
 }
@@ -213,7 +213,7 @@ Errors must be expressed as `ToolFailure`. The runtime catches it and emits a `t
 - Input failed the `parameters` Schema.
 - The handler returned a `ToolFailure`.

-Provider-defined / hosted tools (Anthropic `web_search` / `code_execution` / `web_fetch`, OpenAI Responses `web_search_call` / `file_search_call` / `code_interpreter_call` / `mcp_call` / `local_shell_call` / `image_generation_call` / `computer_use_call`) pass through the runtime untouched:
+Provider-defined / hosted tools (Anthropic `web_search` / `code_execution` / `web_fetch`, OpenAI Responses `web_search_call` / `file_search_call` / `code_interpreter_call` / `mcp_call` / `image_generation_call` / `computer_use_call`) pass through the runtime untouched:

 - Routes surface the model's call as a `tool-call` event with `providerExecuted: true`, and the provider's result as a matching `tool-result` event with `providerExecuted: true`.
 - Callers detect `providerExecuted` on `tool-call` and **skip local dispatch** — no handler is invoked and no `tool-error` is raised for "unknown tool". The provider already executed it.
@@ -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,
+        ),
    })
  }),
 )
@@ -1,4 +1,5 @@
-import { Effect, Schema } from "effect"
+import { Buffer } from "node:buffer"
+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"
@@ -41,6 +42,7 @@ const SSE_EVENTS = new Set([
  "content_block_start",
  "content_block_delta",
  "content_block_stop",
+  "ping",
  "error",
 ])
 export const framing = Framing.sseEvents(SSE_EVENTS)
@@ -53,7 +55,7 @@ export type ThinkingInput =
  | {
      readonly type: "disabled"
    }
-  | ({ readonly type: "enabled" } & (
+  | ({ readonly type: "enabled"; readonly display?: "summarized" | "omitted" } & (
      | { readonly budgetTokens: number; readonly budget_tokens?: number }
      | { readonly budgetTokens?: number; readonly budget_tokens: number }
    ))
@@ -62,6 +64,19 @@ export interface OptionsInput {
  readonly [key: string]: unknown
  readonly thinking?: ThinkingInput
  readonly effort?: string
+  readonly service_tier?: "auto" | "standard_only"
+  readonly serviceTier?: "auto" | "standard_only"
+  // 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 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 }
 }

 export type ProviderOptionsInput = OptionsInput
@@ -81,32 +96,68 @@ const AnthropicTextBlock = Schema.Struct({
 })
 type AnthropicTextBlock = Schema.Schema.Type<typeof AnthropicTextBlock>

+// SDK: Base64ImageSource:201 {type:"base64", media_type:"image/jpeg"|... , data}, URLImageSource:3817 {type:"url", url}, FileImageSource:2350 {type:"file", file_id}
+// SDK: ImageBlockParam:2356 {source: Base64|URL|File, cache_control, transformations:2381 {oversized_image?}}
+const AnthropicBase64ImageSource = Schema.Struct({
+  type: Schema.tag("base64"),
+  media_type: Schema.String,
+  data: Schema.String,
+})
+const AnthropicURLImageSource = Schema.Struct({ type: Schema.tag("url"), url: Schema.String })
+const AnthropicFileImageSource = Schema.Struct({ type: Schema.tag("file"), file_id: Schema.String })
+const AnthropicImageSource = Schema.Union([
+  AnthropicBase64ImageSource,
+  AnthropicURLImageSource,
+  AnthropicFileImageSource,
+])
+const AnthropicImageTransformations = Schema.Struct({
+  oversized_image: Schema.optional(Schema.Literals(["downsize", "error"])),
+})
+
 const AnthropicImageBlock = Schema.Struct({
  type: Schema.tag("image"),
-  source: Schema.Struct({
-    type: Schema.tag("base64"),
-    media_type: Schema.String,
-    data: Schema.String,
-  }),
+  source: AnthropicImageSource,
  cache_control: Schema.optional(AnthropicCacheControl),
+  transformations: Schema.optional(AnthropicImageTransformations),
 })
 type AnthropicImageBlock = Schema.Schema.Type<typeof AnthropicImageBlock>

+// SDK: Base64PDFSource:209 {type:"base64", media_type:"application/pdf", data}, PlainTextSource:2716 {type:"text", media_type:"text/plain", data},
+// SDK: URLPDFSource:3823 {type:"url", url}, FileDocumentSource:2344 {type:"file", file_id}, ContentBlockSource:2266 {type:"content", content}
+// SDK: DocumentBlockParam:2297 {source: 5-way union, cache_control, citations, context, title}
+const AnthropicBase64PDFSource = Schema.Struct({
+  type: Schema.tag("base64"),
+  media_type: Schema.Literal("application/pdf"),
+  data: Schema.String,
+})
+const AnthropicPlainTextSource = Schema.Struct({
+  type: Schema.tag("text"),
+  media_type: Schema.Literal("text/plain"),
+  data: Schema.String,
+})
+const AnthropicURLPDFSource = Schema.Struct({ type: Schema.tag("url"), url: Schema.String })
+const AnthropicFileDocumentSource = Schema.Struct({ type: Schema.tag("file"), file_id: Schema.String })
+const AnthropicDocumentSource = Schema.Union([
+  AnthropicBase64PDFSource,
+  AnthropicPlainTextSource,
+  AnthropicURLPDFSource,
+  AnthropicFileDocumentSource,
+])
+
 const AnthropicDocumentBlock = Schema.Struct({
  type: Schema.tag("document"),
-  source: Schema.Struct({
-    type: Schema.tag("base64"),
-    media_type: Schema.Literal("application/pdf"),
-    data: Schema.String,
-  }),
+  source: AnthropicDocumentSource,
  cache_control: Schema.optional(AnthropicCacheControl),
+  title: Schema.optional(Schema.String),
+  context: Schema.optional(Schema.String),
+  citations: Schema.optional(Schema.Struct({ enabled: Schema.Boolean })),
 })
 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),
 })

@@ -204,14 +255,18 @@ const AnthropicTool = Schema.Struct({
 type AnthropicTool = Schema.Schema.Type<typeof AnthropicTool>

 const AnthropicToolChoice = Schema.Union([
-  Schema.Struct({ type: Schema.Literals(["auto", "any", "none"]) }),
-  Schema.Struct({ type: Schema.tag("tool"), name: Schema.String }),
+  Schema.Struct({
+    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) }),
 ])

 const AnthropicThinking = Schema.Union([
  Schema.Struct({
    type: Schema.tag("enabled"),
    budget_tokens: Schema.Number,
+    display: Schema.optional(Schema.Literals(["summarized", "omitted"])),
  }),
  Schema.Struct({
    type: Schema.tag("adaptive"),
@@ -222,10 +277,28 @@ const AnthropicThinking = Schema.Union([
  }),
 ])

+// SDK OutputConfig:2684 {effort?: "low"|"medium"|"high"|"xhigh"|"max"|null, format?: JSONOutputFormat:2399}
+const AnthropicJsonOutputFormat = Schema.Struct({
+  type: Schema.Literal("json_schema"),
+  schema: JsonObject,
+})
 const AnthropicOutputConfig = Schema.Struct({
  effort: Schema.optional(Schema.String),
+  format: Schema.optional(Schema.NullOr(AnthropicJsonOutputFormat)),
 })

+// SDK Metadata:2649 {user_id?: string|null}
+const AnthropicMetadata = Schema.Struct({ user_id: optionalNull(Schema.String) })
+
+// SDK MessageCreateParamsContainer:2596 ContainerParams|string; ContainerParams:2172 {id?, skills?}
+const AnthropicContainer = Schema.Union([
+  Schema.String,
+  Schema.Struct({
+    id: optionalNull(Schema.String),
+    skills: optionalNull(Schema.Array(JsonObject)),
+  }),
+])
+
 const AnthropicBodyFields = {
  model: Schema.String,
  system: optionalArray(AnthropicTextBlock),
@@ -240,6 +313,12 @@ const AnthropicBodyFields = {
  stop_sequences: optionalArray(Schema.String),
  thinking: Schema.optional(AnthropicThinking),
  output_config: Schema.optional(AnthropicOutputConfig),
+  // SDK top-level passthrough: cache_control:4638, container:4643, inference_geo:4649, metadata:4654, service_tier:4670
+  cache_control: Schema.optional(AnthropicCacheControl),
+  container: Schema.optional(Schema.NullOr(AnthropicContainer)),
+  inference_geo: Schema.optional(Schema.NullOr(Schema.String)),
+  metadata: Schema.optional(AnthropicMetadata),
+  service_tier: Schema.optional(Schema.Literals(["auto", "standard_only"])),
 }
 export const AnthropicMessagesBody = Schema.Struct(AnthropicBodyFields)
 export type AnthropicMessagesBody = Schema.Schema.Type<typeof AnthropicMessagesBody>
@@ -282,6 +361,8 @@ 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),
@@ -292,13 +373,15 @@ 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
@@ -368,10 +451,26 @@ const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition, inputSc

 const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
  ProviderShared.matchToolChoice("Anthropic Messages", toolChoice, {
-    auto: () => ({ type: "auto" as const }),
+    auto: () => ({
+      type: "auto" as const,
+      ...(toolChoice.disableParallelToolUse === undefined
+        ? {}
+        : { disable_parallel_tool_use: toolChoice.disableParallelToolUse }),
+    }),
    none: () => ({ type: "none" as const }),
-    required: () => ({ type: "any" as const }),
-    tool: (name) => ({ type: "tool" as const, name }),
+    required: () => ({
+      type: "any" as const,
+      ...(toolChoice.disableParallelToolUse === undefined
+        ? {}
+        : { disable_parallel_tool_use: toolChoice.disableParallelToolUse }),
+    }),
+    tool: (name) => ({
+      type: "tool" as const,
+      name,
+      ...(toolChoice.disableParallelToolUse === undefined
+        ? {}
+        : { disable_parallel_tool_use: toolChoice.disableParallelToolUse }),
+    }),
  })

 const scrubToolCallID = (id: string) => id.replace(/[^a-zA-Z0-9_-]/g, "_")
@@ -410,7 +509,152 @@ const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult
  return { type: wireType, tool_use_id: scrubToolCallID(part.id), content: payload } satisfies AnthropicServerToolResultBlock
 })

-const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (part: MediaPart) {
+const fileIdFromMetadata = (metadata: MediaPart["metadata"]): string | undefined => {
+  if (!ProviderShared.isRecord(metadata)) return undefined
+  const anthropic = metadata.anthropic
+  if (ProviderShared.isRecord(anthropic)) {
+    if (typeof anthropic.file_id === "string") return anthropic.file_id
+    if (typeof anthropic.fileId === "string") return anthropic.fileId
+  }
+  if (typeof metadata.file_id === "string") return metadata.file_id
+  if (typeof metadata.fileId === "string") return metadata.fileId
+  return undefined
+}
+
+const transformationsFromMetadata = (
+  metadata: MediaPart["metadata"],
+): AnthropicImageBlock["transformations"] | undefined => {
+  if (!ProviderShared.isRecord(metadata)) return undefined
+  const anthropic = ProviderShared.isRecord(metadata.anthropic) ? metadata.anthropic : undefined
+  const raw = anthropic?.transformations ?? metadata.transformations
+  if (ProviderShared.isRecord(raw)) {
+    const value = raw.oversized_image
+    if (value === "downsize" || value === "error") return { oversized_image: value }
+  }
+  if (anthropic && (anthropic.oversized_image === "downsize" || anthropic.oversized_image === "error"))
+    return { oversized_image: anthropic.oversized_image }
+  return undefined
+}
+
+const documentTitleFromPart = (part: MediaPart): string | undefined => {
+  if (ProviderShared.isRecord(part.metadata)) {
+    const anthropic = part.metadata.anthropic
+    if (ProviderShared.isRecord(anthropic) && typeof anthropic.title === "string") return anthropic.title
+    if (typeof part.metadata.title === "string") return part.metadata.title
+  }
+  if (typeof part.filename === "string" && part.filename.length > 0) return part.filename
+  return undefined
+}
+
+const documentContextFromMetadata = (metadata: MediaPart["metadata"]): string | undefined => {
+  if (!ProviderShared.isRecord(metadata)) return undefined
+  const anthropic = ProviderShared.isRecord(metadata.anthropic) ? metadata.anthropic : undefined
+  if (anthropic && typeof anthropic.context === "string") return anthropic.context
+  if (typeof metadata.context === "string") return metadata.context
+  return 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)
+    : metadata.citations
+  if (ProviderShared.isRecord(raw) && typeof raw.enabled === "boolean") return { enabled: raw.enabled }
+  return undefined
+}
+
+const isHttpUrl = (value: string) => /^https?:\/\//i.test(value.trim())
+
+const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (
+  part: MediaPart,
+  breakpoints?: Cache.Breakpoints,
+) {
+  const mime = part.mediaType.toLowerCase()
+  const cacheControlValue = breakpoints ? cacheControl(breakpoints, part.cache) : undefined
+  const fileId = fileIdFromMetadata(part.metadata)
+
+  // SDK file sources: FileImageSource:2350 / FileDocumentSource:2344 {type:"file", file_id}
+  if (fileId) {
+    if (mime.startsWith("image/"))
+      return {
+        type: "image" as const,
+        source: { type: "file" as const, file_id: fileId },
+        ...(cacheControlValue === undefined ? {} : { cache_control: cacheControlValue }),
+        ...(transformationsFromMetadata(part.metadata) === undefined
+          ? {}
+          : { transformations: transformationsFromMetadata(part.metadata)! }),
+      } satisfies AnthropicImageBlock
+    return {
+      type: "document" as const,
+      source: { type: "file" as const, file_id: fileId },
+      ...(cacheControlValue === undefined ? {} : { cache_control: cacheControlValue }),
+      ...(documentTitleFromPart(part) === undefined ? {} : { title: documentTitleFromPart(part)! }),
+      ...(documentContextFromMetadata(part.metadata) === undefined
+        ? {}
+        : { context: documentContextFromMetadata(part.metadata)! }),
+      ...(citationsFromMetadata(part.metadata) === undefined
+        ? {}
+        : { citations: citationsFromMetadata(part.metadata)! }),
+    } satisfies AnthropicDocumentBlock
+  }
+
+  const rawString = typeof part.data === "string" ? part.data.trim() : undefined
+  // SDK URL sources: URLImageSource:3817 / URLPDFSource:3823 {type:"url", url}
+  if (rawString && isHttpUrl(rawString) && !rawString.startsWith("data:")) {
+    if (mime.startsWith("image/"))
+      return {
+        type: "image" as const,
+        source: { type: "url" as const, url: rawString },
+        ...(cacheControlValue === undefined ? {} : { cache_control: cacheControlValue }),
+        ...(transformationsFromMetadata(part.metadata) === undefined
+          ? {}
+          : { transformations: transformationsFromMetadata(part.metadata)! }),
+      } satisfies AnthropicImageBlock
+    if (mime === "application/pdf")
+      return {
+        type: "document" as const,
+        source: { type: "url" as const, url: rawString },
+        ...(cacheControlValue === undefined ? {} : { cache_control: cacheControlValue }),
+        ...(documentTitleFromPart(part) === undefined ? {} : { title: documentTitleFromPart(part)! }),
+        ...(documentContextFromMetadata(part.metadata) === undefined
+          ? {}
+          : { context: documentContextFromMetadata(part.metadata)! }),
+        ...(citationsFromMetadata(part.metadata) === undefined
+          ? {}
+          : { citations: citationsFromMetadata(part.metadata)! }),
+      } satisfies AnthropicDocumentBlock
+  }
+
+  // SDK PlainTextSource:2716 {type:"text", media_type:"text/plain", data}
+  if (mime === "text/plain") {
+    const textData =
+      typeof part.data !== "string"
+        ? Buffer.from(part.data).toString("utf8")
+        : part.data.startsWith("data:")
+          ? (() => {
+              const comma = part.data.indexOf(",")
+              const payload = comma >= 0 ? part.data.slice(comma + 1) : part.data
+              return part.data.includes(";base64")
+                ? Buffer.from(payload, "base64").toString("utf8")
+                : decodeURIComponent(payload)
+            })()
+          : part.data
+    return {
+      type: "document" as const,
+      source: { type: "text" as const, media_type: "text/plain" as const, data: textData },
+      ...(cacheControlValue === undefined ? {} : { cache_control: cacheControlValue }),
+      ...(documentTitleFromPart(part) === undefined ? {} : { title: documentTitleFromPart(part)! }),
+      ...(documentContextFromMetadata(part.metadata) === undefined
+        ? {}
+        : { context: documentContextFromMetadata(part.metadata)! }),
+      ...(citationsFromMetadata(part.metadata) === undefined
+        ? {}
+        : { citations: citationsFromMetadata(part.metadata)! }),
+    } satisfies AnthropicDocumentBlock
+  }
+
  const media = ProviderShared.normalizeMedia(part)
  if (media.mime === "application/pdf")
    return {
@@ -420,6 +664,14 @@ const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (part: Me
        media_type: "application/pdf" as const,
        data: media.base64,
      },
+      ...(cacheControlValue === undefined ? {} : { cache_control: cacheControlValue }),
+      ...(documentTitleFromPart(part) === undefined ? {} : { title: documentTitleFromPart(part)! }),
+      ...(documentContextFromMetadata(part.metadata) === undefined
+        ? {}
+        : { context: documentContextFromMetadata(part.metadata)! }),
+      ...(citationsFromMetadata(part.metadata) === undefined
+        ? {}
+        : { citations: citationsFromMetadata(part.metadata)! }),
    } satisfies AnthropicDocumentBlock
  if (!media.mime.startsWith("image/"))
    return yield* invalid(`Anthropic Messages does not support media type ${part.mediaType}`)
@@ -430,6 +682,10 @@ const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (part: Me
      media_type: media.mime,
      data: media.base64,
    },
+    ...(cacheControlValue === undefined ? {} : { cache_control: cacheControlValue }),
+    ...(transformationsFromMetadata(part.metadata) === undefined
+      ? {}
+      : { transformations: transformationsFromMetadata(part.metadata)! }),
  } satisfies AnthropicImageBlock
 })

@@ -449,6 +705,26 @@ 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.
@@ -538,7 +814,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
          continue
        }
        if (part.type === "media") {
-          content.push(yield* lowerMedia(part))
+          content.push(yield* lowerMedia(part, breakpoints))
          continue
        }
        return yield* ProviderShared.unsupportedContent("Anthropic Messages", "user", ["text", "media"])
@@ -555,15 +831,30 @@ 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.
+          // 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)
          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
        }
@@ -605,24 +896,74 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
 })

 const resolveOptions = Effect.fn("AnthropicMessages.resolveOptions")(function* (request: LLMRequest) {
-  const input = request.providerOptions
+  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 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)
+      ? { 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 })
+      : undefined
+  const rawInferenceGeo =
+    (input as Record<string, unknown> | undefined)?.inference_geo ??
+    (input as Record<string, unknown> | undefined)?.inferenceGeo
+  const inference_geo = typeof rawInferenceGeo === "string" ? rawInferenceGeo : undefined
+  const rawCacheControl =
+    (input as Record<string, unknown> | undefined)?.cache_control ??
+    (input as Record<string, unknown> | undefined)?.cacheControl
+  const cache_control =
+    ProviderShared.isRecord(rawCacheControl) && rawCacheControl.type === "ephemeral"
+      ? (rawCacheControl as { type: "ephemeral"; ttl?: "5m" | "1h" })
+      : undefined
+  const rawOutputConfig =
+    (input as Record<string, unknown> | undefined)?.output_config ??
+    (input as Record<string, unknown> | undefined)?.outputConfig
+  const outputConfigEffort =
+    typeof (input as Record<string, unknown> | undefined)?.effort === "string"
+      ? ((input as Record<string, unknown>).effort as string)
+      : ProviderShared.isRecord(rawOutputConfig) && typeof rawOutputConfig.effort === "string"
+        ? (rawOutputConfig.effort as string)
+        : undefined
+  const outputConfigFormat =
+    ProviderShared.isRecord(rawOutputConfig) && ProviderShared.isRecord(rawOutputConfig.format)
+      ? (rawOutputConfig.format as { type: "json_schema"; schema: Record<string, unknown> })
+      : undefined
+  const output_config =
+    outputConfigEffort === undefined && outputConfigFormat === undefined
+      ? undefined
+      : {
+          ...(outputConfigEffort === undefined ? {} : { effort: outputConfigEffort }),
+          ...(outputConfigFormat === undefined ? {} : { format: outputConfigFormat }),
+        }
  return {
    thinking: yield* resolveThinking(input?.thinking),
-    effort: typeof input?.effort === "string" ? input.effort : undefined,
+    effort: outputConfigEffort,
+    output_config,
+    service_tier,
+    metadata,
+    container,
+    inference_geo,
+    cache_control,
  }
 })

 const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function* (input: unknown) {
  if (!ProviderShared.isRecord(input)) return undefined
-  if (input.type === "adaptive") {
-    const display =
-      input.display === "summarized"
-        ? ("summarized" as const)
-        : input.display === "omitted"
-          ? ("omitted" as const)
-          : undefined
+  const display =
+    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 === "disabled") return { type: "disabled" as const }
  if (input.type !== "enabled") return undefined
  const budget =
@@ -633,7 +974,7 @@ const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function*
        : undefined
  if (budget === undefined)
    return yield* ProviderShared.invalidRequest("Anthropic thinking provider option requires budgetTokens")
-  return { type: "enabled" as const, budget_tokens: budget }
+  return { type: "enabled" as const, budget_tokens: budget, ...(display === undefined ? {} : { display }) }
 })

 const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (request: LLMRequest) {
@@ -683,7 +1024,13 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    top_k: generation?.topK,
    stop_sequences: generation?.stop,
    thinking: options.thinking,
-    output_config: options.effort === undefined ? undefined : { effort: options.effort },
+    output_config: options.output_config,
+    // top-level passthrough per SDK MessageCreateParamsBase:4638,4643,4649,4654,4670
+    cache_control: options.cache_control,
+    container: options.container,
+    inference_geo: options.inference_geo,
+    metadata: options.metadata,
+    service_tier: options.service_tier,
  }
 })

@@ -763,7 +1110,7 @@ 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): LLMEvent | undefined => {
  if (!block.type || !isServerToolResultType(block.type)) return undefined
  const errorPayload =
    typeof block.content === "object" && block.content !== null && "type" in block.content
@@ -790,11 +1137,15 @@ const onMessageStart = (state: ParserState, event: AnthropicEvent): StepResult =
  return [usage ? { ...state, usage: mergeUsage(state.usage, usage) } : 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]

-  if ((block.type === "tool_use" || block.type === "server_tool_use") && event.index !== undefined) {
+  if (block.type === "tool_use" || block.type === "server_tool_use") {
+    if (event.index === undefined || !block.id) return [state, NO_EVENTS]
    const events: LLMEvent[] = []
    const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
    return [
@@ -802,7 +1153,7 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
        ...state,
        lifecycle,
        tools: ToolStream.start(state.tools, event.index, {
-          id: block.id ?? String(event.index),
+          id: block.id,
          name: block.name ?? "",
          input:
            block.input !== undefined && (!ProviderShared.isRecord(block.input) || Object.keys(block.input).length > 0)
@@ -814,7 +1165,7 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
      [
        ...events,
        LLMEvent.toolInputStart({
-          id: block.id ?? String(event.index),
+          id: block.id,
          name: block.name ?? "",
          providerExecuted: block.type === "server_tool_use" ? true : undefined,
        }),
@@ -880,11 +1231,12 @@ 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) },
@@ -893,6 +1245,7 @@ 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 [
      {
@@ -905,6 +1258,7 @@ 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,
@@ -957,7 +1311,10 @@ 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 onMessageDelta = (
+  state: ParserState,
+  event: AnthropicEvent & { readonly delta?: AnthropicStreamDelta },
+): StepResult => {
  const usage = mergeUsage(state.usage, mapUsage(event.usage))
  return [
    {
@@ -1004,18 +1361,80 @@ const providerErrorMessage = (event: AnthropicEvent): string => {
 }

 const onError = (event: AnthropicEvent) =>
-  new AIError({
-    module: ADAPTER,
-    method: "stream",
-    reason: classifyProviderFailure({ message: providerErrorMessage(event), code: event.error?.type }),
-  })
+  Effect.fail(
+    new AIError({
+      module: ADAPTER,
+      method: "stream",
+      reason: classifyProviderFailure({ message: providerErrorMessage(event), code: event.error?.type }),
+    }),
+  )
+
+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") return Effect.succeed(onContentBlockStart(state, event))
-  if (event.type === "content_block_delta") return onContentBlockDelta(state, event)
+  if (event.type === "content_block_start") {
+    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 ${event.index}`),
+        )
+    }
+    return Effect.succeed(onContentBlockStart(state, { ...event, content_block: block }))
+  }
+  if (event.type === "content_block_delta") {
+    if (!ProviderShared.isRecord(event.delta)) return invalidStreamEvent(event)
+    if (typeof event.delta.type === "string" && !isKnownStreamDeltaType(event.delta.type))
+      return Effect.succeed<StepResult>([state, NO_EVENTS])
+    const decoded = decodeAnthropicStreamDelta(event.delta)
+    if (Option.isNone(decoded)) return invalidStreamEvent(event)
+    return onContentBlockDelta(state, { ...event, delta: decoded.value })
+  }
  if (event.type === "content_block_stop") return onContentBlockStop(state, event)
-  if (event.type === "message_delta") return Effect.succeed(onMessageDelta(state, event))
+  if (event.type === "message_delta") {
+    const decoded = decodeAnthropicStreamDelta(event.delta)
+    if (Option.isNone(decoded)) return invalidStreamEvent(event)
+    return Effect.succeed(onMessageDelta(state, { ...event, delta: decoded.value }))
+  }
  if (event.type === "message_stop") return onMessageStop(state)
  if (event.type === "error") return onError(event)
  return Effect.succeed<StepResult>([state, NO_EVENTS])
@@ -1051,7 +1470,10 @@ export const route = Route.make({
  provider: "anthropic",
  providerMetadataKey: "anthropic",
  protocol,
-  endpoint: Endpoint.path(PATH, { baseURL: DEFAULT_BASE_URL }),
+  endpoint: Endpoint.path(
+    (input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH),
+    { baseURL: DEFAULT_BASE_URL },
+  ),
  auth: Auth.none,
  framing,
  headers: () => ({ "anthropic-version": "2023-06-01" }),
@@ -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>

@@ -650,22 +646,14 @@ 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) {
      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:
+            event.exception.details.message ?? event.exception.details.originalMessage ?? "Bedrock Converse stream error",
+          code: event.exception.type,
        }),
      })
    }
@@ -716,7 +704,7 @@ export const protocol = Protocol.make({
      reasoningSignatures: {},
    }),
    step,
-    onHalt,
+    onHalt: (state) => Effect.succeed(onHalt(state)),
  },
 })

@@ -82,7 +82,7 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
        "Failed to parse Bedrock Converse event-stream payload",
      )) as Record<string, unknown>
      delete parsed.p
-      out.push({ [eventType]: parsed })
+      out.push(messageType === "exception" ? { exception: { type: eventType, details: parsed } } : { [eventType]: parsed })
    }
    return [cursor, out] as const
  })
@@ -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"
@@ -17,7 +17,7 @@ import {
  type ToolCallPart,
  type ToolDefinition,
 } from "../schema/index.js"
-import { JsonObject, optionalArray, ProviderShared } from "./shared.js"
+import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
 import { GeminiToolSchema } from "./utils/gemini-tool-schema.js"
 import { Lifecycle } from "./utils/lifecycle.js"
 import { ToolSchemaProjection } from "./utils/tool-schema.js"
@@ -41,6 +41,13 @@ const requiresThoughtSignatureFallback = (modelID: string) => {
 // so their tool-result attachments lower as a separate user turn instead.
 const routesLegacyToolMedia = (modelID: string) => /gemini-2[.-]5(?:[.-]|$)/i.test(modelID)

+// Blacklist: Gemini 1.x/2.x ignore or reject explicit function call ids.
+// Every other model id (Gemini 3+, gemma, anything unrecognized) gets them.
+const omitsFunctionCallIds = (modelID: string) => {
+  const match = /^gemini(?:-live)?-(\d+)/i.exec(modelID)
+  return match !== null && Number(match[1]) < 3
+}
+
 export interface OptionsInput {
  readonly [key: string]: unknown
  readonly cachedContent?: string
@@ -75,10 +82,15 @@ export type ProviderOptionsInput = OptionsInput
 // =============================================================================
 // Request Body Schema
 // =============================================================================
+// Gemini is known to send explicit `null` for optional streaming fields
+// (usage counts, flags, whole subtrees), so every response-side optional uses
+// `optionalNull` instead of bare `Schema.optional`. The same part/content
+// schemas lower the outbound request body; encoding drops `undefined` keys,
+// so the shared schemas stay safe there.
 const GeminiTextPart = Schema.Struct({
  text: Schema.String,
-  thought: Schema.optional(Schema.Boolean),
-  thoughtSignature: Schema.optional(Schema.String),
+  thought: optionalNull(Schema.Boolean),
+  thoughtSignature: optionalNull(Schema.String),
 })

 const GeminiInlineDataPart = Schema.Struct({
@@ -91,11 +103,11 @@ type GeminiInlineDataPart = Schema.Schema.Type<typeof GeminiInlineDataPart>

 const GeminiFunctionCallPart = Schema.Struct({
  functionCall: Schema.Struct({
-    id: Schema.optional(Schema.String),
+    id: optionalNull(Schema.String),
    name: Schema.String,
    args: Schema.optional(Schema.Unknown),
  }),
-  thoughtSignature: Schema.optional(Schema.String),
+  thoughtSignature: optionalNull(Schema.String),
 })

 const GeminiFunctionResponsePart = Schema.Struct({
@@ -113,13 +125,19 @@ const GeminiContentPart = Schema.Union([
  GeminiFunctionCallPart,
  GeminiFunctionResponsePart,
 ])
+const decodeGeminiContentPart = Schema.decodeUnknownOption(GeminiContentPart)

 const GeminiContent = Schema.Struct({
-  role: Schema.Literals(["user", "model"]),
-  parts: Schema.Array(GeminiContentPart),
+  role: optionalNull(Schema.Literals(["user", "model"])),
+  parts: optionalNull(Schema.Array(GeminiContentPart)),
 })
 type GeminiContent = Schema.Schema.Type<typeof GeminiContent>

+const GeminiResponseContent = Schema.Struct({
+  role: optionalNull(Schema.Literals(["user", "model"])),
+  parts: optionalNull(Schema.Array(Schema.Unknown)),
+})
+
 const GeminiSystemInstruction = Schema.Struct({
  parts: Schema.Array(Schema.Struct({ text: Schema.String })),
 })
@@ -179,45 +197,46 @@ const GeminiBody = Schema.Struct(GeminiBodyFields)
 export type GeminiBody = Schema.Schema.Type<typeof GeminiBody>

 const GeminiUsage = Schema.Struct({
-  cachedContentTokenCount: Schema.optional(Schema.Number),
-  thoughtsTokenCount: Schema.optional(Schema.Number),
-  promptTokenCount: Schema.optional(Schema.Number),
-  candidatesTokenCount: Schema.optional(Schema.Number),
-  totalTokenCount: Schema.optional(Schema.Number),
+  cachedContentTokenCount: optionalNull(Schema.Number),
+  thoughtsTokenCount: optionalNull(Schema.Number),
+  promptTokenCount: optionalNull(Schema.Number),
+  candidatesTokenCount: optionalNull(Schema.Number),
+  totalTokenCount: optionalNull(Schema.Number),
 })
 type GeminiUsage = Schema.Schema.Type<typeof GeminiUsage>

 const GeminiCandidate = Schema.Struct({
-  content: Schema.optional(GeminiContent),
-  finishReason: Schema.optional(Schema.String),
+  content: optionalNull(GeminiResponseContent),
+  finishReason: optionalNull(Schema.String),
 })

 const GeminiPromptFeedback = Schema.StructWithRest(
  Schema.Struct({
-    blockReason: Schema.optional(Schema.String),
-    blockReasonMessage: Schema.optional(Schema.String),
-    safetyRatings: Schema.optional(Schema.Unknown),
+    blockReason: optionalNull(Schema.String),
+    blockReasonMessage: optionalNull(Schema.String),
+    safetyRatings: optionalNull(Schema.Unknown),
  }),
  [Schema.Record(Schema.String, Schema.Unknown)],
 )
 type GeminiPromptFeedback = Schema.Schema.Type<typeof GeminiPromptFeedback>

 const GeminiEvent = Schema.Struct({
-  candidates: optionalArray(GeminiCandidate),
-  promptFeedback: Schema.optional(GeminiPromptFeedback),
-  usageMetadata: Schema.optional(GeminiUsage),
+  candidates: optionalNull(Schema.Array(GeminiCandidate)),
+  promptFeedback: optionalNull(GeminiPromptFeedback),
+  usageMetadata: optionalNull(GeminiUsage),
 })
 type GeminiEvent = Schema.Schema.Type<typeof GeminiEvent>

 interface ParserState {
+  readonly route: string
  readonly finishReason?: string
  readonly hasToolCalls: boolean
-  readonly nextToolCallId: number
  readonly promptFeedback?: GeminiPromptFeedback
  readonly usage?: Usage
  readonly lifecycle: Lifecycle.State
  readonly reasoningSignature?: string
  readonly textSignature?: string
+  readonly seenCallIds?: ReadonlySet<string>
 }

 // =============================================================================
@@ -275,20 +294,14 @@ const thoughtSignature = (providerMetadata: ProviderMetadata | undefined) => {
    : undefined
 }

-const functionCallId = (providerMetadata: ProviderMetadata | undefined) => {
-  const google = providerMetadata?.google
-  return ProviderShared.isRecord(google) && typeof google.functionCallId === "string"
-    ? google.functionCallId
-    : undefined
-}
-
-const lowerToolCall = (part: ToolCallPart) => ({
-  functionCall: { id: functionCallId(part.providerMetadata), name: part.name, args: part.input },
+const lowerToolCall = (part: ToolCallPart, omitIds: boolean) => ({
+  functionCall: { ...(omitIds ? {} : { id: part.id }), name: part.name, args: part.input },
  thoughtSignature: thoughtSignature(part.providerMetadata),
 })

 const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMRequest) {
  const contents: GeminiContent[] = []
+  const omitCallIds = omitsFunctionCallIds(request.model.id)
  const legacyToolMedia = routesLegacyToolMedia(request.model.id)
  let pendingMedia: GeminiInlineDataPart[] | undefined
  const flushMedia = () => {
@@ -304,8 +317,8 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
      const previous = contents.at(-1)
      // Gemini rejects a continuation whose function-response turn carries extra
      // parts, so an update after a tool result starts its own user turn.
-      if (previous?.role === "user" && !previous.parts.some((item) => "functionResponse" in item))
-        contents[contents.length - 1] = { role: "user", parts: [...previous.parts, { text: part.text }] }
+      if (previous?.role === "user" && !(previous.parts ?? []).some((item) => "functionResponse" in item))
+        contents[contents.length - 1] = { role: "user", parts: [...(previous.parts ?? []), { text: part.text }] }
      else contents.push({ role: "user", parts: [{ text: part.text }] })
      continue
    }
@@ -337,7 +350,7 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
          continue
        }
        if (part.type === "tool-call") {
-          const lowered = lowerToolCall(part)
+          const lowered = lowerToolCall(part, omitCallIds)
          const signature = lowered.thoughtSignature
          parts.push({
            ...lowered,
@@ -362,7 +375,7 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
      if (part.result.type !== "content") {
        parts.push({
          functionResponse: {
-            id: functionCallId(part.providerMetadata),
+            ...(omitCallIds ? {} : { id: part.id }),
            name: part.name,
            response: {
              name: part.name,
@@ -383,7 +396,7 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
      if (legacyToolMedia && media.length > 0) (pendingMedia ??= []).push(...media)
      parts.push({
        functionResponse: {
-          id: functionCallId(part.providerMetadata),
+          ...(omitCallIds ? {} : { id: part.id }),
          name: part.name,
          response: {
            name: part.name,
@@ -396,8 +409,8 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
    // Gemini requires every response to a parallel call batch in one user turn,
    // so consecutive tool results join the open function-response turn.
    const previous = contents.at(-1)
-    if (previous?.role === "user" && previous.parts.some((item) => "functionResponse" in item))
-      contents[contents.length - 1] = { role: "user", parts: [...previous.parts, ...parts] }
+    if (previous?.role === "user" && (previous.parts ?? []).some((item) => "functionResponse" in item))
+      contents[contents.length - 1] = { role: "user", parts: [...(previous.parts ?? []), ...parts] }
    else contents.push({ role: "user", parts })
  }

@@ -487,21 +500,25 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
 // to produce the inclusive `outputTokens` the rest of the contract expects.
 const mapUsage = (usage: GeminiUsage | undefined) => {
  if (!usage) return undefined
-  const cached = usage.cachedContentTokenCount
-  const nonCached = ProviderShared.subtractTokens(usage.promptTokenCount, cached)
+  // Explicit provider nulls decode as `null`; normalize to `undefined` so the
+  // token arithmetic below treats them like absent counts.
+  const promptTokens = usage.promptTokenCount ?? undefined
+  const cached = usage.cachedContentTokenCount ?? undefined
+  const thoughts = usage.thoughtsTokenCount ?? undefined
+  const visible = usage.candidatesTokenCount ?? undefined
+  const nonCached = ProviderShared.subtractTokens(promptTokens, cached)
  // `candidatesTokenCount` is visible-only; sum with thoughts to produce the
  // inclusive `outputTokens` the contract expects. Only compute the total
  // when the visible component is reported — otherwise we'd fabricate an
  // inclusive number from a partial breakdown.
-  const outputTokens =
-    usage.candidatesTokenCount !== undefined ? usage.candidatesTokenCount + (usage.thoughtsTokenCount ?? 0) : undefined
+  const outputTokens = visible !== undefined ? visible + (thoughts ?? 0) : undefined
  return new Usage({
-    inputTokens: usage.promptTokenCount,
+    inputTokens: promptTokens,
    outputTokens,
    nonCachedInputTokens: nonCached,
    cacheReadInputTokens: cached,
-    reasoningTokens: usage.thoughtsTokenCount,
-    totalTokens: ProviderShared.totalTokens(usage.promptTokenCount, outputTokens, usage.totalTokenCount),
+    reasoningTokens: thoughts,
+    totalTokens: ProviderShared.totalTokens(promptTokens, outputTokens, usage.totalTokenCount ?? undefined),
    providerMetadata: { google: usage },
  })
 }
@@ -536,7 +553,10 @@ const mapFinishReason = (finishReason: string | undefined, hasToolCalls: boolean
 }

 const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
-  const promptBlockReason = state.finishReason === undefined ? state.promptFeedback?.blockReason : undefined
+  // `?? undefined` normalizes an explicit `null` blockReason back to absent so
+  // the "nothing to finish" check below keeps its meaning.
+  const promptBlockReason =
+    state.finishReason === undefined ? (state.promptFeedback?.blockReason ?? undefined) : undefined
  const finishReason = state.finishReason ?? promptBlockReason
  if (finishReason === undefined && state.usage === undefined) return []

@@ -580,11 +600,26 @@ const step = (state: ParserState, event: GeminiEvent) => {
  const events: LLMEvent[] = []
  let hasToolCalls = nextState.hasToolCalls
  let lifecycle = nextState.lifecycle
-  let nextToolCallId = nextState.nextToolCallId
  let reasoningSignature = nextState.reasoningSignature
  let textSignature = nextState.textSignature
+  // 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.
@@ -620,11 +655,14 @@ const step = (state: ParserState, event: GeminiEvent) => {

    if ("functionCall" in part) {
      const input = part.functionCall.args === undefined ? {} : part.functionCall.args
-      const id = `tool_${nextToolCallId++}`
-      const metadata = {
-        ...(part.functionCall.id === undefined ? {} : { functionCallId: part.functionCall.id }),
-        ...(part.thoughtSignature === undefined ? {} : { thoughtSignature: part.thoughtSignature }),
-      }
+      // Gemini 2.0+ supplies a unique function call ID on the part; when omitted (e.g. Gemini 1.5),
+      // generate a globally unique ID rather than a per-request counter to prevent cross-request collisions in downstream registries.
+      // A repeated supplier id would replay as two identical calls, so only the first occurrence keeps it.
+      // A `null` supplier id normalizes to absent so the generated-id fallback applies.
+      const supplied = part.functionCall.id ?? undefined
+      const duplicate = supplied !== undefined && seenCallIds.has(supplied)
+      if (supplied !== undefined) seenCallIds.add(supplied)
+      const id = supplied !== undefined && !duplicate ? supplied : `tool_${crypto.randomUUID().replaceAll("-", "")}`
      lifecycle = Lifecycle.reasoningEnd(
        lifecycle,
        events,
@@ -637,7 +675,8 @@ const step = (state: ParserState, event: GeminiEvent) => {
          id,
          name: part.functionCall.name,
          input,
-          providerMetadata: Object.keys(metadata).length > 0 ? googleMetadata(metadata) : undefined,
+          providerMetadata:
+            part.thoughtSignature ? googleMetadata({ thoughtSignature: part.thoughtSignature }) : undefined,
        }),
      )
      hasToolCalls = true
@@ -649,9 +688,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
      ...nextState,
      hasToolCalls,
      lifecycle,
-      nextToolCallId,
      reasoningSignature,
      textSignature,
+      seenCallIds,
      finishReason: candidate.finishReason ?? nextState.finishReason,
    },
    events,
@@ -673,9 +712,13 @@ export const protocol = Protocol.make({
  },
  stream: {
    event: Protocol.jsonEvent(GeminiEvent),
-    initial: () => ({ hasToolCalls: false, nextToolCallId: 0, lifecycle: Lifecycle.initial() }),
+    initial: (request) => ({
+      route: `${request.model.provider}/${request.model.route.id}`,
+      hasToolCalls: false,
+      lifecycle: Lifecycle.initial(),
+    }),
    step,
-    onHalt: finish,
+    onHalt: (state) => Effect.succeed(finish(state)),
  },
 })

@@ -8,3 +8,4 @@ export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
 export * as OpenAIResponses from "./openai-responses.js"
 export * as OpenResponses from "./open-responses.js"
 export * as OpenResponsesChannel from "./open-responses-channel.js"
+export * as XAIResponses from "./xai-responses.js"
@@ -10,6 +10,7 @@ import {
 } from "../route/transport/index.js"
 import * as ProviderShared from "./shared.js"
 import { OpenResponses } from "./open-responses.js"
+import { OpenResponsesContinuation } from "./open-responses-continuation.js"

 const WebSocketResponseCreate = Schema.StructWithRest(Schema.Struct({ type: Schema.tag("response.create") }), [
  Schema.Record(Schema.String, Schema.Unknown),
@@ -25,11 +26,6 @@ export interface Options {
  readonly enabled?: (url: string) => boolean
  readonly url?: (url: string) => string
  readonly headers?: (headers: Headers.Headers) => Headers.Headers
-  readonly driver?: (input: {
-    readonly request: Readonly<Record<string, unknown>>
-    readonly message: string
-    readonly base: WebSocketChannelDriver
-  }) => WebSocketChannelDriver
 }

 export interface Prepared {
@@ -116,6 +112,8 @@ const driver = (options: Options, body: string): WebSocketChannelDriver => {
          responseID = created
          return { type: "frame", frame }
        }
+        // Keepalives carry no response state and may arrive before response.created.
+        if (event.type === "keepalive") return { type: "frame", frame }
        if (!responseID)
          return yield* ProviderShared.eventError(
            options.id,
@@ -158,7 +156,13 @@ export const transport = <Body>(options: Options): Transport<Body, Prepared, str
                  url: yield* WebSocketTransport.toWebSocketUrl(options.url?.(parts.url) ?? parts.url),
                  headers,
                  rotateAfterMs: options.rotateAfterMs,
-                  driver: options.driver?.({ request: create.request, message: create.message, base }) ?? base,
+                  driver: OpenResponsesContinuation.driver({
+                    id: options.id,
+                    name: options.name,
+                    request: create.request,
+                    message: create.message,
+                    base,
+                  }),
                }
              })
            : undefined
@@ -4,7 +4,7 @@ import { Effect, Option, Schema } from "effect"
 import * as ProviderShared from "./shared.js"
 import { OpenResponses } from "./open-responses.js"

-const PROTOCOL = "openai-responses.websocket.v1"
+const PROTOCOL = "open-responses.websocket.v1"
 const VERSION = 1
 const decodeEvent = Schema.decodeUnknownEffect(OpenResponses.protocol.stream.event)

@@ -154,11 +154,16 @@ 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,
          },
        }
      }),
  }
 }

-export const OpenAIResponsesChannel = { driver } as const
+export const OpenResponsesContinuation = { driver } as const
@@ -5,6 +5,7 @@ import { Protocol } from "../route/protocol.js"
 import {
  AIError,
  LLMEvent,
+  ProviderInternalReason,
  Usage,
  type FinishReason,
  type JsonSchema,
@@ -42,8 +43,12 @@ const OpenResponsesInputImage = Schema.Struct({
 const OpenResponsesInputFile = Schema.Struct({
  type: Schema.tag("input_file"),
  filename: Schema.String,
-  file_data: Schema.String,
-  mime_type: Schema.optional(Schema.String),
+  file_data: Schema.optional(Schema.String),
+  file_url: Schema.optional(Schema.String),
+})
+const OpenResponsesInputVideo = Schema.Struct({
+  type: Schema.tag("input_video"),
+  video_url: Schema.String,
 })
 const MediaInput = Schema.Union([OpenResponsesInputImage, OpenResponsesInputFile])
 export type MediaInput = Schema.Schema.Type<typeof MediaInput>
@@ -54,9 +59,14 @@ const OpenResponsesOutputText = Schema.Struct({
  text: Schema.String,
 })

-export const MessagePhase = Schema.Literals(["commentary", "final_answer"])
+export const MessagePhase = Schema.NullOr(Schema.Literals(["commentary", "final_answer"]))
 type MessagePhase = Schema.Schema.Type<typeof MessagePhase>

+const messagePhase = (value: unknown): MessagePhase | undefined => {
+  if (value === null || value === "commentary" || value === "final_answer") return value
+  return undefined
+}
+
 const OpenResponsesReasoningSummaryText = Schema.Struct({
  type: Schema.tag("summary_text"),
  text: Schema.String,
@@ -81,6 +91,7 @@ const OpenResponsesFunctionCallOutputContent = Schema.Union([
  OpenResponsesInputText,
  OpenResponsesInputImage,
  OpenResponsesInputFile,
+  OpenResponsesInputVideo,
 ])

 const OpenResponsesFunctionCallOutput = Schema.Union([
@@ -237,6 +248,7 @@ const OpenResponsesErrorPayload = Schema.Struct({
  message: optionalNull(Schema.String),
  param: optionalNull(Schema.String),
 })
+type OpenResponsesErrorPayload = Schema.Schema.Type<typeof OpenResponsesErrorPayload>

 const WebSocketErrorHeader = Schema.Union([Schema.String, Schema.Number, Schema.Boolean])
 export const WebSocketErrorEvent = Schema.StructWithRest(
@@ -273,8 +285,10 @@ export const Event = Schema.StructWithRest(
  Schema.Struct({
    type: Schema.String,
    delta: Schema.optional(Schema.String),
+    arguments: Schema.optional(Schema.String),
    text: Schema.optional(Schema.String),
    item_id: Schema.optional(Schema.String),
+    output_index: Schema.optional(Schema.Number),
    summary_index: Schema.optional(Schema.Number),
    item: Schema.optional(StreamItem),
    response: Schema.optional(
@@ -283,6 +297,7 @@ export const Event = Schema.StructWithRest(
          id: Schema.optional(Schema.String),
          service_tier: optionalNull(Schema.String),
          incomplete_details: optionalNull(Schema.Struct({ reason: Schema.optional(Schema.String) })),
+          output: Schema.optional(Schema.Array(StreamItem)),
          usage: optionalNull(OpenResponsesUsage),
          error: optionalNull(OpenResponsesErrorPayload),
        }),
@@ -301,19 +316,8 @@ export const Event = Schema.StructWithRest(
 )
 export type Event = Schema.Schema.Type<typeof Event>

-const RefusalEvent = Schema.Union([
-  Schema.Struct({
-    type: Schema.tag("response.refusal.delta"),
-    item_id: Schema.String,
-    delta: Schema.String,
-  }),
-  Schema.Struct({
-    type: Schema.tag("response.refusal.done"),
-    item_id: Schema.String,
-    refusal: Schema.String,
-  }),
-])
-const isRefusalEvent = Schema.is(RefusalEvent)
+// Which lowered input item a persisted item id is about to be attached to.
+export type ItemKind = "message" | "reasoning" | "function-call" | "reference"

 export interface Extension {
  readonly id: string
@@ -323,7 +327,10 @@ export interface Extension {
    readonly media: ProviderShared.NormalizedMedia
    readonly request: LLMRequest
  }) => MediaInput | undefined
-  readonly messagePhase?: (value: unknown) => MessagePhase | null | undefined
+  // Optional grammar check applied before a persisted item id is resent as
+  // part of replayed history. Returning false drops the id; every lowered
+  // item treats a dropped id the same as an absent one.
+  readonly acceptsItemID?: (kind: ItemKind, id: string) => boolean
 }

 const BASE: Extension = { id: ADAPTER, name: NAME }
@@ -335,8 +342,8 @@ export interface ParserState {
  readonly tools: ToolStream.State<string>
  readonly hasFunctionCall: boolean
  readonly lifecycle: Lifecycle.State
+  readonly outputItems: Readonly<Record<number, string>>
  readonly messageItems: ReadonlySet<string>
-  readonly messagePhase: (value: unknown) => MessagePhase | null | undefined
  readonly messagePhases: Readonly<Record<string, MessagePhase | null>>
  readonly reasoningItems: Readonly<Record<string, ReasoningStreamItem>>
  readonly store: boolean | undefined
@@ -350,6 +357,10 @@ interface ReasoningStreamItem {
  // strings, but typing the map as `Record<number, ...>` documents intent
  // and matches the wire field.
  readonly summaryParts: Readonly<Record<number, ReasoningSummaryStatus>>
+  // Summary indexes that received at least one streamed delta. The `:0` block
+  // is started eagerly when the item opens, so block existence cannot tell
+  // whether a `.done` final would duplicate streamed text.
+  readonly deltaIndexes: ReadonlySet<number>
 }

 // =============================================================================
@@ -366,7 +377,7 @@ export const lowerTool = Effect.fn("OpenResponses.lowerTool")(function* (
    type: "function" as const,
    name: tool.name,
    description: tool.description,
-    parameters: ToolSchemaProjection.responses(inputSchema),
+    parameters: inputSchema,
    // The common tool definition does not currently express Responses strict-schema policy.
    strict: false,
  }
@@ -380,28 +391,46 @@ export const lowerToolChoice = (protocolName: string, toolChoice: NonNullable<LL
    tool: (toolName) => ({ type: "function" as const, name: toolName }),
  })

+// Servers validate item ids on replayed history, and a malformed or oversized
+// id can fail an otherwise valid request. Only server-issued tokens are worth
+// resending; anything else is treated as absent so the item is resent without
+// an id (or skipped, for items that cannot be expressed without one).
+const ITEM_ID_PATTERN = /^[A-Za-z0-9_-]{1,64}$/
 const itemID = (providerMetadata: ProviderMetadata | undefined, providerMetadataKey: string) => {
  const metadata = providerMetadata?.[providerMetadataKey]
-  return ProviderShared.isRecord(metadata) && typeof metadata.itemId === "string" && metadata.itemId.length > 0
+  return ProviderShared.isRecord(metadata) &&
+    typeof metadata.itemId === "string" &&
+    ITEM_ID_PATTERN.test(metadata.itemId)
    ? metadata.itemId
    : undefined
 }

-const lowerToolCall = (part: ToolCallPart, providerMetadataKey: string): OpenResponsesInputItem => {
+const acceptsItemID = (extension: Extension, kind: ItemKind, id: string | undefined): id is string =>
+  id !== undefined && (extension.acceptsItemID?.(kind, id) ?? true)
+
+const lowerToolCall = (
+  part: ToolCallPart,
+  providerMetadataKey: string,
+  extension: Extension,
+): OpenResponsesInputItem => {
  const id = itemID(part.providerMetadata, providerMetadataKey)
  return {
    type: "function_call",
-    ...(id ? { id } : {}),
+    ...(acceptsItemID(extension, "function-call", id) ? { id } : {}),
    call_id: part.id,
    name: part.name,
    arguments: ProviderShared.encodeJson(part.input),
  }
 }

-const lowerReasoning = (part: ReasoningPart, providerMetadataKey: string): OpenResponsesReasoningInput | undefined => {
+const lowerReasoning = (
+  part: ReasoningPart,
+  providerMetadataKey: string,
+  extension: Extension,
+): OpenResponsesReasoningInput | undefined => {
  const metadata = part.providerMetadata?.[providerMetadataKey]
  const id = itemID(part.providerMetadata, providerMetadataKey)
-  if (!ProviderShared.isRecord(metadata) || !id) return undefined
+  if (!ProviderShared.isRecord(metadata) || !acceptsItemID(extension, "reasoning", id)) return undefined
  const encryptedContent =
    typeof metadata.reasoningEncryptedContent === "string" || metadata.reasoningEncryptedContent === null
      ? metadata.reasoningEncryptedContent
@@ -414,26 +443,29 @@ const lowerReasoning = (part: ReasoningPart, providerMetadataKey: string): OpenR
  }
 }

-const hostedToolItemID = (part: ToolResultPart, providerMetadataKey: string) => {
-  return itemID(part.providerMetadata, providerMetadataKey)
-}
-
 const lowerMedia = Effect.fn("OpenResponses.lowerMedia")(function* (
  part: MediaPart,
  request: LLMRequest,
  extension: Extension,
+  target: "message" | "tool-result",
 ) {
  const media = ProviderShared.normalizeMedia(part)
  const extended = extension.lowerMedia?.({ part, media, request })
  if (extended) return extended
+  const url =
+    typeof part.data === "string" && (part.data.startsWith("https://") || part.data.startsWith("http://"))
+      ? part.data
+      : undefined
  if (!media.mime.startsWith("image/")) {
+    if (target === "tool-result" && media.mime.startsWith("video/"))
+      return { type: "input_video" as const, video_url: url ?? media.dataUrl }
    return {
      type: "input_file" as const,
      filename: part.filename ?? (media.mime === "application/pdf" ? "document.pdf" : "file"),
-      file_data: media.dataUrl,
+      ...(url ? { file_url: url } : { file_data: media.dataUrl }),
    }
  }
-  return { type: "input_image" as const, image_url: media.dataUrl }
+  return { type: "input_image" as const, image_url: url ?? media.dataUrl }
 })

 const lowerUserContent = Effect.fnUntraced(function* (
@@ -442,10 +474,17 @@ const lowerUserContent = Effect.fnUntraced(function* (
  extension: Extension,
 ) {
  if (part.type === "text") return { type: "input_text" as const, text: part.text }
-  if (part.type === "media") return yield* lowerMedia(part, request, extension)
+  if (part.type === "media") return yield* lowerMessageMedia(part, request, extension)
  return yield* ProviderShared.unsupportedContent(extension.name, "user", ["text", "media"])
 })

+const lowerMessageMedia = Effect.fnUntraced(function* (part: MediaPart, request: LLMRequest, extension: Extension) {
+  const lowered = yield* lowerMedia(part, request, extension, "message")
+  if (lowered.type === "input_video")
+    return yield* ProviderShared.invalidRequest(`${extension.name} user messages do not support input_video`)
+  return lowered
+})
+
 // Tool results may carry structured text, images, and files. Keep media as provider-native
 // content instead of JSON-stringifying base64 into a prompt string.
 const lowerToolResultContentItem = Effect.fnUntraced(function* (
@@ -458,6 +497,20 @@ const lowerToolResultContentItem = Effect.fnUntraced(function* (
    { type: "media", mediaType: item.mime, data: item.uri, filename: item.name },
    request,
    extension,
+    "tool-result",
+  )
+})
+
+const lowerHostedToolResultContentItem = Effect.fnUntraced(function* (
+  item: Content,
+  request: LLMRequest,
+  extension: Extension,
+) {
+  if (item.type === "text") return { type: "input_text" as const, text: item.text }
+  return yield* lowerMessageMedia(
+    { type: "media", mediaType: item.mime, data: item.uri, filename: item.name },
+    request,
+    extension,
  )
 })

@@ -509,8 +562,9 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
          Array<{ id: string | undefined; phase: MessagePhase | null | undefined; parts: TextPart[] }>
        >((groups, part) => {
          const metadata = part.providerMetadata?.[providerMetadataKey]
-          const id = itemID(part.providerMetadata, providerMetadataKey)
-          const phase = ProviderShared.isRecord(metadata) ? messagePhase(metadata.phase, extension) : undefined
+          const rawID = itemID(part.providerMetadata, providerMetadataKey)
+          const id = acceptsItemID(extension, "message", rawID) ? rawID : undefined
+          const phase = ProviderShared.isRecord(metadata) ? messagePhase(metadata.phase) : undefined
          const group = groups.at(-1)
          if (group && group.id === id && group.phase === phase) group.parts.push(part)
          else groups.push({ id, phase, parts: [part] })
@@ -534,7 +588,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        }
        if (part.type === "reasoning") {
          flushText()
-          const reasoning = lowerReasoning(part, providerMetadataKey)
+          const reasoning = lowerReasoning(part, providerMetadataKey, extension)
          if (!reasoning) continue
          if (store !== false) {
            if (!reasoningReferences.has(reasoning.id)) input.push({ type: "item_reference", id: reasoning.id })
@@ -555,22 +609,30 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        if (part.type === "tool-call") {
          flushText()
          if (part.providerExecuted === true) continue
-          input.push(lowerToolCall(part, providerMetadataKey))
+          input.push(lowerToolCall(part, providerMetadataKey, extension))
          continue
        }
        if (part.type === "tool-result" && part.providerExecuted === true) {
          flushText()
-          const itemID = hostedToolItemID(part, providerMetadataKey)
-          if (store !== false && itemID && !hostedToolReferences.has(itemID))
-            input.push({ type: "item_reference", id: itemID })
-          if (store === false && part.result.type === "content") {
-            const content: ReadonlyArray<Content> = part.result.value
+          const id = itemID(part.providerMetadata, providerMetadataKey)
+          const reference = acceptsItemID(extension, "reference", id) ? id : undefined
+          if (store !== false && reference && !hostedToolReferences.has(reference))
+            input.push({ type: "item_reference", id: reference })
+          if (store === false) {
+            // The server is not storing this exchange, so the tool outcome has to
+            // travel in the input. Non-content results degrade to their text form.
+            const content: ReadonlyArray<Content> =
+              part.result.type === "content"
+                ? part.result.value
+                : [{ type: "text", text: ProviderShared.toolResultText(part) }]
            input.push({
              role: "user",
-              content: yield* Effect.forEach(content, (item) => lowerToolResultContentItem(item, request, extension)),
+              content: yield* Effect.forEach(content, (item) =>
+                lowerHostedToolResultContentItem(item, request, extension),
+              ),
            })
          }
-          if (itemID) hostedToolReferences.add(itemID)
+          if (reference) hostedToolReferences.add(reference)
          continue
        }
        return yield* ProviderShared.unsupportedContent(extension.name, "assistant", [
@@ -607,6 +669,8 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques

 const lowerOptions = (request: LLMRequest) => {
  const options = OpenResponsesOptions.resolve(request)
+  const cacheKey = ProviderShared.promptCacheKey(request)
+  const parallelToolCalls = resolveParallelToolCalls(request)
  return {
    ...(options.instructions ? { instructions: options.instructions } : {}),
    ...(options.store !== undefined ? { store: options.store } : {}),
@@ -616,7 +680,7 @@ const lowerOptions = (request: LLMRequest) => {
      ? { stream_options: { include_obfuscation: options.streamOptions.includeObfuscation } }
      : {}),
    ...(options.topLogprobs !== undefined ? { top_logprobs: options.topLogprobs } : {}),
-    ...(request.promptCacheKey ? { prompt_cache_key: request.promptCacheKey } : {}),
+    ...(cacheKey ? { prompt_cache_key: cacheKey } : {}),
    ...(options.include ? { include: options.include } : {}),
    ...(options.reasoningEffort || options.reasoningSummary
      ? { reasoning: { effort: options.reasoningEffort, summary: options.reasoningSummary } }
@@ -624,11 +688,18 @@ const lowerOptions = (request: LLMRequest) => {
    ...(options.textVerbosity ? { text: { verbosity: options.textVerbosity } } : {}),
    ...(options.serviceTier ? { service_tier: options.serviceTier } : {}),
    ...(options.maxToolCalls !== undefined ? { max_tool_calls: options.maxToolCalls } : {}),
-    ...(options.parallelToolCalls !== undefined ? { parallel_tool_calls: options.parallelToolCalls } : {}),
+    ...(parallelToolCalls !== undefined ? { parallel_tool_calls: parallelToolCalls } : {}),
    ...(options.truncation ? { truncation: options.truncation } : {}),
  }
 }

+export const resolveParallelToolCalls = (request: LLMRequest) => {
+  const configured = OpenResponsesOptions.resolve(request).parallelToolCalls
+  if (configured !== undefined) return configured
+  const disabled = request.toolChoice?.disableParallelToolUse
+  return disabled === undefined ? undefined : !disabled
+}
+
 const allowedToolChoice = (request: LLMRequest) => {
  const allowed = OpenResponsesOptions.resolve(request).allowedTools
  if (!allowed) return undefined
@@ -732,7 +803,7 @@ const TERMINAL_TYPES = new Set(["error", "response.completed", "response.incompl
 export const terminal = (event: Event) => TERMINAL_TYPES.has(event.type)

 const onOutputTextDelta = (state: ParserState, event: Event, id: string): StepResult => {
-  if (!event.delta) return [state, NO_EVENTS]
+  if (!event.delta || !state.messageItems.has(id)) return [state, NO_EVENTS]
  const events: LLMEvent[] = []
  const phase = state.messagePhases[id]
  const metadata = providerMetadata(state, { itemId: id, ...(phase === undefined ? {} : { phase }) })
@@ -749,21 +820,37 @@ const onOutputTextDone = (state: ParserState, event: Event, id: string): StepRes
  return [{ ...state, lifecycle: Lifecycle.textEnd(state.lifecycle, events, id) }, events]
 }

+export const outputItemID = (state: ParserState, event: Event) =>
+  event.output_index === undefined ? event.item_id : (state.outputItems[event.output_index] ?? event.item_id)
+
 export const onReasoningDelta = (state: ParserState, event: Event, itemID: string): StepResult => {
-  if (!event.delta) return [state, NO_EVENTS]
+  const item = state.reasoningItems[itemID]
+  if (!event.delta || !item) return [state, NO_EVENTS]
+  const index = event.summary_index ?? 0
  const events: LLMEvent[] = []
-  const id =
-    event.summary_index !== undefined || state.reasoningItems[itemID] ? `${itemID}:${event.summary_index ?? 0}` : itemID
  return [
    {
      ...state,
-      lifecycle: Lifecycle.reasoningDelta(state.lifecycle, events, id, event.delta),
+      lifecycle: Lifecycle.reasoningDelta(state.lifecycle, events, `${itemID}:${index}`, event.delta),
+      reasoningItems: {
+        ...state.reasoningItems,
+        [itemID]: { ...item, deltaIndexes: new Set([...item.deltaIndexes, index]) },
+      },
    },
    events,
  ]
 }

-export const onReasoningDone = (state: ParserState, _event: Event): StepResult => [state, NO_EVENTS]
+// Some compatible gateways emit a reasoning final without streaming any
+// deltas, mirroring `response.output_text.done`. Reconcile the complete text
+// as a single delta unless that summary index already streamed one.
+export const onReasoningDone = (state: ParserState, event: Event, itemID: string): StepResult => {
+  const item = state.reasoningItems[itemID]
+  if (!item || typeof event.text !== "string") return [state, NO_EVENTS]
+  const index = event.summary_index ?? 0
+  if (item.deltaIndexes.has(index)) return [state, NO_EVENTS]
+  return onReasoningDelta(state, { ...event, delta: event.text }, itemID)
+}

 const reasoningMetadata = (state: ParserState, item: StreamItem & { id: string }) =>
  providerMetadata(state, { itemId: item.id, reasoningEncryptedContent: item.encrypted_content ?? null })
@@ -782,18 +869,17 @@ const reasoningMetadata = (state: ParserState, item: StreamItem & { id: string }
 // best-effort, not guaranteed.
 const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
  const item = event.item
-  if (item?.type === "message" && item.id)
+  if (item?.type === "message" && item.id) {
+    const phase = messagePhase(item.phase)
    return [
      {
        ...state,
        messageItems: new Set([...state.messageItems, item.id]),
-        messagePhases: (() => {
-          const phase = state.messagePhase(item.phase)
-          return phase === undefined ? state.messagePhases : { ...state.messagePhases, [item.id]: phase }
-        })(),
+        messagePhases: phase === undefined ? state.messagePhases : { ...state.messagePhases, [item.id]: phase },
      },
      NO_EVENTS,
    ]
+  }
  if (item && isReasoningItem(item)) {
    const events: LLMEvent[] = []
    return [
@@ -802,7 +888,11 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
        lifecycle: Lifecycle.reasoningStart(state.lifecycle, events, `${item.id}:0`, reasoningMetadata(state, item)),
        reasoningItems: {
          ...state.reasoningItems,
-          [item.id]: { encryptedContent: item.encrypted_content, summaryParts: { 0: "active" } },
+          [item.id]: {
+            encryptedContent: item.encrypted_content,
+            summaryParts: { 0: "active" },
+            deltaIndexes: new Set(),
+          },
        },
      },
      events,
@@ -832,27 +922,9 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {

 const onReasoningSummaryPartAdded = (state: ParserState, event: Event): StepResult => {
  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
-  const item = state.reasoningItems[event.item_id] ?? { encryptedContent: undefined, summaryParts: {} }
-  if (event.summary_index === 0) {
-    if (state.reasoningItems[event.item_id]) return [state, NO_EVENTS]
-    const events: LLMEvent[] = []
-    return [
-      {
-        ...state,
-        lifecycle: Lifecycle.reasoningStart(
-          state.lifecycle,
-          events,
-          `${event.item_id}:0`,
-          providerMetadata(state, { itemId: event.item_id, reasoningEncryptedContent: null }),
-        ),
-        reasoningItems: {
-          ...state.reasoningItems,
-          [event.item_id]: { ...item, summaryParts: { 0: "active" } },
-        },
-      },
-      events,
-    ]
-  }
+  const item = state.reasoningItems[event.item_id]
+  if (!item) return [state, NO_EVENTS]
+  if (event.summary_index === 0) return [state, NO_EVENTS]

  const events: LLMEvent[] = []
  const closed = Object.entries(item.summaryParts)
@@ -931,12 +1003,24 @@ const onFunctionCallArgumentsDelta = Effect.fn("OpenResponses.onFunctionCallArgu
  state: ParserState,
  event: Event,
 ) {
-  if (!event.item_id || !event.delta) return [state, NO_EVENTS] satisfies StepResult
+  if (!event.item_id) return [state, NO_EVENTS] satisfies StepResult
+  const tool = state.tools[event.item_id]
+  if (!tool) return [state, NO_EVENTS] satisfies StepResult
+  const final = event.type === "response.function_call_arguments.done" ? event.arguments : undefined
+  if (event.type === "response.function_call_arguments.done" && final === undefined)
+    return [state, NO_EVENTS] satisfies StepResult
+  if (final !== undefined && !final.startsWith(tool.input))
+    return [
+      { ...state, tools: ToolStream.start(state.tools, event.item_id, { ...tool, input: final }) },
+      NO_EVENTS,
+    ] satisfies StepResult
+  const delta = final === undefined ? event.delta : final.slice(tool.input.length)
+  if (!delta) return [state, NO_EVENTS] satisfies StepResult
  const result = ToolStream.appendExisting(
    state.id,
    state.tools,
    event.item_id,
-    event.delta,
+    delta,
    `${state.name} tool argument delta is missing its tool call`,
  )
  if (ToolStream.isError(result)) return yield* result
@@ -951,7 +1035,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
  if (!item) return [state, NO_EVENTS] satisfies StepResult

  if (item.type === "message" && item.id) {
-    const itemPhase = state.messagePhase(item.phase)
+    const itemPhase = messagePhase(item.phase)
    const phase = itemPhase === undefined ? state.messagePhases[item.id] : itemPhase
    const events: LLMEvent[] = []
    const messageItems = new Set(state.messageItems)
@@ -1033,64 +1117,100 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
 })

 const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (state: ParserState, event: Event) {
+  const reconciled =
+    event.type === "response.completed"
+      ? yield* Effect.reduce(
+          event.response?.output ?? [],
+          () => [state, NO_EVENTS] satisfies StepResult,
+          ([current, events], item) => {
+            if (
+              !item.id ||
+              ((item.type !== "function_call" || !current.tools[item.id]) &&
+                (item.type !== "reasoning" || !current.reasoningItems[item.id]))
+            )
+              return Effect.succeed([current, events] satisfies StepResult)
+            return onOutputItemDone(current, { type: "response.output_item.done", item }).pipe(
+              Effect.map(([next, emitted]) => [next, [...events, ...emitted]] satisfies StepResult),
+            )
+          },
+        )
+      : ([state, NO_EVENTS] satisfies StepResult)
+  const current = reconciled[0]
  // Some compatible providers omit output_item.done even after completing the response.
  const pending =
    event.type === "response.completed"
-      ? yield* ToolStream.finishAll(state.id, state.tools)
-      : { tools: state.tools, events: NO_EVENTS }
-  const events: LLMEvent[] = [...pending.events]
+      ? yield* ToolStream.finishAll(current.id, current.tools)
+      : { tools: current.tools, events: NO_EVENTS }
+  const events: LLMEvent[] = [...reconciled[1], ...pending.events]
  const hasFunctionCall =
    pending.events.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
-    state.hasFunctionCall
-  const lifecycle = Lifecycle.finish(state.lifecycle, events, {
+    current.hasFunctionCall
+  const lifecycle = Lifecycle.finish(current.lifecycle, events, {
    reason: {
      normalized: mapFinishReason(event, hasFunctionCall),
      raw: event.response?.incomplete_details?.reason,
    },
-    usage: mapUsage(event.response?.usage, state.providerMetadataKey),
+    usage: mapUsage(event.response?.usage, current.providerMetadataKey),
    providerMetadata:
      event.response?.id || event.response?.service_tier
-        ? providerMetadata(state, {
+        ? providerMetadata(current, {
            responseId: event.response.id,
            serviceTier: event.response.service_tier,
          })
        : undefined,
  })
-  return [{ ...state, lifecycle, hasFunctionCall, tools: pending.tools }, events] satisfies StepResult
+  return [{ ...current, lifecycle, hasFunctionCall, tools: pending.tools }, events] satisfies StepResult
 })

-// Build a single human-readable message from whatever the provider supplied.
+// Build the prettiest summary available from whatever the provider supplied.
 // When both code and message are present, prefix the code so consumers see
 // the failure mode (e.g. `rate_limit_exceeded: Slow down`) instead of just
 // the bare message — production rate limits and context-length failures used
-// to be indistinguishable from generic stream drops.
-const providerErrorMessage = (event: Event, fallback: string): string => {
-  const nested = event.error ?? event.response?.error ?? undefined
+// to be indistinguishable from generic stream drops. Returns undefined when
+// the payload carries no usable summary.
+const providerErrorMessage = (event: Event, nested: OpenResponsesErrorPayload | undefined): string | undefined => {
  const message = event.message || nested?.message || undefined
  const code = event.code || nested?.code || undefined
  if (message && code) return `${code}: ${message}`
-  return message || code || fallback
+  return message || code
 }

 export const providerFailure = (id: string, event: Event, fallback: string) => {
-  const code = event.code || event.error?.code || event.response?.error?.code || undefined
-  const message = providerErrorMessage(event, fallback)
+  const nested = event.error ?? event.response?.error ?? undefined
+  const code = event.code || nested?.code || undefined
+  // Keep the full raw payload on the error even when the message is a summary.
+  const body = JSON.stringify(nested ?? event) ?? ""
+  const summary = providerErrorMessage(event, nested)
+  const message = summary ?? (body === "{}" ? fallback : body)
  const status =
    typeof event.status === "number"
      ? event.status
      : typeof event.status_code === "number"
        ? event.status_code
        : undefined
+  const reason =
+    event.type === "error" &&
+    event.error === undefined &&
+    event.response === undefined &&
+    summary === undefined &&
+    status === undefined
+      ? new ProviderInternalReason({ message })
+      : classifyProviderFailure({ message, code, status, rawBody: body })
  return new AIError({
    module: id,
    method: "stream",
-    reason: classifyProviderFailure({ message, code, status }),
+    body,
+    reason,
  })
 }

 const providerError = (state: ParserState, event: Event, fallback: string) => providerFailure(state.id, event, fallback)

-export const step = (state: ParserState, event: Event) => {
+export const step = (state: ParserState, input: Event) => {
+  const event =
+    input.item_id && outputItemID(state, input) !== input.item_id
+      ? { ...input, item_id: outputItemID(state, input) }
+      : input
  if (event.type === "response.output_text.delta" || event.type === "response.output_text.done") {
    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(
@@ -1100,20 +1220,26 @@ export const step = (state: ParserState, event: Event) => {
    )
  }
  if (event.type === "response.refusal.delta" || event.type === "response.refusal.done") {
-    if (!isRefusalEvent(event)) return ProviderShared.eventError(state.id, `${event.type} is malformed`)
+    const value = event.type === "response.refusal.delta" ? event.delta : event.refusal
+    if (!event.item_id || typeof value !== "string")
+      return ProviderShared.eventError(state.id, `${event.type} is malformed`)
    return Effect.succeed(
      event.type === "response.refusal.delta"
        ? onOutputTextDelta(state, event, event.item_id)
-        : onOutputTextDone(state, { ...event, text: event.refusal }, event.item_id),
+        : onOutputTextDone(state, { ...event, text: value }, event.item_id),
    )
  }
  if (event.type === "response.reasoning.delta" || event.type === "response.reasoning_summary_text.delta") {
    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(onReasoningDelta(state, event, event.item_id))
  }
-  if (event.type === "response.reasoning.done" || event.type === "response.reasoning_summary_text.done") {
+  if (
+    event.type === "response.reasoning.done" ||
+    event.type === "response.reasoning_summary_text.done" ||
+    event.type === "response.reasoning_text.done"
+  ) {
    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
-    return Effect.succeed(onReasoningDone(state, event))
+    return Effect.succeed(onReasoningDone(state, event, event.item_id))
  }
  if (event.type === "response.reasoning_summary_part.added")
    return event.item_id
@@ -1126,9 +1252,19 @@ export const step = (state: ParserState, event: Event) => {
  if (event.type === "response.output_item.added") {
    if (event.item?.type === "message" && !event.item.id)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
-    return Effect.succeed(onOutputItemAdded(state, event))
+    return Effect.succeed(
+      onOutputItemAdded(
+        event.output_index !== undefined && event.item?.id
+          ? { ...state, outputItems: { ...state.outputItems, [event.output_index]: event.item.id } }
+          : state,
+        event,
+      ),
+    )
  }
-  if (event.type === "response.function_call_arguments.delta") return onFunctionCallArgumentsDelta(state, event)
+  if (event.type === "response.function_call_arguments.delta" || event.type === "response.function_call_arguments.done")
+    return event.item_id
+      ? onFunctionCallArgumentsDelta(state, event)
+      : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
  if (event.type === "response.output_item.done") {
    if (event.item?.type === "message" && !event.item.id)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
@@ -1158,18 +1294,13 @@ export const initial = (request: LLMRequest, extension: Extension = BASE): Parse
  hasFunctionCall: false,
  tools: ToolStream.empty<string>(),
  lifecycle: Lifecycle.initial(),
+  outputItems: {},
  messageItems: new Set<string>(),
-  messagePhase: (value) => messagePhase(value, extension),
  messagePhases: {},
  reasoningItems: {},
  store: OpenResponsesOptions.resolve(request).store,
 })

-const messagePhase = (value: unknown, extension: Extension): MessagePhase | null | undefined => {
-  if (value === "commentary" || value === "final_answer") return value
-  return extension.messagePhase?.(value)
-}
-
 export const protocol = Protocol.make({
  id: ADAPTER,
  body: {
@@ -7,7 +7,10 @@ import { HttpTransport } from "../route/transport/index.js"
 import { Protocol } from "../route/protocol.js"
 import {
  AIError,
+  InvalidProviderOutputReason,
  LLMEvent,
+  ProviderInternalReason,
+  UnknownProviderReason,
  Usage,
  type FinishReason,
  type FinishReasonDetails,
@@ -51,7 +54,12 @@ 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 +141,7 @@ 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),
@@ -156,6 +165,9 @@ const OpenAIChatUsage = Schema.StructWithRest(
    prompt_tokens: optionalNull(Schema.Number),
    completion_tokens: optionalNull(Schema.Number),
    total_tokens: optionalNull(Schema.Number),
+    // Zai reports cache hits as top-level `cached_tokens`; DeepSeek uses `prompt_cache_hit_tokens`.
+    cached_tokens: optionalNull(Schema.Number),
+    prompt_cache_hit_tokens: optionalNull(Schema.Number),
    prompt_tokens_details: optionalNull(
      Schema.StructWithRest(
        Schema.Struct({
@@ -204,22 +216,33 @@ const OpenAIChatDelta = Schema.StructWithRest(
  [Schema.Record(Schema.String, Schema.Unknown)],
 )

-const OpenAIChatChoice = Schema.Struct({
-  delta: optionalNull(OpenAIChatDelta),
-  finish_reason: optionalNull(Schema.String),
-  native_finish_reason: optionalNull(Schema.String),
-})
+const OpenAIChatChoice = Schema.StructWithRest(
+  Schema.Struct({
+    delta: optionalNull(OpenAIChatDelta),
+    finish_reason: optionalNull(Schema.String),
+    native_finish_reason: optionalNull(Schema.String),
+    // Moonshot streams usage on `choice.usage` instead of top-level `usage`.
+    usage: optionalNull(OpenAIChatUsage),
+  }),
+  [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>
 type OpenAIChatRequestMessage = LLMRequest["messages"][number]

@@ -242,6 +265,7 @@ export interface ParserState {
  readonly reasoningEmitted: boolean
  readonly latestToolIndex?: number
  readonly nextToolIndex: number
+  readonly requireFinishReason: boolean
 }

 // =============================================================================
@@ -256,12 +280,18 @@ interface LoweringOptions {
  ) => Schema.Schema.Type<typeof OpenAIChatCacheControl> | undefined
 }

-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: ToolSchemaProjection.openAI(inputSchema),
+    parameters: inputSchema,
+    ...(supportsStrictMode ? { strict: false } : {}),
  },
  cache_control: options.cacheControl?.(tool.cache),
 })
@@ -509,11 +539,134 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
  return messages
 })

-const lowerOptions = (request: LLMRequest) => {
+// Anthropic via LiteLLM and Amazon Bedrock require `tools` to be present
+// whenever the conversation history contains tool calls/results. Send an
+// explicit empty array when we have history but no active tools.
+const hasToolHistory = (messages: ReadonlyArray<LLMRequest["messages"][number]>) => {
+  for (const message of messages) {
+    if (message.role === "tool") return true
+    if (message.role === "assistant" && message.content.some((part) => part.type === "tool-call")) return true
+  }
+  return false
+}
+
+// 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.promptCacheKey(request)
  return {
-    ...(options.store !== undefined ? { store: options.store } : {}),
-    ...(request.promptCacheKey ? { prompt_cache_key: request.promptCacheKey } : {}),
+    ...(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 } : {}),
  }
 }
@@ -531,23 +684,39 @@ 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),
    tools:
      request.tools.length === 0
-        ? undefined
+        ? hasHistory
+          ? []
+          : undefined
        : request.tools.map((tool) =>
            lowerTool(
              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 }),
@@ -557,7 +726,7 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
    presence_penalty: generation?.presencePenalty,
    seed: generation?.seed,
    stop: generation?.stop,
-    ...lowerOptions(request),
+    ...lowerOptions(request, supportsStore),
  }
 })

@@ -567,25 +736,58 @@ 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 finishReasonError = (event: OpenAIChatEvent, reason: AIError["reason"]) =>
+  new AIError({
+    module: ADAPTER,
+    method: "stream",
+    body: ProviderShared.encodeJson(event),
+    reason,
+  })
+
+const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event: OpenAIChatEvent, reason: string) {
+  switch (reason) {
+    case "error":
+      return yield* finishReasonError(
+        event,
+        new UnknownProviderReason({ message: "Provider reported an error (finish_reason: error)" }),
+      )
+    case "network_error":
+      return yield* finishReasonError(
+        event,
+        new ProviderInternalReason({ message: "Provider reported a network error (finish_reason: network_error)" }),
+      )
+    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
 // total) with a `reasoning_tokens` subset. We pass the inclusive totals
 // through and derive the non-cached breakdown so the `AI.Usage` contract is
 // satisfied on both sides.
+// 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 => {
  if (!usage) return undefined
  const input = usage.prompt_tokens ?? undefined
  const output = usage.completion_tokens ?? undefined
-  const cached = usage.prompt_tokens_details?.cached_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 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))
@@ -680,23 +882,33 @@ 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",
+        body,
        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 usage = mapUsage(event.usage) ?? state.usage
    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 rawFinishReason = choice?.finish_reason
    const finishReason =
-      rawFinishReason !== undefined && rawFinishReason !== null
-        ? { normalized: mapFinishReason(rawFinishReason), raw: choice?.native_finish_reason ?? rawFinishReason }
+      rawFinishReason
+        ? {
+            normalized: yield* mapFinishReason(event, rawFinishReason),
+            raw: choice?.native_finish_reason ?? rawFinishReason,
+          }
        : state.finishReason
    const delta = choice?.delta
    const toolDeltas = delta?.tool_calls ?? []
@@ -716,7 +928,11 @@ 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
    }

@@ -788,14 +1004,19 @@ 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* ProviderShared.eventError(ADAPTER, result.reason.message, 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.
@@ -818,16 +1039,27 @@ 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({
+      module: ADAPTER,
+      method: "stream",
+      reason: new InvalidProviderOutputReason({
+        classification: "incomplete-stream",
+        message: "OpenAI Chat stream ended without finish_reason",
+        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
@@ -836,7 +1068,7 @@ 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) }
  const metadata = reasoningMetadata(
    state.reasoningField,
    state.reasoningDetailsObserved ? state.reasoningDetails : undefined,
@@ -850,7 +1082,7 @@ const finishEvents = (state: ParserState): ReadonlyArray<LLMEvent> => {
  events.push(...toolCallEvents)
  Lifecycle.finish(lifecycle, events, { reason, usage: state.usage })
  return events
-}
+})

 // =============================================================================
 // Protocol And OpenAI Route
@@ -879,6 +1111,7 @@ export const protocol = Protocol.make({
      reasoningDetailsObserved: false,
      reasoningEmitted: false,
      nextToolIndex: 0,
+      requireFinishReason: request.model.compatibility?.requireFinishReason ?? true,
    }),
    step,
    onHalt: finishEvents,
@@ -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"
@@ -110,7 +110,7 @@ export const model = (input: ModelInput) => {
      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)
-        return Effect.succeed(undefined)
+        return Effect.undefined
      })
      const multipartMask =
        mask === undefined
@@ -5,14 +5,13 @@ import { Auth } from "../route/auth.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Protocol } from "../route/protocol.js"
 import { HttpTransport } from "../route/transport/index.js"
-import { LLMEvent, LLMRequest, type JsonSchema, type ToolDefinition } from "../schema/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 { Lifecycle } from "./utils/lifecycle.js"
 import { OpenAIImage } from "./utils/openai-image.js"
+import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
 import { ToolSchemaProjection } from "./utils/tool-schema.js"
-import { OpenResponsesChannel, type Options } from "./open-responses-channel.js"
-import { OpenAIResponsesChannel } from "./openai-responses-channel.js"
+import { OpenResponsesChannel } from "./open-responses-channel.js"

 const ADAPTER = "openai-responses"
 const NAME = "OpenAI Responses"
@@ -40,20 +39,8 @@ const OpenAIResponsesToolChoice = Schema.Union([
  Schema.Struct({ type: Schema.tag("image_generation") }),
 ])

-const OpenAIResponsesInputItem = Schema.Union([
-  Schema.Struct({
-    type: Schema.tag("message"),
-    id: Schema.optionalKey(Schema.String),
-    role: Schema.tag("assistant"),
-    content: Schema.Array(Schema.Struct({ type: Schema.tag("output_text"), text: Schema.String })),
-    phase: Schema.optionalKey(Schema.NullOr(OpenResponses.MessagePhase)),
-  }),
-  OpenResponses.InputItem,
-])
-
 const OpenAIResponsesCoreFields = {
  ...OpenResponses.coreFields,
-  input: Schema.Array(OpenAIResponsesInputItem),
  tools: optionalArray(OpenAIResponsesTools),
  tool_choice: Schema.optional(OpenAIResponsesToolChoice),
 }
@@ -64,18 +51,27 @@ 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,
-  messagePhase: (value: unknown) => (value === null ? null : undefined),
-  lowerMedia: ({ part, media, request }) => {
-    if (request.model.provider !== "xai" || media.mime !== "application/pdf") return undefined
-    return {
-      type: "input_file",
-      filename: part.filename ?? "document.pdf",
-      file_data: media.base64,
-      mime_type: media.mime,
-    }
+  acceptsItemID: (kind: OpenResponses.ItemKind, id: string) => {
+    const prefixes = ITEM_ID_PREFIXES[kind]
+    return prefixes.length === 0 || prefixes.some((prefix) => id.startsWith(prefix))
  },
 } satisfies OpenResponses.Extension

@@ -115,8 +111,10 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
    extension,
  )
  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
+  const parallelToolCalls = OpenResponses.resolveParallelToolCalls(request)
  return {
    ...body,
+    ...(parallelToolCalls === undefined ? {} : { parallel_tool_calls: parallelToolCalls }),
    tools:
      request.tools.length === 0
        ? undefined
@@ -128,46 +126,7 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
  } satisfies OpenAIResponsesBody
 })

-type HostedToolData = OpenResponses.StreamItem & {
-  readonly id: string
-  readonly status?: string
-  readonly action?: unknown
-  readonly queries?: unknown
-  readonly results?: unknown
-  readonly code?: string
-  readonly container_id?: string
-  readonly outputs?: unknown
-  readonly server_label?: string
-  readonly output?: unknown
-  readonly result?: string
-  readonly output_format?: "png" | "jpeg" | "webp"
-  readonly error?: unknown
-}
-
-const HOSTED_TOOLS = {
-  web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
-  web_search_preview_call: { name: "web_search_preview", input: (item) => item.action ?? {} },
-  file_search_call: { name: "file_search", input: (item) => ({ queries: item.queries ?? [] }) },
-  code_interpreter_call: {
-    name: "code_interpreter",
-    input: (item) => ({ code: item.code, container_id: item.container_id }),
-  },
-  computer_use_call: { name: "computer_use", input: (item) => item.action ?? {} },
-  image_generation_call: { name: "image_generation", input: () => ({}) },
-  mcp_call: {
-    name: "mcp",
-    input: (item) => ({ server_label: item.server_label, name: item.name, arguments: item.arguments }),
-  },
-  local_shell_call: { name: "local_shell", input: (item) => item.action ?? {} },
-} as const satisfies Record<string, { readonly name: string; readonly input: (item: HostedToolData) => unknown }>
-
-type HostedToolType = keyof typeof HOSTED_TOOLS
-type HostedToolItem = HostedToolData & { readonly type: HostedToolType }
-
-const isHostedToolItem = (item: OpenResponses.StreamItem): item is HostedToolItem =>
-  item.type in HOSTED_TOOLS && typeof item.id === "string" && item.id.length > 0
-
-const hostedToolResult = Effect.fn("OpenAIResponses.hostedToolResult")(function* (item: HostedToolItem) {
+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(
@@ -188,44 +147,31 @@ const hostedToolResult = Effect.fn("OpenAIResponses.hostedToolResult")(function*
  return isError ? { type: "error" as const, value: item.error } : { type: "json" as const, value: item }
 })

-const onHostedToolDone = Effect.fn("OpenAIResponses.onHostedToolDone")(function* (
-  state: OpenResponses.ParserState,
-  item: HostedToolItem,
-) {
-  const tool = HOSTED_TOOLS[item.type]
-  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: yield* hostedToolResult(item),
-      providerExecuted: true,
-      providerMetadata,
-    }),
-  )
-  return [{ ...state, lifecycle }, events] satisfies OpenResponses.StepResult
-})
+const HOSTED_TOOLS = {
+  web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
+  web_search_preview_call: { name: "web_search_preview", input: (item) => item.action ?? {} },
+  file_search_call: { name: "file_search", input: (item) => ({ queries: item.queries ?? [] }) },
+  code_interpreter_call: {
+    name: "code_interpreter",
+    input: (item) => ({ code: item.code, container_id: item.container_id }),
+  },
+  computer_call: { name: "computer_use", input: (item) => item.action ?? {} },
+  image_generation_call: { name: "image_generation", input: () => ({}), result: hostedToolResult },
+  mcp_call: {
+    name: "mcp",
+    input: (item) => ({ server_label: item.server_label, name: item.name, arguments: item.arguments }),
+  },
+} as const satisfies ResponsesHostedTools.Definitions

 const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
-  if (event.type === "response.reasoning_text.delta" || event.type === "response.reasoning_summary.delta")
+  if (event.type === "response.reasoning_text.delta")
    return event.item_id
-      ? Effect.succeed(OpenResponses.onReasoningDelta(state, event, event.item_id))
+      ? Effect.succeed(
+          OpenResponses.onReasoningDelta(state, event, OpenResponses.outputItemID(state, event) ?? event.item_id),
+        )
      : ProviderShared.eventError(ADAPTER, `${event.type} is missing item_id`)
-  if (event.type === "response.reasoning_text.done" || event.type === "response.reasoning_summary.done")
-    return event.item_id
-      ? Effect.succeed(OpenResponses.onReasoningDone(state, event))
-      : ProviderShared.eventError(ADAPTER, `${event.type} is missing item_id`)
-  if (event.type === "response.output_item.done" && event.item && isHostedToolItem(event.item))
-    return onHostedToolDone(state, event.item)
+  if (event.type === "response.output_item.done" && event.item && ResponsesHostedTools.isItem(event.item, HOSTED_TOOLS))
+    return ResponsesHostedTools.onDone(state, event.item, HOSTED_TOOLS)
  return OpenResponses.step(state, event)
 }

@@ -247,11 +193,7 @@ const endpoint = Endpoint.path<OpenAIResponsesBody>(PATH, { baseURL: DEFAULT_BAS
 const auth = Auth.none

 export const httpTransport = HttpTransport.sseJson.with<OpenAIResponsesBody>()
-export const channelTransport = (options: Omit<Options, "driver">) =>
-  OpenResponsesChannel.transport<OpenAIResponsesBody>({
-    ...options,
-    driver: (input) => OpenAIResponsesChannel.driver({ id: options.id, name: options.name, ...input }),
-  })
+export const channelTransport = OpenResponsesChannel.transport<OpenAIResponsesBody>
 export const transport = channelTransport({
  id: ADAPTER,
  name: NAME,
@@ -267,7 +209,7 @@ 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"
@@ -24,6 +24,17 @@ export const JsonObject = Schema.Record(Schema.String, Schema.Unknown)
 export const optionalArray = <const S extends Schema.Top>(schema: S) => Schema.optional(Schema.Array(schema))
 export const optionalNull = <const S extends Schema.Top>(schema: S) => Schema.optional(Schema.NullOr(schema))

+export const OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH = 64
+
+// OpenAI limits `prompt_cache_key` to 64 chars; DeepSeek and Zai inherit the same
+// limit via their OpenAI-compatible APIs. Clamp with unicode-aware slicing.
+export const promptCacheKey = (request: LLMRequest): string | undefined => {
+  if (request.cache === "none" || request.promptCacheKey === undefined) return undefined
+  const chars = Array.from(request.promptCacheKey)
+  if (chars.length <= OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH) return request.promptCacheKey
+  return chars.slice(0, OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH).join("")
+}
+
 /**
 * Streaming tool-call accumulator. Adapters that build a tool call across
 * multiple `tool-input-delta` chunks store the partial JSON input string here
@@ -16,7 +16,7 @@ export const decodeDataUrl = (
  url: string,
  module: string,
 ): Effect.Effect<{ readonly mediaType: string; readonly data: Uint8Array } | undefined, AIError> => {
-  if (!url.startsWith("data:")) return Effect.succeed(undefined)
+  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"))
  return Effect.fromResult(Encoding.decodeBase64(match[2])).pipe(
@@ -28,7 +28,11 @@ export const ResponseIncludables = [
 export type ResponseIncludable = (typeof ResponseIncludables)[number] | (string & {})

 export const ServiceTiers = ["auto", "default", "flex", "priority"] as const
-export type ServiceTier = (typeof ServiceTiers)[number]
+export type ServiceTier = (typeof ServiceTiers)[number] | (string & {})
+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]
@@ -38,7 +42,7 @@ export const ResponseIncludableSchema = Schema.declare<ResponseIncludable>(
  (value): value is ResponseIncludable => typeof value === "string",
  { title: "ResponseIncludable" },
 )
-export const ServiceTierSchema = Schema.Literals(ServiceTiers)
+export const ServiceTierSchema = ServiceTier
 export const TruncationSchema = Schema.Literals(Truncations)

 export const AllowedTools = Schema.Struct({
@@ -9,8 +9,8 @@ export type OpenAITextVerbosity = OpenResponsesOptions.TextVerbosity
 // in lockstep with `openai-node/src/resources/responses/responses.ts`.
 export const OpenAIResponseIncludables = OpenResponsesOptions.ResponseIncludables
 export type OpenAIResponseIncludable = OpenResponsesOptions.ResponseIncludable
-export const OpenAIServiceTiers = OpenResponsesOptions.ServiceTiers
-export type OpenAIServiceTier = OpenResponsesOptions.ServiceTier
+export const OpenAIServiceTiers = [...OpenResponsesOptions.ServiceTiers, "scale"] as const
+export type OpenAIServiceTier = (typeof OpenAIServiceTiers)[number] | (string & {})

 export const OpenAIReasoningEffort = OpenResponsesOptions.ReasoningEffort
 export const OpenAITextVerbosity = OpenResponsesOptions.TextVerbosity
@@ -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
@@ -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
@@ -0,0 +1,70 @@
+import { Effect } from "effect"
+import { LLMEvent, type AIError, type ToolResultPart } from "../../schema/index.js"
+import { OpenResponses } from "../open-responses.js"
+import { Lifecycle } from "./lifecycle.js"
+
+export type Item = OpenResponses.StreamItem & {
+  readonly id: string
+  readonly status?: string
+  readonly action?: unknown
+  readonly queries?: unknown
+  readonly results?: unknown
+  readonly code?: string
+  readonly container_id?: string
+  readonly outputs?: unknown
+  readonly server_label?: string
+  readonly output?: unknown
+  readonly result?: string
+  readonly output_format?: "png" | "jpeg" | "webp"
+  readonly error?: unknown
+}
+
+export interface Definition {
+  readonly name: string
+  readonly input: (item: Item) => unknown
+  readonly result?: (item: Item) => Effect.Effect<ToolResultPart["result"], AIError>
+}
+
+export type Definitions = Readonly<Record<string, Definition>>
+
+export const isItem = <Tools extends Definitions>(item: OpenResponses.StreamItem, tools: Tools): item is Item =>
+  item.type in tools && typeof item.id === "string" && item.id.length > 0
+
+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
+})
+
+export * as ResponsesHostedTools from "./responses-hosted-tools.js"
@@ -2,20 +2,6 @@ import type { JsonSchema, LanguageModelToolSchemaCompatibility } from "../../sch
 import { isRecord } from "../../utils/record.js"
 import { GeminiToolSchema } from "./gemini-tool-schema.js"

-const removeNullSchemas = (value: unknown): unknown => {
-  if (Array.isArray(value)) return value.map(removeNullSchemas)
-  if (!isRecord(value)) return value
-  const fields = Object.fromEntries(
-    Object.entries(value)
-      .filter(([key]) => key !== "anyOf")
-      .map(([key, field]) => [key, removeNullSchemas(field)]),
-  )
-  if (!Array.isArray(value.anyOf)) return fields
-  const variants = value.anyOf.filter((variant) => !isRecord(variant) || variant.type !== "null").map(removeNullSchemas)
-  if (variants.length === 1 && isRecord(variants[0])) return { ...fields, ...variants[0] }
-  return { ...fields, anyOf: variants }
-}
-
 const tupleItemsSchema = (items: ReadonlyArray<unknown>) => {
  const projected = items.map(moonshotNode)
  if (projected.length === 0) return {}
@@ -45,24 +31,7 @@ const moonshot = (schema: JsonSchema): JsonSchema => {
  return isRecord(projected) ? projected : {}
 }

-const openAI = (schema: JsonSchema): JsonSchema => {
-  const variants = Array.isArray(schema.anyOf) ? schema.anyOf.filter(isRecord) : []
-  const flattened =
-    variants.length === 0
-      ? { ...schema, type: "object" }
-      : {
-          ...Object.fromEntries(Object.entries(schema).filter(([key]) => key !== "anyOf")),
-          type: "object",
-          properties: variants.reduce(
-            (properties, variant) => ({ ...(isRecord(variant.properties) ? variant.properties : {}), ...properties }),
-            {},
-          ),
-          additionalProperties: false,
-        }
-  const normalized = removeNullSchemas(flattened)
-  return isRecord(normalized) ? normalized : { type: "object" }
-}
-
+const openAI = (schema: JsonSchema): JsonSchema => schema
 const responses = openAI

 const gemini = (schema: JsonSchema): JsonSchema => GeminiToolSchema.convert(schema) ?? {}
@@ -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.
 */
@@ -0,0 +1,47 @@
+import { Protocol } from "../route/protocol.js"
+import { OpenResponses } from "./open-responses.js"
+import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
+
+const ADAPTER = "xai-responses"
+const NAME = "xAI Responses"
+
+const extension = {
+  id: ADAPTER,
+  name: NAME,
+} satisfies OpenResponses.Extension
+
+const HOSTED_TOOLS = {
+  web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
+  x_search_call: { name: "x_search", input: (item) => item.action ?? {} },
+  file_search_call: { name: "file_search", input: (item) => ({ queries: item.queries ?? [] }) },
+  code_interpreter_call: {
+    name: "code_interpreter",
+    input: (item) => ({ code: item.code, container_id: item.container_id }),
+  },
+  image_generation_call: { name: "image_generation", input: () => ({}) },
+  mcp_call: {
+    name: "mcp",
+    input: (item) => ({ server_label: item.server_label, name: item.name, arguments: item.arguments }),
+  },
+} as const satisfies ResponsesHostedTools.Definitions
+
+// Grok speaks the standard Responses reasoning dialect (`reasoning_summary_text.*`,
+// handled by the baseline); only its hosted tool vocabulary differs.
+const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
+  if (event.type === "response.output_item.done" && event.item && ResponsesHostedTools.isItem(event.item, HOSTED_TOOLS))
+    return ResponsesHostedTools.onDone(state, event.item, HOSTED_TOOLS)
+  return OpenResponses.step(state, event)
+}
+
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: OpenResponses.protocol.body,
+  stream: {
+    event: OpenResponses.protocol.stream.event,
+    initial: (request) => OpenResponses.initial(request, extension),
+    step,
+    terminal: OpenResponses.terminal,
+  },
+})
+
+export * as XAIResponses from "./xai-responses.js"
@@ -80,6 +80,9 @@ 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
  readonly retryAfterMs?: number | undefined
  readonly rateLimit?: HttpRateLimitDetails | undefined
  readonly http?: HttpContext | undefined
@@ -89,11 +92,13 @@ export interface ProviderFailure {
 // Keep HTTP failures and provider-reported stream failures on one typed path so
 // session retry policy never needs provider-specific string matching.
 export function classifyProviderFailure(input: ProviderFailure): AIError["reason"] {
-  const body = input.http?.body ?? ""
+  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 text = body || input.message
+  // 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)

@@ -38,13 +38,23 @@ export type Settings = ProviderPackage.Settings &

 const fromRequest = Effect.fn("GoogleVertex.fromRequest")(function* (request: LLMRequest) {
  const body = yield* Gemini.protocol.body.from(request)
+  // Vertex's native REST schema rejects `id` on FunctionCall/FunctionResponse parts with HTTP 400,
+  // unlike AI Studio, so history minted there cannot be lowered verbatim.
+  const contents = body.contents.map((content) => ({
+    ...content,
+    parts: (content.parts ?? []).map((part) => {
+      if ("functionCall" in part) return { ...part, functionCall: { ...part.functionCall, id: undefined } }
+      if ("functionResponse" in part) return { ...part, functionResponse: { ...part.functionResponse, id: undefined } }
+      return part
+    }),
+  }))
  const value = request.providerOptions?.labels
  const labels = ProviderShared.isRecord(value)
    ? Object.fromEntries(
        Object.entries(value).filter((entry): entry is [string, string] => typeof entry[1] === "string"),
      )
    : undefined
-  return { ...body, labels }
+  return { ...body, contents, labels }
 })

 const protocol = {
@@ -1,9 +1,13 @@
 import { mergeProviderOptions, type ProviderOptions } from "../schema/index.js"
-import type { OpenResponsesOptionsInput } from "./open-responses-options.js"
+import type { OpenAIServiceTier } from "../protocols/utils/openai-options.js"
+import type { Options } from "../protocols/utils/open-responses-options.js"

 export type { OpenAIResponseIncludable, OpenAIServiceTier } from "../protocols/utils/openai-options.js"

-export type OpenAIOptionsInput = OpenResponsesOptionsInput
+export type OpenAIOptionsInput = Omit<Options, "serviceTier"> & {
+  readonly serviceTier?: OpenAIServiceTier
+  readonly [key: string]: unknown
+}

 export type OpenAIProviderOptionsInput = OpenAIOptionsInput

@@ -119,7 +119,6 @@ export const protocol = Protocol.make({
            ...body,
            messages,
            ...bodyOptions(request.providerOptions),
-            ...(request.promptCacheKey ? { prompt_cache_key: request.promptCacheKey } : {}),
          } as OpenRouterBody
        }),
      ),
@@ -5,7 +5,8 @@ import { HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
 import * as OpenAICompatibleProfiles from "./openai-compatible-profile.js"
 import * as OpenAICompatibleChat from "../protocols/openai-compatible-chat.js"
 import * as OpenAIChat from "../protocols/openai-chat.js"
-import * as OpenAIResponses from "../protocols/openai-responses.js"
+import { OpenResponsesChannel } from "../protocols/open-responses-channel.js"
+import { XAIResponses } from "../protocols/xai-responses.js"
 import { XAIImages } from "../protocols/xai-images.js"
 import type { OpenAIOptionsInput } from "./openai-options.js"
 import type { ProviderPackage } from "../provider-package.js"
@@ -34,14 +35,14 @@ const responsesRoute = Route.make({
  id: "openai-responses",
  provider: id,
  providerMetadataKey: "xai",
-  protocol: OpenAIResponses.protocol,
+  protocol: XAIResponses.protocol,
  endpoint: Endpoint.path("/responses", { baseURL: OpenAICompatibleProfiles.profiles.xai.baseURL }),
-  transport: OpenAIResponses.channelTransport({
+  transport: OpenResponsesChannel.transport({
    id: "openai-responses",
    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({
@@ -7,6 +7,7 @@ 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 {
@@ -321,12 +322,30 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
                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,
-                ),
+              const stream = Stream.suspend(() => {
+                let state = protocol.stream.initial(request)
+                const parsed = events.pipe(
+                  Stream.mapEffect((event) =>
+                    protocol.stream.step(state, event).pipe(
+                      Effect.map(([next, output]) => {
+                        state = next
+                        return output
+                      }),
+                    ),
+                  ),
+                  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),
              )
@@ -382,7 +401,8 @@ 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
@@ -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>
 }

 /**
@@ -153,6 +153,9 @@ 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

@@ -152,6 +152,8 @@ 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>

@@ -51,6 +51,7 @@ export const MediaPart = Schema.Struct({
  mediaType: Schema.String,
  data: Schema.Union([Schema.String, Schema.Uint8Array]),
  filename: Schema.optional(Schema.String),
+  cache: Schema.optional(CacheHint),
  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
 }).annotate({ identifier: "LLM.Content.Media" })
 export type MediaPart = Schema.Schema.Type<typeof MediaPart>
@@ -255,6 +256,7 @@ export namespace ToolDefinition {
 export class ToolChoice extends Schema.Class<ToolChoice>("LLM.ToolChoice")({
  type: Schema.Literals(["auto", "none", "required", "tool"]),
  name: Schema.optional(Schema.String),
+  disableParallelToolUse: Schema.optional(Schema.Boolean),
 }) {}

 export namespace ToolChoice {
@@ -155,6 +155,11 @@ export class LanguageModelCompatibility extends Schema.Class<LanguageModelCompat
  reasoningField: Schema.optional(Schema.String),
  maxTokensField: Schema.optional(LanguageModelMaxTokensFieldCompatibility),
  requireFinishReason: 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 {
@@ -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
+}
@@ -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({
@@ -0,0 +1,31 @@
+{
+  "version": 1,
+  "metadata": {
+    "tags": [
+      "prefix:azure",
+      "provider:azure"
+    ],
+    "name": "azure/chat-streams-text",
+    "recordedAt": "2026-08-23T17:21:53.198Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://aiden-azury-group.openai.azure.com/openai/v1/chat/completions?api-version=v1",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-5.6-luna\",\"messages\":[{\"role\":\"user\",\"content\":\"Reply with exactly one word: hello\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"store\":false,\"reasoning_effort\":\"medium\"}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"choices\":[],\"created\":0,\"id\":\"\",\"model\":\"\",\"object\":\"\",\"prompt_filter_results\":[{\"prompt_index\":0,\"content_filter_results\":{}}]}\n\ndata: {\"choices\":[{\"content_filter_results\":{},\"delta\":{\"content\":\"\",\"refusal\":null,\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1787505712,\"id\":\"chatcmpl-EG6BEiYSfrcTSI2WX8PqNzERZDcPc\",\"model\":\"gpt-5.6-luna-2026-07-09\",\"obfuscation\":\"Mxr\",\"object\":\"chat.completion.chunk\",\"service_tier\":\"default\",\"system_fingerprint\":null,\"usage\":null}\n\ndata: {\"choices\":[{\"content_filter_results\":{},\"delta\":{\"content\":\"hello\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1787505712,\"id\":\"chatcmpl-EG6BEiYSfrcTSI2WX8PqNzERZDcPc\",\"model\":\"gpt-5.6-luna-2026-07-09\",\"obfuscation\":\"\",\"object\":\"chat.completion.chunk\",\"service_tier\":\"default\",\"system_fingerprint\":null,\"usage\":null}\n\ndata: {\"choices\":[{\"content_filter_results\":{},\"delta\":{},\"finish_reason\":\"stop\",\"index\":0,\"logprobs\":null}],\"created\":1787505712,\"id\":\"chatcmpl-EG6BEiYSfrcTSI2WX8PqNzERZDcPc\",\"model\":\"gpt-5.6-luna-2026-07-09\",\"obfuscation\":\"WyZa5AY1CaCeFdS\",\"object\":\"chat.completion.chunk\",\"service_tier\":\"default\",\"system_fingerprint\":null,\"usage\":null}\n\ndata: {\"choices\":[],\"created\":1787505712,\"id\":\"chatcmpl-EG6BEiYSfrcTSI2WX8PqNzERZDcPc\",\"latency_checkpoint\":{\"engine_tbt_ms\":20,\"engine_ttft_ms\":106,\"engine_ttlt_ms\":206,\"pre_inference_ms\":89,\"service_tbt_ms\":20,\"service_ttft_ms\":480,\"service_ttlt_ms\":576,\"total_duration_ms\":491,\"user_visible_ttft_ms\":391},\"model\":\"gpt-5.6-luna-2026-07-09\",\"obfuscation\":\"6\",\"object\":\"chat.completion.chunk\",\"service_tier\":\"default\",\"system_fingerprint\":null,\"usage\":{\"completion_tokens\":5,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"audio_tokens\":0,\"reasoning_tokens\":0,\"rejected_prediction_tokens\":0},\"prompt_tokens\":13,\"prompt_tokens_details\":{\"audio_tokens\":0,\"cache_write_tokens\":0,\"cached_tokens\":0},\"total_tokens\":18}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -7,7 +7,13 @@
    "route": "cloudflare-workers-ai",
    "transport": "http",
    "model": "@cf/openai/gpt-oss-20b",
-    "tags": ["prefix:cloudflare-workers-ai", "provider:cloudflare-workers-ai", "tool", "tool-call", "golden"]
+    "tags": [
+      "prefix:cloudflare-workers-ai",
+      "provider:cloudflare-workers-ai",
+      "tool",
+      "tool-call",
+      "golden"
+    ]
  },
  "interactions": [
    {
@@ -18,7 +24,7 @@
        "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,
@@ -29,4 +35,4 @@
      }
    }
  ]
-}
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "tags": [
+      "prefix:google-vertex",
+      "provider:google-vertex",
+      "protocol:gemini"
+    ],
+    "name": "google-vertex/calls-a-tool",
+    "recordedAt": "2026-08-23T17:21:51.036Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://aiplatform.googleapis.com/v1/publishers/google/models/gemini-3.5-flash:streamGenerateContent?alt=sse",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"contents\":[{\"role\":\"user\",\"parts\":[{\"text\":\"What is the weather in Paris? Use the lookup_weather tool.\"}]}],\"tools\":[{\"functionDeclarations\":[{\"name\":\"lookup_weather\",\"description\":\"Look up the current weather for a city\",\"parameters\":{\"required\":[\"city\"],\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}}}}]}]}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"candidates\": [{\"content\": {\"role\": \"model\",\"parts\": [{\"functionCall\": {\"name\": \"lookup_weather\",\"args\": {\"city\": \"Paris\"},\"id\": \"call_425130\"},\"thoughtSignature\": \"AY89a1+1fXnLgYhHMuN3Ak6LBhT6PcrYOW7iPav4LfsacvG/Z6l1yJ+AsU7vWhFj/JyPIbsJJQ+GjohM9sCIZ6nqUOIg3reo/7osmrCvFrVHedTHQcwiPzoz2Kp3gb+uWjFAXxk1EX4IRAKcu0ox1W/Z9PpuZvHkTerGO2a82e02N6MAF1YhhtbXFvSdqLRih2Os68rdOk5/Bcld7ol8qUgeyIZ3CtI3OJ5jwRcD8LjvK33A7ZFzH5Bxp/peUmXvqnu5iNhnGBxZaJy/vupCtxRZxjaS+ojG0/UhyrnRiKIpbzQ0FBkxePPn8GCX/LOe2y3GUc98co8lN8OOuCd9ZmEdx5AjHmQkPO9fAV9SxG6Bda6SDWVL8o/Uz3WSQYoUEfAdoajEWIBvcisoeCJjb7zgmRRZ9VQSPl3RXj5LFRvX8jn0YKV1CahYbc24jA==\"}]}}],\"usageMetadata\": {\"trafficType\": \"ON_DEMAND\"},\"modelVersion\": \"gemini-3.5-flash\",\"createTime\": \"2026-08-23T17:21:50.308576Z\",\"responseId\": \"LiyLauDqErCErb8Pj8aWkAs\"}\r\n\r\ndata: {\"candidates\": [{\"content\": {\"role\": \"model\",\"parts\": [{\"text\": \"\"}]},\"finishReason\": \"STOP\"}],\"usageMetadata\": {\"promptTokenCount\": 39,\"candidatesTokenCount\": 16,\"totalTokenCount\": 102,\"trafficType\": \"ON_DEMAND\",\"promptTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 39}],\"candidatesTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 16}],\"thoughtsTokenCount\": 47},\"modelVersion\": \"gemini-3.5-flash\",\"createTime\": \"2026-08-23T17:21:50.308576Z\",\"responseId\": \"LiyLauDqErCErb8Pj8aWkAs\"}\r\n\r\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "tags": [
+      "prefix:google-vertex",
+      "provider:google-vertex",
+      "protocol:gemini"
+    ],
+    "name": "google-vertex/continues-after-a-tool-result",
+    "recordedAt": "2026-08-23T17:21:51.853Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://aiplatform.googleapis.com/v1/publishers/google/models/gemini-3.5-flash:streamGenerateContent?alt=sse",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"contents\":[{\"role\":\"user\",\"parts\":[{\"text\":\"What is the weather in Paris?\"}]},{\"role\":\"model\",\"parts\":[{\"functionCall\":{\"name\":\"lookup_weather\",\"args\":{\"city\":\"Paris\"}},\"thoughtSignature\":\"skip_thought_signature_validator\"}]},{\"role\":\"user\",\"parts\":[{\"functionResponse\":{\"name\":\"lookup_weather\",\"response\":{\"name\":\"lookup_weather\",\"content\":\"18C, light rain\"}}}]}],\"tools\":[{\"functionDeclarations\":[{\"name\":\"lookup_weather\",\"description\":\"Look up the current weather for a city\",\"parameters\":{\"required\":[\"city\"],\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}}}}]}]}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"candidates\": [{\"content\": {\"role\": \"model\",\"parts\": [{\"text\": \"The weather in Paris is currently 18°C with light rain.\"}]}}],\"usageMetadata\": {\"trafficType\": \"ON_DEMAND\"},\"modelVersion\": \"gemini-3.5-flash\",\"createTime\": \"2026-08-23T17:21:51.220919Z\",\"responseId\": \"LyyLave9DbWnrb8P1IjLmQQ\"}\r\n\r\ndata: {\"candidates\": [{\"content\": {\"role\": \"model\",\"parts\": [{\"text\": \"\",\"thoughtSignature\": \"AY89a197c+fpHJftPtcufnqMAyoRQKVEQK+KeG+RVHVx2wKil3L4jP4YWvfVbcuOFr2jio4Kre/hCrDANAoMFSvaZrdaPeo1b5bXQSmJKMH03yM5M6q6ME6JiBvXym143U4exIde4UbOh2tMeyXMvB3aWxcavIHd78g5G5QPLreo6A3LO5871cYYVeRwteY+/zbEdqfaAq1hlk6WYpWkNljYpjMyKwr15YC8rFLh3HYayS9tTN++GGrk/reZn6C3OEPlzPou/pXRATzcEAGVl/TW\"}]},\"finishReason\": \"STOP\"}],\"usageMetadata\": {\"promptTokenCount\": 59,\"candidatesTokenCount\": 15,\"totalTokenCount\": 98,\"trafficType\": \"ON_DEMAND\",\"promptTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 59}],\"candidatesTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 15}],\"thoughtsTokenCount\": 24},\"modelVersion\": \"gemini-3.5-flash\",\"createTime\": \"2026-08-23T17:21:51.220919Z\",\"responseId\": \"LyyLave9DbWnrb8P1IjLmQQ\"}\r\n\r\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "tags": [
+      "prefix:google-vertex",
+      "provider:google-vertex",
+      "protocol:gemini"
+    ],
+    "name": "google-vertex/streams-text",
+    "recordedAt": "2026-08-23T17:21:50.112Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://aiplatform.googleapis.com/v1/publishers/google/models/gemini-3.5-flash:streamGenerateContent?alt=sse",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"contents\":[{\"role\":\"user\",\"parts\":[{\"text\":\"Reply with exactly one word: hello\"}]}]}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"candidates\": [{\"content\": {\"role\": \"model\",\"parts\": [{\"text\": \"Hello\"}]}}],\"usageMetadata\": {\"trafficType\": \"ON_DEMAND\"},\"modelVersion\": \"gemini-3.5-flash\",\"createTime\": \"2026-08-23T17:21:48.528714Z\",\"responseId\": \"LCyLasqiIO6crb8P1sDboQc\"}\r\n\r\ndata: {\"candidates\": [{\"content\": {\"role\": \"model\",\"parts\": [{\"text\": \"\",\"thoughtSignature\": \"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\"}]},\"finishReason\": \"STOP\"}],\"usageMetadata\": {\"promptTokenCount\": 7,\"candidatesTokenCount\": 1,\"totalTokenCount\": 150,\"trafficType\": \"ON_DEMAND\",\"promptTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 7}],\"candidatesTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 1}],\"thoughtsTokenCount\": 142},\"modelVersion\": \"gemini-3.5-flash\",\"createTime\": \"2026-08-23T17:21:48.528714Z\",\"responseId\": \"LCyLasqiIO6crb8P1sDboQc\"}\r\n\r\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,58 @@
+import { describe, expect, test } from "bun:test"
+import { Allow, MalformedJSON, PartialJSON, parse } from "../src/protocols/utils/partial-json.js"
+
+describe("partial JSON", () => {
+  test("parses complete JSON", () => {
+    expect(parse('{"key":"value","items":[1,true,null]}')).toEqual({
+      key: "value",
+      items: [1, true, null],
+    })
+
+    const object = parse('{"__proto__":{"safe":true}}') as Record<string, unknown>
+    expect(Object.hasOwn(object, "__proto__")).toBe(true)
+  })
+
+  test("parses partial strings", () => {
+    expect(parse('"hello')).toBe("hello")
+    expect(parse('"hello \\u12')).toBe("hello ")
+    expect(() => parse('"hello', ~Allow.STR)).toThrow(PartialJSON)
+  })
+
+  test("repairs invalid escapes and raw control characters", () => {
+    expect(parse('{"path":"A\\H","text":"first\tsecond"}')).toEqual({
+      path: "A\\H",
+      text: "first\tsecond",
+    })
+  })
+
+  test("preserves prototype keys in partial objects", () => {
+    const object = parse('{"__proto__":{"safe":true}') as Record<string, unknown>
+
+    expect(Object.hasOwn(object, "__proto__")).toBe(true)
+    expect(Object.getPrototypeOf(object)).toBe(Object.prototype)
+    expect(object.__proto__).toEqual({ safe: true })
+  })
+
+  test("controls partial collection values independently", () => {
+    expect(parse('["', Allow.ARR)).toEqual([])
+    expect(parse('["', Allow.ARR | Allow.STR)).toEqual([""])
+    expect(parse('{"key":"', Allow.OBJ)).toEqual({})
+    expect(parse('{"key":"', Allow.OBJ | Allow.STR)).toEqual({ key: "" })
+  })
+
+  test("parses partial literals and numbers", () => {
+    expect(parse("nu", Allow.NULL)).toBeNull()
+    expect(parse("tr", Allow.BOOL)).toBe(true)
+    expect(parse("fa", Allow.BOOL)).toBe(false)
+    expect(parse("1e", Allow.NUM)).toBe(1)
+  })
+
+  test("distinguishes disallowed partial values from malformed values", () => {
+    expect(() => parse("[", Allow.STR)).toThrow(PartialJSON)
+    expect(() => parse("n", ~Allow.NULL)).toThrow(MalformedJSON)
+  })
+
+  test("rejects empty input", () => {
+    expect(() => parse("  ")).toThrow("is empty")
+  })
+})
@@ -106,3 +106,20 @@ describe("provider error classification", () => {
    expect(classifyProviderFailure({ message: "not-json" })._tag).toBe("UnknownProvider")
  })
 })
+
+describe("provider error rawBody classification", () => {
+  test("classifies overflow signals buried in the raw payload when the summary is vague", () => {
+    const reason = classifyProviderFailure({
+      message: "Request failed",
+      rawBody: '{"error":{"message":"This model\'s maximum context length is 40960 tokens"}}',
+    })
+    expect(reason._tag).toBe("InvalidRequest")
+    expect(reason).toMatchObject({ classification: "context-overflow" })
+  })
+
+  test("extracts nested codes from the raw payload", () => {
+    expect(
+      classifyProviderFailure({ message: "Request failed", rawBody: '{"error":{"code":"insufficient_quota"}}' })._tag,
+    ).toBe("QuotaExceeded")
+  })
+})
@@ -4,10 +4,4 @@ import { GoogleVertexChat } from "../../src/providers.js"
 const model = GoogleVertexChat.configure({ accessToken: "test", project: "project" }).model("gemini")

 LLM.request({ model, prompt: "Hello", providerOptions: { serviceTier: "priority" } })
-
-LLM.request({
-  model,
-  prompt: "Hello",
-  // @ts-expect-error Vertex OpenAI-compatible service tiers use the OpenAI union.
-  providerOptions: { serviceTier: "premium" },
-})
+LLM.request({ model, prompt: "Hello", providerOptions: { serviceTier: "future-tier" } })
@@ -8,6 +8,8 @@ LLM.request({ model: selected, prompt: "Hello", providerOptions: { reasoningEffo
 LLM.request({ model: selected, prompt: "Hello", providerOptions: { reasoningEffort: "experimental" } })
 LLM.request({ model: selected, prompt: "Hello", providerOptions: { textVerbosity: "low" } })
 LLM.request({ model: selected, prompt: "Hello", providerOptions: { textVerbosity: "verbose" } })
+LLM.request({ model: selected, prompt: "Hello", providerOptions: { serviceTier: "scale" } })
+LLM.request({ model: selected, prompt: "Hello", providerOptions: { serviceTier: "future-tier" } })
 LLM.request({ model: chat, prompt: "Hello", providerOptions: { reasoningEffort: "max" } })
 LLM.request({ model: chat, prompt: "Hello", providerOptions: { reasoningEffort: "experimental" } })

@@ -95,7 +95,11 @@ describe("provider package entrypoints", () => {
    })
    expect(selected.route.defaults.headers).toEqual({ "x-application": "opencode" })
    expect(selected.route.defaults.http?.body).toEqual({ service_tier: "priority" })
-    expect(selected.route.defaults.providerOptions).toEqual({ reasoningEffort: "low", store: true })
+    expect(selected.route.defaults.providerOptions).toEqual({
+      reasoningEffort: "low",
+      store: true,
+      include: ["reasoning.encrypted_content"],
+    })
  })

  test("maps Anthropic-compatible settings onto the executable model", async () => {
@@ -285,7 +289,10 @@ describe("provider package entrypoints", () => {
      baseURL: "https://aiplatform.googleapis.com/v1/projects/vertex-project/locations/global/endpoints/openapi",
      path: "/responses",
    })
-    expect(responses.route.defaults.providerOptions).toEqual({ store: false })
+    expect(responses.route.defaults.providerOptions).toEqual({
+      store: false,
+      include: ["reasoning.encrypted_content"],
+    })
  })

  test("rejects conflicting Vertex auth settings at runtime", async () => {
@@ -18,6 +18,15 @@ const opus48 = AnthropicMessages.route
  .with({ endpoint: { baseURL: "https://api.anthropic.test/v1/" }, auth: Auth.header("x-api-key", "test") })
  .model({ id: "claude-opus-4-8" })

+const compileUnsignedReasoning = (model: LLMRequest["model"]) =>
+  compileRequest(
+    LLM.request({
+      model,
+      messages: [Message.assistant([{ type: "reasoning", text: "unsigned reasoning" }])],
+      cache: "none",
+    }),
+  )
+
 const request = LLM.request({
  id: "req_1",
  model,
@@ -564,6 +573,65 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("demotes unsigned reasoning when signatures are required", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileUnsignedReasoning(model)
+
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: [{ type: "text", text: "unsigned reasoning" }] },
+      ])
+    }),
+  )
+
+  it.effect("infers empty-signature compatibility across Kimi providers", () =>
+    Effect.gen(function* () {
+      const coding = AnthropicMessages.route.with({
+        provider: "kimi-for-coding",
+        endpoint: { baseURL: "https://compatible.test/v1/" },
+        auth: Auth.header("x-api-key", "test"),
+      })
+      const moonshot = AnthropicMessages.route
+        .with({
+          provider: "moonshotai",
+          endpoint: { baseURL: "https://api.moonshot.ai/anthropic" },
+          auth: Auth.bearer("test"),
+        })
+        .model({ id: "kimi-k2.6" })
+      const codingPrepared = yield* compileUnsignedReasoning(coding.model({ id: "k3" }))
+      const moonshotPrepared = yield* compileUnsignedReasoning(moonshot)
+
+      expect(codingPrepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "thinking", thinking: "unsigned reasoning", signature: "" }],
+        },
+      ])
+      expect(moonshotPrepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "thinking", thinking: "unsigned reasoning", signature: "" }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("lets an explicit signature requirement override inference", () =>
+    Effect.gen(function* () {
+      const compatible = AnthropicMessages.route
+        .with({
+          provider: "kimi-for-coding",
+          endpoint: { baseURL: "https://api.kimi.com/coding/v1/" },
+          auth: Auth.header("x-api-key", "test"),
+        })
+        .model({ id: "k3", compatibility: { requireSignature: true } })
+      const prepared = yield* compileUnsignedReasoning(compatible)
+
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: [{ type: "text", text: "unsigned reasoning" }] },
+      ])
+    }),
+  )
+
  it.effect("round-trips redacted thinking as redacted_thinking blocks", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -702,6 +770,108 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("ignores unknown content block and delta variants", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              { type: "future_event", content_block: 42, delta: 42 },
+              { type: "content_block_start", index: 0, content_block: { type: "future_block", text: 42 } },
+              { type: "content_block_delta", index: 0, delta: { text: "ignored" } },
+              { type: "content_block_delta", index: 0, delta: { type: "future_delta", text: 42 } },
+              { type: "content_block_delta", index: 0, delta: { type: "text_delta", text: "hidden" } },
+              { type: "content_block_delta", index: 0, delta: { type: "thinking_delta", thinking: "hidden" } },
+              { type: "content_block_delta", index: 0, delta: { type: "signature_delta", signature: "hidden" } },
+              { type: "content_block_stop", index: 0 },
+              { type: "content_block_start", index: 1, content_block: { type: "text", text: "" } },
+              { type: "content_block_delta", index: 1, delta: { type: "text_delta", text: "Hello" } },
+              { type: "content_block_stop", index: 1 },
+              { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 1 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([{ type: "text", text: "Hello" }])
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "end_turn" })
+    }),
+  )
+
+  it.effect("rejects malformed recognized content block variants", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              { type: "content_block_start", index: 0, content_block: { type: "text", text: 42 } },
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidProviderOutput",
+        message: "Invalid anthropic/anthropic-messages stream event",
+      })
+    }),
+  )
+
+  it.effect("rejects malformed recognized content delta variants", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              { type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
+              { type: "content_block_delta", index: 0, delta: { type: "text_delta", text: 42 } },
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidProviderOutput",
+        message: "Invalid anthropic/anthropic-messages stream event",
+      })
+    }),
+  )
+
+  it.effect("rejects malformed payloads on unrelated stream events", () =>
+    Effect.gen(function* () {
+      const events = [
+        { type: "message_start", message: { usage: { input_tokens: 1 } }, delta: 42 },
+        { type: "content_block_start", index: 0 },
+        { type: "content_block_delta", index: 0 },
+        { type: "content_block_stop", index: 0, content_block: { type: "text", text: 42 } },
+        { type: "message_delta" },
+        { type: "message_delta", delta: { stop_reason: 42 } },
+        { type: "message_stop", delta: { text: 42 } },
+        { type: "error", error: { type: "overloaded_error", message: "busy" }, content_block: 42 },
+      ]
+
+      yield* Effect.forEach(events, (event) =>
+        Effect.gen(function* () {
+          const error = yield* LLMClient.generate(request).pipe(
+            Effect.provide(fixedResponse(sseEvents(event))),
+            Effect.flip,
+          )
+
+          expect(error.reason).toMatchObject({
+            _tag: "InvalidProviderOutput",
+            message: "Invalid anthropic/anthropic-messages stream event",
+          })
+        }),
+      )
+    }),
+  )
+
  it.effect("rejects malformed recognized SSE events", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
@@ -1060,8 +1230,14 @@ describe("Anthropic Messages route", () => {
      expect(response.events).toEqual([
        { type: "step-start", index: 0 },
        { type: "tool-input-start", id: "call_1", name: "lookup" },
-        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"' },
-        { type: "tool-input-delta", id: "call_1", name: "lookup", text: ':"weather"}' },
+        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"', input: {} },
+        {
+          type: "tool-input-delta",
+          id: "call_1",
+          name: "lookup",
+          text: ':"weather"}',
+          input: { query: "weather" },
+        },
        { type: "tool-input-end", id: "call_1", name: "lookup", providerMetadata: undefined },
        {
          type: "tool-call",
@@ -0,0 +1,93 @@
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, Message, ToolDefinition, ToolCallPart } from "../../src/index.js"
+import { Azure } from "../../src/providers.js"
+import { LLMClient } from "../../src/route.js"
+import { recordedTests } from "../recorded-test.js"
+
+const resourceName = process.env.AZURE_OPENAI_RESOURCE_NAME ?? "aiden-azury-group"
+
+const chatModel = Azure.configure({
+  resourceName,
+  apiKey: process.env.AZURE_OPENAI_API_KEY ?? "fixture",
+}).chat("gpt-5.6-luna")
+
+const responsesModel = Azure.configure({
+  resourceName,
+  apiKey: process.env.AZURE_OPENAI_API_KEY ?? "fixture",
+}).responses("gpt-5.6-luna")
+
+const lookupWeather = ToolDefinition.make({
+  name: "lookup_weather",
+  description: "Look up the current weather for a city",
+  inputSchema: { type: "object", properties: { city: { type: "string" } }, required: ["city"] },
+})
+
+const recorded = recordedTests({
+  prefix: "azure",
+  provider: "azure",
+  requires: ["AZURE_OPENAI_API_KEY"],
+})
+
+describe("Azure OpenAI recorded", () => {
+  recorded.effect("chat streams text", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({ model: chatModel, prompt: "Reply with exactly one word: hello" }),
+      )
+
+      expect(response.text.toLowerCase()).toContain("hello")
+    }),
+  )
+
+  recorded.effect("responses streams text", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({ model: responsesModel, prompt: "Reply with exactly one word: bonjour" }),
+      )
+
+      expect(response.text.toLowerCase()).toContain("bonjour")
+    }),
+  )
+
+  recorded.effect("responses calls a tool", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({
+          model: responsesModel,
+          prompt: "What is the weather in Paris? Use the lookup_weather tool.",
+          tools: [lookupWeather],
+        }),
+      )
+
+      const call = response.toolCalls.find((part) => part.name === "lookup_weather")
+      expect(call).toBeDefined()
+      expect(call?.input).toMatchObject({ city: "Paris" })
+    }),
+  )
+
+  recorded.effect("responses continues after a tool result", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({
+          model: responsesModel,
+          messages: [
+            Message.user("What is the weather in Paris?"),
+            Message.assistant([
+              ToolCallPart.make({ id: "call_paris_1", name: "lookup_weather", input: { city: "Paris" } }),
+            ]),
+            Message.tool({
+              id: "call_paris_1",
+              name: "lookup_weather",
+              result: "18C, light rain",
+              resultType: "text",
+            }),
+          ],
+          tools: [lookupWeather],
+        }),
+      )
+
+      expect(response.text.length).toBeGreaterThan(0)
+    }),
+  )
+})
@@ -475,8 +475,14 @@ describe("Bedrock Converse route", () => {
      ])
      const events = response.events.filter((event) => event.type === "tool-input-delta")
      expect(events).toEqual([
-        { type: "tool-input-delta", id: "tool_1", name: "lookup", text: '{"query"' },
-        { type: "tool-input-delta", id: "tool_1", name: "lookup", text: ':"weather"}' },
+        { type: "tool-input-delta", id: "tool_1", name: "lookup", text: '{"query"', input: {} },
+        {
+          type: "tool-input-delta",
+          id: "tool_1",
+          name: "lookup",
+          text: ':"weather"}',
+          input: { query: "weather" },
+        },
      ])
      expect(response.events.at(-1)).toMatchObject({
        type: "finish",
@@ -485,7 +491,7 @@ describe("Bedrock Converse route", () => {
    }),
  )

-  it.effect("emits malformed tool input as an unexecuted tool error", () =>
+  it.effect("recovers incomplete tool input at finalization", () =>
    Effect.gen(function* () {
      const body = eventStreamBody(
        ["messageStart", { role: "assistant" }],
@@ -502,10 +508,10 @@ describe("Bedrock Converse route", () => {
      )
      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))

-      expect(response.events.find((event) => event.type === "tool-input-error")).toMatchObject({
+      expect(response.events.find((event) => event.type === "tool-call")).toMatchObject({
        id: "tool_1",
        name: "lookup",
-        raw: '{"query":"partial',
+        input: { query: "partial" },
      })
      expect(response.finishReason).toEqual({ normalized: "tool-calls", raw: "end_turn" })
    }),
@@ -710,6 +716,32 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("ignores unknown normal stream events", () =>
+    Effect.gen(function* () {
+      const body = concat([
+        eventFrame("messageStart", { role: "assistant" }),
+        eventFrame("futureEvent", { message: "Ignore this" }),
+        eventFrame("messageStop", { stopReason: "end_turn" }),
+      ])
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "end_turn" })
+    }),
+  )
+
+  it.effect("fails unknown stream exceptions after message stop", () =>
+    Effect.gen(function* () {
+      const body = concat([
+        eventFrame("messageStart", { role: "assistant" }),
+        eventFrame("messageStop", { stopReason: "end_turn" }),
+        exceptionFrame("futureException", { message: "A future provider failure" }),
+      ])
+      const error = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)), Effect.flip)
+
+      expect(error.reason).toMatchObject({ _tag: "UnknownProvider", message: "A future provider failure" })
+    }),
+  )
+
  it.effect("classifies throttlingException as a rate limit", () =>
    Effect.gen(function* () {
      const body = concat([
@@ -6,6 +6,7 @@ import { AmazonBedrockMantle } from "../../src/providers.js"
 import { compileRequest, LLMClient } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
 import { dynamicResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
 import { recordedTests } from "../recorded-test.js"

 const credentials = {
@@ -71,7 +72,9 @@ describe("Amazon Bedrock Mantle provider", () => {
            Effect.gen(function* () {
              const request = yield* HttpClientRequest.toWeb(input.request)
              seen.push({ url: request.url, authorization: request.headers.get("authorization") ?? undefined })
-              return input.respond("", { headers: { "content-type": "text/event-stream" } })
+              return input.respond(sseEvents({ choices: [{ delta: {}, finish_reason: "stop" }] }), {
+                headers: { "content-type": "text/event-stream" },
+              })
            }),
          ),
        ),
@@ -156,14 +156,13 @@ describe("Gemini route", () => {
      expect(prepared.body.contents).toEqual([
        {
          role: "model",
-          parts: [{ functionCall: { id: undefined, name: "lookup", args: { query: "weather" } } }],
+          parts: [{ functionCall: { name: "lookup", args: { query: "weather" } } }],
        },
        {
          role: "user",
          parts: [
            {
              functionResponse: {
-                id: undefined,
                name: "lookup",
                response: { name: "lookup", content: "done" },
              },
@@ -201,8 +200,8 @@ describe("Gemini route", () => {
        {
          role: "model",
          parts: [
-            { functionCall: { id: undefined, name: "lookup", args: { query: "weather" } } },
-            { functionCall: { id: undefined, name: "lookup", args: { query: "time" } } },
+            { functionCall: { name: "lookup", args: { query: "weather" } } },
+            { functionCall: { name: "lookup", args: { query: "time" } } },
          ],
        },
        {
@@ -210,14 +209,12 @@ describe("Gemini route", () => {
          parts: [
            {
              functionResponse: {
-                id: undefined,
                name: "lookup",
                response: { name: "lookup", content: "sunny" },
              },
            },
            {
              functionResponse: {
-                id: undefined,
                name: "lookup",
                response: { name: "lookup", content: "noon" },
              },
@@ -228,6 +225,104 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("lowers function call ids for gemini 3 models", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: gemini3,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "done", resultType: "text" }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          parts: [
+            {
+              functionCall: { id: "call_1", name: "lookup", args: { query: "weather" } },
+              thoughtSignature: "skip_thought_signature_validator",
+            },
+          ],
+        },
+        {
+          role: "user",
+          parts: [
+            {
+              functionResponse: {
+                id: "call_1",
+                name: "lookup",
+                response: { name: "lookup", content: "done" },
+              },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("omits function call ids entirely for pre-gemini-3 models", () =>
+    Effect.gen(function* () {
+      const messages = [
+        Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } })]),
+        Message.tool({ id: "call_1", name: "lookup", result: "done", resultType: "text" }),
+      ]
+      const legacy = yield* compileRequest(LLM.request({ model, messages }))
+      const older = yield* compileRequest(
+        LLM.request({
+          model: Gemini.route
+            .with({
+              endpoint: { baseURL: "https://generativelanguage.test/v1beta/" },
+              auth: Auth.header("x-goog-api-key", "test"),
+            })
+            .model({ id: "gemini-1.5-flash" }),
+          messages,
+        }),
+      )
+
+      expect(legacy.body.contents).toEqual([
+        { role: "model", parts: [{ functionCall: { name: "lookup", args: { query: "weather" } } }] },
+        {
+          role: "user",
+          parts: [{ functionResponse: { name: "lookup", response: { name: "lookup", content: "done" } } }],
+        },
+      ])
+      expect(JSON.stringify(legacy.body.contents)).not.toContain('"id"')
+      expect(JSON.stringify(older.body.contents)).not.toContain('"id"')
+    }),
+  )
+
+  it.effect("includes function call ids for non-gemini model ids", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: Gemini.route
+            .with({
+              endpoint: { baseURL: "https://generativelanguage.test/v1beta/" },
+              auth: Auth.header("x-goog-api-key", "test"),
+            })
+            .model({ id: "gemma-3-27b-it" }),
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "done", resultType: "text" }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        { role: "model", parts: [{ functionCall: { id: "call_1", name: "lookup", args: { query: "weather" } } }] },
+        {
+          role: "user",
+          parts: [
+            { functionResponse: { id: "call_1", name: "lookup", response: { name: "lookup", content: "done" } } },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("prepares multimodal user input and tool history", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -419,13 +514,16 @@ describe("Gemini route", () => {
      expect(prepared.body.contents).toEqual([
        {
          role: "model",
-          parts: [{ functionCall: { name: "read", args: { path: "pixel.png" } }, thoughtSignature: "sig_1" }],
+          parts: [
+            { functionCall: { id: "call_image", name: "read", args: { path: "pixel.png" } }, thoughtSignature: "sig_1" },
+          ],
        },
        {
          role: "user",
          parts: [
            {
              functionResponse: {
+                id: "call_image",
                name: "read",
                response: { name: "read", content: "Image read successfully" },
                parts: [{ inlineData: { mimeType: "image/png", data: "AAECAw==" } }],
@@ -808,6 +906,54 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("ignores unknown response parts", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents({
+              candidates: [
+                {
+                  content: {
+                    role: "model",
+                    parts: [
+                      { text: "Hello " },
+                      { executableCode: { language: "PYTHON", code: "print('ignored')" } },
+                      { text: "world" },
+                    ],
+                  },
+                  finishReason: "STOP",
+                },
+              ],
+            }),
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("Hello world")
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "STOP" })
+    }),
+  )
+
+  it.effect("rejects malformed recognized response parts", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents({
+              candidates: [{ content: { role: "model", parts: [{ text: 42 }] } }],
+            }),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error).toBeInstanceOf(AIError)
+      expect(error.reason).toMatchObject({ _tag: "InvalidProviderOutput" })
+      expect(error.message).toContain("Invalid google/gemini stream event")
+    }),
+  )
+
  it.effect("preserves thoughtSignature for reasoning and tool-call continuation", () =>
    Effect.gen(function* () {
      const body = sseEvents({
@@ -848,8 +994,8 @@ describe("Gemini route", () => {
        providerMetadata: { google: { thoughtSignature: "thought_sig" } },
      })
      expect(toolCall).toMatchObject({
-        id: "tool_0",
-        providerMetadata: { google: { functionCallId: "provider_call", thoughtSignature: "tool_sig" } },
+        id: "provider_call",
+        providerMetadata: { google: { thoughtSignature: "tool_sig" } },
      })
      expect(response.events.findIndex((event) => event.type === "reasoning-end")).toBeLessThan(
        response.events.findIndex((event) => event.type === "tool-call"),
@@ -857,23 +1003,22 @@ describe("Gemini route", () => {

      const prepared = yield* compileRequest(
        LLM.request({
-          model,
+          model: gemini3,
          messages: [
            Message.assistant([
              { type: "reasoning", text: "thinking", providerMetadata: reasoningEnd?.providerMetadata },
              ToolCallPart.make({
-                id: "tool_0",
+                id: "provider_call",
                name: "lookup",
                input: { query: "weather" },
                providerMetadata: toolCall?.providerMetadata,
              }),
            ]),
            Message.tool({
-              id: "tool_0",
+              id: "provider_call",
              name: "lookup",
              result: "done",
              resultType: "text",
-              providerMetadata: toolCall?.providerMetadata,
            }),
          ],
        }),
@@ -977,7 +1122,7 @@ describe("Gemini route", () => {
          role: "model",
          parts: [
            {
-              functionCall: { id: undefined, name: "lookup", args: { query: "weather" } },
+              functionCall: { id: "tool_0", name: "lookup", args: { query: "weather" } },
              thoughtSignature: "skip_thought_signature_validator",
            },
          ],
@@ -987,7 +1132,7 @@ describe("Gemini route", () => {
          parts: [
            {
              functionResponse: {
-                id: undefined,
+                id: "tool_0",
                name: "lookup",
                response: { name: "lookup", content: "done" },
              },
@@ -1023,15 +1168,15 @@ describe("Gemini route", () => {
          role: "model",
          parts: [
            {
-              functionCall: { id: undefined, name: "lookup", args: { query: "weather" } },
+              functionCall: { id: "tool_0", name: "lookup", args: { query: "weather" } },
              thoughtSignature: "parallel_signature",
            },
            {
-              functionCall: { id: undefined, name: "lookup", args: { query: "news" } },
+              functionCall: { id: "tool_1", name: "lookup", args: { query: "news" } },
              thoughtSignature: undefined,
            },
            {
-              functionCall: { id: undefined, name: "lookup", args: { query: "sports" } },
+              functionCall: { id: "tool_2", name: "lookup", args: { query: "sports" } },
              thoughtSignature: undefined,
            },
          ],
@@ -1059,11 +1204,11 @@ describe("Gemini route", () => {
          role: "model",
          parts: [
            {
-              functionCall: { id: undefined, name: "lookup", args: { query: "weather" } },
+              functionCall: { id: "tool_0", name: "lookup", args: { query: "weather" } },
              thoughtSignature: "skip_thought_signature_validator",
            },
            {
-              functionCall: { id: undefined, name: "lookup", args: { query: "news" } },
+              functionCall: { id: "tool_1", name: "lookup", args: { query: "news" } },
              thoughtSignature: "skip_thought_signature_validator",
            },
          ],
@@ -1101,21 +1246,17 @@ describe("Gemini route", () => {
        providerMetadata: { google: { promptTokenCount: 5, candidatesTokenCount: 1 } },
      })

-      expect(response.toolCalls).toEqual([
-        {
-          type: "tool-call",
-          id: "tool_0",
-          name: "lookup",
-          input: { query: "weather" },
-          providerExecuted: undefined,
-          providerMetadata: undefined,
-        },
-      ])
+      expect(response.toolCalls[0].id).toMatch(/^tool_[0-9a-zA-Z]+$/)
+      expect(response.toolCalls[0]).toMatchObject({
+        type: "tool-call",
+        name: "lookup",
+        input: { query: "weather" },
+      })
      expect(response.events).toEqual([
        { type: "step-start", index: 0 },
        {
          type: "tool-call",
-          id: "tool_0",
+          id: response.toolCalls[0].id,
          name: "lookup",
          input: { query: "weather" },
          providerExecuted: undefined,
@@ -1158,7 +1299,8 @@ describe("Gemini route", () => {
        ),
      )

-      expect(response.toolCalls).toEqual([{ type: "tool-call", id: "tool_0", name: "ping", input: {} }])
+      expect(response.toolCalls[0].id).toMatch(/^tool_[0-9a-zA-Z]+$/)
+      expect(response.toolCalls).toMatchObject([{ type: "tool-call", name: "ping", input: {} }])
    }),
  )

@@ -1198,7 +1340,7 @@ describe("Gemini route", () => {
            content: {
              role: "model",
              parts: [
-                { functionCall: { id: "tool_0", name: "lookup", args: { query: "weather" } } },
+                { functionCall: { id: "call_0", name: "lookup", args: { query: "weather" } } },
                { functionCall: { name: "lookup", args: { query: "news" } } },
              ],
            },
@@ -1212,16 +1354,19 @@ describe("Gemini route", () => {
        }),
      ).pipe(Effect.provide(fixedResponse(body)))

-      expect(response.toolCalls).toEqual([
-        {
-          type: "tool-call",
-          id: "tool_0",
-          name: "lookup",
-          input: { query: "weather" },
-          providerMetadata: { google: { functionCallId: "tool_0" } },
-        },
-        { type: "tool-call", id: "tool_1", name: "lookup", input: { query: "news" } },
-      ])
+      expect(response.toolCalls[0]).toMatchObject({
+        type: "tool-call",
+        id: "call_0",
+        name: "lookup",
+        input: { query: "weather" },
+      })
+      expect(response.toolCalls[1]).toMatchObject({
+        type: "tool-call",
+        name: "lookup",
+        input: { query: "news" },
+      })
+      expect(response.toolCalls[1].id).toMatch(/^tool_[0-9a-zA-Z]+$/)
+      expect(response.toolCalls[0].id).not.toBe(response.toolCalls[1].id)
      expect(response.events.at(-1)).toMatchObject({
        type: "finish",
        reason: { normalized: "tool-calls", raw: "STOP" },
@@ -1229,6 +1374,62 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("replaces repeated supplier ids with fresh fallback ids", () =>
+    Effect.gen(function* () {
+      const body = sseEvents({
+        candidates: [
+          {
+            content: {
+              role: "model",
+              parts: [
+                { functionCall: { id: "dup_call", name: "lookup", args: { query: "weather" } } },
+                { functionCall: { id: "dup_call", name: "lookup", args: { query: "news" } } },
+              ],
+            },
+            finishReason: "STOP",
+          },
+        ],
+      })
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.toolCalls[0]).toMatchObject({
+        id: "dup_call",
+        providerMetadata: undefined,
+      })
+      expect(response.toolCalls[1].id).toMatch(/^tool_[0-9a-zA-Z]+$/)
+      expect(response.toolCalls[1].id).not.toBe(response.toolCalls[0].id)
+    }),
+  )
+
+  it.effect("assigns distinct unique fallback ids across separate requests", () =>
+    Effect.gen(function* () {
+      const body = sseEvents({
+        candidates: [
+          {
+            content: {
+              role: "model",
+              parts: [{ functionCall: { name: "lookup", args: { query: "weather" } } }],
+            },
+            finishReason: "STOP",
+          },
+        ],
+      })
+      const req = LLMRequest.update(request, {
+        tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+      })
+      const first = yield* LLMClient.generate(req).pipe(Effect.provide(fixedResponse(body)))
+      const second = yield* LLMClient.generate(req).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(first.toolCalls[0].id).toMatch(/^tool_[0-9a-zA-Z]+$/)
+      expect(second.toolCalls[0].id).toMatch(/^tool_[0-9a-zA-Z]+$/)
+      expect(first.toolCalls[0].id).not.toBe(second.toolCalls[0].id)
+    }),
+  )
+
  it.effect("maps length and content-filter finish reasons", () =>
    Effect.gen(function* () {
      const length = yield* LLMClient.generate(request).pipe(
@@ -1339,6 +1540,73 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("survives explicit null usage counts", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { candidates: [{ content: { role: "model", parts: [{ text: "Hi" }] } }] },
+              { usageMetadata: { promptTokenCount: null, candidatesTokenCount: 5 } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("Hi")
+      expect(response.usage).toMatchObject({ outputTokens: 5, totalTokens: 5 })
+      expect(response.usage?.inputTokens).toBeUndefined()
+      expect(response.usage?.nonCachedInputTokens).toBeUndefined()
+      expect(response.usage?.cacheReadInputTokens).toBeUndefined()
+      expect(response.usage?.reasoningTokens).toBeUndefined()
+    }),
+  )
+
+  it.effect("survives null candidates, content, parts, and finish reason", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { candidates: null },
+              { candidates: [{ content: { role: "model", parts: null } }] },
+              { candidates: [{ content: null, finishReason: null }] },
+              {
+                candidates: [
+                  { content: { role: "model", parts: [{ text: "Hello" }] }, finishReason: "STOP" as const },
+                ],
+              },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("Hello")
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "STOP" })
+    }),
+  )
+
+  it.effect("treats a null thought flag on a text part as visible output", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents({
+              candidates: [
+                { content: { role: "model", parts: [{ text: "Visible", thought: null }] }, finishReason: "STOP" },
+              ],
+            }),
+          ),
+        ),
+      )
+      const reasoningStart = response.events.find((event) => event.type === "reasoning-start")
+
+      expect(reasoningStart).toBeUndefined()
+      expect(response.reasoning ?? "").toBe("")
+      expect(response.text).toBe("Visible")
+    }),
+  )
+
  it.effect("fails invalid stream events", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
@@ -0,0 +1,77 @@
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+
+import { LLM, Message, ToolDefinition, ToolCallPart } from "../../src/index.js"
+import { GoogleVertex } from "../../src/providers.js"
+import { LLMClient } from "../../src/route.js"
+import { recordedTests } from "../recorded-test.js"
+
+const model = GoogleVertex.configure({
+  apiKey: process.env.GOOGLE_VERTEX_API_KEY ?? "fixture",
+}).model("gemini-3.5-flash")
+
+const lookupWeather = ToolDefinition.make({
+  name: "lookup_weather",
+  description: "Look up the current weather for a city",
+  inputSchema: { type: "object", properties: { city: { type: "string" } }, required: ["city"] },
+})
+
+const recorded = recordedTests({
+  prefix: "google-vertex",
+  provider: "google-vertex",
+  protocol: "gemini",
+  requires: ["GOOGLE_VERTEX_API_KEY"],
+})
+
+describe("Google Vertex Gemini recorded", () => {
+  recorded.effect("streams text", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({ model, prompt: "Reply with exactly one word: hello" }),
+      )
+
+      expect(response.text.toLowerCase()).toContain("hello")
+    }),
+  )
+
+  recorded.effect("calls a tool", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({
+          model,
+          prompt: "What is the weather in Paris? Use the lookup_weather tool.",
+          tools: [lookupWeather],
+        }),
+      )
+
+      const call = response.toolCalls.find((part) => part.name === "lookup_weather")
+      expect(call).toBeDefined()
+      expect(call?.input).toMatchObject({ city: "Paris" })
+    }),
+  )
+
+  recorded.effect("continues after a tool result", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("What is the weather in Paris?"),
+            Message.assistant([
+              ToolCallPart.make({ id: "call_paris_1", name: "lookup_weather", input: { city: "Paris" } }),
+            ]),
+            Message.tool({
+              id: "call_paris_1",
+              name: "lookup_weather",
+              result: "18C, light rain",
+              resultType: "text",
+            }),
+          ],
+          tools: [lookupWeather],
+        }),
+      )
+
+      expect(response.text.length).toBeGreaterThan(0)
+    }),
+  )
+})
@@ -1,7 +1,7 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
-import { LLM } from "../../src/index.js"
+import { LLM, Message, ToolCallPart } from "../../src/index.js"
 import { GoogleVertex, GoogleVertexChat, GoogleVertexMessages, GoogleVertexResponses } from "../../src/providers.js"
 import { LLMClient } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
@@ -75,6 +75,53 @@ describe("Google Vertex providers", () => {
    }),
  )

+  it.effect("strips function call ids Vertex does not accept from lowered bodies", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: GoogleVertex.configure({
+            accessToken: "vertex-token",
+            project: "vertex-project",
+          }).model("gemini-3.5-flash"),
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({
+                id: "call_1",
+                name: "lookup",
+                input: { query: "weather" },
+                providerMetadata: { google: { functionCallId: "provider_call_1" } },
+              }),
+            ]),
+            Message.tool({
+              id: "call_1",
+              name: "lookup",
+              result: "sunny",
+              resultType: "text",
+              providerMetadata: { google: { functionCallId: "provider_call_1" } },
+            }),
+          ],
+        }),
+      )
+
+      expect(JSON.stringify(prepared.body.contents)).not.toContain('"id"')
+      expect(prepared.body.contents).toMatchObject([
+        { role: "model", parts: [{ functionCall: { id: undefined, name: "lookup", args: { query: "weather" } } }] },
+        {
+          role: "user",
+          parts: [
+            {
+              functionResponse: {
+                id: undefined,
+                name: "lookup",
+                response: { name: "lookup", content: "sunny" },
+              },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("projects Anthropic Messages onto the Vertex raw-predict API", () =>
    Effect.gen(function* () {
      const model = GoogleVertexMessages.configure({
@@ -190,6 +237,7 @@ describe("Google Vertex providers", () => {
              })
              return input.respond(
                sseEvents(
+                  { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
                  { type: "response.output_text.delta", item_id: "msg_1", delta: "Hello." },
                  { type: "response.completed", response: { id: "resp_1" } },
                ),
@@ -47,7 +47,7 @@ describe("OpenAI Chat route", () => {
    Effect.gen(function* () {
      const prepared = yield* compileRequest(request)

-      expect(prepared.body).toEqual({
+      expect(prepared.body).toMatchObject({
        model: "gpt-4o-mini",
        messages: [
          { role: "system", content: "You are concise." },
@@ -55,7 +55,8 @@ describe("OpenAI Chat route", () => {
        ],
        stream: true,
        stream_options: { include_usage: true },
-        max_tokens: 20,
+        store: false,
+        max_completion_tokens: 20,
        temperature: 0,
      })
    }),
@@ -191,6 +192,21 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("omits the prompt cache key when caching is disabled", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          prompt: "Hello",
+          promptCacheKey: "session_123",
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body).not.toHaveProperty("prompt_cache_key")
+    }),
+  )
+
  it.effect("maps the xAI Chat prompt cache key to conversation affinity", () =>
    LLMClient.generate(
      LLM.request({
@@ -325,7 +341,7 @@ describe("OpenAI Chat route", () => {
        }),
      )

-      expect(prepared.body).toEqual({
+      expect(prepared.body).toMatchObject({
        model: "gpt-4o-mini",
        messages: [
          { role: "user", content: "What is the weather?" },
@@ -342,8 +358,10 @@ describe("OpenAI Chat route", () => {
          },
          { role: "tool", tool_call_id: "call_1", content: encodeJson({ forecast: "sunny" }) },
        ],
+        tools: [],
        stream: true,
        stream_options: { include_usage: true },
+        store: false,
      })
    }),
  )
@@ -1161,8 +1179,14 @@ describe("OpenAI Chat route", () => {
      expect(response.events).toEqual([
        { type: "step-start", index: 0 },
        { type: "tool-input-start", id: "call_1", name: "lookup", providerMetadata: undefined },
-        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"' },
-        { type: "tool-input-delta", id: "call_1", name: "lookup", text: ':"weather"}' },
+        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"', input: {} },
+        {
+          type: "tool-input-delta",
+          id: "call_1",
+          name: "lookup",
+          text: ':"weather"}',
+          input: { query: "weather" },
+        },
        { type: "tool-input-end", id: "call_1", name: "lookup", providerMetadata: undefined },
        {
          type: "tool-call",
@@ -1242,6 +1266,11 @@ describe("OpenAI Chat route", () => {
      ).pipe(Effect.provide(fixedResponse(body)), Effect.flip)

      expect(error.message).toContain("OpenAI Chat tool call delta is missing id or name")
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      if (error.reason._tag !== "InvalidProviderOutput") return
+      expect(decodeJson(error.reason.raw ?? "")).toMatchObject({
+        choices: [{ finish_reason: "tool_calls" }],
+      })
    }),
  )

@@ -1255,6 +1284,7 @@ describe("OpenAI Chat route", () => {
        deltaChunk({ tool_calls: [{ index: 0, function: { arguments: ':"weather"}' } }] }),
      )
      const input = LLMRequest.update(request, {
+        model: LanguageModel.update(model, { compatibility: { requireFinishReason: false } }),
        tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
      })
      const response = yield* LLMClient.generate(input).pipe(Effect.provide(fixedResponse(body)))
@@ -1262,8 +1292,14 @@ describe("OpenAI Chat route", () => {
      expect(response.events).toEqual([
        { type: "step-start", index: 0 },
        { type: "tool-input-start", id: "call_1", name: "lookup", providerMetadata: undefined },
-        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"' },
-        { type: "tool-input-delta", id: "call_1", name: "lookup", text: ':"weather"}' },
+        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"', input: {} },
+        {
+          type: "tool-input-delta",
+          id: "call_1",
+          name: "lookup",
+          text: ':"weather"}',
+          input: { query: "weather" },
+        },
        { type: "tool-input-end", id: "call_1", name: "lookup", providerMetadata: undefined },
        {
          type: "tool-call",
@@ -70,7 +70,7 @@ describe("OpenAI-compatible Chat route", () => {
        baseURL: "https://api.deepseek.test/v1/",
        query: { "api-version": "2026-01-01" },
      })
-      expect(prepared.body).toEqual({
+      expect(prepared.body).toMatchObject({
        model: "deepseek-chat",
        messages: [
          { role: "system", content: "You are concise." },
@@ -79,7 +79,7 @@ describe("OpenAI-compatible Chat route", () => {
        tools: [
          {
            type: "function",
-            function: { name: "lookup", description: "Lookup data", parameters: { type: "object" } },
+            function: { name: "lookup", description: "Lookup data", parameters: { type: "object" }, strict: false },
          },
        ],
        tool_choice: "required",
@@ -130,7 +130,7 @@ describe("OpenAI-compatible Chat route", () => {
    Effect.gen(function* () {
      const prepared = yield* compileRequest(request)

-      expect(prepared.body).toEqual({
+      expect(prepared.body).toMatchObject({
        model: "deepseek-chat",
        messages: [
          { role: "system", content: "You are concise." },
@@ -158,6 +158,29 @@ describe("OpenAI-compatible Chat route", () => {
    }),
  )

+  it.effect("enables ZAI tool streaming except for GLM 4.5 models", () =>
+    Effect.gen(function* () {
+      const prepare = (provider: string, baseURL: string, id: string) =>
+        compileRequest(
+          LLM.request({
+            model: OpenAICompatibleChat.route.with({ provider, endpoint: { baseURL } }).model({ id }),
+            prompt: "Use a tool.",
+            tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: {} })],
+          }),
+        )
+
+      const current = yield* prepare("zai", "https://api.z.ai/api/paas/v4", "glm-4.7")
+      expect(current.body).toMatchObject({ tool_stream: true })
+
+      const legacy = yield* Effect.all(
+        ["glm-4.5", "glm-4.5-air", "glm-4.5-flash", "glm-4.5v"].map((id) =>
+          prepare("zhipuai", "https://open.bigmodel.cn/api/paas/v4", id),
+        ),
+      )
+      legacy.forEach((item) => expect(item.body).not.toHaveProperty("tool_stream"))
+    }),
+  )
+
  it.effect("matches AI SDK compatible tool request body fixture", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -180,7 +203,7 @@ describe("OpenAI-compatible Chat route", () => {
        }),
      )

-      expect(prepared.body).toEqual({
+      expect(prepared.body).toMatchObject({
        model: "deepseek-chat",
        messages: [
          { role: "user", content: "What is the weather?" },
@@ -204,6 +227,7 @@ describe("OpenAI-compatible Chat route", () => {
              name: "lookup",
              description: "Lookup data",
              parameters: { type: "object", properties: { query: { type: "string" } }, required: ["query"] },
+              strict: false,
            },
          },
        ],
@@ -329,13 +353,106 @@ describe("OpenAI-compatible Chat route", () => {
    }),
  )

-  it.effect("treats an empty finish reason as terminal", () =>
+  it.effect("rejects a stream without a required finish reason", () =>
    Effect.gen(function* () {
-      const response = yield* LLMClient.generate(request).pipe(
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(deltaChunk({ content: "Hello" }), deltaChunk({}, "")))),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidProviderOutput",
+        classification: "incomplete-stream",
+        message: "OpenAI Chat stream ended without finish_reason",
+      })
+    }),
+  )
+
+  it.effect("infers stop when finish reasons are optional", () =>
+    Effect.gen(function* () {
+      const compatible = OpenAICompatibleChat.route
+        .with({ provider: "custom", endpoint: { baseURL: "https://api.custom.test/v1" } })
+        .model({ id: "custom-model", compatibility: { requireFinishReason: false } })
+      const response = yield* LLMClient.generate(LLMRequest.update(request, { model: compatible })).pipe(
        Effect.provide(fixedResponse(sseEvents(deltaChunk({ content: "Hello" }), deltaChunk({}, "")))),
      )

-      expect(response.finishReason).toEqual({ normalized: "unknown", raw: "" })
+      expect(response.finishReason).toEqual({ normalized: "stop" })
+    }),
+  )
+
+  it.effect("normalizes the end finish reason to stop", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(deltaChunk({ content: "Hello" }), deltaChunk({}, "end")))),
+      )
+
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "end" })
+    }),
+  )
+
+  it.effect("classifies provider error finish reasons", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(deltaChunk({}, "network_error")))),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "ProviderInternal",
+        message: "Provider reported a network error (finish_reason: network_error)",
+      })
+      expect(decodeJson(error.body ?? "")).toMatchObject({
+        id: "chatcmpl_fixture",
+        choices: [{ finish_reason: "network_error" }],
+      })
+
+      const generic = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(deltaChunk({}, "error")))),
+        Effect.flip,
+      )
+      expect(generic.reason).toMatchObject({
+        _tag: "UnknownProvider",
+        message: "Provider reported an error (finish_reason: error)",
+      })
+    }),
+  )
+
+  it.effect("preserves explicit provider error events", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents({
+              id: "chatcmpl_error",
+              error: { code: 502, message: "Provider disconnected", details: { upstream: "vendor" } },
+              trace_id: "trace_1",
+            }),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", message: "Provider disconnected", status: 502 })
+      expect(decodeJson(error.body ?? "")).toMatchObject({
+        id: "chatcmpl_error",
+        error: { code: 502, message: "Provider disconnected", details: { upstream: "vendor" } },
+        trace_id: "trace_1",
+      })
+    }),
+  )
+
+  it.effect("preserves provider finish outcomes in the common reason algebra", () =>
+    Effect.gen(function* () {
+      const filtered = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(deltaChunk({}, "content_filter")))),
+      )
+      const future = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(deltaChunk({}, "future_reason")))),
+      )
+
+      expect(filtered.finishReason).toEqual({ normalized: "content-filter", raw: "content_filter" })
+      expect(future.finishReason).toEqual({ normalized: "unknown", raw: "future_reason" })
    }),
  )

@@ -355,6 +472,11 @@ describe("OpenAI-compatible Chat route", () => {
      )

      expect(error.message).toContain("OpenAI Chat received content after the finish reason")
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      if (error.reason._tag !== "InvalidProviderOutput") return
+      expect(decodeJson(error.reason.raw ?? "")).toMatchObject({
+        choices: [{ delta: { tool_calls: [{ id: "call_1" }] } }],
+      })
    }),
  )
 })
@@ -52,10 +52,28 @@ describe("Open Responses-compatible route", () => {
          { role: "user", content: [{ type: "input_text", text: "Say hello." }] },
        ],
        stream: true,
+        store: false,
+        include: ["reasoning.encrypted_content"],
      })
    }),
  )

+  it.effect("allows callers to override stateless encrypted reasoning defaults", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+        provider: "example",
+      }).model("example-model")
+      const prepared = yield* compileRequest(
+        LLM.request({ model, prompt: "Say hello.", providerOptions: { store: true, include: [] } }),
+      )
+
+      expect(prepared.body.store).toBe(true)
+      expect(prepared.body.include).toBeUndefined()
+    }),
+  )
+
  it.effect("lowers chronological system updates as standard developer messages", () =>
    Effect.gen(function* () {
      const model = configure({
@@ -78,6 +96,45 @@ describe("Open Responses-compatible route", () => {
    }),
  )

+  it.effect("uses data URLs for embedded PDF messages and tool results", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+        provider: "example",
+      }).model("example-model")
+      const pdf = "data:application/pdf;base64,JVBERi0xLjQ="
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user([{ type: "media", mediaType: "application/pdf", data: pdf, filename: "input.pdf" }]),
+            Message.assistant({ type: "tool-call", id: "call_1", name: "read", input: {} }),
+            Message.tool({
+              id: "call_1",
+              name: "read",
+              resultType: "content",
+              result: [{ type: "file", uri: pdf, mime: "application/pdf", name: "result.pdf" }],
+            }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        {
+          role: "user",
+          content: [{ type: "input_file", filename: "input.pdf", file_data: pdf }],
+        },
+        { type: "function_call", call_id: "call_1", name: "read", arguments: "{}" },
+        {
+          type: "function_call_output",
+          call_id: "call_1",
+          output: [{ type: "input_file", filename: "result.pdf", file_data: pdf }],
+        },
+      ])
+    }),
+  )
+
  it.effect("rejects OpenAI-native tools", () =>
    Effect.gen(function* () {
      const model = configure({
@@ -93,7 +150,212 @@ describe("Open Responses-compatible route", () => {
    }),
  )

-  it.effect("omits OpenAI-only nullable phases from the Open Responses baseline", () =>
+  it.effect("lowers canonical parallel tool control", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+      }).model("example-model")
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          prompt: "Read the file.",
+          tools: [
+            ToolDefinition.make({
+              name: "read",
+              description: "Read a file.",
+              inputSchema: { type: "object" },
+            }),
+          ],
+          toolChoice: { type: "auto", disableParallelToolUse: true },
+        }),
+      )
+
+      expect(prepared.body.parallel_tool_calls).toBe(false)
+      expect(prepared.body.tools).toEqual([
+        {
+          type: "function",
+          name: "read",
+          description: "Read a file.",
+          parameters: { type: "object" },
+          strict: false,
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps foreign item id grammars but drops malformed ids", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+      }).model("example-model")
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              // The baseline does not enforce a provider id grammar, so a
+              // non-OpenAI but well-formed token is resent as-is.
+              { type: "text", text: "Kept.", providerMetadata: { openresponses: { itemId: "history_1" } } },
+              // Shape violations are dropped even without a grammar policy.
+              {
+                type: "text",
+                text: "Dropped.",
+                providerMetadata: { openresponses: { itemId: `m${"a".repeat(64)}` } },
+              },
+            ]),
+          ],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        {
+          type: "message",
+          id: "history_1",
+          role: "assistant",
+          content: [{ type: "output_text", text: "Kept." }],
+        },
+        {
+          type: "message",
+          role: "assistant",
+          content: [{ type: "output_text", text: "Dropped." }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("routes response deltas by output index", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+      }).model("example-model")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Say hello." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 2, item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", output_index: 2, item_id: "wrong_message", delta: "Indexed" },
+              { type: "response.output_item.done", output_index: 2, item: { type: "message", id: "msg_1" } },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        { type: "text", text: "Indexed", providerMetadata: { openresponses: { itemId: "msg_1" } } },
+      ])
+    }),
+  )
+
+  it.effect("finalizes pending function calls from completed response output", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+        provider: "example",
+      }).model("example-model")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Look it up." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                item: { type: "function_call", id: "item_1", call_id: "call_1", name: "lookup", arguments: "" },
+              },
+              { type: "response.function_call_arguments.delta", item_id: "item_1", delta: '{"query":"par' },
+              {
+                type: "response.completed",
+                response: {
+                  output: [
+                    {
+                      type: "function_call",
+                      id: "item_1",
+                      call_id: "call_1",
+                      name: "lookup",
+                      arguments: '{"query":"complete"}',
+                    },
+                  ],
+                },
+              },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.find(LLMEvent.is.toolCall)).toMatchObject({
+        input: { query: "complete" },
+        providerMetadata: { openresponses: { itemId: "item_1" } },
+      })
+    }),
+  )
+
+  it.effect("preserves terminal reasoning metadata when item completion is missing", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+      }).model("example-model")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think it through." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                item: { type: "reasoning", id: "rs_raw", encrypted_content: null },
+              },
+              { type: "response.reasoning_summary_text.delta", item_id: "rs_raw", delta: "Thinking" },
+              {
+                type: "response.completed",
+                response: {
+                  output: [{ type: "reasoning", id: "rs_raw", encrypted_content: "raw-state" }],
+                },
+              },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
+        providerMetadata: { openresponses: { itemId: "rs_raw", reasoningEncryptedContent: "raw-state" } },
+      })
+    }),
+  )
+
+  it.effect("reconciles raw reasoning finals without streamed deltas", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+      }).model("example-model")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think it through." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                item: { type: "reasoning", id: "rs_raw", encrypted_content: null },
+              },
+              // Raw reasoning finals carry no summary index; they reconcile
+              // into the item's first block.
+              { type: "response.reasoning.done", item_id: "rs_raw", text: "Raw chain of thought." },
+              {
+                type: "response.output_item.done",
+                item: { type: "reasoning", id: "rs_raw", encrypted_content: "raw-state" },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Raw chain of thought.")
+    }),
+  )
+
+  it.effect("preserves nullable phases in the forgiving Open Responses baseline", () =>
    Effect.gen(function* () {
      const model = configure({
        apiKey: "test-key",
@@ -113,7 +375,14 @@ describe("Open Responses-compatible route", () => {
      )

      expect(prepared.body).toMatchObject({
-        input: [{ type: "message", role: "assistant", content: [{ type: "output_text", text: "Unclassified." }] }],
+        input: [
+          {
+            type: "message",
+            role: "assistant",
+            content: [{ type: "output_text", text: "Unclassified." }],
+            phase: null,
+          },
+        ],
      })
    }),
  )
@@ -187,6 +456,7 @@ describe("Open Responses-compatible route", () => {
          streamOptions: { includeObfuscation: false },
          topLogprobs: 3,
          truncation: "auto",
+          serviceTier: "provider-tier",
          allowedTools: { toolNames: ["lookup"] },
          maxToolCalls: 2,
          parallelToolCalls: false,
@@ -211,6 +481,7 @@ describe("Open Responses-compatible route", () => {
        presence_penalty: 0.2,
        frequency_penalty: -0.1,
        truncation: "auto",
+        service_tier: "provider-tier",
        tool_choice: {
          type: "allowed_tools",
          mode: "auto",
@@ -190,6 +190,21 @@ describe("OpenRouter", () => {
    }),
  )

+  it.effect("omits the prompt cache key when caching is disabled", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: OpenRouter.configure({ apiKey: "test-key" }).model("openai/gpt-4o-mini"),
+          prompt: "Hello",
+          promptCacheKey: "session_123",
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body).not.toHaveProperty("prompt_cache_key")
+    }),
+  )
+
  it.effect("filters invalid known OpenRouter options while preserving extensions", () =>
    Effect.gen(function* () {
      const invalid: Record<string, unknown> = {
@@ -64,7 +64,7 @@ const targets: ReadonlyArray<{
    id: "xai",
    name: "xAI Grok 4.5",
    provider: "xai",
-    protocol: "openai-responses",
+    protocol: "xai-responses",
    requires: "XAI_API_KEY",
    filename: "verification.pdf",
    maxTokens: 40,
@@ -0,0 +1,133 @@
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMEvent } from "../../src/index.js"
+import { XAI } from "../../src/providers.js"
+import { OpenResponses } from "../../src/protocols/open-responses.js"
+import { OpenAIResponses } from "../../src/protocols/openai-responses.js"
+import { XAIResponses } from "../../src/protocols/xai-responses.js"
+import { LLMClient } from "../../src/route.js"
+import { compileRequest } from "../../src/route/client.js"
+import { it } from "../lib/effect.js"
+import { fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+const model = XAI.configure({ apiKey: "test", baseURL: "https://api.x.ai/v1" }).responses("grok-4.6")
+
+describe("xAI Responses route", () => {
+  it.effect("extends the Open Responses baseline directly", () =>
+    Effect.gen(function* () {
+      expect(XAIResponses.protocol.body).toBe(OpenResponses.protocol.body)
+      expect(XAIResponses.protocol.body).not.toBe(OpenAIResponses.protocol.body)
+
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Hello" }))
+      expect(prepared.protocol).toBe("xai-responses")
+      expect(prepared.body.store).toBe(false)
+      expect(prepared.body.include).toEqual(["reasoning.encrypted_content"])
+    }),
+  )
+
+  it.effect("allows callers to opt out of encrypted reasoning", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Hello", providerOptions: { include: [] } }))
+
+      expect(prepared.body.store).toBe(false)
+      expect(prepared.body.include).toBeUndefined()
+    }),
+  )
+
+  it.effect("parses xAI reasoning summaries", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think" })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                item: { type: "reasoning", id: "reasoning_1" },
+              },
+              // Grok streams reasoning with the standard summary event name.
+              {
+                type: "response.reasoning_summary_text.delta",
+                item_id: "reasoning_1",
+                summary_index: 0,
+                delta: "Considering.",
+              },
+              {
+                type: "response.output_item.done",
+                item: { type: "reasoning", id: "reasoning_1", encrypted_content: "opaque" },
+              },
+              { type: "response.completed", response: { id: "response_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content.find((part) => part.type === "reasoning")).toMatchObject({
+        type: "reasoning",
+        text: "Considering.",
+        providerMetadata: { xai: { itemId: "reasoning_1", reasoningEncryptedContent: "opaque" } },
+      })
+    }),
+  )
+
+  it.effect("routes xAI reasoning summaries by output index", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think" })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                output_index: 3,
+                item: { type: "reasoning", id: "reasoning_1" },
+              },
+              {
+                type: "response.reasoning_summary_text.delta",
+                output_index: 3,
+                item_id: "wrong_reasoning",
+                summary_index: 0,
+                delta: "Considering.",
+              },
+              {
+                type: "response.output_item.done",
+                output_index: 3,
+                item: { type: "reasoning", id: "reasoning_1", encrypted_content: "opaque" },
+              },
+              { type: "response.completed", response: { id: "response_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Considering.")
+      expect(response.message.content.find((part) => part.type === "reasoning")).toMatchObject({
+        providerMetadata: { xai: { itemId: "reasoning_1", reasoningEncryptedContent: "opaque" } },
+      })
+    }),
+  )
+
+  it.effect("parses xAI hosted tool items", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Search X" })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.done",
+                item: { type: "x_search_call", id: "x_search_1", status: "completed", action: { query: "news" } },
+              },
+              { type: "response.completed", response: { id: "response_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.find(LLMEvent.is.toolCall)).toMatchObject({
+        id: "x_search_1",
+        name: "x_search",
+        input: { query: "news" },
+        providerExecuted: true,
+      })
+    }),
+  )
+})
@@ -78,6 +78,31 @@ describe("Z.ai Images", () => {
    ),
  )

+  it.effect("sanitizes unpaired surrogates in outbound image requests", () =>
+    Image.generate({
+      model: ZAI.configure({ apiKey: "test", http: { body: { metadata: { source: "default\uDC00" } } } }).image("model"),
+      prompt: "A red circle \uD800 on a white background \u{1F600}",
+    }).pipe(
+      Effect.provide(
+        ImageClient.layer.pipe(
+          Layer.provide(
+            dynamicResponse((input) => {
+              expect(JSON.parse(input.text)).toMatchObject({
+                prompt: "A red circle \uFFFD on a white background \u{1F600}",
+                metadata: { source: "default\uFFFD" },
+              })
+              return Effect.succeed(
+                input.respond(JSON.stringify({ data: [{ url: "https://example.test/image.jpg" }] }), {
+                  headers: { "content-type": "application/json" },
+                }),
+              )
+            }),
+          ),
+        ),
+      ),
+    ),
+  )
+
  it.effect("lets raw native options override aliases", () =>
    Image.generate({
      model: ZAI.configure({ apiKey: "test" }).image("model"),
@@ -101,7 +101,7 @@ describe("LLMClient tools", () => {
      const messages = Reflect.get(second, "messages")
      const tools = Reflect.get(second, "tools")

-      expect(Reflect.get(second, "max_tokens")).toBe(50)
+      expect(Reflect.get(second, "max_completion_tokens")).toBe(50)
      expect(Reflect.get(second, "tool_choice")).toBe("auto")
      expect(tools).toHaveLength(1)
      expect(
@@ -191,7 +191,7 @@ describe("LLMClient tools", () => {
        success: Schema.String,
        execute: () => Effect.succeed("hello"),
      })
-      const providerMetadata = { google: { functionCallId: "provider_call" } }
+      const providerMetadata = { google: { thoughtSignature: "provider_sig" } }
      const dispatched = yield* ToolRuntime.dispatch(
        { tool },
        LLMEvent.toolCall({ id: "call_1", name: "tool", input: {}, providerMetadata }),
@@ -50,32 +50,7 @@ describe("tool schema projections", () => {
    })
  })

-  test("openai keeps one flat object top-level schema", () => {
-    expect(
-      ToolSchemaProjection.openAI({
-        anyOf: [
-          {
-            type: "object",
-            properties: {
-              path: { type: "string" },
-              maybe: { anyOf: [{ type: "string" }, { type: "null" }] },
-            },
-          },
-          { type: "object", properties: { resource: { type: "string" } } },
-        ],
-      }),
-    ).toEqual({
-      type: "object",
-      properties: {
-        path: { type: "string" },
-        maybe: { type: "string" },
-        resource: { type: "string" },
-      },
-      additionalProperties: false,
-    })
-  })
-
-  it.effect("applies model compatibility before protocol projection", () =>
+  it.effect("applies model compatibility without changing schema semantics", () =>
    Effect.gen(function* () {
      const model = OpenAIChat.route
        .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
@@ -107,11 +82,15 @@ describe("tool schema projections", () => {

      expect(prepared.body.tools?.[0]?.function.parameters).toEqual({
        type: "object",
-        properties: {
-          tuple: { type: "array", items: { anyOf: [{ type: "string" }, { type: "number" }] } },
-          linked: { $ref: "#/$defs/Linked" },
-        },
-        additionalProperties: false,
+        anyOf: [
+          {
+            type: "object",
+            properties: {
+              tuple: { type: "array", items: { anyOf: [{ type: "string" }, { type: "number" }] } },
+              linked: { $ref: "#/$defs/Linked" },
+            },
+          },
+        ],
      })
    }),
  )
@@ -23,9 +23,11 @@ describe("ToolStream", () => {

      expect(first.events).toEqual([
        { type: "tool-input-start", id: "call_1", name: "lookup" },
-        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"' },
+        { type: "tool-input-delta", id: "call_1", name: "lookup", text: '{"query"', input: {} },
+      ])
+      expect(second.events).toEqual([
+        { type: "tool-input-delta", id: "call_1", name: "lookup", text: ':"weather"}', input: { query: "weather" } },
      ])
-      expect(second.events).toEqual([{ type: "tool-input-delta", id: "call_1", name: "lookup", text: ':"weather"}' }])
      expect(finished).toEqual({
        tools: {},
        events: [
@@ -36,6 +38,45 @@ describe("ToolStream", () => {
    }),
  )

+  it.effect("exposes cumulative partial string values", () =>
+    Effect.gen(function* () {
+      const result = ToolStream.appendOrStart(
+        ADAPTER,
+        ToolStream.empty<number>(),
+        0,
+        { id: "call_1", name: "lookup", text: '{"query":"wea' },
+        "missing tool",
+      )
+      if (ToolStream.isError(result)) return yield* result
+
+      expect(result.events.at(-1)).toEqual({
+        type: "tool-input-delta",
+        id: "call_1",
+        name: "lookup",
+        text: '{"query":"wea',
+        input: { query: "wea" },
+      })
+    }),
+  )
+
+  it.effect("defaults partial input to an empty object when the accumulated value cannot be parsed", () =>
+    Effect.gen(function* () {
+      const result = ToolStream.appendOrStart(
+        ADAPTER,
+        ToolStream.empty<number>(),
+        0,
+        { id: "call_1", name: "lookup", text: "x" },
+        "missing tool",
+      )
+      if (ToolStream.isError(result)) return yield* result
+
+      expect(result.events).toEqual([
+        { type: "tool-input-start", id: "call_1", name: "lookup" },
+        { type: "tool-input-delta", id: "call_1", name: "lookup", text: "x", input: {} },
+      ])
+    }),
+  )
+
  it.effect("keeps accumulated identity when later deltas contain empty strings", () =>
    Effect.gen(function* () {
      const first = ToolStream.appendOrStart(
@@ -91,7 +132,7 @@ describe("ToolStream", () => {
    }),
  )

-  it.effect("finalizes malformed local input as a non-executable tool error", () =>
+  it.effect("finalizes incomplete local input using the partial JSON parser", () =>
    Effect.gen(function* () {
      const tools = ToolStream.start(ToolStream.empty<string>(), "item_1", {
        id: "call_1",
@@ -103,18 +144,46 @@ describe("ToolStream", () => {
      expect(finished).toEqual({
        tools: {},
        events: [
-          {
-            type: "tool-input-error",
-            id: "call_1",
-            name: "lookup",
-            raw: '{"query":"partial',
-          },
+          { type: "tool-input-end", id: "call_1", name: "lookup" },
+          { type: "tool-call", id: "call_1", name: "lookup", input: { query: "partial" } },
        ],
      })
    }),
  )

-  it.effect("preserves valid siblings when one parallel input is malformed", () =>
+  it.effect("repairs malformed string escapes in final local input", () =>
+    Effect.gen(function* () {
+      const tools = ToolStream.start(ToolStream.empty<string>(), "item_1", {
+        id: "call_1",
+        name: "lookup",
+        input: '{"path":"A\\H","text":"first\tsecond"}',
+      })
+      const finished = yield* ToolStream.finish(ADAPTER, tools, "item_1")
+
+      expect(finished.events).toEqual([
+        { type: "tool-input-end", id: "call_1", name: "lookup" },
+        { type: "tool-call", id: "call_1", name: "lookup", input: { path: "A\\H", text: "first\tsecond" } },
+      ])
+    }),
+  )
+
+  it.effect("defaults unrecoverable local input to an empty object", () =>
+    Effect.gen(function* () {
+      const tools = ToolStream.start(ToolStream.empty<string>(), "item_1", {
+        id: "call_1",
+        name: "lookup",
+        input: "invalid",
+      })
+      const finished = yield* ToolStream.finish(ADAPTER, tools, "item_1")
+
+      expect(finished.events).toEqual([
+        { type: "tool-input-end", id: "call_1", name: "lookup" },
+        { type: "tool-call", id: "call_1", name: "lookup", input: {} },
+      ])
+    }),
+  )
+
+  it.effect("recovers incomplete input alongside valid parallel tool calls", () =>
    Effect.gen(function* () {
      const valid = ToolStream.start(ToolStream.empty<number>(), 0, {
        id: "call_valid",
@@ -133,12 +202,8 @@ describe("ToolStream", () => {
        events: [
          { type: "tool-input-end", id: "call_valid", name: "lookup" },
          { type: "tool-call", id: "call_valid", name: "lookup", input: { query: "weather" } },
-          {
-            type: "tool-input-error",
-            id: "call_invalid",
-            name: "lookup",
-            raw: '{"query":"partial',
-          },
+          { type: "tool-input-end", id: "call_invalid", name: "lookup" },
+          { type: "tool-call", id: "call_invalid", name: "lookup", input: { query: "partial" } },
        ],
      })
    }),
@@ -19,6 +19,12 @@

 - Always prefer `createStore` over multiple `createSignal` calls

+## Typography
+
+- Use `--line-height-compact` (`16px`) for `13px` compact UI text and `--line-height-base` (`20px`) for body text.
+- Do not use `leading-none`, `line-height: 1`, or a `13px` line height for normal text. Inter descenders clip inside truncation and overflow containers.
+- Keep control and row heights explicit. Fix font metrics directly rather than using transforms, negative margins, or clip-padding compensation.
+
 ## Localization

 - NEVER hardcode user-visible English strings in production code. ALWAYS use an i18n key for visible copy, placeholders, accessible labels, tooltips, menus, dialogs, toasts, empty states, and displayed errors.
@@ -259,6 +259,34 @@ export function event(
  return makeEvent(type, data)
 }

+export function compactionStarted(data: Extract<OpenCodeEvent, { type: "session.compaction.started" }>["data"]) {
+  return makeEvent("session.compaction.started", data)
+}
+
+export function compactionDelta(data: Extract<OpenCodeEvent, { type: "session.compaction.delta" }>["data"]) {
+  return makeEvent("session.compaction.delta", data)
+}
+
+export function compactionEnded(data: Extract<OpenCodeEvent, { type: "session.compaction.ended" }>["data"]) {
+  return makeEvent("session.compaction.ended", data)
+}
+
+export function compactionFailed(data: Extract<OpenCodeEvent, { type: "session.compaction.failed" }>["data"]) {
+  return makeEvent("session.compaction.failed", data)
+}
+
+export function toolInputStarted(data: Extract<OpenCodeEvent, { type: "session.tool.input.started" }>["data"]) {
+  return makeEvent("session.tool.input.started", data)
+}
+
+export function toolInputEnded(data: Extract<OpenCodeEvent, { type: "session.tool.input.ended" }>["data"]) {
+  return makeEvent("session.tool.input.ended", data)
+}
+
+export function toolCalled(data: Extract<OpenCodeEvent, { type: "session.tool.called" }>["data"]) {
+  return makeEvent("session.tool.called", data)
+}
+
 export function validateTimelineEvent(input: unknown): OpenCodeEvent {
  if (!input || typeof input !== "object") throw new Error("Timeline event must be an object")
  if (!("type" in input) || typeof input.type !== "string") throw new Error("Timeline event requires a type")
@@ -871,8 +899,26 @@ function provider() {
        name: "OpenCode",
        models: { "claude-opus-4-6": { id: "claude-opus-4-6", name: "Claude Opus 4.6", limit: { context: 200_000 } } },
      },
+      {
+        id: "company-gateway",
+        name: "Company Gateway",
+        models: {
+          "fast-nano": {
+            id: "fast-nano",
+            api: { id: "openai/gpt-5.4-nano" },
+            name: "GPT-5.4 nano",
+            limit: { context: 128_000 },
+          },
+          "long-context": {
+            id: "long-context",
+            api: { id: "company/long-context" },
+            name: "Company Gateway Extra Long Context Model for Narrow Timeline Layouts",
+            limit: { context: 128_000 },
+          },
+        },
+      },
    ],
-    connected: ["opencode"],
+    connected: ["opencode", "company-gateway"],
    default: { providerID: "opencode", modelID: "claude-opus-4-6" },
  }
 }
@@ -1,4 +1,5 @@
 import { expect, test } from "@playwright/test"
+import { createTwoFilesPatch } from "diff"
 import {
  defineVisualRegions,
  reportVisualStability,
@@ -13,10 +14,63 @@ import {
  setupTimeline,
  shell,
  textPart,
+  toolPart,
  userMessage,
  type TimelineMessage,
 } from "./fixture"

+test("follows an expanded patch that arrives as the user reaches the bottom", async ({ page }) => {
+  const toolID = "prt_bottom_follow_patch"
+  const input = { patchText: "Update src/edit.ts" }
+  const timeline = await setupTimeline(page, {
+    messages: [
+      ...history(20),
+      userMessage(),
+      assistantMessage([textPart("prt_bottom_follow_text", "Working")], { completed: false }),
+    ],
+    settings: { editToolPartsExpanded: true },
+    reducedMotion: true,
+  })
+  const scroller = page.locator(".scroll-view__viewport", { has: page.locator("[data-timeline-row]") })
+  await scroller.evaluate((element) => {
+    element.scrollTop = Math.max(0, element.scrollHeight - element.clientHeight - 300)
+    element.dispatchEvent(new WheelEvent("wheel", { bubbles: true, cancelable: true, deltaY: 300 }))
+    element.scrollTop = element.scrollHeight
+  })
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(1)
+
+  await timeline.send(partUpdated(toolPart(toolID, "patch", "running", input)))
+  await timeline.send(
+    partUpdated(
+      toolPart(toolID, "patch", "completed", input, {
+        metadata: {
+          files: [
+            {
+              file: "src/edit.ts",
+              status: "modified",
+              patch: createTwoFilesPatch(
+                "a/src/edit.ts",
+                "b/src/edit.ts",
+                Array.from({ length: 40 }, (_, index) => `export const value${index} = ${index}\n`).join(""),
+                Array.from({ length: 40 }, (_, index) => `export const value${index} = ${index + 1}\n`).join(""),
+              ),
+              additions: 40,
+              deletions: 40,
+            },
+          ],
+        },
+      }),
+    ),
+  )
+
+  await timeline.waitForPart(toolID)
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(1)
+})
+
 test("does not reverse visible rows when the user wheels during shell remeasurement", async ({ page }, testInfo) => {
  const shellID = "prt_wheel_01_shell"
  const followingID = "prt_wheel_02_following"
@@ -53,6 +53,37 @@ benchmark.describe("performance: first navigation paint", () => {
    expect(result.summary.unknownSamples).toBe(0)
  })

+  benchmark("opens a session from the new session page without a blank frame", async ({ page, report }) => {
+    await mockStressTimeline(page)
+    await installTimelineSettings(page)
+    await installStressSessionTabs(page, { draftID })
+    await page.goto("/")
+
+    const draftHref = stressDraftHref(draftID)
+    const draftTab = page.locator(`[data-slot="titlebar-tabs"] a[href="${draftHref}"]`)
+    await expect(draftTab).toHaveCount(1)
+    await draftTab.click()
+    await expect(page.locator('[data-component="new-session"]')).toBeVisible()
+
+    const href = stressSessionHref(fixture.targetID)
+    const sessionTab = page.locator(`[data-slot="titlebar-tabs"] a[href="${href}"]`)
+    await expect(sessionTab).toHaveCount(1)
+    const result = await measureFirstNavigation(page, {
+      href,
+      destinationPath: href,
+      sourceSelector: '[data-component="new-session"]',
+      destinationSelector: messageSelector(fixture.expected.targetMessageIDs.at(-1)!),
+      contentSelector,
+      navigate: async () => {
+        await sessionTab.click()
+        await expectSessionTitle(page, fixture.expected.targetTitle)
+      },
+    })
+    report(result)
+    expect(result.summary.blankSamples).toBe(0)
+    expect(result.summary.unknownSamples).toBe(0)
+  })
+
  benchmark("opens a child session without a blank frame", async ({ page, report }) => {
    await setup(page)
    const href = stressSessionHref(fixture.childID)
@@ -0,0 +1,106 @@
+import { expect, test } from "@playwright/test"
+import { fixture, pageMessages } from "../smoke/session-timeline.fixture"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+test("renames, exports, and deletes a home session from its context menu", async ({ page }) => {
+  const sessions = fixture.sessions.map((session) => ({ ...session }))
+  await mockOpenCodeServer(page, {
+    sessions,
+    provider: fixture.provider,
+    directory: fixture.directory,
+    project: fixture.project,
+    pageMessages,
+  })
+  await page.route("**/api/session/*/rename", async (route) => {
+    const sessionID = new URL(route.request().url()).pathname.split("/").at(-2)
+    const session = sessions.find((item) => item.id === sessionID)
+    const payload: unknown = route.request().postDataJSON()
+    if (!payload || typeof payload !== "object" || !("title" in payload) || typeof payload.title !== "string")
+      throw new Error("Invalid rename payload")
+    if (session) session.title = payload.title
+    await route.fulfill({ status: 204, headers: { "access-control-allow-origin": "*" } })
+  })
+  await page.addInitScript((directory) => {
+    localStorage.setItem(
+      "opencode.global.dat:server",
+      JSON.stringify({
+        projects: { local: [{ worktree: directory, expanded: true }] },
+        lastProject: { local: directory },
+      }),
+    )
+  }, fixture.directory)
+
+  await page.goto("/")
+  const row = page.locator('[data-component="home-session-row"]').filter({ hasText: fixture.expected.targetTitle })
+  await expect(row).toBeVisible()
+  const container = page.locator(`[data-component="home-session-row-container"][data-session-id="${fixture.targetID}"]`)
+  const titleBox = await container.locator('[data-component="home-session-title"]').boundingBox()
+  const avatarBox = await container.locator('[data-component="project-avatar-v2"]').boundingBox()
+  await expect(container.getByRole("button", { name: "More options" })).toHaveCount(0)
+
+  await row.focus()
+  await row.press("Shift+F10")
+  await expect(page.getByRole("menuitem", { name: "Rename" })).toBeVisible()
+  await page.keyboard.press("Escape")
+  await expect(page.getByRole("menuitem", { name: "Rename" })).toBeHidden()
+  await expect(row).toBeFocused()
+
+  const rowBox = await row.boundingBox()
+  await row.click({ button: "right", position: { x: 48, y: 12 } })
+  await expect(page).toHaveURL("/")
+  await expect(page.getByRole("menuitem", { name: "Rename" })).toBeVisible()
+  await expect(page.getByRole("menuitem", { name: "Export..." })).toBeVisible()
+  await expect(page.getByRole("menuitem", { name: "Delete..." })).toBeVisible()
+  const menuBox = await page.locator('[data-component="menu-v2-content"]').boundingBox()
+  expect(Math.abs((menuBox?.x ?? 0) - (rowBox?.x ?? 0) - 48)).toBeLessThan(4)
+
+  await page.getByRole("menuitem", { name: "Rename" }).click()
+  const title = page.locator('[data-component="home-session-rename"]')
+  await expect(title).toBeFocused()
+  await expect(title).toHaveValue(fixture.expected.targetTitle)
+  const editorBox = await title.boundingBox()
+  const editingAvatarBox = await container.locator('[data-component="project-avatar-v2"]').boundingBox()
+  expect(editorBox?.x).toBe(titleBox?.x)
+  expect(editingAvatarBox).toEqual(avatarBox)
+  expect(
+    await title.evaluate((element) => ({
+      outline: getComputedStyle(element).outlineStyle,
+      shadow: getComputedStyle(element).boxShadow,
+    })),
+  ).toEqual({ outline: "none", shadow: "none" })
+  expect(await container.evaluate((element) => getComputedStyle(element).outlineStyle)).toBe("none")
+  await title.fill("Renamed from Home")
+  const renamed = page.waitForRequest(
+    (request) => request.method() === "POST" && new URL(request.url()).pathname.endsWith("/rename"),
+  )
+  await title.press("Enter")
+  expect((await renamed).postDataJSON()).toEqual({ title: "Renamed from Home" })
+  let renamedRow = page.locator('[data-component="home-session-row"]').filter({ hasText: "Renamed from Home" })
+  await expect(renamedRow).toBeVisible()
+
+  await renamedRow.click()
+  await expect(page).toHaveURL(new RegExp(`/session/${fixture.targetID}$`))
+  await expect(page.locator('[data-slot="titlebar-tabs"] a').filter({ hasText: "Renamed from Home" })).toBeVisible()
+  await page.getByRole("button", { name: "Home" }).click()
+  await expect(page).toHaveURL("/")
+  renamedRow = page.locator('[data-component="home-session-row"]').filter({ hasText: "Renamed from Home" })
+  await expect(renamedRow).toBeVisible()
+
+  await renamedRow.click({ button: "right" })
+  const download = page.waitForEvent("download")
+  const exportItem = page.getByRole("menuitem", { name: "Export..." })
+  await exportItem.click()
+  expect((await download).suggestedFilename()).toBe("renamed-from-home.json")
+  await expect(exportItem).toBeHidden()
+
+  await renamedRow.click({ button: "right" })
+  await page.getByRole("menuitem", { name: "Delete..." }).click()
+  const dialog = page.getByRole("dialog")
+  await expect(dialog).toContainText('Delete session "Renamed from Home"?')
+  const removed = page.waitForRequest(
+    (request) => request.method() === "DELETE" && new URL(request.url()).pathname.endsWith(`/${fixture.targetID}`),
+  )
+  await dialog.getByRole("button", { name: "Delete session" }).click()
+  await removed
+  await expect(renamedRow).toBeHidden()
+})
@@ -0,0 +1,53 @@
+import { expect, test } from "@playwright/test"
+import { mockOpenCodeServer } from "../utils/mock-server"
+import { expectAppVisible } from "../utils/waits"
+
+const draftID = "draft_new_session_workspace_branch"
+const directory = "C:/OpenCode/WorkspaceBranch"
+const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
+
+test("selects a base branch for a new workspace", async ({ page }) => {
+  await mockOpenCodeServer(page, {
+    directory,
+    project: {
+      id: "proj_new_session_workspace_branch",
+      worktree: directory,
+      vcs: "git",
+      name: "workspace-branch",
+      time: { created: 1700000000000, updated: 1700000000000 },
+      sandboxes: [],
+    },
+    provider: { all: [], connected: [], default: {} },
+    sessions: [],
+    pageMessages: () => ({ items: [] }),
+    vcsBranches: ["feature/api", "main", "origin/release"],
+  })
+  await page.addInitScript(
+    ({ directory, draftID, server }) => {
+      localStorage.setItem(
+        "opencode.global.dat:server",
+        JSON.stringify({
+          projects: { local: [{ worktree: directory, expanded: true }] },
+          lastProject: { local: directory },
+        }),
+      )
+      localStorage.setItem(
+        "opencode.window.browser.dat:tabs",
+        JSON.stringify([{ type: "draft", draftID, server, directory }]),
+      )
+    },
+    { directory, draftID, server },
+  )
+
+  await page.goto(`/new-session?draftId=${draftID}`)
+  await expectAppVisible(page.locator('[data-component="composer-editor"]'))
+  await page.getByRole("button", { name: "Local", exact: true }).click()
+  await page.getByRole("menuitem", { name: "New workspace", exact: true }).click()
+  await page.getByRole("button", { name: "from main", exact: true }).click()
+  await page.getByRole("menuitemradio", { name: "feature/api", exact: true }).click()
+
+  const selected = page.getByRole("button", { name: "from feature/api", exact: true })
+  await expect(selected).toBeVisible()
+  await selected.click()
+  await expect(page.getByRole("menuitemradio", { name: "feature/api", exact: true })).toBeChecked()
+})
@@ -13,9 +13,14 @@ const sessionB = session("ses_server_b", directoryB, "Server B session")

 test("session settings use the remote server context", async ({ page }) => {
  const permissionRequests: string[] = []
+  const permissionResponses: PermissionResponse[] = []
  await installSseTransport(page, { server: serverA })
  await installSseTransport(page, { server: serverB })
-  await mockServers(page, permissionRequests)
+  // Server A has no tab and is never visited: a pending request there proves
+  // one toggle sweeps every connected server, not just the focused one.
+  await mockServers(page, permissionRequests, permissionResponses, {
+    pending: { [serverA]: [pendingPermission("permission-pending-a", sessionA.id)] },
+  })
  await configureServers(page)

  await page.goto(`/server/${base64Encode(serverB)}/session/${sessionB.id}`)
@@ -38,7 +43,17 @@ test("session settings use the remote server context", async ({ page }) => {
      }),
    )
    .toBe(true)
-  expect(permissionRequests.every((request) => new URL(request).origin === serverB)).toBe(true)
+  await expect
+    .poll(() => permissionResponses)
+    .toEqual([
+      {
+        origin: serverA,
+        directory: undefined,
+        sessionID: sessionA.id,
+        permissionID: "permission-pending-a",
+        body: { reply: "once" },
+      },
+    ])

  await dialog.getByRole("tab", { name: "Models" }).click()
  await expect(dialog.getByRole("switch", { name: "Server B Model" })).toBeEnabled()
@@ -143,6 +158,99 @@ test("auto-accept responds for an unfocused server session", async ({ page }) =>
    ])
 })

+test("auto-accept sweeps again after a reconnect", async ({ page }) => {
+  const permissionRequests: string[] = []
+  const permissionResponses: PermissionResponse[] = []
+  const pendingA: MockPermission[] = []
+  const listFailures: Record<string, number> = {}
+  const sessionGets: string[] = []
+  await installSseTransport(page, { server: serverB })
+  const transport = await installSseTransport(page, { server: serverA, retry: 20 })
+  await mockServers(page, permissionRequests, permissionResponses, {
+    pending: { [serverA]: pendingA },
+    listFailures,
+    sessionGets,
+  })
+  await configureServers(page, [{ type: "session", server: serverA, sessionId: sessionA.id }])
+
+  await page.goto(`/server/${base64Encode(serverA)}/session/${sessionA.id}`)
+  await expect(page.getByRole("heading", { name: sessionA.title, exact: true })).toBeVisible()
+  const first = await transport.waitForConnection()
+
+  await page.keyboard.press("Control+,")
+  const autoAccept = page.locator(".settings-dialog").locator('[data-action="settings-auto-accept-permissions"]')
+  await autoAccept.locator('[data-slot="switch-control"]').click()
+  await expect(autoAccept.getByRole("switch")).toBeChecked()
+  await expect
+    .poll(() =>
+      permissionRequests.some((request) => {
+        const url = new URL(request)
+        return url.origin === serverA && url.searchParams.get("location[directory]") === directoryA
+      }),
+    )
+    .toBe(true)
+  await page.keyboard.press("Escape")
+
+  // This request is asked while the client is disconnected, so it is never
+  // delivered as an event and only a reconnect sweep can find it. The first
+  // listing after the reconnect fails, so only the bounded sweep retry can
+  // deliver the reply.
+  pendingA.push(pendingPermission("permission-offline-a", sessionA.id))
+  listFailures[serverA] = 1
+  const syncsBeforeReconnect = sessionGets.length
+  await transport.disconnect()
+  await transport.waitForConnection({ after: first.id })
+
+  await expect
+    .poll(() => permissionResponses)
+    .toEqual([
+      {
+        origin: serverA,
+        directory: undefined,
+        sessionID: sessionA.id,
+        permissionID: "permission-offline-a",
+        body: { reply: "once" },
+      },
+    ])
+  // The reconnect sweep must resync active sessions instead of trusting
+  // cached locations, since another client may have moved them meanwhile.
+  expect(sessionGets.slice(syncsBeforeReconnect)).toContain(sessionA.id)
+})
+
+test("auto-accept approves a request discovered by opening a session", async ({ page }) => {
+  const permissionRequests: string[] = []
+  const permissionResponses: PermissionResponse[] = []
+  await installSseTransport(page, { server: serverA })
+  await installSseTransport(page, { server: serverB })
+  // The request is only served from the per-session permission list, so it
+  // reaches the client through the store sync when the session view opens,
+  // never through a location sweep or an event.
+  await mockServers(page, permissionRequests, permissionResponses, {
+    sessionPending: { [sessionA.id]: [pendingPermission("permission-synced-a", sessionA.id)] },
+  })
+  await configureServers(page, [{ type: "session", server: serverA, sessionId: sessionA.id }])
+
+  await page.goto(`/server/${base64Encode(serverA)}/session/${sessionA.id}`)
+  await expect(page.getByRole("heading", { name: sessionA.title, exact: true })).toBeVisible()
+
+  await page.keyboard.press("Control+,")
+  const autoAccept = page.locator(".settings-dialog").locator('[data-action="settings-auto-accept-permissions"]')
+  await autoAccept.locator('[data-slot="switch-control"]').click()
+  await expect(autoAccept.getByRole("switch")).toBeChecked()
+
+  await expect
+    .poll(() => permissionResponses)
+    .toEqual([
+      {
+        origin: serverA,
+        directory: undefined,
+        sessionID: sessionA.id,
+        permissionID: "permission-synced-a",
+        body: { reply: "once" },
+      },
+    ])
+})
+
 type PermissionResponse = {
  origin: string
  directory?: string
@@ -151,6 +259,19 @@ type PermissionResponse = {
  body: unknown
 }

+type MockPermission = {
+  id: string
+  sessionID: string
+  action: string
+  resources: string[]
+  metadata: Record<string, unknown>
+  save: unknown[]
+}
+
+function pendingPermission(id: string, sessionID: string): MockPermission {
+  return { id, sessionID, action: "shell", resources: ["git status"], metadata: {}, save: [] }
+}
+
 async function configureServers(page: Page, tabs: { type: "session"; server: string; sessionId: string }[] = []) {
  await page.addInitScript(
    ({ serverB, tabs }) => {
@@ -161,7 +282,23 @@ async function configureServers(page: Page, tabs: { type: "session"; server: str
  )
 }

-async function mockServers(page: Page, permissionRequests: string[], permissionResponses: PermissionResponse[] = []) {
+type MockServerOptions = {
+  // Pending requests served from /api/permission/request, keyed by origin.
+  pending?: Record<string, MockPermission[]>
+  // Pending requests served from /api/session/:id/permission, keyed by session ID.
+  sessionPending?: Record<string, MockPermission[]>
+  // Counts of /api/permission/request calls to fail with a 500, keyed by origin.
+  listFailures?: Record<string, number>
+  // Records /api/session/:id GETs so tests can assert session resyncs.
+  sessionGets?: string[]
+}
+
+async function mockServers(
+  page: Page,
+  permissionRequests: string[],
+  permissionResponses: PermissionResponse[] = [],
+  options: MockServerOptions = {},
+) {
  await page.route("**/*", async (route) => {
    const url = new URL(route.request().url())
    if (url.origin !== serverA && url.origin !== serverB) return route.fallback()
@@ -178,8 +315,12 @@ async function mockServers(page: Page, permissionRequests: string[], permissionR
        permissionID: response[2]!,
        body: route.request().postDataJSON(),
      })
-      return json(route, true)
+      // The generated client requires exactly 204 for a successful reply.
+      return route.fulfill({ status: 204, headers: { "access-control-allow-origin": "*" } })
    }
+    const sessionPermission = url.pathname.match(/^\/api\/session\/([^/]+)\/permission$/)
+    if (route.request().method() === "GET" && sessionPermission)
+      return json(route, { data: options.sessionPending?.[sessionPermission[1]!] ?? [] })
    if (requestDirectory && requestDirectory !== directory) return json(route, { name: "InvalidDirectory" }, 500)
    if (url.pathname === "/api/provider")
      return json(route, {
@@ -197,7 +338,12 @@ async function mockServers(page: Page, permissionRequests: string[], permissionR
    if (url.pathname === "/api/agent") return json(route, { location: { directory }, data: [] })
    if (url.pathname === "/api/permission/request") {
      permissionRequests.push(url.toString())
-      return json(route, { location: { directory }, data: [] })
+      const failures = options.listFailures?.[url.origin] ?? 0
+      if (failures > 0) {
+        options.listFailures![url.origin] = failures - 1
+        return json(route, { name: "Internal" }, 500)
+      }
+      return json(route, { location: { directory }, data: options.pending?.[url.origin] ?? [] })
    }
    if (["/api/command", "/api/reference", "/api/question/request"].includes(url.pathname))
      return json(route, { location: { directory }, data: [] })
@@ -219,9 +365,13 @@ async function mockServers(page: Page, permissionRequests: string[], permissionR
      return json(route, { id: remote ? sessionB.projectID : "project-server-a", directory, canonical: directory })
    if (url.pathname === "/api/session")
      return json(route, { data: sessions.map((session) => currentSession(session)), cursor: {} })
-    if (url.pathname === "/api/session/active") return json(route, { data: {} })
+    if (url.pathname === "/api/session/active")
+      return json(route, { data: Object.fromEntries(sessions.map((session) => [session.id, { type: "running" }])) })
    const currentSessionInfo = sessions.find((session) => url.pathname === `/api/session/${session.id}`)
-    if (currentSessionInfo) return json(route, { data: currentSession(currentSessionInfo) })
+    if (currentSessionInfo) {
+      options.sessionGets?.push(currentSessionInfo.id)
+      return json(route, { data: currentSession(currentSessionInfo) })
+    }
    if (sessions.some((session) => url.pathname === `/api/session/${session.id}/message`))
      return json(route, { data: [], cursor: {} })
    if (sessions.some((session) => url.pathname === `/api/session/${session.id}/inbox`))
@@ -45,21 +45,17 @@ test("opens the comment editor for a line number range", async ({ page }) => {
  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on lines 1-3")
 })

-test("shows a comment button when a line number is hovered", async ({ page }) => {
+test("shows a comment button when a diff line is hovered", async ({ page }) => {
  const review = page.locator('[data-component="session-review"]')
-  const lineNumber = review.locator('[data-column-number="1"]').last()
-  await expectAppVisible(lineNumber)
+  const line = review.getByText("export const first = 1", { exact: true })
+  await expectAppVisible(line)

-  const comment = review.getByRole("button", { name: "Comment", exact: true })
-  await expect(async () => {
-    await lineNumber.hover()
-    await expect(lineNumber).toHaveAttribute("data-hovered", "")
-    await expect(comment).toHaveCount(1)
-    await expect(comment).toHaveCSS("pointer-events", "auto")
-    await comment.focus()
-    await expect(comment).toBeFocused()
-  }).toPass({ timeout: 10_000 })
-  await comment.press("Enter")
+  const comment = review.getByRole("button", { name: "Comment", exact: true, includeHidden: true })
+  await expect(comment).toHaveCount(1)
+  await line.dispatchEvent("pointermove", { pointerType: "mouse", bubbles: true, composed: true })
+  await expect(comment).toBeVisible()
+  await expect(comment).toHaveCSS("pointer-events", "auto")
+  await comment.dispatchEvent("click")
  await expect(review.getByRole("textbox")).toBeVisible()
  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on line 1")
 })
@@ -18,7 +18,7 @@ const branchDiffs = [
  ),
 ]

-test("keeps the review tree and terminal sized when both panels are open", async ({ page }) => {
+test("uses side placement by default and supports the terminal across the bottom", async ({ page }) => {
  test.setTimeout(120_000)
  await page.setViewportSize({ width: 1400, height: 900 })
  await mockOpenCodeServer(page, {
@@ -27,7 +27,7 @@ test("keeps the review tree and terminal sized when both panels are open", async
      id: projectID,
      worktree: directory,
      vcs: "git",
-      name: "review-terminal-stacked",
+      name: "review-terminal-bottom",
      time: { created: 1700000000000, updated: 1700000000000 },
      sandboxes: [],
    },
@@ -45,7 +45,7 @@ test("keeps the review tree and terminal sized when both panels are open", async
    sessions: [
      {
        id: sessionID,
-        slug: "review-terminal-stacked",
+        slug: "review-terminal-bottom",
        projectID,
        directory,
        title,
@@ -138,7 +138,22 @@ test("keeps the review tree and terminal sized when both panels are open", async
  await page.keyboard.press("Control+Backquote")
  await expect(page.locator("#terminal-panel")).toBeVisible()
  await expectTree(page, 2_773, "action.yml")
-  await expectStackGeometry(page)
+  await expectSideGeometry(page)
+
+  await page.evaluate(() => {
+    const settings = JSON.parse(localStorage.getItem("settings.v3") ?? "{}")
+    localStorage.setItem(
+      "settings.v3",
+      JSON.stringify({ ...settings, general: { ...settings.general, terminalPlacement: "bottom" } }),
+    )
+  })
+  await page.reload()
+  await expectSessionReady(page, { server, sessionID, title })
+  await expect(page.locator("#review-panel")).toBeVisible()
+  await page.keyboard.press("Control+Backquote")
+  await expect(page.locator("#terminal-panel")).toBeVisible()
+  await expectTree(page, 2_773, "action.yml")
+  await expectBottomGeometry(page)
 })

 async function expectTree(page: Page, total: number, file: string) {
@@ -163,26 +178,84 @@ async function expectMountedTree(page: Page, total: number) {
  expect(state.rows).toBeLessThanOrEqual(60)
 }

-async function expectStackGeometry(page: Page) {
+async function expectSideGeometry(page: Page) {
+  await expectPanelGap(page, 8)
+  const geometry = await page.evaluate(() => {
+    const review = document.querySelector<HTMLElement>("#review-panel")!.getBoundingClientRect()
+    const terminal = document.querySelector<HTMLElement>("#terminal-panel")!.getBoundingClientRect()
+    return {
+      reviewLeft: review.left,
+      reviewRight: review.right,
+      terminalLeft: terminal.left,
+      terminalRight: terminal.right,
+      terminalTop: terminal.top,
+      terminalBottom: terminal.bottom,
+      reviewTop: review.top,
+      reviewBottom: review.bottom,
+    }
+  })
+  expect(Math.abs(geometry.terminalLeft - geometry.reviewLeft)).toBeLessThanOrEqual(1)
+  expect(Math.abs(geometry.terminalRight - geometry.reviewRight)).toBeLessThanOrEqual(1)
+  expect(geometry.terminalTop).toBeGreaterThan(geometry.reviewTop)
+  expect(geometry.terminalTop - geometry.reviewBottom).toBeGreaterThanOrEqual(7)
+  expect(geometry.terminalTop - geometry.reviewBottom).toBeLessThanOrEqual(9)
+}
+
+async function expectBottomGeometry(page: Page) {
+  await expectPanelGap(page, 8)
  const geometry = await page.evaluate(() => {
    const review = document.querySelector<HTMLElement>("#review-panel")!
    const terminal = document.querySelector<HTMLElement>("#terminal-panel")!
+    const terminalRect = terminal.getBoundingClientRect()
    const reviewParent = review.parentElement!.getBoundingClientRect()
    const terminalParent = terminal.parentElement!.getBoundingClientRect()
    const sidebar = review.querySelector<HTMLElement>('[data-slot="session-review-v2-sidebar"]')!
    return {
      review: review.getBoundingClientRect().height,
+      reviewBottom: review.getBoundingClientRect().bottom,
      reviewParent: reviewParent.height,
-      terminal: terminal.getBoundingClientRect().height,
+      terminal: terminalRect.height,
+      terminalLeft: terminalRect.left,
+      terminalRight: terminalRect.right,
+      terminalTop: terminalRect.top,
      terminalParent: terminalParent.height,
      sidebar: sidebar.getBoundingClientRect().width,
+      viewport: window.innerWidth,
    }
  })
  expect(Math.abs(geometry.review - geometry.reviewParent)).toBeLessThanOrEqual(1)
  expect(Math.abs(geometry.terminal - geometry.terminalParent)).toBeLessThanOrEqual(1)
+  expect(geometry.terminalTop - geometry.reviewBottom).toBeGreaterThanOrEqual(7)
+  expect(geometry.terminalTop - geometry.reviewBottom).toBeLessThanOrEqual(9)
+  expect(geometry.terminalLeft).toBeLessThanOrEqual(9)
+  expect(geometry.terminalRight).toBeGreaterThanOrEqual(geometry.viewport - 9)
  expect(geometry.sidebar).toBeGreaterThanOrEqual(240)
 }

+async function expectPanelGap(page: Page, expected: number) {
+  await expect
+    .poll(() => {
+      return page.evaluate(() => {
+        const review = document.querySelector<HTMLElement>("#review-panel")?.getBoundingClientRect()
+        const terminal = document.querySelector<HTMLElement>("#terminal-panel")?.getBoundingClientRect()
+        if (!review || !terminal) return Number.NEGATIVE_INFINITY
+        const gap = terminal.top - review.bottom
+        return gap
+      })
+    })
+    .toBeGreaterThanOrEqual(expected - 1)
+  await expect
+    .poll(() => {
+      return page.evaluate(() => {
+        const review = document.querySelector<HTMLElement>("#review-panel")?.getBoundingClientRect()
+        const terminal = document.querySelector<HTMLElement>("#terminal-panel")?.getBoundingClientRect()
+        if (!review || !terminal) return Number.POSITIVE_INFINITY
+        return terminal.top - review.bottom
+      })
+    })
+    .toBeLessThanOrEqual(expected + 1)
+}
+
 function base64Encode(value: string) {
  return Buffer.from(value, "utf8").toString("base64").replace(/\+/g, "-").replace(/\//g, "_").replace(/=/g, "")
 }
@@ -0,0 +1,48 @@
+import { expect, test, type Route } from "@playwright/test"
+
+const server = "http://127.0.0.1:4097"
+
+test("nested server dialog keeps focus inside the top layer", async ({ page }) => {
+  await page.addInitScript((server) => {
+    localStorage.setItem("opencode.global.dat:server", JSON.stringify({ list: [server] }))
+  }, server)
+  await page.route("**/*", async (route) => {
+    const url = new URL(route.request().url())
+    if (url.origin !== server) return route.fallback()
+    if (url.pathname === "/api/event") {
+      return route.fulfill({
+        status: 200,
+        contentType: "text/event-stream",
+        body: 'data: {"id":"evt_connected","type":"server.connected","data":{}}\n\n',
+      })
+    }
+    if (url.pathname === "/api/global/health" || url.pathname === "/api/health") {
+      return json(route, { healthy: true, version: "2.0.0" })
+    }
+    return json(route, {})
+  })
+
+  await page.goto("/")
+  await page.keyboard.press("Control+,")
+  const settings = page.locator(".settings-dialog")
+  await expect(settings).toBeVisible()
+  await settings.getByRole("tab", { name: "Servers" }).click()
+  await settings.getByRole("button", { name: "Add server" }).click()
+
+  const editor = page.getByRole("dialog", { name: "Add server" })
+  await expect(editor.getByPlaceholder("http://localhost:4096")).toBeFocused()
+  const username = editor.getByPlaceholder("username")
+  const password = editor.getByPlaceholder("password")
+  await username.click()
+  await expect(username).toBeFocused()
+  await username.fill("kit")
+  await expect(username).toHaveValue("kit")
+  await page.keyboard.press("Tab")
+  await expect(password).toBeFocused()
+  await password.fill("secret")
+  await expect(password).toHaveValue("secret")
+})
+
+function json(route: Route, body: unknown, status = 200) {
+  return route.fulfill({ status, contentType: "application/json", body: JSON.stringify(body) })
+}
@@ -20,45 +20,47 @@ const messages = [
  },
  { id: "msg_second", type: "user", text: "Second prompt", time: { created: 4 } },
 ] satisfies SessionMessageInfo[]
+const session = {
+  id: sessionID,
+  slug: "session-message-revert",
+  projectID,
+  directory,
+  title: "Session message revert",
+  agent: "build",
+  model: { id: "test", providerID: "opencode" },
+  version: "dev",
+  time: { created: 1, updated: 4 },
+}
+const fixture = {
+  directory,
+  project: {
+    id: projectID,
+    worktree: directory,
+    canonical: directory,
+    vcs: "git",
+    name: "session-message-revert",
+    time: { created: 1, updated: 1 },
+    sandboxes: [],
+  },
+  provider: {
+    all: [
+      {
+        id: "opencode",
+        name: "OpenCode",
+        models: { test: { id: "test", name: "Test", variants: {}, limit: { context: 200_000 } } },
+      },
+    ],
+    connected: ["opencode"],
+    default: { providerID: "opencode", modelID: "test" },
+  },
+  pageMessages: () => ({ items: messages }),
+}

 test("reverts directly to the selected user message", async ({ page }) => {
  const staged: { sessionID: string; messageID: string }[] = []
  await mockOpenCodeServer(page, {
-    directory,
-    project: {
-      id: projectID,
-      worktree: directory,
-      canonical: directory,
-      vcs: "git",
-      name: "session-message-revert",
-      time: { created: 1, updated: 1 },
-      sandboxes: [],
-    },
-    provider: {
-      all: [
-        {
-          id: "opencode",
-          name: "OpenCode",
-          models: { test: { id: "test", name: "Test", variants: {}, limit: { context: 200_000 } } },
-        },
-      ],
-      connected: ["opencode"],
-      default: { providerID: "opencode", modelID: "test" },
-    },
-    sessions: [
-      {
-        id: sessionID,
-        slug: "session-message-revert",
-        projectID,
-        directory,
-        title: "Session message revert",
-        agent: "build",
-        model: { id: "test", providerID: "opencode" },
-        version: "dev",
-        time: { created: 1, updated: 4 },
-      },
-    ],
-    pageMessages: () => ({ items: messages }),
+    ...fixture,
+    sessions: [session],
    onRevertStage: (input) => staged.push(input),
  })
  await page.goto(`/server/${base64Encode(server)}/session/${sessionID}`)
@@ -77,3 +79,19 @@ test("reverts directly to the selected user message", async ({ page }) => {
  await expect(page.getByRole("textbox", { name: "Prompt" })).toHaveText("Second prompt")
  expect(staged).toEqual([{ sessionID, messageID: "msg_second" }])
 })
+
+test("hides revert actions in a child session", async ({ page }) => {
+  await mockOpenCodeServer(page, {
+    ...fixture,
+    sessions: [
+      { ...session, id: "ses_parent", slug: "parent", title: "Parent session" },
+      { ...session, parentID: "ses_parent" },
+    ],
+  })
+  await page.goto(`/server/${base64Encode(server)}/session/${sessionID}`)
+  await expectSessionTitle(page, "Session message revert")
+
+  const message = page.locator('[data-message-id="msg_second"]')
+  await message.hover()
+  await expect(message.getByRole("button", { name: "Revert message" })).toHaveCount(0)
+})
@@ -0,0 +1,226 @@
+import { expect, test, type Page } from "@playwright/test"
+import type { OpenCodeEvent } from "@opencode-ai/client/promise"
+import { base64Encode } from "@opencode-ai/util/encode"
+import { mockOpenCodeServer } from "../utils/mock-server"
+import { expectAppVisible } from "../utils/waits"
+
+const directory = "C:/OpenCode/SessionQueueRegression"
+const projectID = "proj_session_queue_regression"
+const sessionID = "ses_session_queue_regression"
+const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
+
+type InboxRow = {
+  id: string
+  sessionID: string
+  timeCreated: number
+  type: "user"
+  payload: { text: string; metadata?: Record<string, unknown> }
+  delivery: "steer" | "queue"
+}
+
+function createQueueMock(seed: string[]) {
+  const rows: InboxRow[] = seed.map((text, index) => ({
+    id: `inb_seed_${index + 1}`,
+    sessionID,
+    timeCreated: 1700000000000 + index,
+    type: "user",
+    payload: { text },
+    delivery: "queue",
+  }))
+  const events: OpenCodeEvent[] = []
+  const prompts: Record<string, unknown>[] = []
+  const changes: { inboxID: string; action: "cancel" | "steer" }[] = []
+  const log: string[] = []
+  let sequence = 0
+  const emit = (type: OpenCodeEvent["type"], data: OpenCodeEvent["data"]) => {
+    sequence += 1
+    events.push({
+      id: `evt_queue_${sequence}`,
+      type,
+      created: Date.now(),
+      durable: { aggregateID: sessionID, seq: sequence, version: 1 },
+      data,
+    } as OpenCodeEvent)
+  }
+  return {
+    rows,
+    prompts,
+    changes,
+    log,
+    events: () => events.splice(0),
+    onPrompt: (input: { sessionID: string; body: Record<string, unknown> }) => {
+      prompts.push(input.body)
+      log.push(`prompt:${String(input.body.delivery ?? "steer")}`)
+      const row: InboxRow = {
+        id: typeof input.body.id === "string" ? input.body.id : `inb_mock_${sequence}`,
+        sessionID: input.sessionID,
+        timeCreated: Date.now(),
+        type: "user",
+        payload: {
+          text: typeof input.body.text === "string" ? input.body.text : "",
+          ...(input.body.metadata === undefined ? {} : { metadata: input.body.metadata as Record<string, unknown> }),
+        },
+        delivery: input.body.delivery === "queue" ? "queue" : "steer",
+      }
+      rows.push(row)
+      emit("session.inbox.enqueued", {
+        sessionID: input.sessionID,
+        inboxID: row.id,
+        item: { type: "user", payload: row.payload, delivery: row.delivery },
+      })
+    },
+    onInboxChange: (input: { sessionID: string; inboxID: string; action: "cancel" | "steer" }) => {
+      changes.push({ inboxID: input.inboxID, action: input.action })
+      log.push(`${input.action}:${input.inboxID}`)
+      const index = rows.findIndex((row) => row.id === input.inboxID)
+      const row = rows[index]
+      if (!row) return
+      if (input.action === "cancel") {
+        rows.splice(index, 1)
+        emit("session.inbox.cancelled", { sessionID: input.sessionID, inboxID: input.inboxID })
+        return
+      }
+      row.delivery = "steer"
+      emit("session.inbox.delivery.changed", {
+        sessionID: input.sessionID,
+        inboxID: input.inboxID,
+        delivery: "steer",
+      })
+    },
+  }
+}
+
+async function openSession(page: Page, mock: ReturnType<typeof createQueueMock>, followUpBehavior?: "queue" | "steer") {
+  if (followUpBehavior) {
+    await page.addInitScript(
+      (behavior) => localStorage.setItem("settings.v3", JSON.stringify({ general: { followUpBehavior: behavior } })),
+      followUpBehavior,
+    )
+  }
+  await mockOpenCodeServer(page, {
+    directory,
+    project: {
+      id: projectID,
+      worktree: directory,
+      vcs: "git",
+      name: "session-queue-regression",
+      time: { created: 1700000000000, updated: 1700000000000 },
+      sandboxes: [],
+    },
+    provider: {
+      all: [
+        {
+          id: "opencode",
+          name: "OpenCode",
+          models: { "queue-model": { id: "queue-model", name: "Queue Model", limit: { context: 200_000 } } },
+        },
+      ],
+      connected: ["opencode"],
+      default: { providerID: "opencode", modelID: "queue-model" },
+    },
+    sessions: [
+      {
+        id: sessionID,
+        slug: "session-queue-regression",
+        projectID,
+        directory,
+        title: "Session queue regression",
+        version: "dev",
+        time: { created: 1700000000000, updated: 1700000000000 },
+      },
+    ],
+    pageMessages: () => ({ items: [] }),
+    sessionStatus: () => ({ [sessionID]: { type: "running" } }),
+    inbox: () => mock.rows.map((row) => ({ ...row, payload: { ...row.payload } })),
+    onPrompt: mock.onPrompt,
+    onInboxChange: mock.onInboxChange,
+    events: mock.events,
+  })
+  await page.goto(`/server/${base64Encode(server)}/session/${sessionID}`)
+  const composer = page.locator('[data-component="composer"]')
+  await expectAppVisible(composer)
+  return {
+    composer,
+    input: composer.locator('[data-component="composer-editor"]'),
+    rows: page.locator('[data-component="session-queue-row"]'),
+  }
+}
+
+test("follow-up preference controls Enter while Mod+Enter uses the alternate delivery", async ({ page }) => {
+  const mock = createQueueMock([])
+  const view = await openSession(page, mock, "queue")
+
+  await view.input.fill("queue this follow-up")
+  await expect(view.composer.locator('[data-action="composer-alternate-delivery"]')).toContainText("Steer")
+  await view.input.press("Enter")
+  await expect(view.rows.getByText("queue this follow-up", { exact: true })).toBeVisible()
+
+  await view.input.fill("steer this correction")
+  await view.input.press("ControlOrMeta+Enter")
+  await expect.poll(() => mock.prompts.map((prompt) => prompt.delivery)).toEqual(["queue", "steer"])
+  await expect(view.input).toHaveText("")
+})
+
+test("dragging reorders queued prompts", async ({ page }) => {
+  const mock = createQueueMock(["first queued prompt", "second queued prompt", "third queued prompt"])
+  const view = await openSession(page, mock)
+  await expect(view.rows).toHaveCount(3)
+
+  const first = view.rows.filter({ hasText: "first queued prompt" })
+  const third = view.rows.filter({ hasText: "third queued prompt" })
+  await first.getByRole("button", { name: "Reorder queued prompt" }).hover()
+  await page.mouse.down()
+  const target = await third.boundingBox()
+  if (!target) throw new Error("The target queue row is not visible")
+  await page.mouse.move(target.x + target.width / 2, target.y + target.height / 2, { steps: 10 })
+  await page.mouse.up()
+
+  await expect(view.rows.locator('[data-action="session-queue-edit"]')).toHaveText([
+    "second queued prompt",
+    "third queued prompt",
+    "first queued prompt",
+  ])
+  expect(mock.prompts.map((prompt) => prompt.text)).toEqual([
+    "second queued prompt",
+    "third queued prompt",
+    "first queued prompt",
+  ])
+  expect(mock.changes).toEqual([
+    { inboxID: "inb_seed_1", action: "cancel" },
+    { inboxID: "inb_seed_2", action: "cancel" },
+    { inboxID: "inb_seed_3", action: "cancel" },
+  ])
+})
+
+test("editing restores the existing draft and replaces only the original queue position", async ({ page }) => {
+  const mock = createQueueMock(["first queued prompt", "tighten the error copy", "third queued prompt"])
+  const view = await openSession(page, mock)
+  const original = view.rows.getByText("tighten the error copy", { exact: true })
+  await expect(original).toBeVisible()
+
+  await view.input.fill("my in-progress draft")
+  await original.click()
+  await expect(view.input).toHaveText("tighten the error copy")
+  await view.input.press("Escape")
+  await expect(view.input).toHaveText("my in-progress draft")
+
+  await original.click()
+  await expect(view.input).toHaveText("tighten the error copy")
+  await view.input.fill("tighten the error copy and add a retry hint")
+  await view.input.press("Enter")
+
+  await expect(view.rows.locator('[data-action="session-queue-edit"]')).toHaveText([
+    "first queued prompt",
+    "tighten the error copy and add a retry hint",
+    "third queued prompt",
+  ])
+  await expect(view.input).toHaveText("my in-progress draft")
+  expect(mock.prompts.map((prompt) => prompt.text)).toEqual([
+    "tighten the error copy and add a retry hint",
+    "tighten the error copy and add a retry hint",
+    "third queued prompt",
+  ])
+  expect(mock.prompts.every((prompt) => prompt.delivery === "queue" && prompt.resume === false)).toBe(true)
+  expect(mock.changes.map((change) => change.action)).toEqual(["cancel", "cancel", "cancel"])
+  expect(mock.log[0]).toBe("prompt:queue")
+})
@@ -99,7 +99,7 @@ test.describe("regression: session timeline local row state", () => {
    await wrapper.evaluate((element) => {
      ;(element as HTMLElement).dataset.regressionMarker = "before-stream"
    })
-    await wrapper.locator('[data-slot="collapsible-trigger"]').first().click()
+    await wrapper.locator('[data-scope="apply-patch"] button').click()
    await expectExpanded(wrapper, false)

    events.push(...textEvents())
@@ -139,7 +139,7 @@ test.describe("regression: session timeline local row state", () => {
    expect(siblingProbe).toEqual({
      fileMarker: "before",
      frameMarker: "before",
-      rowKey: `assistant-part:${userMessageID}:part:${assistantMessageID}:${editPartID}`,
+      rowKey: `assistant-part:file:part:${assistantMessageID}:${editPartID}`,
      rowMarker: "before",
      shadowRoots: 0,
      toolMarker: "before",
@@ -179,8 +179,8 @@ test.describe("regression: session timeline local row state", () => {
    await expectSessionTitle(page, title)

    const wrapper = page.locator(`[data-timeline-part-id="${editPartID}"]`).first()
-    const trigger = wrapper.locator('[data-slot="collapsible-trigger"]').first()
-    const diff = wrapper.locator('[data-component="edit-content"]').first()
+    const trigger = wrapper.locator('[data-component="sticky-accordion-header"]')
+    const diff = wrapper.locator('[data-component="apply-patch-file-diff"]').first()
    await expectAppVisible(diff)
    await expect.poll(() => wrapper.evaluate((element) => element.getBoundingClientRect().height)).toBeGreaterThan(500)
    const samples = await wrapper.evaluate(async (element) => {
@@ -190,8 +190,8 @@ test.describe("regression: session timeline local row state", () => {
      for (const offset of [0, 120, 240, 360, 480]) {
        root.scrollBy(0, offset - (result.at(-1)?.offset ?? 0))
        await new Promise(requestAnimationFrame)
-        const trigger = element.querySelector<HTMLElement>('[data-slot="collapsible-trigger"]')!
-        const diff = element.querySelector<HTMLElement>('[data-component="edit-content"]')!
+        const trigger = element.querySelector<HTMLElement>('[data-component="sticky-accordion-header"]')!
+        const diff = element.querySelector<HTMLElement>('[data-component="apply-patch-file-diff"]')!
        result.push({
          offset,
          trigger: trigger.getBoundingClientRect().y,
@@ -202,7 +202,7 @@ test.describe("regression: session timeline local row state", () => {
      return result
    })

-    expect(samples[0]!.trigger).toBeLessThan(samples[0]!.diff)
+    expect(samples[0]!.trigger).toBeGreaterThanOrEqual(samples[0]!.diff)
    expect(samples.every((sample) => Math.abs(sample.trigger - samples[0]!.trigger) <= 1)).toBe(true)
    expect(samples.every((sample) => sample.trigger < sample.bottom)).toBe(true)
  })
@@ -234,7 +234,9 @@ async function readToolState(page: Page) {
    .evaluate(
      (element, textPartID) => ({
        expanded: (() => {
-          const trigger = element.querySelector('[data-slot="collapsible-trigger"]')
+          const trigger =
+            element.querySelector('[data-scope="apply-patch"] button') ??
+            element.querySelector('[data-slot="collapsible-trigger"]')
          const aria = trigger?.getAttribute("aria-expanded")
          if (aria === "true") return true
          if (aria === "false") return false
@@ -409,7 +411,9 @@ function eventValue<Type extends OpenCodeEvent["type"]>(
 }

 function readExpanded(element: Element) {
-  const trigger = element.querySelector('[data-slot="collapsible-trigger"]')
+  const trigger =
+    element.querySelector('[data-scope="apply-patch"] button') ??
+    element.querySelector('[data-slot="collapsible-trigger"]')
  const aria = trigger?.getAttribute("aria-expanded")
  if (aria === "true") return true
  if (aria === "false") return false
@@ -1,5 +1,12 @@
 import { expect, test } from "@playwright/test"
-import { assistantMessage, setupTimeline, toolPart, userMessage } from "../performance/timeline-stability/fixture"
+import { createTwoFilesPatch } from "diff"
+import {
+  assistantMessage,
+  setupTimeline,
+  toolPart,
+  userMessage,
+  userText,
+} from "../performance/timeline-stability/fixture"

 test("renders completed write content", async ({ page }) => {
  const id = "prt_file_projection_write"
@@ -47,5 +54,144 @@ test("renders a completed single-file patch", async ({ page }) => {
    settings: { editToolPartsExpanded: true },
  })

-  await expect(page.locator(`[data-timeline-part-id="${id}"] [data-component="apply-patch-file-diff"]`)).toBeVisible()
+  const wrapper = page.locator(`[data-timeline-part-id="${id}"]`)
+  const file = wrapper.locator('[data-scope="apply-patch"]')
+  const scroller = page.locator(".scroll-view__viewport", { has: page.locator("[data-timeline-row]") })
+  await expect(file.getByRole("button")).toHaveAttribute("aria-expanded", "false")
+  await expect(wrapper.locator('[data-component="apply-patch-file-diff"]')).toHaveCount(0)
+  await file.getByRole("button").click()
+  await expect(wrapper.locator('[data-component="apply-patch-file-diff"]')).toBeVisible()
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(1)
+
+  await file.getByRole("button").click()
+  await expect(wrapper.locator('[data-component="apply-patch-file-diff"]')).toHaveCount(0)
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(1)
+
+  await file.getByRole("button").click()
+  await expect(wrapper.locator('[data-component="apply-patch-file-diff"]')).toBeVisible()
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(1)
+})
+
+test("keeps an expanded file diff header at the same viewport position", async ({ page }) => {
+  const id = "prt_file_projection_anchored_patch"
+  const before = Array.from({ length: 80 }, (_, index) => `export const value${index} = ${index}\n`).join("")
+  const after = before.replaceAll(" = ", " = compute(").replaceAll("\n", ")\n")
+  await setupTimeline(page, {
+    messages: [
+      userMessage([userText("Preceding context ".repeat(120))]),
+      assistantMessage([
+        toolPart(
+          id,
+          "patch",
+          "completed",
+          { patchText: "Update src/anchored.ts" },
+          {
+            metadata: {
+              files: [
+                {
+                  file: "src/anchored.ts",
+                  status: "modified",
+                  patch: createTwoFilesPatch("a/src/anchored.ts", "b/src/anchored.ts", before, after),
+                  additions: 80,
+                  deletions: 80,
+                },
+              ],
+            },
+          },
+        ),
+      ]),
+    ],
+    viewport: { width: 1200, height: 600 },
+  })
+
+  const scroller = page.locator(".scroll-view__viewport", { has: page.locator("[data-timeline-row]") })
+  const wrapper = page.locator(`[data-timeline-part-id="${id}"]`)
+  const row = page.locator("[data-timeline-key]", { has: wrapper })
+  const trigger = wrapper.getByRole("button")
+  await expect
+    .poll(() =>
+      row.evaluate((element) => {
+        const measured = element.querySelector<HTMLElement>("[data-index]")
+        return measured
+          ? Math.abs(element.getBoundingClientRect().height - measured.getBoundingClientRect().height)
+          : Number.POSITIVE_INFINITY
+      }),
+    )
+    .toBeLessThanOrEqual(1)
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight))
+    .toBeGreaterThan(1)
+  await scroller.evaluate((element) => {
+    element.scrollTop = element.scrollHeight - element.clientHeight - 0.25
+  })
+  await expect(trigger).toBeInViewport()
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(0.5)
+  const bottomScrollTop = await scroller.evaluate((element) => element.scrollTop)
+  await scroller.hover()
+  await page.mouse.wheel(0, -20)
+  await expect
+    .poll(() => scroller.evaluate((element, bottom) => bottom - element.scrollTop, bottomScrollTop))
+    .toBeGreaterThan(0)
+  const y = await trigger.evaluate((element) => element.getBoundingClientRect().y)
+  const collapsedHeight = await row.evaluate((element) => element.getBoundingClientRect().height)
+  await trigger.click()
+  await expect(wrapper.locator('[data-component="apply-patch-file-diff"]')).toBeVisible()
+  await expect
+    .poll(() =>
+      row.evaluate((element, collapsed) => {
+        const measured = element.querySelector<HTMLElement>("[data-index]")
+        const allocatedHeight = element.getBoundingClientRect().height
+        return {
+          grew: allocatedHeight > collapsed + 1,
+          measured: measured ? Math.abs(allocatedHeight - measured.getBoundingClientRect().height) <= 1 : false,
+        }
+      }, collapsedHeight),
+    )
+    .toEqual({ grew: true, measured: true })
+  await expect
+    .poll(() => trigger.evaluate((element, initialY) => Math.abs(element.getBoundingClientRect().y - initialY), y))
+    .toBeLessThanOrEqual(5)
+
+  const scrollTop = await scroller.evaluate((element) => element.scrollTop)
+  await scroller.hover()
+  await page.mouse.wheel(0, 200)
+  await expect
+    .poll(() => scroller.evaluate((element, initial) => element.scrollTop - initial, scrollTop))
+    .toBeGreaterThan(50)
+  const scrolled = await scroller.evaluate((element, initial) => element.scrollTop - initial, scrollTop)
+  expect(scrolled).toBeLessThan(400)
+
+  const expandedY = await trigger.evaluate((element) => element.getBoundingClientRect().y)
+  await trigger.click()
+  await expect(wrapper.locator('[data-component="apply-patch-file-diff"]')).toHaveCount(0)
+  await expect
+    .poll(() =>
+      scroller.evaluate((element) => Math.abs(element.scrollHeight - element.clientHeight - element.scrollTop)),
+    )
+    .toBeLessThanOrEqual(1)
+  await expect.poll(() => trigger.evaluate((element) => element.getBoundingClientRect().y)).toBeGreaterThan(expandedY)
+
+  await trigger.click()
+  await expect(wrapper.locator('[data-component="apply-patch-file-diff"]')).toBeVisible()
+  await expect
+    .poll(() =>
+      row.evaluate((element) => {
+        const measured = element.querySelector<HTMLElement>("[data-index]")
+        return measured
+          ? Math.abs(element.getBoundingClientRect().height - measured.getBoundingClientRect().height)
+          : Number.POSITIVE_INFINITY
+      }),
+    )
+    .toBeLessThanOrEqual(1)
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(1)
 })
--- a/Show More
+++ b/Show More