fix(desktop): use configured notification sounds

fix(app): suppress notifications while focused
fix(ai): reconcile reasoning finals that arrive without deltas (#44595 )
2026-08-24 10:36:19 +00:00 · 2026-08-24 13:36:41 +08:00 · 2026-08-24 13:31:29 +08:00 · 2026-08-24 00:29:23 -05:00 · 2026-08-24 05:26:57 +00:00 · 2026-08-24 12:59:50 +08:00
1335 changed files with 54833 additions and 46824 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 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Nested AGENTS.md instructions are re-injected after compaction. Previously the in-memory dedup claim outlived the synthetic message that compaction dropped from model-visible history, so nested instructions were silently lost for the rest of the process lifetime. The claim now only guards in-flight loads; the synthetic message metadata in durable history is the sole lasting ledger, so any history truncation (compaction, revert) self-heals on the next read in that subtree.
@@ -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.
@@ -0,0 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Title generation and compaction summaries now build their model requests through the shared session request boundary, gaining unsupported-media filtering and image bounds while explicitly opting out of session context hooks: plugins that shape the agent conversation do not observe title or compaction requests. Title requests gain the fork-aware session prompt cache key, and compaction summaries in forked sessions reuse the fork root's prompt cache key instead of the fork's own.
@@ -0,0 +1,37 @@
+name: deploy-posts
+
+on:
+  push:
+    branches:
+      - v2
+    paths:
+      - packages/posts/**
+      - bun.lock
+      - .github/workflows/deploy-posts.yml
+  workflow_dispatch:
+
+concurrency:
+  group: deploy-posts-${{ github.ref_name }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+
+jobs:
+  deploy:
+    if: github.repository == 'anomalyco/opencode' && github.ref_name == 'v2'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+
+      - name: Build
+        working-directory: packages/posts
+        run: bun run build
+
+      - name: Deploy
+        working-directory: packages/posts
+        run: bun run deploy
+        env:
+          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}
@@ -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
@@ -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
@@ -50,6 +50,14 @@ jobs:
      - name: Setup Bun
        uses: ./.github/actions/setup-bun

+      - name: Test Effect simplification rules
+        if: runner.os == 'Linux'
+        run: bun run test:effect-simplification-rules
+
+      - name: Check Effect simplifications
+        if: runner.os == 'Linux'
+        run: bun run lint:effect-simplifications
+
      - name: Configure git identity
        run: |
          git config --global user.email "bot@opencode.ai"
@@ -76,6 +84,11 @@ jobs:
        env:
          OPENCODE_EXPERIMENTAL_DISABLE_FILEWATCHER: ${{ runner.os == 'Windows' && 'true' || 'false' }}

+      - name: Verify published codemode package
+        if: runner.os == 'Linux'
+        working-directory: packages/codemode
+        run: bun run script/publish.ts --dry-run
+
      - name: Verify compiled service lifecycle
        if: always()
        timeout-minutes: 10
@@ -1,19 +1,3 @@
 {
-  "$schema": "https://opencode.ai/tui.json",
-  "plugin": [
-    [
-      "./plugins/tui-smoke.tsx",
-      {
-        "enabled": false,
-        "label": "workspace",
-        "keybinds": {
-          "smoke_modal": "ctrl+alt+m",
-          "smoke_screen": "ctrl+alt+o",
-          "smoke_screen_home": "escape,ctrl+shift+h",
-          "smoke_screen_modal": "ctrl+alt+m",
-          "smoke_dialog_close": "escape,q"
-        }
-      }
-    ]
-  ]
+  "$schema": "https://opencode.ai/tui.json"
 }
@@ -1,5 +1,5 @@
 - After changing the public Protocol or Server `HttpApi`, run `bun run generate` from `packages/client`. Do not edit generated client files directly.
- Keep runtime dependencies directed from Schema to Core and Protocol, then from Core and Protocol to Server. Client runtime code may depend on Schema and Protocol but never Core or Server; `sdk-next` composes Client, Core, and Server.
+- Keep runtime dependencies directed from Schema to Core and Protocol, then from Core and Protocol to Server. Client runtime code may depend on Schema and Protocol but never Core or Server; `sdk` composes Client, Core, and Server.
 - Current implementation changes belong in `packages/core`, `packages/cli`, `packages/server`, `packages/protocol`, `packages/schema`, and related generated client surfaces when required.
 - The default branch in this repo is `v2`.
 - Base all new branches and worktrees on `v2`, or `origin/v2` when the local `v2` ref is unavailable. Do not base them on `dev`.
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-JEqi00PCle+o5OfBlJJaZtXd+4sYB3o+rvYiESlN4dY=",
-    "aarch64-linux": "sha256-zk3Uk1SQyeRrQ7BuFwlOnQAptUHIkq+oPdfd+sTEq5U=",
-    "aarch64-darwin": "sha256-3BOd3EcqimoG3rTI6lTHe91YVlCoEi8/68eT1lbOi0c=",
-    "x86_64-darwin": "sha256-X7wGmjiMloF5Zhuc20kAxLC+tl613YNXRgA+dQjP2WM="
+    "x86_64-linux": "sha256-LvDHCOm8OAZfvb0I0L6AbdOevRoQmEJnnqrSgAiNHv8=",
+    "aarch64-linux": "sha256-O0L0iHjb4cwl9xWHIna8VFHyQzoAKzfY8oMpVNayMOg=",
+    "aarch64-darwin": "sha256-ETP8FE71NqufYDUbR7tBdsMEOVQ44wLmsZBeZiSRBRY=",
+    "x86_64-darwin": "sha256-WUcoLldDriT3QxcdlnBQhuPrxDNub0EDvvZXk/pDMpY="
  }
 }
@@ -18,40 +18,41 @@
    "bench:devex": "bun run --cwd packages/app test:bench:devex",
    "lint": "oxlint",
    "lint:effect-patterns": "ast-grep scan -c script/ast-grep/sgconfig.yml packages/util/src packages/core/src packages/server/src packages/protocol/src packages/cli/src",
+    "lint:effect-simplifications": "ast-grep scan -c script/ast-grep/effect-simplifications/sgconfig.yml --off=unused-suppression packages",
    "test:lint-rules": "ast-grep test -c script/ast-grep/sgconfig.yml",
+    "test:effect-simplification-rules": "ast-grep test -c script/ast-grep/effect-simplifications/sgconfig.yml",
    "typecheck": "bun turbo typecheck --concurrency=3",
    "typecheck:profile": "bun script/profile-typecheck.ts",
    "typecheck:profile:packages": "bun script/profile-typecheck-packages.ts",
    "upgrade-opentui": "bun run script/upgrade-opentui.ts",
    "postinstall": "bun run --cwd packages/core fix-node-pty",
    "prepare": "husky",
+    "reserve-packages": "bun script/reserve-package-names.ts",
    "random": "echo 'Random script'",
    "sso": "aws sso login --sso-session=opencode --no-browser",
-    "translate:app": "bun run script/translate-app.ts",
    "test": "echo 'do not run tests from root' && exit 1"
  },
  "workspaces": {
    "packages": [
      "packages/*",
      "packages/console/*",
-      "packages/stats/*",
-      "packages/slack"
+      "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": "0.5.4",
-      "@opentui/keymap": "0.5.4",
-      "@opentui/solid": "0.5.4",
-      "@tanstack/solid-virtual": "3.13.32",
+      "@opentui/core": "0.5.7",
+      "@opentui/keymap": "0.5.7",
+      "@opentui/solid": "0.5.7",
+      "@tanstack/solid-virtual": "3.13.37",
      "@shikijs/stream": "4.2.0",
      "@standard-schema/spec": "1.1.0",
      "ulid": "3.0.1",
@@ -70,9 +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",
@@ -128,7 +129,6 @@
    "@aws-sdk/client-s3": "3.933.0",
    "@opencode-ai/plugin": "workspace:*",
    "@opencode-ai/script": "workspace:*",
-    "@opencode-ai/sdk": "1.18.5",
    "heap-snapshot-toolkit": "1.1.3",
    "typescript": "catalog:"
  },
@@ -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",
-    "@ff-labs/fff-bun@0.10.1": "patches/@ff-labs%2Ffff-bun@0.10.1.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.
@@ -370,7 +370,7 @@ const responseError = Effect.fn("RecordingEnv.responseError")(function* (
  response: HttpClientResponse.HttpClientResponse,
 ) {
  if (response.status >= 200 && response.status < 300) return undefined
-  const body = yield* response.text.pipe(Effect.catch(() => Effect.succeed("")))
+  const body = yield* response.text.pipe(Effect.orElseSucceed(() => ""))
  return `${response.status}${body ? `: ${body.slice(0, 180)}` : ""}`
 })

@@ -1,3 +1,4 @@
+import { Buffer } from "node:buffer"
 import { Effect, Schema } from "effect"
 import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client.js"
@@ -33,6 +34,19 @@ export const DEFAULT_BASE_URL = "https://api.anthropic.com/v1"
 export const PATH = "/messages"
 export const DEFAULT_MAX_TOKENS = 32_000

+const SSE_EVENTS = new Set([
+  "message",
+  "message_start",
+  "message_delta",
+  "message_stop",
+  "content_block_start",
+  "content_block_delta",
+  "content_block_stop",
+  "ping",
+  "error",
+])
+export const framing = Framing.sseEvents(SSE_EVENTS)
+
 export type ThinkingInput =
  | {
      readonly type: "adaptive"
@@ -41,7 +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 }
    ))
@@ -50,6 +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
@@ -69,25 +96,61 @@ 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>

@@ -192,14 +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"),
@@ -210,10 +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),
@@ -228,13 +313,19 @@ 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>

 const AnthropicUsage = Schema.StructWithRest(
  Schema.Struct({
-    input_tokens: Schema.optional(Schema.Number),
+    input_tokens: optionalNull(Schema.Number),
    output_tokens: Schema.optional(Schema.Number),
    cache_creation_input_tokens: optionalNull(Schema.Number),
    cache_read_input_tokens: optionalNull(Schema.Number),
@@ -356,22 +447,40 @@ 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, "_")
+
 const lowerToolCall = (part: ToolCallPart): AnthropicToolUseBlock => ({
  type: "tool_use",
-  id: part.id,
+  id: scrubToolCallID(part.id),
  name: part.name,
  input: part.input,
 })

 const lowerServerToolCall = (part: ToolCallPart): AnthropicServerToolUseBlock => ({
  type: "server_tool_use",
-  id: part.id,
+  id: scrubToolCallID(part.id),
  name: part.name,
  input: part.input,
 })
@@ -393,10 +502,155 @@ const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult
  // Prefer the provider-owned replay payload; fall back to the result value for
  // histories constructed directly from provider events.
  const payload = part.providerMetadata?.anthropic?.["result"] ?? part.result.value
-  return { type: wireType, tool_use_id: part.id, content: payload } satisfies AnthropicServerToolResultBlock
+  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 {
@@ -406,6 +660,14 @@ 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}`)
@@ -416,6 +678,10 @@ 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
 })

@@ -524,7 +790,7 @@ 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"])
@@ -575,7 +841,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
        return yield* ProviderShared.unsupportedContent("Anthropic Messages", "tool", ["tool-result"])
      content.push({
        type: "tool_result",
-        tool_use_id: part.id,
+        tool_use_id: scrubToolCallID(part.id),
        content: yield* lowerToolResultContent(part),
        is_error: part.result.type === "error" ? true : undefined,
        cache_control: cacheControl(breakpoints, part.cache),
@@ -591,24 +857,74 @@ 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 =
@@ -619,7 +935,7 @@ 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) {
@@ -669,7 +985,13 @@ 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,
  }
 })

@@ -692,7 +1014,7 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // expose that subset through `output_tokens_details.thinking_tokens`.
 const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
  if (!usage) return undefined
-  const nonCached = usage.input_tokens
+  const nonCached = usage.input_tokens ?? undefined
  const cacheRead = usage.cache_read_input_tokens ?? undefined
  const cacheWrite = usage.cache_creation_input_tokens ?? undefined
  const inputTokens = ProviderShared.sumTokens(nonCached, cacheRead, cacheWrite)
@@ -780,7 +1102,8 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
  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 [
@@ -788,7 +1111,7 @@ 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)
@@ -800,7 +1123,7 @@ 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,
        }),
@@ -990,15 +1313,28 @@ 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 step = (state: ParserState, event: AnthropicEvent) => {
  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_start") {
+    const block = event.content_block
+    if (block && (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 ${event.index}`),
+        )
+    }
+    return Effect.succeed(onContentBlockStart(state, event))
+  }
  if (event.type === "content_block_delta") return onContentBlockDelta(state, event)
  if (event.type === "content_block_stop") return onContentBlockStop(state, event)
  if (event.type === "message_delta") return Effect.succeed(onMessageDelta(state, event))
@@ -1037,9 +1373,12 @@ export const route = Route.make({
  provider: "anthropic",
  providerMetadataKey: "anthropic",
  protocol,
-  endpoint: Endpoint.path(PATH, { baseURL: DEFAULT_BASE_URL }),
+  endpoint: Endpoint.path(
+    (input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH),
+    { baseURL: DEFAULT_BASE_URL },
+  ),
  auth: Auth.none,
-  framing: Framing.sse,
+  framing,
  headers: () => ({ "anthropic-version": "2023-06-01" }),
 })

@@ -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"
@@ -37,6 +37,17 @@ const requiresThoughtSignatureFallback = (modelID: string) => {
  return !/(^|\/)gemini-robotics-er-1\.5(?:[.-]|$)/i.test(modelID)
 }

+// Gemini 3 accepts media nested inside function responses; matched Gemini 2.5 variants reject it,
+// so their tool-result attachments lower as a separate user turn instead.
+const routesLegacyToolMedia = (modelID: string) => /gemini-2[.-]5(?:[.-]|$)/i.test(modelID)
+
+// Blacklist: Gemini 1.x/2.x ignore or reject explicit function call ids.
+// Every other model id (Gemini 3+, gemma, anything unrecognized) gets them.
+const omitsFunctionCallIds = (modelID: string) => {
+  const match = /^gemini(?:-live)?-(\d+)/i.exec(modelID)
+  return match !== null && Number(match[1]) < 3
+}
+
 export interface OptionsInput {
  readonly [key: string]: unknown
  readonly cachedContent?: string
@@ -71,10 +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({
@@ -87,11 +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({
@@ -111,8 +127,8 @@ const GeminiContentPart = Schema.Union([
 ])

 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>

@@ -175,44 +191,45 @@ 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(GeminiContent),
+  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 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>
 }

 // =============================================================================
@@ -270,29 +287,31 @@ 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 = () => {
+    if (!pendingMedia) return
+    contents.push({ role: "user", parts: [{ text: "Attached media from tool result:" }, ...pendingMedia] })
+    pendingMedia = undefined
+  }

  for (const message of request.messages) {
+    if (message.role !== "tool") flushMedia()
    if (message.role === "system") {
      const part = yield* ProviderShared.wrappedSystemUpdate("Gemini", message)
      const previous = contents.at(-1)
      // Gemini rejects a continuation whose function-response turn carries extra
      // parts, so an update after a tool result starts its own user turn.
-      if (previous?.role === "user" && !previous.parts.some((item) => "functionResponse" in item))
-        contents[contents.length - 1] = { role: "user", parts: [...previous.parts, { text: part.text }] }
+      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
    }
@@ -316,7 +335,7 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
        if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call"]))
          return yield* ProviderShared.unsupportedContent("Gemini", "assistant", ["text", "reasoning", "tool-call"])
        if (part.type === "text") {
-          parts.push({ text: part.text })
+          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata) })
          continue
        }
        if (part.type === "reasoning") {
@@ -324,7 +343,7 @@ 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,
@@ -349,7 +368,7 @@ 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,
@@ -367,21 +386,28 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
        const value = ProviderShared.normalizeToolFile(item)
        media.push({ inlineData: { mimeType: value.mime, data: value.base64 } })
      }
+      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,
            content: text.join("\n"),
          },
-          parts: media.length > 0 ? media : undefined,
+          parts: legacyToolMedia || media.length === 0 ? undefined : media,
        },
      })
    }
-    contents.push({ role: "user", parts })
+    // Gemini requires every response to a parallel call batch in one user turn,
+    // so consecutive tool results join the open function-response turn.
+    const previous = contents.at(-1)
+    if (previous?.role === "user" && (previous.parts ?? []).some((item) => "functionResponse" in item))
+      contents[contents.length - 1] = { role: "user", parts: [...(previous.parts ?? []), ...parts] }
+    else contents.push({ role: "user", parts })
  }

+  flushMedia()
  return contents
 })

@@ -467,21 +493,25 @@ 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 },
  })
 }
@@ -516,19 +546,24 @@ 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 []

  const events: LLMEvent[] = []
-  const lifecycle = state.reasoningSignature
-    ? Lifecycle.reasoningEnd(
-        state.lifecycle,
-        events,
-        "reasoning-0",
-        googleMetadata({ thoughtSignature: state.reasoningSignature }),
-      )
-    : state.lifecycle
+  let lifecycle = state.lifecycle
+  if (state.reasoningSignature !== undefined)
+    lifecycle = Lifecycle.reasoningEnd(
+      lifecycle,
+      events,
+      "reasoning-0",
+      googleMetadata({ thoughtSignature: state.reasoningSignature }),
+    )
+  if (state.textSignature !== undefined)
+    lifecycle = Lifecycle.textEnd(lifecycle, events, "text-0", googleMetadata({ thoughtSignature: state.textSignature }))
  Lifecycle.finish(lifecycle, events, {
    reason: {
      normalized:
@@ -558,12 +593,17 @@ 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) {
-    if ("thoughtSignature" in part && part.thoughtSignature && "thought" in part && part.thought)
-      reasoningSignature = part.thoughtSignature
+  for (const part of candidate.content.parts ?? []) {
+    const signature = "thoughtSignature" in part && part.thoughtSignature ? part.thoughtSignature : undefined
+    // Gemini attaches replay signatures to thought parts, visible text, or function calls;
+    // each block kind must retain the signature attached to its own parts.
+    if (signature !== undefined && "thought" in part && part.thought) reasoningSignature = signature
+    else if (signature !== undefined && "text" in part) textSignature = signature
    if ("text" in part && part.text.length > 0) {
      if (part.thought) {
        lifecycle = Lifecycle.reasoningDelta(
@@ -571,7 +611,7 @@ const step = (state: ParserState, event: GeminiEvent) => {
          events,
          "reasoning-0",
          part.text,
-          part.thoughtSignature ? googleMetadata({ thoughtSignature: part.thoughtSignature }) : undefined,
+          signature ? googleMetadata({ thoughtSignature: signature }) : undefined,
        )
        continue
      }
@@ -581,17 +621,27 @@ const step = (state: ParserState, event: GeminiEvent) => {
        "reasoning-0",
        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
      )
-      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", part.text)
+      lifecycle = Lifecycle.textDelta(
+        lifecycle,
+        events,
+        "text-0",
+        part.text,
+        textSignature ? googleMetadata({ thoughtSignature: textSignature }) : undefined,
+      )
+      textSignature = undefined
      continue
    }

    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,
@@ -604,7 +654,8 @@ 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
@@ -616,8 +667,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
      ...nextState,
      hasToolCalls,
      lifecycle,
-      nextToolCallId,
      reasoningSignature,
+      textSignature,
+      seenCallIds,
      finishReason: candidate.finishReason ?? nextState.finishReason,
    },
    events,
@@ -639,7 +691,7 @@ export const protocol = Protocol.make({
  },
  stream: {
    event: Protocol.jsonEvent(GeminiEvent),
-    initial: () => ({ hasToolCalls: false, nextToolCallId: 0, lifecycle: Lifecycle.initial() }),
+    initial: () => ({ hasToolCalls: false, lifecycle: Lifecycle.initial() }),
    step,
    onHalt: finish,
  },
@@ -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),
@@ -22,12 +23,9 @@ export interface Options {
  readonly id: string
  readonly name: string
  readonly rotateAfterMs?: number
+  readonly enabled?: (url: string) => boolean
+  readonly url?: (url: string) => string
  readonly headers?: (headers: Headers.Headers) => Headers.Headers
-  readonly driver?: (input: {
-    readonly request: Readonly<Record<string, unknown>>
-    readonly message: string
-    readonly base: WebSocketChannelDriver
-  }) => WebSocketChannelDriver
 }

 export interface Prepared {
@@ -114,6 +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,
@@ -147,18 +147,25 @@ export const transport = <Body>(options: Options): Transport<Body, Prepared, str
      Effect.gen(function* () {
        const parts = yield* HttpTransport.jsonRequestParts(input)
        const headers = Headers.remove(options.headers?.(parts.headers) ?? parts.headers, "content-length")
-        const channel = input.webSocket
-          ? yield* Effect.gen(function* () {
-              const create = yield* message(parts.jsonBody)
-              const base = driver(options, create.message)
-              return {
-                url: yield* WebSocketTransport.toWebSocketUrl(parts.url),
-                headers,
-                rotateAfterMs: options.rotateAfterMs,
-                driver: options.driver?.({ request: create.request, message: create.message, base }) ?? base,
-              }
-            })
-          : undefined
+        const channel =
+          input.webSocket && (options.enabled?.(parts.url) ?? true)
+            ? yield* Effect.gen(function* () {
+                const create = yield* message(parts.jsonBody)
+                const base = driver(options, create.message)
+                return {
+                  url: yield* WebSocketTransport.toWebSocketUrl(options.url?.(parts.url) ?? parts.url),
+                  headers,
+                  rotateAfterMs: options.rotateAfterMs,
+                  driver: OpenResponsesContinuation.driver({
+                    id: options.id,
+                    name: options.name,
+                    request: create.request,
+                    message: create.message,
+                    base,
+                  }),
+                }
+              })
+            : undefined
        return {
          http: {
            request: ProviderShared.jsonPost({ url: parts.url, body: parts.bodyText, headers: parts.headers }),
@@ -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)

@@ -161,4 +161,4 @@ export const driver = (input: DriverInput): WebSocketChannelDriver => {
  }
 }

-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(
@@ -301,19 +313,8 @@ 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 +324,10 @@ 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 }
@@ -336,7 +340,6 @@ export interface ParserState {
  readonly hasFunctionCall: boolean
  readonly lifecycle: Lifecycle.State
  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 +353,10 @@ 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 +373,7 @@ 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 +387,46 @@ 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 +439,29 @@ 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 +470,17 @@ 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 +493,20 @@ 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 +558,9 @@ 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 +584,7 @@ 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 +605,30 @@ 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 +665,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques

 const lowerOptions = (request: LLMRequest) => {
  const options = OpenResponsesOptions.resolve(request)
+  const cacheKey = ProviderShared.clampPromptCacheKey(request.promptCacheKey)
  return {
    ...(options.instructions ? { instructions: options.instructions } : {}),
    ...(options.store !== undefined ? { store: options.store } : {}),
@@ -616,7 +675,7 @@ 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 } }
@@ -732,7 +791,7 @@ 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 }) })
@@ -750,20 +809,33 @@ const onOutputTextDone = (state: ParserState, event: Event, id: string): StepRes
 }

 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 +854,17 @@ 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 +873,11 @@ 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 +907,9 @@ 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)
@@ -931,7 +988,7 @@ 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 || !event.delta || !state.tools[event.item_id]) return [state, NO_EVENTS] satisfies StepResult
  const result = ToolStream.appendExisting(
    state.id,
    state.tools,
@@ -951,7 +1008,7 @@ 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)
@@ -1059,32 +1116,45 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
  return [{ ...state, 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,
  })
 }

@@ -1100,20 +1170,27 @@ 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, `${event.type} is missing item_id`)
    return Effect.succeed(onReasoningDelta(state, event, event.item_id))
  }
-  if (event.type === "response.reasoning.done" || event.type === "response.reasoning_summary_text.done") {
+  if (
+    event.type === "response.reasoning.done" ||
+    event.type === "response.reasoning_summary_text.done" ||
+    event.type === "response.reasoning_summary.done" ||
+    event.type === "response.reasoning_text.done"
+  ) {
    if (!event.item_id) return ProviderShared.eventError(state.id, `${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
@@ -1128,7 +1205,10 @@ export const step = (state: ParserState, event: Event) => {
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
    return Effect.succeed(onOutputItemAdded(state, event))
  }
-  if (event.type === "response.function_call_arguments.delta") return onFunctionCallArgumentsDelta(state, event)
+  if (event.type === "response.function_call_arguments.delta")
+    return event.item_id
+      ? onFunctionCallArgumentsDelta(state, event)
+      : ProviderShared.eventError(state.id, `${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`)
@@ -1159,17 +1239,11 @@ export const initial = (request: LLMRequest, extension: Extension = BASE): Parse
  tools: ToolStream.empty<string>(),
  lifecycle: Lifecycle.initial(),
  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: {
@@ -156,6 +156,9 @@ 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({
@@ -204,11 +207,16 @@ 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])),
@@ -261,7 +269,7 @@ const lowerTool = (tool: ToolDefinition, inputSchema: JsonSchema, options: Lower
  function: {
    name: tool.name,
    description: tool.description,
-    parameters: ToolSchemaProjection.openAI(inputSchema),
+    parameters: inputSchema,
  },
  cache_control: options.cacheControl?.(tool.cache),
 })
@@ -509,11 +517,23 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
  return messages
 })

+// 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
+}
+
 const lowerOptions = (request: LLMRequest) => {
  const options = OpenAIOptions.resolve(request)
+  const cacheKey = ProviderShared.clampPromptCacheKey(request.promptCacheKey)
  return {
    ...(options.store !== undefined ? { store: options.store } : {}),
-    ...(request.promptCacheKey ? { prompt_cache_key: request.promptCacheKey } : {}),
+    ...(cacheKey ? { prompt_cache_key: cacheKey } : {}),
    ...(options.reasoningEffort ? { reasoning_effort: options.reasoningEffort } : {}),
  }
 }
@@ -532,12 +552,15 @@ 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 hasHistory = hasToolHistory(request.messages)
  return {
    model: request.model.id,
    messages: yield* lowerMessages(request, options),
    tools:
      request.tools.length === 0
-        ? undefined
+        ? hasHistory
+          ? []
+          : undefined
        : request.tools.map((tool) =>
            lowerTool(
              tool,
@@ -581,11 +604,18 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // 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))
@@ -691,8 +721,11 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
        }),
      })
    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
@@ -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 } from "./open-responses-channel.js"
-import { OpenAIResponsesChannel } from "./openai-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

@@ -128,46 +124,7 @@ 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 +145,29 @@ 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")
    return event.item_id
      ? Effect.succeed(OpenResponses.onReasoningDelta(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)
 }

@@ -247,12 +189,12 @@ const endpoint = Endpoint.path<OpenAIResponsesBody>(PATH, { baseURL: DEFAULT_BAS
 const auth = Auth.none

 export const httpTransport = HttpTransport.sseJson.with<OpenAIResponsesBody>()
-export const transport = OpenResponsesChannel.transport<OpenAIResponsesBody>({
+export const channelTransport = OpenResponsesChannel.transport<OpenAIResponsesBody>
+export const transport = channelTransport({
  id: ADAPTER,
  name: NAME,
  rotateAfterMs: WEBSOCKET_ROTATE_AFTER_MS,
  headers: (headers) => Headers.set(headers, "openai-beta", headers["openai-beta"] ?? WEBSOCKET_PROTOCOL_HEADER),
-  driver: (input) => OpenAIResponsesChannel.driver({ id: ADAPTER, name: NAME, ...input }),
 })

 export const route = Route.make({
@@ -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 clampPromptCacheKey = (key: string | undefined): string | undefined => {
+  if (key === undefined) return undefined
+  const chars = Array.from(key)
+  if (chars.length <= OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH) return key
+  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
@@ -197,19 +208,28 @@ export const errorText = (error: unknown) => {

 /**
 * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
- * decoder, and drops empty / `[DONE]` keep-alive events so the protocol event
- * schema sees one JSON string per element. The SSE channel emits a
+ * decoder, optionally filters named events, and drops empty / `[DONE]`
+ * keep-alive events so the protocol event schema sees one JSON string per
+ * element. The SSE channel emits a
 * `Retry` control event on its error channel; we drop it here (we don't
 * implement client-driven retries). Decoder failures become provider output
 * errors so the public error channel stays `AIError`.
 */
-export const sseFraming = (bytes: Stream.Stream<Uint8Array, AIError>): Stream.Stream<string, AIError> =>
+export const sseFraming = (
+  bytes: Stream.Stream<Uint8Array, AIError>,
+  events?: ReadonlySet<string>,
+): Stream.Stream<string, AIError> =>
  bytes.pipe(
    Stream.decodeText(),
    Stream.pipeThroughChannel(Sse.decode()),
    Stream.catchTag("Retry", () => Stream.empty),
    Stream.catchTag("SseError", (error) => Stream.fail(eventError("sse", error.message))),
-    Stream.filter((event) => event.data.length > 0 && event.data !== "[DONE]"),
+    Stream.filter(
+      (event) =>
+        (events === undefined || events.has(event.event)) &&
+        event.data.length > 0 &&
+        (event.data !== "[DONE]" || (events !== undefined && event.event !== "message")),
+    ),
    Stream.map((event) => event.data),
  )

@@ -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(
@@ -21,9 +21,15 @@ export const textStart = (state: State, events: LLMEvent[], id: string, provider
  return { ...stepped, text: new Set([...stepped.text, id]) }
 }

-export const textDelta = (state: State, events: LLMEvent[], id: string, text: string): State => {
+export const textDelta = (
+  state: State,
+  events: LLMEvent[],
+  id: string,
+  text: string,
+  providerMetadata?: ProviderMetadata,
+): State => {
  const started = textStart(state, events, id)
-  events.push(LLMEvent.textDelta({ id, text }))
+  events.push(LLMEvent.textDelta({ id, text, providerMetadata }))
  return started
 }

@@ -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 +1,70 @@
+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) ?? {}
@@ -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"
@@ -74,11 +74,15 @@ const INVALID_REQUEST_CODES = new Set(["invalid_prompt", "invalid_request_error"
 const RATE_LIMIT_TEXT = /rate increased too quickly|rate[-_\s]?limit|too[_\s]?many[_\s]?requests/i
 const QUOTA_TEXT = /insufficient[-_\s]?quota|quota[-_\s]?exceeded/i
 const CONTENT_POLICY_TEXT = /content[-_\s]?policy|content_filter|safety/i
+const NETWORK_ERROR_TEXT = /network[-_\s]error/i

 export interface ProviderFailure {
  readonly message: string
  readonly status?: number | undefined
  readonly code?: string | undefined
+  // Raw wire payload, scanned for failure signals (codes, overflow phrases)
+  // that the summary message does not carry. Not shown to users.
+  readonly rawBody?: string | undefined
  readonly retryAfterMs?: number | undefined
  readonly rateLimit?: HttpRateLimitDetails | undefined
  readonly http?: HttpContext | undefined
@@ -88,11 +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)

@@ -127,6 +133,7 @@ export function classifyProviderFailure(input: ProviderFailure): AIError["reason
      retryAfterMs: input.retryAfterMs,
      rateLimit: input.rateLimit,
    })
+  if (NETWORK_ERROR_TEXT.test(text)) return new ProviderInternalReason({ ...common, status: input.status })
  if (codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable")))
    return new ProviderInternalReason({
      ...common,
@@ -1,3 +1,4 @@
+import { Headers } from "effect/unstable/http"
 import { Auth } from "../route/auth.js"
 import { type AtLeastOne, type ProviderAuthOption } from "../route/auth-options.js"
 import type { Route as RouteDef, RouteDefaultsInput } from "../route/client.js"
@@ -10,6 +11,7 @@ import { withOpenAIOptions, type OpenAIProviderOptionsInput } from "./openai-opt

 export const id = ProviderID.make("azure")
 const routeAuth = Auth.remove("authorization")
+const RESPONSES_WEBSOCKET_ROTATE_AFTER_MS = 55 * 60 * 1000

 // Azure needs the customer's resource URL; supply either `resourceName`
 // (helper builds the URL) or `baseURL` directly.
@@ -40,6 +42,30 @@ const responsesRoute = OpenAIResponses.route.with({
  id: "azure-openai-responses",
  provider: id,
  auth: routeAuth,
+  transport: OpenAIResponses.channelTransport({
+    id: "azure-openai-responses",
+    name: "Azure OpenAI Responses",
+    rotateAfterMs: RESPONSES_WEBSOCKET_ROTATE_AFTER_MS,
+    enabled: (value) => {
+      const url = new URL(value)
+      return (
+        url.protocol === "https:" &&
+        url.hostname.endsWith(".openai.azure.com") &&
+        url.pathname.endsWith("/openai/v1/responses") &&
+        url.searchParams.get("api-version") === "v1"
+      )
+    },
+    url: (value) => {
+      const url = new URL(value)
+      url.searchParams.delete("api-version")
+      return url.toString()
+    },
+    headers: (headers) => {
+      const apiKey = headers["api-key"]
+      if (!apiKey) return headers
+      return Headers.remove(Headers.set(headers, "authorization", `Bearer ${apiKey}`), "api-key")
+    },
+  }),
 })

 const chatRoute = OpenAIChat.route.with({
@@ -4,7 +4,6 @@ import { AnthropicMessages } from "../protocols/anthropic-messages.js"
 import { Auth } from "../route/auth.js"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
 import { ProviderID, type ModelID } from "../schema/index.js"
 import { GoogleVertexShared } from "./google-vertex-shared.js"
@@ -14,6 +13,7 @@ export type AnthropicProviderOptionsInput = AnthropicMessages.ProviderOptionsInp
 export type AnthropicThinkingInput = AnthropicMessages.ThinkingInput

 const VERSION = "vertex-2023-10-16" as const
+const HEADER_VERSION = "2023-06-01" as const

 export const id = ProviderID.make("google-vertex")

@@ -57,7 +57,8 @@ const route = Route.make({
  }),
  endpoint: Endpoint.path(({ request }) => `/${request.model.id}:streamRawPredict`),
  auth: Auth.none,
-  framing: Framing.sse,
+  framing: AnthropicMessages.framing,
+  headers: () => ({ "anthropic-version": HEADER_VERSION }),
 })

 export const routes = [route]
@@ -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

@@ -9,7 +9,7 @@ import type { ProviderPackage } from "../provider-package.js"
 import * as OpenAICompatibleProfiles from "./openai-compatible-profile.js"
 import * as OpenAIChat from "../protocols/openai-chat.js"
 import { newBreakpoints, ttlBucket } from "../protocols/utils/cache.js"
-import { isRecord } from "../protocols/shared.js"
+import { isRecord, ProviderShared } from "../protocols/shared.js"

 export const profile = OpenAICompatibleProfiles.profiles.openrouter
 export const id = ProviderID.make(profile.provider)
@@ -115,11 +115,12 @@ export const protocol = Protocol.make({
              reasoning_details: reasoningDetails,
            }
          })
+          const cacheKey = ProviderShared.clampPromptCacheKey(request.promptCacheKey)
          return {
            ...body,
            messages,
            ...bodyOptions(request.providerOptions),
-            ...(request.promptCacheKey ? { prompt_cache_key: request.promptCacheKey } : {}),
+            ...(cacheKey ? { prompt_cache_key: cacheKey } : {}),
          } 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"
@@ -28,13 +29,19 @@ export interface Settings extends ProviderPackage.Settings {

 export type { XAIImageOptions } from "../protocols/xai-images.js"

+const RESPONSES_WEBSOCKET_ROTATE_AFTER_MS = 24 * 60 * 1000
+
 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.httpTransport,
+  transport: OpenResponsesChannel.transport({
+    id: "openai-responses",
+    name: "xAI Responses",
+    rotateAfterMs: RESPONSES_WEBSOCKET_ROTATE_AFTER_MS,
+  }),
  defaults: { providerOptions: { store: false } },
 })

@@ -24,4 +24,10 @@ export interface Definition<Frame> {
 /** Server-Sent Events framing. Used by every JSON-streaming HTTP provider. */
 export const sse: Definition<string> = { id: "sse", frame: ProviderShared.sseFraming }

+/** SSE framing restricted to protocol-recognized event names. */
+export const sseEvents = (events: ReadonlySet<string>): Definition<string> => ({
+  id: "sse",
+  frame: (bytes) => ProviderShared.sseFraming(bytes, events),
+})
+
 export * as Framing from "./framing.js"
@@ -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

@@ -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 {
@@ -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"
+      }
+    }
+  ]
+}
@@ -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 +1,100 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "openai",
+    "protocol": "openai-responses",
+    "transport": "websocket",
+    "model": "gpt-5.5",
+    "tags": [
+      "prefix:openai-responses-websocket",
+      "provider:openai",
+      "protocol:openai-responses",
+      "transport:websocket",
+      "tool",
+      "continuation"
+    ],
+    "name": "openai-responses-websocket/continues-a-tool-call-over-one-socket",
+    "recordedAt": "2026-08-20T00:00:00.000Z"
+  },
+  "interactions": [
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 0,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Call get_weather once, then reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_tool_1\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"function_call\",\"id\":\"fc_ws_weather\",\"call_id\":\"call_ws_weather\",\"name\":\"get_weather\",\"arguments\":\"\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.delta\",\"item_id\":\"fc_ws_weather\",\"delta\":\"{\\\"city\\\":\\\"Paris\\\"}\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"function_call\",\"id\":\"fc_ws_weather\",\"call_id\":\"call_ws_weather\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_tool_1\"}}"
+        },
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"type\":\"function_call_output\",\"call_id\":\"call_ws_weather\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"previous_response_id\":\"resp_ws_tool_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_tool_2\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_tool_2\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_tool_2\",\"delta\":\"Paris is sunny.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_tool_2\",\"text\":\"Paris is sunny.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_tool_2\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Paris is sunny.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_tool_2\"}}"
+        }
+      ]
+    }
+  ]
+}
@@ -0,0 +1,119 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "openai",
+    "protocol": "openai-responses",
+    "transport": "websocket",
+    "model": "gpt-5.5",
+    "tags": [
+      "prefix:openai-responses-websocket",
+      "provider:openai",
+      "protocol:openai-responses",
+      "transport:websocket",
+      "reconnect",
+      "full-context"
+    ],
+    "name": "openai-responses-websocket/reconstructs-full-context-after-reconnect",
+    "recordedAt": "2026-08-20T00:00:00.000Z"
+  },
+  "interactions": [
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 0,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_reconnect_1\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_reconnect_1\",\"delta\":\"Alpha.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_reconnect_1\",\"text\":\"Alpha.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_reconnect_1\"}}"
+        }
+      ]
+    },
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 1,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Beta.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_reconnect_2\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_2\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_reconnect_2\",\"delta\":\"Beta.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_reconnect_2\",\"text\":\"Beta.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_2\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Beta.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_reconnect_2\"}}"
+        }
+      ]
+    }
+  ]
+}
@@ -0,0 +1,129 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "openai",
+    "protocol": "openai-responses",
+    "transport": "websocket",
+    "model": "gpt-5.5",
+    "tags": [
+      "prefix:openai-responses-websocket",
+      "provider:openai",
+      "protocol:openai-responses",
+      "transport:websocket",
+      "continuation",
+      "recovery"
+    ],
+    "name": "openai-responses-websocket/recovers-from-explicit-continuation-rejection",
+    "recordedAt": "2026-08-20T00:00:00.000Z"
+  },
+  "interactions": [
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 0,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_rejection_1\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_rejection_1\",\"delta\":\"Ready.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_rejection_1\",\"text\":\"Ready.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_rejection_1\"}}"
+        }
+      ]
+    },
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 1,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"previous_response_id\":\"resp_ws_rejection_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"error\",\"error\":{\"code\":\"previous_response_not_found\",\"message\":\"Previous response not found\"}}"
+        },
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_rejection_2\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_2\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_rejection_2\",\"delta\":\"Recovered.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_rejection_2\",\"text\":\"Recovered.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_2\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Recovered.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_rejection_2\"}}"
+        }
+      ]
+    }
+  ]
+}
@@ -10,6 +10,9 @@ export const sseEvents = (...chunks: ReadonlyArray<unknown>): string =>

 const formatChunk = (chunk: unknown) => `data: ${typeof chunk === "string" ? chunk : JSON.stringify(chunk)}\n\n`

+export const sseNamedEvent = (event: string, data: unknown): string =>
+  `event: ${event}\ndata: ${typeof data === "string" ? data : JSON.stringify(data)}`
+
 /**
 * Build an SSE body from already-serialized strings (used when the chunk shape
 * itself is part of what's being tested, e.g. malformed chunks).
@@ -82,6 +82,14 @@ describe("provider error classification", () => {
    ])
  })

+  test("classifies network error text as provider internal", () => {
+    expect(
+      ["network error", "network-error", "network_error"].map(
+        (message) => classifyProviderFailure({ message })._tag,
+      ),
+    ).toEqual(["ProviderInternal", "ProviderInternal", "ProviderInternal"])
+  })
+
  test("classifies nested provider codes when a top-level code is also present", () => {
    expect(
      [
@@ -98,3 +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" } })

@@ -8,7 +8,7 @@ import * as AnthropicMessages from "../../src/protocols/anthropic-messages.js"
 import { continuationRequest, nativeAnthropicMessagesContinuation } from "../continuation-scenarios.js"
 import { it } from "../lib/effect.js"
 import { dynamicResponse, fixedResponse } from "../lib/http.js"
-import { sseEvents } from "../lib/sse.js"
+import { sseEvents, sseNamedEvent, sseRaw } from "../lib/sse.js"

 const model = AnthropicMessages.route
  .with({ endpoint: { baseURL: "https://api.anthropic.test/v1/" }, auth: Auth.header("x-api-key", "test") })
@@ -327,6 +327,29 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("scrubs outbound tool call IDs without truncating them", () =>
+    Effect.gen(function* () {
+      const id = `functions.lookup:1|${"x".repeat(64)}`
+      const scrubbed = `functions_lookup_1_${"x".repeat(64)}`
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id, name: "lookup", input: {} })]),
+            Message.tool({ id, name: "lookup", result: "done" }),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toMatchObject([
+        { role: "assistant", content: [{ type: "tool_use", id: scrubbed, name: "lookup", input: {} }] },
+        { role: "user", content: [{ type: "tool_result", tool_use_id: scrubbed }] },
+      ])
+      expect(scrubbed.length).toBeGreaterThan(64)
+    }),
+  )
+
  it.effect("batches parallel tool results into one Anthropic user message", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -640,7 +663,60 @@ describe("Anthropic Messages route", () => {
    }),
  )

-  it.effect("maps thinking tokens and preserves unknown Anthropic usage fields", () =>
+  it.effect("ignores unknown named SSE events", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseRaw(
+              sseNamedEvent("message_start", {
+                type: "message_start",
+                message: { usage: { input_tokens: 5 } },
+              }),
+              sseNamedEvent("proxy.stats", "not json"),
+              sseNamedEvent("content_block_start", {
+                type: "content_block_start",
+                index: 0,
+                content_block: { type: "text", text: "" },
+              }),
+              sseNamedEvent("content_block_delta", {
+                type: "content_block_delta",
+                index: 0,
+                delta: { type: "text_delta", text: "Hello" },
+              }),
+              sseNamedEvent("content_block_stop", { type: "content_block_stop", index: 0 }),
+              sseNamedEvent("message_delta", {
+                type: "message_delta",
+                delta: { stop_reason: "end_turn" },
+                usage: { output_tokens: 1 },
+              }),
+              sseNamedEvent("message_stop", { type: "message_stop" }),
+              sseNamedEvent("proxy.done", "still not json"),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([{ type: "text", text: "Hello" }])
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "end_turn" })
+    }),
+  )
+
+  it.effect("rejects malformed recognized SSE events", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseRaw(sseNamedEvent("message_start", "[DONE]")))),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidProviderOutput",
+        message: "Invalid anthropic/anthropic-messages stream event",
+      })
+    }),
+  )
+
+  it.effect("maps nullable input tokens and preserves unknown Anthropic usage fields", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
        Effect.provide(
@@ -663,6 +739,7 @@ describe("Anthropic Messages route", () => {
                type: "message_delta",
                delta: { stop_reason: "end_turn" },
                usage: {
+                  input_tokens: null,
                  output_tokens: 8,
                  server_tool_use: { web_search_requests: 2, terminal_counter: 3 },
                  output_tokens_details: { terminal_detail: "preserved" },
@@ -682,7 +759,7 @@ describe("Anthropic Messages route", () => {
        totalTokens: 15,
        providerMetadata: {
          anthropic: {
-            input_tokens: 5,
+            input_tokens: null,
            cache_read_input_tokens: 2,
            service_tier: "standard",
            cache_creation: { ephemeral_5m_input_tokens: 1 },
@@ -1339,14 +1416,14 @@ describe("Anthropic Messages route", () => {
            Message.assistant([
              {
                type: "tool-call",
-                id: "srvtoolu_abc",
+                id: "srvtoolu.abc",
                name: "web_search",
                input: { query: "effect 4" },
                providerExecuted: true,
              },
              {
                type: "tool-result",
-                id: "srvtoolu_abc",
+                id: "srvtoolu.abc",
                name: "web_search",
                result: { type: "json", value: [{ url: "https://example.com" }] },
                providerExecuted: true,
@@ -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)
+    }),
+  )
+})
@@ -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" },
              },
@@ -181,6 +180,149 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("merges parallel tool results into one function-response turn", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } }),
+              ToolCallPart.make({ id: "call_2", name: "lookup", input: { query: "time" } }),
+            ]),
+            Message.tool({ id: "call_1", name: "lookup", result: "sunny", resultType: "text" }),
+            Message.tool({ id: "call_2", name: "lookup", result: "noon", resultType: "text" }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          parts: [
+            { functionCall: { name: "lookup", args: { query: "weather" } } },
+            { functionCall: { name: "lookup", args: { query: "time" } } },
+          ],
+        },
+        {
+          role: "user",
+          parts: [
+            {
+              functionResponse: {
+                name: "lookup",
+                response: { name: "lookup", content: "sunny" },
+              },
+            },
+            {
+              functionResponse: {
+                name: "lookup",
+                response: { name: "lookup", content: "noon" },
+              },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  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(
@@ -281,14 +423,18 @@ describe("Gemini route", () => {
              functionResponse: {
                name: "read",
                response: { name: "read", content: "Image read successfully" },
-                parts: [
-                  { inlineData: { mimeType: "image/png", data: "AAECAw==" } },
-                  { inlineData: { mimeType: "application/pdf", data: "JVBERi0xLjQ=" } },
-                ],
              },
            },
          ],
        },
+        {
+          role: "user",
+          parts: [
+            { text: "Attached media from tool result:" },
+            { inlineData: { mimeType: "image/png", data: "AAECAw==" } },
+            { inlineData: { mimeType: "application/pdf", data: "JVBERi0xLjQ=" } },
+          ],
+        },
      ])
      expect(JSON.stringify(prepared.body.contents)).not.toContain('"content":"AAECAw=="')
    }),
@@ -321,11 +467,164 @@ describe("Gemini route", () => {
              functionResponse: {
                name: "read",
                response: { name: "read", content: "" },
-                parts: [{ inlineData: { mimeType: "image/jpeg", data: "/9j/" } }],
              },
            },
          ],
        },
+        {
+          role: "user",
+          parts: [
+            { text: "Attached media from tool result:" },
+            { inlineData: { mimeType: "image/jpeg", data: "/9j/" } },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("nests media inside function responses for gemini 3", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: gemini3,
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({
+                id: "call_image",
+                name: "read",
+                input: { path: "pixel.png" },
+                providerMetadata: { google: { thoughtSignature: "sig_1" } },
+              }),
+            ]),
+            Message.tool({
+              id: "call_image",
+              name: "read",
+              result: {
+                type: "content",
+                value: [
+                  { type: "text", text: "Image read successfully" },
+                  { type: "file", uri: "data:image/png;base64,AAECAw==", mime: "image/png", name: "pixel.png" },
+                ],
+              },
+            }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          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==" } }],
+              },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("flushes pending media before system update text", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "shot", input: {} })]),
+            Message.tool({
+              id: "call_1",
+              name: "shot",
+              result: {
+                type: "content",
+                value: [{ type: "file", uri: "data:image/png;base64,AAEC", mime: "image/png" }],
+              },
+            }),
+            Message.system("Update."),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        { role: "model", parts: [{ functionCall: { name: "shot", args: {} } }] },
+        {
+          role: "user",
+          parts: [{ functionResponse: { name: "shot", response: { name: "shot", content: "" } } }],
+        },
+        {
+          role: "user",
+          parts: [
+            { text: "Attached media from tool result:" },
+            { inlineData: { mimeType: "image/png", data: "AAEC" } },
+            { text: "<system-update>\nUpdate.\n</system-update>" },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("collects legacy tool media into one turn after merged responses", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({ id: "call_1", name: "shot", input: {} }),
+              ToolCallPart.make({ id: "call_2", name: "shot", input: {} }),
+            ]),
+            Message.tool({
+              id: "call_1",
+              name: "shot",
+              result: {
+                type: "content",
+                value: [{ type: "file", uri: "data:image/png;base64,AAEC", mime: "image/png" }],
+              },
+            }),
+            Message.tool({
+              id: "call_2",
+              name: "shot",
+              result: {
+                type: "content",
+                value: [{ type: "text", text: "no image here" }],
+              },
+            }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          parts: [
+            { functionCall: { name: "shot", args: {} } },
+            { functionCall: { name: "shot", args: {} } },
+          ],
+        },
+        {
+          role: "user",
+          parts: [
+            { functionResponse: { name: "shot", response: { name: "shot", content: "" } } },
+            { functionResponse: { name: "shot", response: { name: "shot", content: "no image here" } } },
+          ],
+        },
+        {
+          role: "user",
+          parts: [
+            { text: "Attached media from tool result:" },
+            { inlineData: { mimeType: "image/png", data: "AAEC" } },
+          ],
+        },
      ])
    }),
  )
@@ -647,8 +946,8 @@ 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"),
@@ -656,23 +955,22 @@ 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,
            }),
          ],
        }),
@@ -704,6 +1002,61 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("preserves thoughtSignature on visible text parts", () =>
+    Effect.gen(function* () {
+      const body = sseEvents({
+        candidates: [
+          {
+            content: { role: "model", parts: [{ text: "All done.", thoughtSignature: "text_sig" }] },
+            finishReason: "STOP",
+          },
+        ],
+      })
+      const response = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)))
+      const delta = response.events.find((event) => event.type === "text-delta")
+      expect(delta).toMatchObject({
+        id: "text-0",
+        text: "All done.",
+        providerMetadata: { google: { thoughtSignature: "text_sig" } },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [Message.assistant([{ type: "text", text: "All done.", providerMetadata: delta?.providerMetadata }])],
+        }),
+      )
+      expect(prepared.body.contents).toEqual([
+        { role: "model", parts: [{ text: "All done.", thoughtSignature: "text_sig" }] },
+      ])
+    }),
+  )
+
+  it.effect("flushes a trailing empty signed text part at block close", () =>
+    Effect.gen(function* () {
+      const body = sseEvents({
+        candidates: [
+          {
+            content: {
+              role: "model",
+              parts: [{ text: "Working." }, { text: "", thoughtSignature: "tail_sig" }],
+            },
+            finishReason: "STOP",
+          },
+        ],
+      })
+      const response = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)))
+      const delta = response.events.find((event) => event.type === "text-delta")
+      const end = response.events.find((event) => event.type === "text-end")
+
+      expect(delta).toMatchObject({ id: "text-0", text: "Working.", providerMetadata: undefined })
+      expect(end).toMatchObject({
+        id: "text-0",
+        providerMetadata: { google: { thoughtSignature: "tail_sig" } },
+      })
+    }),
+  )
+
  it.effect("replays unsigned Gemini 3 tool calls with the validator bypass sentinel", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -721,7 +1074,7 @@ 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",
            },
          ],
@@ -731,7 +1084,7 @@ describe("Gemini route", () => {
          parts: [
            {
              functionResponse: {
-                id: undefined,
+                id: "tool_0",
                name: "lookup",
                response: { name: "lookup", content: "done" },
              },
@@ -767,15 +1120,15 @@ 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,
            },
          ],
@@ -803,11 +1156,11 @@ 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",
            },
          ],
@@ -845,21 +1198,17 @@ 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,
@@ -902,7 +1251,8 @@ 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: {} }])
    }),
  )

@@ -942,7 +1292,7 @@ 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" } } },
              ],
            },
@@ -956,16 +1306,19 @@ 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" },
@@ -973,6 +1326,62 @@ 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(
@@ -1083,6 +1492,73 @@ 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({
@@ -96,7 +143,7 @@ describe("Google Vertex providers", () => {
                "https://aiplatform.eu.rep.googleapis.com/v1/projects/vertex-project/locations/eu/publishers/anthropic/models/claude-sonnet-4-6:streamRawPredict",
              )
              expect(request.headers.get("authorization")).toBe("Bearer vertex-token")
-              expect(request.headers.get("anthropic-version")).toBeNull()
+              expect(request.headers.get("anthropic-version")).toBe("2023-06-01")
              const body = yield* Effect.promise(() => request.json())
              expect(body).toMatchObject({
                anthropic_version: "vertex-2023-10-16",
@@ -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" } },
                ),
@@ -342,6 +342,7 @@ describe("OpenAI Chat route", () => {
          },
          { role: "tool", tool_call_id: "call_1", content: encodeJson({ forecast: "sunny" }) },
        ],
+        tools: [],
        stream: true,
        stream_options: { include_usage: true },
      })
@@ -78,6 +78,45 @@ 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 +132,79 @@ describe("Open Responses-compatible route", () => {
    }),
  )

-  it.effect("omits OpenAI-only nullable phases from the Open Responses baseline", () =>
+  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("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 +224,14 @@ 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 +305,7 @@ describe("Open Responses-compatible route", () => {
          streamOptions: { includeObfuscation: false },
          topLogprobs: 3,
          truncation: "auto",
+          serviceTier: "provider-tier",
          allowedTools: { toolNames: ["lookup"] },
          maxToolCalls: 2,
          parallelToolCalls: false,
@@ -211,6 +330,7 @@ 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",
@@ -0,0 +1,216 @@
+import { describe, expect } from "bun:test"
+import { Effect, Stream } from "effect"
+import { Socket } from "effect/unstable/socket"
+import { LLM, LLMRequest, Message, ToolRuntime } from "../../src/index.js"
+import {
+  LLMClient,
+  WebSocketTransport,
+  type ChannelCheckpoint,
+  type ChannelObservation,
+  type WebSocketChannelExchange,
+  type WebSocketChannelExecutor,
+  type WebSocketConnection,
+} from "../../src/route.js"
+import { configure } from "../../src/providers/openai.js"
+import { decodeJson } from "../../src/protocols/shared.js"
+import { weatherRuntimeTool, weatherTool, weatherToolName } from "../recorded-scenarios.js"
+import { recordedTests } from "../recorded-test.js"
+
+const model = configure({ apiKey: process.env.OPENAI_API_KEY ?? "fixture" }).responses("gpt-5.5")
+const recorded = recordedTests({
+  prefix: "openai-responses-websocket",
+  provider: "openai",
+  protocol: "openai-responses",
+  requires: ["OPENAI_API_KEY"],
+  tags: ["transport:websocket"],
+  metadata: { transport: "websocket", model: model.id },
+})
+
+const observationFrame = (observation: ChannelObservation) => {
+  if (observation.type === "frame" || observation.type === "completed" || observation.type === "incomplete")
+    return Effect.succeed(observation.frame)
+  return Effect.fail(observation.error)
+}
+
+const terminal = (observation: ChannelObservation) => observation.type !== "frame"
+
+// This deliberately models only sequential test traffic. Core owns production connection pooling and recovery.
+const makeChannel = Effect.gen(function* () {
+  const constructor = yield* Socket.WebSocketConstructor
+  let connection: WebSocketConnection | undefined
+  let checkpoint: ChannelCheckpoint | undefined
+  let pending: ChannelCheckpoint | undefined
+  let opens = 0
+  const sent: unknown[] = []
+
+  const close = Effect.suspend(() => {
+    const current = connection
+    connection = undefined
+    return current ? current.close : Effect.void
+  })
+  yield* Effect.addFinalizer(() => close)
+
+  const executor: WebSocketChannelExecutor = {
+    execute: (exchange: WebSocketChannelExchange) =>
+      Effect.gen(function* () {
+        if (!connection) {
+          connection = yield* WebSocketTransport.open(exchange.connect).pipe(
+            Effect.provideService(Socket.WebSocketConstructor, constructor),
+          )
+          opens += 1
+        }
+        const current = connection
+        const create = yield* exchange.driver.create(checkpoint)
+        if (create.mode === "full") checkpoint = undefined
+        pending = undefined
+        sent.push(decodeJson(create.message))
+        yield* current.sendText(create.message)
+        const decoder = new TextDecoder()
+        return {
+          frames: current.messages.pipe(
+            Stream.map((message) => WebSocketTransport.messageText(message, decoder)),
+            Stream.mapEffect((frame) => exchange.driver.observe(create, frame)),
+            Stream.tap((observation) =>
+              Effect.sync(() => {
+                if (!terminal(observation)) return
+                pending = observation.type === "completed" ? observation.checkpoint : undefined
+                if (observation.type !== "completed") checkpoint = undefined
+              }),
+            ),
+            Stream.takeUntil(terminal),
+            Stream.mapEffect(observationFrame),
+          ),
+          complete: Effect.sync(() => {
+            checkpoint = pending
+            pending = undefined
+          }),
+        }
+      }),
+  }
+
+  return {
+    executor,
+    sent,
+    opens: () => opens,
+    reconnect: (preserveCheckpoint = false) =>
+      close.pipe(
+        Effect.andThen(
+          Effect.sync(() => {
+            pending = undefined
+            if (!preserveCheckpoint) checkpoint = undefined
+          }),
+        ),
+      ),
+  }
+})
+
+describe("OpenAI Responses WebSocket recorded", () => {
+  recorded.effect.with("continues a tool call over one socket", { tags: ["tool", "continuation"] }, () =>
+    Effect.gen(function* () {
+      const channel = yield* makeChannel
+      const request = LLM.request({
+        id: "recorded_openai_responses_websocket_tool",
+        model,
+        system: "Call get_weather once, then reply exactly: Paris is sunny.",
+        prompt: "What is the weather in Paris?",
+        tools: [weatherTool],
+        generation: { maxTokens: 50 },
+        cache: "none",
+      })
+      const first = yield* LLMClient.generate(request, { webSocket: channel.executor })
+      const call = first.toolCalls[0]
+      if (!call) yield* Effect.die("Expected get_weather tool call")
+      const result = yield* ToolRuntime.dispatch({ [weatherToolName]: weatherRuntimeTool }, call)
+      const second = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          messages: [
+            ...request.messages,
+            first.message,
+            Message.tool({ id: call.id, name: call.name, result: result.result }),
+          ],
+        }),
+        { webSocket: channel.executor },
+      )
+
+      expect(second.text).toBe("Paris is sunny.")
+      expect(channel.opens()).toBe(1)
+      expect(channel.sent).toHaveLength(2)
+      expect(channel.sent[1]).toMatchObject({
+        previous_response_id: expect.any(String),
+        input: [{ type: "function_call_output", call_id: call.id, output: expect.any(String) }],
+      })
+    }),
+  )
+
+  recorded.effect.with("reconstructs full context after reconnect", { tags: ["reconnect", "full-context"] }, () =>
+    Effect.gen(function* () {
+      const channel = yield* makeChannel
+      const request = LLM.request({
+        id: "recorded_openai_responses_websocket_reconnect",
+        model,
+        system: "Follow the user's exact reply instruction.",
+        prompt: "Reply exactly: Alpha.",
+        generation: { maxTokens: 30 },
+        cache: "none",
+      })
+      const first = yield* LLMClient.generate(request, { webSocket: channel.executor })
+      yield* channel.reconnect()
+      const second = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          messages: [...request.messages, first.message, Message.user("Reply exactly: Beta.")],
+        }),
+        { webSocket: channel.executor },
+      )
+
+      expect(first.text).toBe("Alpha.")
+      expect(second.text).toBe("Beta.")
+      expect(channel.opens()).toBe(2)
+      expect(channel.sent[1]).not.toHaveProperty("previous_response_id")
+      expect(channel.sent[1]).toMatchObject({
+        input: [
+          { role: "system", content: "Follow the user's exact reply instruction." },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Alpha." }] },
+          { role: "assistant", content: [{ type: "output_text", text: "Alpha." }] },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Beta." }] },
+        ],
+      })
+    }),
+  )
+
+  recorded.effect.with("recovers from explicit continuation rejection", { tags: ["continuation", "recovery"] }, () =>
+    Effect.gen(function* () {
+      const channel = yield* makeChannel
+      const request = LLM.request({
+        id: "recorded_openai_responses_websocket_rejection",
+        model,
+        system: "Follow the user's exact reply instruction.",
+        prompt: "Reply exactly: Ready.",
+        generation: { maxTokens: 30 },
+        cache: "none",
+      })
+      const first = yield* LLMClient.generate(request, { webSocket: channel.executor })
+      const continuation = LLMRequest.update(request, {
+        messages: [...request.messages, first.message, Message.user("Reply exactly: Recovered.")],
+      })
+      yield* channel.reconnect(true)
+      const rejected = yield* LLMClient.generate(continuation, { webSocket: channel.executor }).pipe(Effect.flip)
+      const recovered = yield* LLMClient.generate(continuation, { webSocket: channel.executor })
+
+      expect(rejected).toMatchObject({
+        reason: { _tag: "Transport", delivery: "rejected", recovery: "retry-full" },
+      })
+      expect(recovered.text).toBe("Recovered.")
+      expect(channel.opens()).toBe(2)
+      expect(channel.sent[1]).toHaveProperty("previous_response_id", expect.any(String))
+      expect(channel.sent[2]).not.toHaveProperty("previous_response_id")
+      expect(channel.sent[2]).toMatchObject({
+        input: [
+          { role: "system", content: "Follow the user's exact reply instruction." },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Ready." }] },
+          { role: "assistant", content: [{ type: "output_text", text: "Ready." }] },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Recovered." }] },
+        ],
+      })
+    }),
+  )
+})
@@ -28,7 +28,7 @@ import * as Azure from "../../src/providers/azure.js"
 import * as OpenAI from "../../src/providers/openai.js"
 import * as XAI from "../../src/providers/xai.js"
 import * as OpenAIResponses from "../../src/protocols/openai-responses.js"
-import { OpenAIResponsesChannel } from "../../src/protocols/openai-responses-channel.js"
+import { OpenResponsesContinuation } from "../../src/protocols/open-responses-continuation.js"
 import * as ProviderShared from "../../src/protocols/shared.js"
 import { continuationRequest, nativeOpenAIResponsesContinuation } from "../continuation-scenarios.js"
 import { it } from "../lib/effect.js"
@@ -68,7 +68,7 @@ const baseChannelDriver = (message: string): WebSocketChannelDriver => ({

 const continuationDriver = (request: Readonly<Record<string, unknown>>) => {
  const message = ProviderShared.encodeJson(request)
-  return OpenAIResponsesChannel.driver({
+  return OpenResponsesContinuation.driver({
    id: "openai-responses",
    name: "OpenAI Responses",
    request,
@@ -188,17 +188,15 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("omits unsupported semantic service tiers", () =>
+  it.effect("passes through provider-defined service tiers", () =>
    Effect.gen(function* () {
-      const prepared = yield* compileRequest(
-        LLMRequest.update(request, { providerOptions: { serviceTier: "unsupported" } }),
-      )
+      const prepared = yield* compileRequest(LLMRequest.update(request, { providerOptions: { serviceTier: "scale" } }))

-      expect(prepared.body).not.toHaveProperty("service_tier")
+      expect(prepared.body.service_tier).toBe("scale")
    }),
  )

-  it.effect("flattens top-level object unions in function schemas", () =>
+  it.effect("preserves function schemas", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
        LLMRequest.update(request, {
@@ -238,13 +236,22 @@ describe("OpenAI Responses route", () => {
          strict: false,
          parameters: {
            type: "object",
-            properties: {
-              path: { type: "string" },
-              reference: { type: "string" },
-              limit: { type: "integer", maximum: 2000 },
-              resource: { type: "string" },
-            },
-            additionalProperties: false,
+            anyOf: [
+              {
+                type: "object",
+                properties: {
+                  path: { type: "string" },
+                  reference: { anyOf: [{ type: "string" }, { type: "null" }] },
+                  limit: { type: "integer", maximum: 2000 },
+                },
+                required: ["path"],
+              },
+              {
+                type: "object",
+                properties: { resource: { type: "string" }, limit: { type: "integer", maximum: 51200 } },
+                required: ["resource"],
+              },
+            ],
          },
        },
      ])
@@ -314,6 +321,10 @@ describe("OpenAI Responses route", () => {
              }),
            messages: Stream.fromArray([
              ProviderShared.encodeJson({ type: "response.created", response: { id: "resp_ws" } }),
+              ProviderShared.encodeJson({
+                type: "response.output_item.added",
+                item: { type: "message", id: "msg_1" },
+              }),
              ProviderShared.encodeJson({ type: "response.output_text.delta", item_id: "msg_1", delta: "Hi" }),
              ProviderShared.encodeJson({ type: "response.completed", response: { id: "resp_ws" } }),
            ]),
@@ -394,6 +405,38 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("tolerates keepalive frames before response.created", () =>
+    Effect.gen(function* () {
+      const webSocket = WebSocketTransport.makeDirect({
+        open: () =>
+          Effect.succeed({
+            sendText: () => Effect.void,
+            messages: Stream.fromArray([
+              ProviderShared.encodeJson({ type: "keepalive", sequence_number: 0 }),
+              ProviderShared.encodeJson({ type: "response.created", response: { id: "resp_alive" } }),
+              ProviderShared.encodeJson({
+                type: "response.completed",
+                response: { id: "resp_alive", usage: { input_tokens: 1, output_tokens: 1 } },
+              }),
+            ]),
+            close: Effect.void,
+          }),
+      })
+      const deps = Layer.succeed(
+        RequestExecutor.Service,
+        RequestExecutor.Service.of({ execute: () => Effect.die("unexpected HTTP request") }),
+      )
+      const model = OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses(
+        "gpt-4.1-mini",
+      )
+
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "hi" }), { webSocket }).pipe(
+        Effect.provide(LLMClient.layer.pipe(Layer.provide(deps))),
+      )
+      expect(response.finishReason?.normalized).toBe("stop")
+    }),
+  )
+
  it.effect("continues a tool call with only the new tool output", () =>
    Effect.gen(function* () {
      const firstRequest = {
@@ -691,6 +734,134 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("builds xAI WebSocket requests without OpenAI handshake headers", () =>
+    Effect.gen(function* () {
+      const deps = Layer.succeed(
+        RequestExecutor.Service,
+        RequestExecutor.Service.of({ execute: () => Effect.die("unexpected HTTP request") }),
+      )
+      const response = yield* LLMClient.generate(LLM.request({ model: xaiModel, prompt: "Say hello." }), {
+        webSocket: {
+          execute: (exchange) =>
+            Effect.gen(function* () {
+              expect(exchange.connect.url).toBe("wss://api.x.ai/v1/responses")
+              expect(exchange.connect.rotateAfterMs).toBe(24 * 60 * 1000)
+              expect(exchange.connect.headers.authorization).toBe("Bearer test")
+              expect(exchange.connect.headers["openai-beta"]).toBeUndefined()
+              expect(JSON.parse((yield* exchange.driver.create(undefined)).message)).toMatchObject({
+                type: "response.create",
+                model: "grok-4.5",
+                store: false,
+              })
+              return {
+                frames: Stream.make(
+                  JSON.stringify({ type: "response.created", response: { id: "resp_xai" } }),
+                  JSON.stringify({ type: "response.completed", response: { id: "resp_xai" } }),
+                ),
+                complete: Effect.void,
+              }
+            }),
+        },
+      }).pipe(Effect.provide(LLMClient.layer.pipe(Layer.provide(deps))))
+
+      expect(response.finishReason.normalized).toBe("stop")
+    }),
+  )
+
+  it.effect("builds Azure WebSocket requests with v1 URLs and bearer auth", () =>
+    Effect.gen(function* () {
+      const deps = Layer.succeed(
+        RequestExecutor.Service,
+        RequestExecutor.Service.of({ execute: () => Effect.die("unexpected HTTP request") }),
+      )
+      const cases = [
+        {
+          model: Azure.configure({ resourceName: "opencode-test", apiKey: "azure-key" }).responses("deployment"),
+          authorization: "Bearer azure-key",
+        },
+        {
+          model: Azure.configure({ resourceName: "opencode-test", auth: Auth.bearer("entra-token") }).responses(
+            "deployment",
+          ),
+          authorization: "Bearer entra-token",
+        },
+      ]
+
+      yield* Effect.forEach(cases, (item) =>
+        LLMClient.generate(LLM.request({ model: item.model, prompt: "Say hello." }), {
+          webSocket: {
+            execute: (exchange) =>
+              Effect.gen(function* () {
+                expect(exchange.connect.url).toBe("wss://opencode-test.openai.azure.com/openai/v1/responses")
+                expect(exchange.connect.rotateAfterMs).toBe(55 * 60 * 1000)
+                expect(exchange.connect.headers.authorization).toBe(item.authorization)
+                expect(exchange.connect.headers["api-key"]).toBeUndefined()
+                expect(exchange.connect.headers["openai-beta"]).toBeUndefined()
+                expect(JSON.parse((yield* exchange.driver.create(undefined)).message)).toMatchObject({
+                  type: "response.create",
+                  model: "deployment",
+                  store: false,
+                })
+                return {
+                  frames: Stream.make(
+                    JSON.stringify({ type: "response.created", response: { id: "resp_azure" } }),
+                    JSON.stringify({ type: "response.completed", response: { id: "resp_azure" } }),
+                  ),
+                  complete: Effect.void,
+                }
+              }),
+          },
+        }).pipe(Effect.provide(LLMClient.layer.pipe(Layer.provide(deps)))),
+      )
+    }),
+  )
+
+  it.effect("keeps unsupported Azure endpoints and API versions on HTTP", () =>
+    Effect.gen(function* () {
+      const cases = [
+        {
+          model: Azure.configure({
+            resourceName: "opencode-test",
+            apiKey: "azure-key",
+            apiVersion: "2025-04-01-preview",
+          }).responses("deployment"),
+          url: "https://opencode-test.openai.azure.com/openai/v1/responses?api-version=2025-04-01-preview",
+        },
+        {
+          model: Azure.configure({
+            resourceName: "opencode-test",
+            apiKey: "azure-key",
+            useDeploymentBasedUrls: true,
+          }).responses("deployment"),
+          url: "https://opencode-test.openai.azure.com/openai/deployments/deployment/responses?api-version=v1",
+        },
+        {
+          model: Azure.configure({ baseURL: "https://gateway.example/azure", apiKey: "azure-key" }).responses(
+            "deployment",
+          ),
+          url: "https://gateway.example/azure/responses",
+        },
+      ]
+
+      yield* Effect.forEach(cases, (item) =>
+        LLMClient.generate(LLM.request({ model: item.model, prompt: "Say hello." }), {
+          webSocket: { execute: () => Effect.die("unexpected WebSocket request") },
+        }).pipe(
+          Effect.provide(
+            dynamicResponse((input) =>
+              Effect.gen(function* () {
+                expect(input.request.url).toBe(item.url)
+                return input.respond(sseEvents({ type: "response.completed", response: {} }), {
+                  headers: { "content-type": "text/event-stream" },
+                })
+              }),
+            ),
+          ),
+        ),
+      )
+    }),
+  )
+
  it.effect("uses exactly one HTTP request when no WebSocket executor is supplied", () =>
    Effect.gen(function* () {
      const attempts = yield* Ref.make(0)
@@ -1177,7 +1348,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("uses xAI inline file encoding for PDF tool results", () =>
+  it.effect("uses standard inline file encoding for xAI PDF tool results", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
        LLM.request({
@@ -1205,8 +1376,7 @@ describe("OpenAI Responses route", () => {
        {
          type: "input_file",
          filename: "report.pdf",
-          file_data: "JVBERi0xLjQ=",
-          mime_type: "application/pdf",
+          file_data: "data:application/pdf;base64,JVBERi0xLjQ=",
        },
      ])
    }),
@@ -1236,6 +1406,60 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("lowers remote tool-result media URLs without base64 wrapping", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "fetch", input: {} })]),
+            Message.tool({
+              id: "call_1",
+              name: "fetch",
+              resultType: "content",
+              result: [
+                { type: "file", uri: "https://example.com/image.png", mime: "image/png" },
+                { type: "file", uri: "https://example.com/report.pdf", mime: "application/pdf", name: "report.pdf" },
+              ],
+            }),
+          ],
+        }),
+      )
+
+      expect(expectToolOutput(prepared.body).output).toEqual([
+        { type: "input_image", image_url: "https://example.com/image.png" },
+        { type: "input_file", filename: "report.pdf", file_url: "https://example.com/report.pdf" },
+      ])
+    }),
+  )
+
+  it.effect("lowers tool-result videos as input_video", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "record", input: {} })]),
+            Message.tool({
+              id: "call_1",
+              name: "record",
+              resultType: "content",
+              result: [
+                { type: "file", uri: "data:video/mp4;base64,AAECAw==", mime: "video/mp4" },
+                { type: "file", uri: "https://example.com/demo.mp4", mime: "video/mp4" },
+              ],
+            }),
+          ],
+        }),
+      )
+
+      expect(expectToolOutput(prepared.body).output).toEqual([
+        { type: "input_video", video_url: "data:video/mp4;base64,AAECAw==" },
+        { type: "input_video", video_url: "https://example.com/demo.mp4" },
+      ])
+    }),
+  )
+
  it.effect("prepares the composed native continuation request", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -1445,6 +1669,7 @@ describe("OpenAI Responses route", () => {
  it.effect("parses text and usage stream fixtures", () =>
    Effect.gen(function* () {
      const body = 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.output_text.delta", item_id: "msg_1", delta: "!" },
        {
@@ -1706,6 +1931,67 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("ignores deltas without a matching output item", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_text.delta", item_id: "msg_missing", delta: "orphaned text" },
+              { type: "response.refusal.delta", item_id: "refusal_missing", delta: "orphaned refusal" },
+              {
+                type: "response.reasoning_summary_text.delta",
+                item_id: "rs_missing",
+                summary_index: 0,
+                delta: "orphaned reasoning",
+              },
+              {
+                type: "response.reasoning_summary_part.added",
+                item_id: "rs_still_missing",
+                summary_index: 0,
+              },
+              {
+                type: "response.reasoning_summary_text.delta",
+                item_id: "rs_still_missing",
+                summary_index: 0,
+                delta: "still orphaned reasoning",
+              },
+              {
+                type: "response.function_call_arguments.delta",
+                item_id: "fc_missing",
+                delta: '{"orphaned":true}',
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("")
+      expect(response.message.content).toEqual([])
+      expect(response.events.some(LLMEvent.is.toolCall)).toBeFalse()
+    }),
+  )
+
+  it.effect("rejects function argument deltas without the spec-required item id", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.function_call_arguments.delta", delta: "{}" },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      expect(error.message).toContain("response.function_call_arguments.delta is missing item_id")
+    }),
+  )
+
  it.effect("rejects reasoning events without the spec-required item id", () =>
    Effect.gen(function* () {
      const events = [
@@ -1761,8 +2047,10 @@ describe("OpenAI Responses route", () => {
        Effect.provide(
          fixedResponse(
            sseEvents(
+              { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
              { type: "response.output_text.delta", item_id: "msg_1", delta: "First" },
-              { type: "response.output_text.done", item_id: "msg_1" },
+              { type: "response.output_item.done", item: { type: "message", id: "msg_1" } },
+              { type: "response.output_item.added", item: { type: "message", id: "msg_2" } },
              { type: "response.output_text.delta", item_id: "msg_2", delta: "Second" },
              { type: "response.output_item.done", item: { type: "message", id: "msg_2" } },
              { type: "response.completed", response: { id: "resp_1" } },
@@ -1773,10 +2061,10 @@ describe("OpenAI Responses route", () => {

      expect(response.events.filter((event) => event.type.startsWith("text-"))).toEqual([
        { type: "text-start", id: "msg_1", providerMetadata: { openai: { itemId: "msg_1" } } },
-        { type: "text-delta", id: "msg_1", text: "First" },
-        { type: "text-end", id: "msg_1", providerMetadata: undefined },
+        { type: "text-delta", id: "msg_1", text: "First", providerMetadata: undefined },
+        { type: "text-end", id: "msg_1", providerMetadata: { openai: { itemId: "msg_1" } } },
        { type: "text-start", id: "msg_2", providerMetadata: { openai: { itemId: "msg_2" } } },
-        { type: "text-delta", id: "msg_2", text: "Second" },
+        { type: "text-delta", id: "msg_2", text: "Second", providerMetadata: undefined },
        { type: "text-end", id: "msg_2", providerMetadata: { openai: { itemId: "msg_2" } } },
      ])
    }),
@@ -1785,7 +2073,9 @@ describe("OpenAI Responses route", () => {
  it.effect("parses reasoning summary stream fixtures", () =>
    Effect.gen(function* () {
      const body = sseEvents(
+        { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1" } },
        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", delta: "thinking" },
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
        { type: "response.output_text.delta", item_id: "msg_1", delta: "Hello" },
        { type: "response.reasoning_summary_text.done", item_id: "rs_1" },
        { type: "response.completed", response: { id: "resp_1" } },
@@ -1797,18 +2087,22 @@ describe("OpenAI Responses route", () => {
      expect(response.text).toBe("Hello")
      expect(response.events).toMatchObject([
        { type: "step-start", index: 0 },
-        { type: "reasoning-start", id: "rs_1" },
-        { type: "reasoning-delta", id: "rs_1", text: "thinking" },
+        { type: "reasoning-start", id: "rs_1:0" },
+        { type: "reasoning-delta", id: "rs_1:0", text: "thinking" },
        { type: "text-start", id: "msg_1" },
        { type: "text-delta", id: "msg_1", text: "Hello" },
-        { type: "reasoning-end", id: "rs_1" },
+        { type: "reasoning-end", id: "rs_1:0" },
        { type: "text-end", id: "msg_1" },
        { type: "step-finish", index: 0, reason: { normalized: "stop", raw: undefined } },
        { type: "finish", reason: { normalized: "stop", raw: undefined } },
      ])
      expect(response.events.filter((event) => event.type === "finish")).toHaveLength(1)
      expect(response.message.content).toEqual([
-        { type: "reasoning", text: "thinking" },
+        {
+          type: "reasoning",
+          text: "thinking",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
        { type: "text", text: "Hello", providerMetadata: { openai: { itemId: "msg_1" } } },
      ])
    }),
@@ -1820,6 +2114,7 @@ describe("OpenAI Responses route", () => {
        Effect.provide(
          fixedResponse(
            sseEvents(
+              { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1" } },
              { type: "response.reasoning_summary_text.delta", item_id: "rs_1", delta: "thinking" },
              {
                type: "response.output_item.done",
@@ -1839,7 +2134,7 @@ describe("OpenAI Responses route", () => {
      expect(response.events).toContainEqual(
        expect.objectContaining({
          type: "reasoning-end",
-          id: "rs_1",
+          id: "rs_1:0",
          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
        }),
      )
@@ -1901,6 +2196,105 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("reconciles reasoning summaries that arrive only as finals", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, { providerOptions: { store: false } }),
+      ).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                item: { type: "reasoning", id: "rs_1", encrypted_content: null },
+              },
+              { type: "response.reasoning_summary_part.added", item_id: "rs_1", summary_index: 0 },
+              // No `.delta` events at all: the gateway sends the complete
+              // summary text in the `.done` final.
+              {
+                type: "response.reasoning_summary_text.done",
+                item_id: "rs_1",
+                summary_index: 0,
+                text: "Checked the diff.",
+              },
+              { type: "response.reasoning_summary_part.done", item_id: "rs_1", summary_index: 0 },
+              {
+                type: "response.output_item.done",
+                item: {
+                  type: "reasoning",
+                  id: "rs_1",
+                  summary: [{ type: "summary_text", text: "Checked the diff." }],
+                  encrypted_content: "encrypted-state",
+                },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Checked the diff.")
+      expect(response.events.filter((event) => event.type.startsWith("reasoning-"))).toEqual([
+        {
+          type: "reasoning-start",
+          id: "rs_1:0",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
+        { type: "reasoning-delta", id: "rs_1:0", text: "Checked the diff.", providerMetadata: undefined },
+        {
+          type: "reasoning-end",
+          id: "rs_1:0",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        },
+      ])
+
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [response.message],
+          providerOptions: { store: false, include: ["reasoning.encrypted_content"] },
+        }),
+      )
+      expect(prepared.body.input).toEqual([
+        {
+          type: "reasoning",
+          id: "rs_1",
+          summary: [{ type: "summary_text", text: "Checked the diff." }],
+          encrypted_content: "encrypted-state",
+        },
+      ])
+    }),
+  )
+
+  it.effect("does not duplicate reasoning finals after streamed deltas", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1", encrypted_content: null } },
+              { type: "response.reasoning_summary_part.added", item_id: "rs_1", summary_index: 0 },
+              { type: "response.reasoning_summary_text.delta", item_id: "rs_1", summary_index: 0, delta: "Streamed" },
+              // Repeats the complete text, as the spec allows.
+              { type: "response.reasoning_summary_text.done", item_id: "rs_1", summary_index: 0, text: "Streamed" },
+              { type: "response.reasoning_summary_part.done", item_id: "rs_1", summary_index: 0 },
+              {
+                type: "response.output_item.done",
+                item: { type: "reasoning", id: "rs_1", encrypted_content: "encrypted-state" },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Streamed")
+      expect(response.events.filter((event) => event.type === "reasoning-delta")).toEqual([
+        { type: "reasoning-delta", id: "rs_1:0", text: "Streamed", providerMetadata: undefined },
+      ])
+    }),
+  )
+
  it.effect("closes reasoning summary parts when storage is not disabled", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(LLMRequest.update(request, { providerOptions: { store: true } })).pipe(
@@ -1980,6 +2374,7 @@ describe("OpenAI Responses route", () => {
              })
              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: "Parser now round-trips reasoning." },
                  { type: "response.completed", response: { id: "resp_1" } },
                ),
@@ -2091,6 +2486,134 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("continues stateless hosted tool results with their text form", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("Search."),
+            Message.assistant([
+              ToolCallPart.make({
+                id: "ws_1",
+                name: "web_search",
+                input: { query: "effect 4" },
+                providerExecuted: true,
+                providerMetadata: { openai: { itemId: "ws_1" } },
+              }),
+              {
+                type: "tool-result",
+                id: "ws_1",
+                name: "web_search",
+                result: { type: "json", value: { type: "web_search_call", id: "ws_1", status: "completed" } },
+                providerExecuted: true,
+                providerMetadata: { openai: { itemId: "ws_1" } },
+              },
+            ]),
+            Message.user("Continue."),
+          ],
+          providerOptions: { store: false },
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        { role: "user", content: [{ type: "input_text", text: "Search." }] },
+        {
+          role: "user",
+          content: [{ type: "input_text", text: '{"type":"web_search_call","id":"ws_1","status":"completed"}' }],
+        },
+        { role: "user", content: [{ type: "input_text", text: "Continue." }] },
+      ])
+    }),
+  )
+
+  it.effect("drops replayed item ids outside the server's grammar", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              // Fails the message id prefix.
+              {
+                type: "text",
+                text: "Hello",
+                providerMetadata: { openai: { itemId: "history_1" } },
+              },
+              // Oversized for the Responses item id limit.
+              {
+                type: "text",
+                text: "World",
+                providerMetadata: { openai: { itemId: `m${"a".repeat(64)}` } },
+              },
+              // Fails the reasoning id prefix, so the whole item is unreplayable
+              // statelessly and is skipped rather than sent malformed.
+              {
+                type: "reasoning",
+                text: "Checked the diff.",
+                providerMetadata: { openai: { itemId: "thinking_1", reasoningEncryptedContent: "encrypted-state" } },
+              },
+              ToolCallPart.make({
+                id: "call_1",
+                name: "lookup",
+                input: { query: "weather" },
+                providerMetadata: { openai: { itemId: "toolu_01A" } },
+              }),
+            ]),
+          ],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        {
+          type: "message",
+          role: "assistant",
+          content: [
+            { type: "output_text", text: "Hello" },
+            { type: "output_text", text: "World" },
+          ],
+        },
+        {
+          type: "function_call",
+          call_id: "call_1",
+          name: "lookup",
+          arguments: '{"query":"weather"}',
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps well-formed hosted references and drops malformed ones under storage", () =>
+    Effect.gen(function* () {
+      const hostedResult = (itemId: string) => [
+        ToolCallPart.make({
+          id: itemId,
+          name: "web_search",
+          input: { query: "effect 4" },
+          providerExecuted: true,
+          providerMetadata: { openai: { itemId } },
+        }),
+        {
+          type: "tool-result" as const,
+          id: itemId,
+          name: "web_search",
+          result: { type: "json" as const, value: { status: "completed" } },
+          providerExecuted: true as const,
+          providerMetadata: { openai: { itemId } },
+        },
+      ]
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [Message.assistant(hostedResult("ws_1")), Message.assistant(hostedResult("bad ref"))],
+          providerOptions: { store: true },
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([{ type: "item_reference", id: "ws_1" }])
+    }),
+  )
+
  it.effect("continues stateless hosted image generation with the generated image", () =>
    Effect.gen(function* () {
      const imageTool = OpenAI.imageGeneration({ action: "edit" })
@@ -2214,15 +2737,15 @@ describe("OpenAI Responses route", () => {
      const body = sseEvents(
        {
          type: "response.output_item.added",
-          item: { type: "function_call", id: "item_1", call_id: "call_1", name: "lookup", arguments: "" },
+          item: { type: "function_call", id: "fc_item_1", call_id: "call_1", name: "lookup", arguments: "" },
        },
-        { type: "response.function_call_arguments.delta", item_id: "item_1", delta: '{"query"' },
-        { type: "response.function_call_arguments.delta", item_id: "item_1", delta: ':"weather"}' },
+        { type: "response.function_call_arguments.delta", item_id: "fc_item_1", delta: '{"query"' },
+        { type: "response.function_call_arguments.delta", item_id: "fc_item_1", delta: ':"weather"}' },
        {
          type: "response.output_item.done",
          item: {
            type: "function_call",
-            id: "item_1",
+            id: "fc_item_1",
            call_id: "call_1",
            name: "lookup",
            arguments: '{"query":"weather"}',
@@ -2251,7 +2774,7 @@ describe("OpenAI Responses route", () => {
          type: "tool-input-start",
          id: "call_1",
          name: "lookup",
-          providerMetadata: { openai: { itemId: "item_1" } },
+          providerMetadata: { openai: { itemId: "fc_item_1" } },
        },
        {
          type: "tool-input-delta",
@@ -2269,7 +2792,7 @@ describe("OpenAI Responses route", () => {
          type: "tool-input-end",
          id: "call_1",
          name: "lookup",
-          providerMetadata: { openai: { itemId: "item_1" } },
+          providerMetadata: { openai: { itemId: "fc_item_1" } },
        },
        {
          type: "tool-call",
@@ -2277,7 +2800,7 @@ describe("OpenAI Responses route", () => {
          name: "lookup",
          input: { query: "weather" },
          providerExecuted: undefined,
-          providerMetadata: { openai: { itemId: "item_1" } },
+          providerMetadata: { openai: { itemId: "fc_item_1" } },
        },
        {
          type: "step-finish",
@@ -2298,7 +2821,7 @@ describe("OpenAI Responses route", () => {
      expect(prepared.body.input).toEqual([
        {
          type: "function_call",
-          id: "item_1",
+          id: "fc_item_1",
          call_id: "call_1",
          name: "lookup",
          arguments: '{"query":"weather"}',
@@ -2312,7 +2835,7 @@ describe("OpenAI Responses route", () => {
      const body = sseEvents(
        {
          type: "response.output_item.added",
-          item: { type: "function_call", id: "item_1", call_id: "call_1", name: "lookup", arguments: "" },
+          item: { type: "function_call", id: "fc_item_1", call_id: "call_1", name: "lookup", arguments: "" },
        },
        { type: "response.completed", response: { usage: { input_tokens: 5, output_tokens: 1 } } },
      )
@@ -2324,7 +2847,7 @@ describe("OpenAI Responses route", () => {
            type: "tool-input-end",
            id: "call_1",
            name: "lookup",
-            providerMetadata: { openai: { itemId: "item_1" } },
+            providerMetadata: { openai: { itemId: "fc_item_1" } },
          },
          {
            type: "tool-call",
@@ -2332,7 +2855,7 @@ describe("OpenAI Responses route", () => {
            name: "lookup",
            input: {},
            providerExecuted: undefined,
-            providerMetadata: { openai: { itemId: "item_1" } },
+            providerMetadata: { openai: { itemId: "fc_item_1" } },
          },
        ],
      )
@@ -2345,14 +2868,14 @@ describe("OpenAI Responses route", () => {
      const body = sseEvents(
        {
          type: "response.output_item.added",
-          item: { type: "function_call", id: "item_1", call_id: "call_1", name: "lookup", arguments: "" },
+          item: { type: "function_call", id: "fc_item_1", call_id: "call_1", name: "lookup", arguments: "" },
        },
-        { type: "response.function_call_arguments.delta", item_id: "item_1", delta: '{"query":"streamed"}' },
+        { type: "response.function_call_arguments.delta", item_id: "fc_item_1", delta: '{"query":"streamed"}' },
        {
          type: "response.output_item.done",
          item: {
            type: "function_call",
-            id: "item_1",
+            id: "fc_item_1",
            call_id: "call_1",
            name: "lookup",
            arguments: '{"query":"partial',
@@ -2384,7 +2907,7 @@ describe("OpenAI Responses route", () => {
          type: "response.output_item.done",
          item: {
            type: "function_call",
-            id: "item_1",
+            id: "fc_item_1",
            call_id: "call_1",
            name: "lookup",
            arguments: '{"query":"partial',
@@ -2413,7 +2936,7 @@ describe("OpenAI Responses route", () => {
                type: "response.output_item.done",
                item: {
                  type: "function_call",
-                  id: "item_1",
+                  id: "fc_item_1",
                  call_id: "call_1",
                  name: "lookup",
                  arguments: '{"query":"weather"}',
@@ -2427,7 +2950,7 @@ describe("OpenAI Responses route", () => {

      expect(response.events.find(LLMEvent.is.toolCall)).toMatchObject({
        id: "call_1",
-        providerMetadata: { openai: { itemId: "item_1" } },
+        providerMetadata: { openai: { itemId: "fc_item_1" } },
      })
    }),
  )
@@ -2471,6 +2994,47 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("decodes computer_call as provider-executed tool-call + tool-result", () =>
+    Effect.gen(function* () {
+      const item = {
+        type: "computer_call",
+        id: "computer_1",
+        call_id: "call_1",
+        status: "completed",
+        action: { type: "click", x: 100, y: 200 },
+      }
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.done", item },
+              { type: "response.completed", response: { usage: { input_tokens: 5, output_tokens: 1 } } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter((event) => event.type === "tool-call" || event.type === "tool-result")).toEqual([
+        {
+          type: "tool-call",
+          id: "computer_1",
+          name: "computer_use",
+          input: { type: "click", x: 100, y: 200 },
+          providerExecuted: true,
+          providerMetadata: { openai: { itemId: "computer_1" } },
+        },
+        {
+          type: "tool-result",
+          id: "computer_1",
+          name: "computer_use",
+          result: { type: "json", value: item },
+          providerExecuted: true,
+          providerMetadata: { openai: { itemId: "computer_1" } },
+        },
+      ])
+    }),
+  )
+
  it.effect("decodes image generation output as image content", () =>
    Effect.gen(function* () {
      const item = {
@@ -2592,7 +3156,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("uses xAI inline file encoding for user PDFs", () =>
+  it.effect("uses standard inline file encoding for xAI user PDFs", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
        LLM.request({
@@ -2615,8 +3179,7 @@ describe("OpenAI Responses route", () => {
            {
              type: "input_file",
              filename: "report.pdf",
-              file_data: "JVBERi0xLjQ=",
-              mime_type: "application/pdf",
+              file_data: "data:application/pdf;base64,JVBERi0xLjQ=",
            },
          ],
        },
@@ -2649,6 +3212,37 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("lowers remote user media URLs without base64 wrapping", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user([
+              { type: "media", mediaType: "image/png", data: "https://example.com/image.png" },
+              {
+                type: "media",
+                mediaType: "application/pdf",
+                data: "https://example.com/report.pdf",
+                filename: "report.pdf",
+              },
+            ]),
+          ],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        {
+          role: "user",
+          content: [
+            { type: "input_image", image_url: "https://example.com/image.png" },
+            { type: "input_file", filename: "report.pdf", file_url: "https://example.com/report.pdf" },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("fails with a typed rate limit for provider error frames", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
@@ -2805,36 +3399,42 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("falls back to a stable default when error is null", () =>
+  it.effect("falls back to the raw payload when error is null", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
        Effect.provide(fixedResponse(sseEvents({ type: "error", error: null }))),
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({ _tag: "UnknownProvider", message: "OpenAI Responses stream error" })
+      expect(error.reason).toMatchObject({ _tag: "UnknownProvider" })
+      expect(error.reason.message).toContain('"error":null')
+      expect(error.body).toBe(error.reason.message)
    }),
  )

-  it.effect("falls back to a stable default when both error and response are absent", () =>
+  it.effect("classifies a detail-free error event as a transient provider failure", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
-        Effect.provide(fixedResponse(sseEvents({ type: "error" }))),
+        Effect.provide(fixedResponse(sseEvents({ type: "error", sequence_number: 2 }))),
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({ _tag: "UnknownProvider", message: "OpenAI Responses stream error" })
+      expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
+      expect(error.reason.message).toContain('"type":"error"')
+      expect(error.body).toBe(error.reason.message)
    }),
  )

-  it.effect("falls back to a stable default when response.failed has no error payload", () =>
+  it.effect("keeps the raw response payload when response.failed has no error payload", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
        Effect.provide(fixedResponse(sseEvents({ type: "response.failed", response: { id: "resp_failed_3" } }))),
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({ _tag: "UnknownProvider", message: "OpenAI Responses response failed" })
+      expect(error.reason).toMatchObject({ _tag: "UnknownProvider" })
+      expect(error.reason.message).toContain('"resp_failed_3"')
+      expect(error.body).toBe(error.reason.message)
    }),
  )

@@ -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 +1,86 @@
+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")
+    }),
+  )
+
+  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("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,
+      })
+    }),
+  )
+})
@@ -1,5 +1,7 @@
 import { HttpRecorder } from "@opencode-ai/http-recorder"
+import { NodeSocket } from "@effect/platform-node"
 import { Layer } from "effect"
+import { Socket } from "effect/unstable/socket"
 import * as path from "node:path"
 import { fileURLToPath } from "node:url"
 import { LLMClient, RequestExecutor } from "../src/route.js"
@@ -16,7 +18,7 @@ import {
 const __dirname = path.dirname(fileURLToPath(import.meta.url))
 const FIXTURES_DIR = path.resolve(__dirname, "fixtures", "recordings")

-type RecordedEnv = RequestExecutorService | LLMClientService | ImageClientService
+type RecordedEnv = RequestExecutorService | LLMClientService | ImageClientService | Socket.WebSocketConstructor

 type RecordedTestsOptions = RecordedGroupOptions & {
  readonly options?: HttpRecorder.RecorderOptions
@@ -69,7 +71,7 @@ export const recordedTests = (options: RecordedTestsOptions) =>
        ...metadata,
      }
      if (recording) {
-        if (process.env.CI !== undefined) throw new Error("Unset CI before recording HTTP cassettes")
+        if (process.env.CI !== undefined) throw new Error("Unset CI before recording cassettes")
        HttpRecorder.removeCassetteSync(cassette, { directory: FIXTURES_DIR })
      }
      const requestExecutor = RequestExecutor.layer.pipe(
@@ -81,10 +83,16 @@ export const recordedTests = (options: RecordedTestsOptions) =>
          }),
        ),
      )
+      const webSocket = HttpRecorder.layerWebSocketConstructor(cassette, {
+        ...recorderOptions,
+        directory: FIXTURES_DIR,
+        metadata: recorderMetadata,
+      }).pipe(Layer.provide(NodeSocket.layerWebSocketConstructorWS))
      return Layer.mergeAll(
        requestExecutor,
        LLMClient.layer.pipe(Layer.provide(requestExecutor)),
        ImageClient.layer.pipe(Layer.provide(requestExecutor)),
+        webSocket,
      )
    },
  })
@@ -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" },
+            },
+          },
+        ],
      })
    }),
  )
@@ -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"
@@ -11,7 +11,7 @@ import {
 } from "./timeline-test-helpers"
 import { waitForStableTimeline } from "./session-tab-switch-probe"

-const contentSelector = '[data-message-id], [data-component="prompt-input"]'
+const contentSelector = '[data-message-id], [data-component="composer-editor"]'
 const draftID = "draft_first_navigation"

 benchmark.describe("performance: first navigation paint", () => {
@@ -41,11 +41,42 @@ benchmark.describe("performance: first navigation paint", () => {
      href,
      destinationPath: href,
      sourceSelector: messageSelector(fixture.expected.sourceMessageIDs.at(-1)!),
-      destinationSelector: '[data-component="prompt-input"]',
+      destinationSelector: '[data-component="composer-editor"]',
      contentSelector,
      navigate: async () => {
        await page.locator(`[data-slot="titlebar-tabs"] a[href="${href}"]`).first().click()
-        await expect(page.locator('[data-component="prompt-input"]')).toBeVisible()
+        await expect(page.locator('[data-component="composer-editor"]')).toBeVisible()
+      },
+    })
+    report(result)
+    expect(result.summary.blankSamples).toBe(0)
+    expect(result.summary.unknownSamples).toBe(0)
+  })
+
+  benchmark("opens a session from the new session page without a blank frame", async ({ page, report }) => {
+    await mockStressTimeline(page)
+    await installTimelineSettings(page)
+    await installStressSessionTabs(page, { draftID })
+    await page.goto("/")
+
+    const draftHref = stressDraftHref(draftID)
+    const draftTab = page.locator(`[data-slot="titlebar-tabs"] a[href="${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)
@@ -5,9 +5,11 @@ import { mockOpenCodeServer } from "../../utils/mock-server"
 test("applies message latency after a list response gate is released", async () => {
  const events: string[] = []
  const gate = Promise.withResolvers<void>()
+  const started = Promise.withResolvers<void>()
  let handler: ((route: Route) => Promise<void>) | undefined
  const page = {
    addInitScript: () => Promise.resolve(),
+    on: () => page,
    route: (_url: string, callback: (route: Route) => Promise<void>) => {
      handler = callback
      return Promise.resolve()
@@ -21,6 +23,7 @@ test("applies message latency after a list response gate is released", async ()
    messageDelay: 25,
    beforeMessagesResponse: () => {
      events.push("before")
+      started.resolve()
      return gate.promise
    },
    onMessages: (request) => events.push(request.phase),
@@ -31,12 +34,18 @@ test("applies message latency after a list response gate is released", async ()
  })

  const response = handler!({
-    request: () => ({ url: () => "http://127.0.0.1:4096/api/session/session/message" }),
+    request: () => ({
+      url: () => "http://127.0.0.1:4096/api/session/session/message",
+      method: () => "GET",
+      headers: () => ({}),
+      postDataBuffer: () => null,
+    }),
    fulfill: () => {
      events.push("fulfill")
      return Promise.resolve()
    },
  } as unknown as Route)
+  await started.promise
  expect(events).toEqual(["start", "before"])

  const released = performance.now()
@@ -45,3 +54,42 @@ test("applies message latency after a list response gate is released", async ()
  expect(performance.now() - released).toBeGreaterThanOrEqual(20)
  expect(events).toEqual(["start", "before", "page", "end", "fulfill"])
 })
+
+test("routes requests through the HttpApi contract", async () => {
+  const connected = Promise.withResolvers<{ integrationID: string; body: unknown }>()
+  let handler: ((route: Route) => Promise<void>) | undefined
+  const page = {
+    addInitScript: () => Promise.resolve(),
+    on: () => page,
+    route: (_url: string, callback: (route: Route) => Promise<void>) => {
+      handler = callback
+      return Promise.resolve()
+    },
+  } as unknown as Page
+  await mockOpenCodeServer(page, {
+    provider: {},
+    directory: "C:/OpenCode",
+    project: {},
+    sessions: [],
+    pageMessages: () => ({ items: [] }),
+    onConnectKey: connected.resolve,
+  })
+
+  const body = Buffer.from(JSON.stringify({ key: "secret" }))
+  let status: number | undefined
+  await handler!({
+    request: () => ({
+      url: () => "http://127.0.0.1:4096/api/integration/anthropic/connect/key",
+      method: () => "POST",
+      headers: () => ({ "content-type": "application/json" }),
+      postDataBuffer: () => body,
+    }),
+    fulfill: (response: Parameters<Route["fulfill"]>[0]) => {
+      status = response?.status
+      return Promise.resolve()
+    },
+  } as unknown as Route)
+
+  expect(status).toBe(204)
+  expect(await connected.promise).toEqual({ integrationID: "anthropic", body: { key: "secret" } })
+})
@@ -3,9 +3,9 @@ import { base64Encode } from "@opencode-ai/util/encode"
 import { mockOpenCodeServer } from "../utils/mock-server"
 import { expectAppVisible } from "../utils/waits"

-const directory = "C:/OpenCode/PromptInputV2Editing"
-const projectID = "proj_prompt_input_v2_editing"
-const sessionID = "ses_prompt_input_v2_editing"
+const directory = "C:/OpenCode/ComposerEditing"
+const projectID = "proj_composer_editing"
+const sessionID = "ses_composer_editing"
 const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`

 test("preserves the draft when a populated command menu triggers a built-in", async ({ page }) => {
@@ -15,7 +15,7 @@ test("preserves the draft when a populated command menu triggers a built-in", as
      id: projectID,
      worktree: directory,
      vcs: "git",
-      name: "prompt-input-v2-editing",
+      name: "composer-editing",
      time: { created: 1700000000000, updated: 1700000000000 },
      sandboxes: [],
    },
@@ -23,10 +23,10 @@ test("preserves the draft when a populated command menu triggers a built-in", as
    sessions: [
      {
        id: sessionID,
-        slug: "prompt-input-v2-editing",
+        slug: "composer-editing",
        projectID,
        directory,
-        title: "Prompt input V2 editing",
+        title: "Composer editing",
        version: "dev",
        time: { created: 1700000000000, updated: 1700000000000 },
      },
@@ -34,11 +34,11 @@ test("preserves the draft when a populated command menu triggers a built-in", as
    pageMessages: () => ({ items: [] }),
  })
  await page.goto(`/server/${base64Encode(server)}/session/${sessionID}`)
-  const composer = page.locator('[data-component="prompt-input-v2"]')
-  const input = composer.locator('[data-component="prompt-input"]')
-  await expect.poll(() => input.evaluate((element) => getComputedStyle(element, "::before").content)).toBe(
-    `"${String.fromCodePoint(0x200b)}"`,
-  )
+  const composer = page.locator('[data-component="composer"]')
+  const input = composer.locator('[data-component="composer-editor"]')
+  await expect
+    .poll(() => input.evaluate((element) => getComputedStyle(element, "::before").content))
+    .toBe(`"${String.fromCodePoint(0x200b)}"`)
  await expectAppVisible(composer)

  await input.fill("keep me")
@@ -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()
+})
@@ -46,7 +46,7 @@ test("matches the rounded panel corners to the dark new-session background", asy
  )

  await page.goto(`/new-session?draftId=${draftID}`)
-  await expectAppVisible(page.locator('[data-component="prompt-input"]'))
+  await expectAppVisible(page.locator('[data-component="composer-editor"]'))
  await expect(page.locator("html")).toHaveAttribute("data-color-scheme", "dark")
  const panel = page.locator('main div[class*="rounded-[10px]"][class*="overflow-hidden"]')
  await expect(panel).toHaveCount(1)
@@ -60,7 +60,7 @@ test("expands a folder whose path has a trailing Windows separator", async ({ pa
      if (path) return []
      return [
        {
-          name: "frontend",
+          name: "",
          path: "frontend\\",
          absolute: `${directory}/frontend`,
          type: "directory" as const,
@@ -116,6 +116,7 @@ test("expands a folder whose path has a trailing Windows separator", async ({ pa

  const frontendRow = panel.locator('[data-slot="file-tree-v2-row"][data-path="frontend"]')
  await expect(frontendRow).toBeVisible()
+  await expect(frontendRow.getByText("frontend", { exact: true })).toBeVisible()
  await expect(frontendRow).toHaveAttribute("aria-expanded", "false")
  await frontendRow.click()
  await expect(frontendRow).toHaveAttribute("aria-expanded", "true")
@@ -8,7 +8,7 @@ const projectID = "proj_prompt_thinking_level_regression"
 const sessionID = "ses_prompt_thinking_level_regression"
 const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`

-test("shows the V2 thinking level control while relevant", async ({ page }) => {
+test("shows the thinking level control while relevant", async ({ page }) => {
  await mockOpenCodeServer(page, {
    directory,
    project: {
@@ -51,8 +51,8 @@ test("shows the V2 thinking level control while relevant", async ({ page }) => {
    pageMessages: () => ({ items: [] }),
  })
  await page.goto(`/server/${base64Encode(server)}/session/${sessionID}`)
-  const composer = page.locator('[data-component="prompt-input-v2"]')
-  const input = composer.locator('[data-component="prompt-input"]')
+  const composer = page.locator('[data-component="composer"]')
+  const input = composer.locator('[data-component="composer-editor"]')
  const control = composer.getByRole("button", { name: "Choose model variant" })
  await expectAppVisible(composer)

@@ -22,7 +22,7 @@ test("session settings use the remote server context", async ({ page }) => {
  await expect(page.getByRole("heading", { name: sessionB.title, exact: true })).toBeVisible()
  await page.keyboard.press("Control+,")

-  const dialog = page.locator(".settings-v2-dialog")
+  const dialog = page.locator(".settings-dialog")
  const autoAccept = dialog.locator('[data-action="settings-auto-accept-permissions"]')
  const input = autoAccept.getByRole("switch")
  await expect(autoAccept).toBeVisible()
@@ -63,7 +63,7 @@ test("auto-accept responds for an unfocused server session", async ({ page }) =>
  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-v2-dialog").locator('[data-action="settings-auto-accept-permissions"]')
+  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
@@ -224,6 +224,8 @@ async function mockServers(page: Page, permissionRequests: string[], permissionR
    if (currentSessionInfo) 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`))
+      return json(route, { data: [] })
    if (url.pathname === "/api/location") return json(route, { directory })
    if (url.pathname === "/api/vcs")
      return json(route, { location: { directory }, data: { branch: "main", defaultBranch: "main" } })
@@ -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,
@@ -76,7 +76,7 @@ test("keeps the review tree and terminal sized when both panels are open", async
      }),
    })
  })
-  await page.route("**/pty*", (route) =>
+  await page.route(/\/api\/pty(?:\?.*)?$/, (route) =>
    route.fulfill({
      status: 200,
      contentType: "application/json",
@@ -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) })
+}
@@ -0,0 +1,79 @@
+import type { SessionMessageInfo } from "@opencode-ai/client/promise"
+import { base64Encode } from "@opencode-ai/util/encode"
+import { expect, test } from "@playwright/test"
+import { mockOpenCodeServer } from "../utils/mock-server"
+import { expectSessionTitle } from "../utils/waits"
+
+const directory = "C:/OpenCode/SessionMessageRevert"
+const projectID = "proj_session_message_revert"
+const sessionID = "ses_session_message_revert"
+const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
+const messages = [
+  { id: "msg_first", type: "user", text: "First prompt", time: { created: 1 } },
+  {
+    id: "msg_first_reply",
+    type: "assistant",
+    agent: "build",
+    model: { id: "test", providerID: "opencode" },
+    content: [{ type: "text", text: "First reply" }],
+    time: { created: 2, completed: 3 },
+  },
+  { id: "msg_second", type: "user", text: "Second prompt", time: { created: 4 } },
+] satisfies SessionMessageInfo[]
+
+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 }),
+    onRevertStage: (input) => staged.push(input),
+  })
+  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()
+  const response = page.waitForResponse(
+    (response) =>
+      response.request().method() === "POST" &&
+      new URL(response.url()).pathname === `/api/session/${sessionID}/revert/stage`,
+  )
+  await message.getByRole("button", { name: "Revert message" }).click()
+  expect((await response).ok()).toBe(true)
+
+  await expect(page.getByRole("textbox", { name: "Prompt" })).toHaveText("Second prompt")
+  expect(staged).toEqual([{ sessionID, messageID: "msg_second" }])
+})
@@ -120,7 +120,7 @@ test("restores the draft caret before typing after a request dock closes", async
  await transport.waitForConnection()
  await expectSessionTitle(page, title)

-  const editor = page.locator('[data-component="prompt-input"][contenteditable="true"]')
+  const editor = page.locator('[data-component="composer-editor"][contenteditable="true"]')
  const draft = "keep the caret at the end"
  await editor.fill(draft)
  await page.evaluate(() => new Promise<void>((resolve) => requestAnimationFrame(() => resolve())))
@@ -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}:${editPartID}`,
+      rowKey: `assistant-part:part:${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)
 })
@@ -2,7 +2,7 @@ import { expect, test } from "@playwright/test"
 import { assistantMessage, setupTimeline, toolPart, userMessage } from "../performance/timeline-stability/fixture"
 import { createTwoFilesPatch } from "diff"

-test("preserves nested patch file state through outer collapse and reopen", async ({ page }) => {
+test("keeps patch file disclosures independent", async ({ page }) => {
  const patchID = "prt_nested_patch"
  const files = [patchFile("src/a.ts", "modified"), patchFile("src/b.ts", "added"), patchFile("src/old.ts", "deleted")]
  await setupTimeline(page, {
@@ -21,15 +21,17 @@ test("preserves nested patch file state through outer collapse and reopen", asyn
    settings: { editToolPartsExpanded: true },
  })
  const wrapper = page.locator(`[data-timeline-part-id="${patchID}"]`)
-  const outer = wrapper.locator('[data-slot="collapsible-trigger"]').first()
+  const modified = wrapper.locator('[data-scope="apply-patch"] [data-type="update"]')
  const deleted = wrapper.locator('[data-scope="apply-patch"] [data-type="delete"]')
+  await expect(wrapper.locator('[data-scope="apply-patch"] [aria-expanded="false"]')).toHaveCount(3)
  await deleted.getByRole("button").click()
  await expect(deleted.getByRole("button")).toHaveAttribute("aria-expanded", "true")
-  await outer.click()
-  await expect(outer).toHaveAttribute("aria-expanded", "false")
-  await outer.click()
-  await expect(outer).toHaveAttribute("aria-expanded", "true")
-  await expect(deleted.getByRole("button")).toHaveAttribute("aria-expanded", "true")
+  await expect(modified.getByRole("button")).toHaveAttribute("aria-expanded", "false")
+  await modified.getByRole("button").click()
+  await expect(modified.getByRole("button")).toHaveAttribute("aria-expanded", "true")
+  await deleted.getByRole("button").click()
+  await expect(deleted.getByRole("button")).toHaveAttribute("aria-expanded", "false")
+  await expect(modified.getByRole("button")).toHaveAttribute("aria-expanded", "true")
 })

 function patchFile(file: string, status: "added" | "modified" | "deleted") {
@@ -17,7 +17,7 @@ import { mockOpenCodeServer } from "../utils/mock-server"
 import { installSseTransport } from "../utils/sse-transport"
 import { expectSessionTitle } from "../utils/waits"

-const messagePageSize = 200
+const messagePageSize = 20
 const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const messages = Array.from({ length: messagePageSize / 2 + 1 }, (_, index) => {
  const id = `msg_${String(index + 1001).padStart(4, "0")}_history_root_user`
@@ -1,5 +1,6 @@
 import { expect, test } from "@playwright/test"
 import {
+  assistantID,
  assistantMessage,
  completedAssistantInfo,
  messageUpdated,
@@ -8,9 +9,13 @@ import {
  renderedPartID,
  setupTimeline,
  shell,
+  sessionID,
  status,
  stepStarted,
  textPart,
+  toolCalled,
+  toolInputEnded,
+  toolInputStarted,
  userMessage,
 } from "../performance/timeline-stability/fixture"

@@ -34,7 +39,56 @@ for (const expanded of [false, true]) {
  })
 }

-test("shows and expands a running shell command without shimmering it", async ({ page }) => {
+test("transitions a streaming shell from writing through command execution", async ({ page }) => {
+  const id = "prt_shell_streaming_input"
+  const command = "printf ready"
+  const timeline = await setupTimeline(page, {
+    messages: [userMessage(), assistantMessage([], { completed: false })],
+  })
+  await timeline.send(toolInputStarted({ sessionID, assistantMessageID: assistantID, id, name: "shell" }))
+
+  const tool = page.locator(`[data-timeline-part-id="${id}"]`)
+  const title = tool.locator('[data-slot="basic-tool-tool-title"]')
+  const titleShimmer = title.locator('[data-component="text-shimmer"]')
+  const subtitle = tool.locator('[data-slot="basic-tool-tool-subtitle"]')
+  await expect(titleShimmer).toHaveAttribute("aria-label", "Shell")
+  await expect(titleShimmer).toHaveAttribute("data-active", "true")
+  await expect(subtitle).toHaveText("Writing command...")
+  await expect(subtitle.locator('[data-component="text-shimmer"]')).toHaveCount(0)
+  await expect(tool.locator('[data-component="shell-submessage"]')).toHaveCount(0)
+  await expect(tool.locator('[data-slot="collapsible-trigger"]')).toHaveCSS("height", "28px")
+  await expect(tool.locator('[data-component="tool-trigger"]')).toHaveCSS("gap", "6px")
+  await expect(title).toHaveCSS("font-size", "13px")
+  await expect(title).toHaveCSS("font-family", "Inter, sans-serif")
+  await expect(title).toHaveCSS("font-weight", "530")
+  await expect(title).toHaveCSS("line-height", "16px")
+  await expect(title).toHaveCSS("color", "rgb(22, 22, 22)")
+  await expect(subtitle).toHaveCSS("font-size", "13px")
+  await expect(subtitle).toHaveCSS("font-family", "Inter, sans-serif")
+  await expect(subtitle).toHaveCSS("font-weight", "440")
+  await expect(subtitle).toHaveCSS("line-height", "16px")
+  await expect(subtitle).toHaveCSS("color", "rgb(92, 92, 92)")
+
+  const input = JSON.stringify({ command })
+  await timeline.send(toolInputEnded({ sessionID, assistantMessageID: assistantID, id, text: input }))
+  await expect(titleShimmer).toHaveAttribute("data-active", "true")
+  await expect(subtitle).toHaveText(command)
+  await expect(tool).not.toContainText("Writing command...")
+
+  await timeline.send(
+    toolCalled({
+      sessionID,
+      assistantMessageID: assistantID,
+      id,
+      input: { command },
+      executed: true,
+    }),
+  )
+  await expect(titleShimmer).toHaveAttribute("data-active", "true")
+  await expect(subtitle).toHaveText(command)
+})
+
+test("shimmers and expands a running shell command", async ({ page }) => {
  const id = "prt_shell_running_command"
  const command = "sleep 10 && echo done"
  await setupTimeline(page, {
@@ -44,8 +98,10 @@ test("shows and expands a running shell command without shimmering it", async ({

  const tool = page.locator(`[data-timeline-part-id="${id}"]`)
  await expect(tool.locator('[data-component="text-shimmer"]')).toHaveAttribute("data-active", "true")
+  await expect(tool).not.toContainText("Writing command...")
  await expect(tool.locator('[data-component="shell-submessage"]')).toHaveText(command)
  await expect(tool.locator('[data-component="shell-submessage"] [data-component="text-shimmer"]')).toHaveCount(0)
+  await expect(tool.locator('[data-slot="collapsible-trigger"]')).toHaveCSS("height", "28px")
  await tool.locator('[data-slot="collapsible-trigger"]').click()
  await expect(tool.locator('[data-slot="collapsible-trigger"]')).toHaveAttribute("aria-expanded", "true")
  await expect(tool.locator('[data-slot="bash-pre"]')).toContainText("still running")
@@ -1,6 +1,15 @@
 import { expect, test } from "@playwright/test"
 import type { SessionMessageAssistant, SessionMessageInfo } from "@opencode-ai/client/promise"
-import { event, session, sessionID, setupTimeline } from "../performance/timeline-stability/fixture"
+import {
+  compactionDelta,
+  compactionEnded,
+  compactionFailed,
+  compactionStarted,
+  event,
+  session,
+  sessionID,
+  setupTimeline,
+} from "../performance/timeline-stability/fixture"

 const user = { id: "msg_user", type: "user", text: "Run it", time: { created: 1 } } satisfies SessionMessageInfo

@@ -73,6 +82,160 @@ test("renders current protocol notices in CLI order", async ({ page }) => {
  expect(ownerWarnings).toEqual([])
 })

+test("renders a compaction summary while it streams and after completion", async ({ page }) => {
+  const timeline = await setupTimeline(page, { sessionMessages: [user, assistant(true)] })
+
+  await timeline.send(
+    compactionStarted({
+      sessionID,
+      reason: "manual",
+      recent: "",
+    }),
+  )
+
+  const compaction = page.locator('[data-component="session-compaction-message"]')
+  await expect(compaction.getByText("Session compacted", { exact: true })).toBeVisible()
+
+  await timeline.send(
+    compactionDelta({
+      sessionID,
+      text: "## Checkpoint\n\nStreamed implementation details.",
+    }),
+  )
+  await expect(compaction.getByRole("heading", { name: "Checkpoint" })).toBeVisible()
+  await expect(compaction).toContainText("Streamed implementation details.")
+
+  await timeline.send(
+    compactionEnded({
+      sessionID,
+      reason: "manual",
+      text: "## Checkpoint\n\nFinal implementation details.",
+      recent: "",
+    }),
+  )
+  await expect(compaction).toContainText("Final implementation details.")
+  await expect(compaction).not.toContainText("Streamed implementation details.")
+})
+
+test("updates running compactions to failed and cancelled boundaries", async ({ page }) => {
+  const timeline = await setupTimeline(page, { sessionMessages: [user, assistant(true)] })
+
+  await timeline.send(compactionStarted({ sessionID, reason: "auto", recent: "" }))
+  await timeline.send(compactionDelta({ sessionID, text: "Partial summary that should be discarded." }))
+  await timeline.send(
+    compactionFailed({
+      sessionID,
+      reason: "auto",
+      error: {
+        type: "compaction.failed",
+        message: 'Error: {"error":{"type":"ProviderError","message":"The provider rejected the summary."}}',
+      },
+    }),
+  )
+
+  const compactions = page.locator('[data-component="session-compaction-message"]')
+  const failed = compactions.filter({ hasText: "The provider rejected the summary." })
+  await expect(failed.getByText("Session compacted", { exact: true })).toBeVisible()
+  await expect(failed.getByText("ProviderError: The provider rejected the summary.", { exact: true })).toBeVisible()
+  await expect(failed).not.toContainText("Partial summary that should be discarded.")
+
+  await timeline.send(compactionStarted({ sessionID, reason: "manual", recent: "" }))
+  await timeline.send(
+    compactionFailed({
+      sessionID,
+      reason: "manual",
+      error: { type: "aborted", message: "Cancellation detail should stay hidden." },
+    }),
+  )
+
+  await expect(compactions).toHaveCount(2)
+  const cancelled = compactions.filter({ hasNotText: "The provider rejected the summary." })
+  await expect(cancelled.getByText("Session compacted", { exact: true })).toBeVisible()
+  await expect(cancelled).not.toContainText("Cancellation detail should stay hidden.")
+})
+
+test("shows a delegating row while subagent input streams", async ({ page }) => {
+  await setupTimeline(page, {
+    sessionMessages: [
+      user,
+      {
+        ...assistant(false),
+        content: [
+          {
+            type: "tool",
+            id: "call_subagent",
+            name: "subagent",
+            state: { status: "streaming", input: "" },
+            time: { created: 2 },
+          },
+        ],
+      },
+    ],
+  })
+
+  const delegating = page.locator('[data-component="task-tool-delegating"]')
+  await expect(delegating).toBeVisible()
+  const shimmer = delegating.locator('[data-component="text-shimmer"]')
+  await expect(shimmer).toHaveAttribute("aria-label", "Delegating agent...")
+  await expect(shimmer).toHaveCSS("line-height", "16px")
+  const icon = delegating.locator('[data-slot="icon-svg"]')
+  await expect(icon.locator('use[href="#opencode-v2-icon-subagent"]')).toBeVisible()
+  await expect(icon).toHaveCSS("color", "rgb(174, 174, 174)")
+  await expect(page.locator('[data-component="task-tool-card"]')).toHaveCount(0)
+  await expect(page.locator('[data-timeline-row="Thinking"]')).toHaveCount(0)
+})
+
+test("renders the moved location notice in its compact timeline style", async ({ page }) => {
+  const directory = `/Users/usrnk1/Developer/opencode/${"nested-directory/".repeat(24)}session`
+  await page.setViewportSize({ width: 480, height: 720 })
+  await setupTimeline(page, {
+    sessionMessages: [
+      user,
+      {
+        id: "msg_location",
+        type: "location-switched",
+        location: { directory },
+        time: { created: 2 },
+      },
+    ],
+  })
+
+  const notice = page.locator('[data-slot="session-timeline-notice"][data-type="location-switched"]')
+  const label = notice.locator('[data-slot="session-timeline-notice-label"]')
+  const value = notice.locator('[data-slot="session-timeline-notice-value"]')
+  const tooltipTrigger = notice.locator('[data-component="tooltip-v2-trigger"]')
+
+  await expect(label).toHaveText("Moved to")
+  await expect(value).toHaveText(directory)
+  await expect(notice).not.toContainText("·")
+  await expect(notice.locator("svg")).toHaveCount(0)
+  await expect(notice).toHaveCSS("height", "28px")
+  await expect(notice).toHaveCSS("gap", "8px")
+  await expect(notice).toHaveCSS("padding-top", "4px")
+  await expect(notice).toHaveCSS("padding-bottom", "4px")
+  await expect(label).toHaveCSS("font-size", "13px")
+  await expect(label).toHaveCSS("font-weight", "530")
+  await expect(label).toHaveCSS("line-height", "16px")
+  await expect(label).toHaveCSS("color", "rgb(128, 128, 128)")
+  await expect(value).toHaveCSS("font-size", "13px")
+  await expect(value).toHaveCSS("font-weight", "440")
+  await expect(value).toHaveCSS("line-height", "16px")
+  await expect(value).toHaveCSS("color", "rgb(128, 128, 128)")
+  await expect(value).toHaveCSS("text-overflow", "ellipsis")
+  await expect(value).toHaveCSS("white-space", "nowrap")
+  await expect(value).toHaveAttribute("dir", "ltr")
+  await expect.poll(() => value.evaluate((element) => element.scrollWidth > element.clientWidth)).toBe(true)
+
+  const tooltip = page.getByText("Session working directory changed", { exact: true })
+  await label.hover()
+  await expect(tooltip).toBeVisible()
+  await page.mouse.move(0, 0)
+  await expect(tooltip).toBeHidden()
+  await tooltipTrigger.focus()
+  await expect(tooltipTrigger).toBeFocused()
+  await expect(tooltip).toBeVisible()
+})
+
 test("moves blocking work to the background with Ctrl+B", async ({ page }) => {
  await setupTimeline(page, { sessionMessages: [user, assistant(false, true)] })
  const card = page.locator('[data-component="task-tool-card"]')
@@ -81,7 +244,24 @@ test("moves blocking work to the background with Ctrl+B", async ({ page }) => {
  await expect(card).not.toContainText("(background)")
  await expect(page.getByText("Called `subagent`", { exact: false })).toHaveCount(0)
  await expect(page.locator('[data-component="background-tool-control"]')).toHaveCount(0)
-  await expect(page.locator('[data-action="session-background-toggle"]')).toContainText("Move 1 subagent to background")
+  const hint = page.locator('[data-component="session-background-hint"]')
+  const hintPrefix = hint.locator('[data-slot="session-background-hint-prefix"]')
+  await expect(hint).toBeVisible()
+  await expect(page.locator('[data-timeline-row="Thinking"]')).toHaveCount(0)
+  await expect
+    .poll(async () => {
+      const [cardBox, hintBox, prefixBox] = await Promise.all([
+        card.boundingBox(),
+        hint.boundingBox(),
+        hintPrefix.boundingBox(),
+      ])
+      if (!cardBox || !hintBox || !prefixBox) return undefined
+      return {
+        aligned: Math.abs(cardBox.x - prefixBox.x) < 2,
+        ordered: cardBox.y < hintBox.y,
+      }
+    })
+    .toEqual({ aligned: true, ordered: true })

  const request = page.waitForRequest(
    (request) =>
@@ -104,10 +284,10 @@ test("navigates from a running subagent card and hides background controls in th
    sessionStatus: { [sessionID]: { type: "busy" }, [childID]: { type: "busy" } },
  })

-  await expect(page.locator('[data-action="session-background-toggle"]')).toContainText("Move 1 subagent to background")
+  await expect(page.getByText(/move running work to the background/i)).toBeVisible()
  await page.locator('[data-component="task-tool-card"]').click()
  await expect(page).toHaveURL(new RegExp(`/session/${childID}$`))
-  await expect(page.locator('[data-component="session-background-dock"]')).toHaveCount(0)
+  await expect(page.getByText(/move running work to the background/i)).toHaveCount(0)
 })

 test("shows a badge for active background work", async ({ page }) => {
@@ -118,7 +298,14 @@ test("shows a badge for active background work", async ({ page }) => {
    sessionStatus: { [childID]: { type: "busy" } },
  })

-  await expect(page.locator('[data-component="session-background-dock"]')).toContainText("1 subagent in background")
+  await page.getByRole("button", { name: "Session details" }).click()
+  const summary = page.getByRole("button", { name: "1 item running in background" })
+  await expect(summary).toContainText("1")
+  await expect(summary).toContainText("Running work in background")
+  await summary.click()
+  await expect(
+    page.locator('[data-component="session-background-list"]').getByText("Agent", { exact: true }),
+  ).toBeVisible()
 })

 test("separates blocking and already-backgrounded work into two rows", async ({ page }) => {
@@ -193,10 +380,15 @@ test("separates blocking and already-backgrounded work into two rows", async ({
    },
  })

-  const dock = page.locator('[data-component="session-background-dock"]')
  const backgroundCard = page.locator('[data-timeline-part-id="call_backgrounded"]')
-  await expect(dock).toContainText("Move 1 subagent to background")
-  await expect(dock.getByText("Running 1 shell and 1 subagent in background", { exact: true })).toBeVisible()
+  await expect(page.getByText(/move running work to the background/i)).toBeVisible()
+  await page.getByRole("button", { name: "Session details" }).click()
+  const summary = page.getByRole("button", { name: "2 items running in background" })
+  await expect(summary).toContainText("2")
+  await summary.click()
+  const list = page.locator('[data-component="session-background-list"]')
+  await expect(list).toContainText("Background task")
+  await expect(list).toContainText("sleep 120")
  await expect(backgroundCard).toContainText("Background task (background)")
  await expect(backgroundCard.locator('[data-component="session-progress-indicator-v2"]')).toBeVisible()
  await expect(
@@ -1,6 +1,7 @@
 import { expect, test } from "@playwright/test"
 import {
  assistantMessage,
+  partUpdated,
  setupTimeline,
  status,
  toolPart,
@@ -69,9 +70,126 @@ test.describe("session timeline projection", () => {
    ]) {
      await expect(page.locator(`[data-timeline-part-id="${id}"]`).first(), id).toBeVisible()
    }
+    const patch = page.locator('[data-timeline-part-id="prt_patch"]')
+    await expect(patch.getByText("1 file", { exact: true })).toBeVisible()
+    await expect(patch.getByRole("button", { name: "Patch 1 file", exact: true })).toHaveCount(0)
+    await expect(patch.getByRole("button")).toHaveCount(1)
+    await expect(patch.locator('[data-scope="apply-patch"] button[aria-expanded="false"]')).toHaveCount(1)
+    await expect(patch.locator('[data-slot="message-part-title-filename"]')).toHaveCount(0)
+    await expect(patch.locator('[data-slot="message-part-actions"]')).toHaveCount(0)
+    const edit = page.locator('[data-timeline-part-id="prt_edit"]')
+    await expect(edit.locator('[data-component="apply-patch-tool"]')).toBeVisible()
+    await expect(edit.locator('[data-slot="basic-tool-tool-title"]')).toContainText("Edit")
    await expect(page.locator('[data-timeline-part-id="prt_todo"]')).toHaveCount(0)
  })

+  test("combines adjacent patch calls and repeated files into one group", async ({ page }) => {
+    const first = "prt_patch_first"
+    const second = "prt_patch_second"
+    const timeline = await setupTimeline(page, {
+      messages: [
+        userMessage(),
+        assistantMessage([
+          toolPart(
+            first,
+            "patch",
+            "completed",
+            { patchText: "Update src/first.ts" },
+            {
+              metadata: { files: [patchFile("src/first.ts", "modified")] },
+            },
+          ),
+        ]),
+      ],
+    })
+
+    const initial = page.locator(`[data-timeline-part-id="${first}"]`)
+    const initialFile = initial.locator('[data-scope="apply-patch"] [data-type="update"]')
+    await expect(initialFile).toBeVisible()
+    await initialFile.getByRole("button").click()
+    await expect(initialFile.getByRole("button")).toHaveAttribute("aria-expanded", "true")
+    await initial.evaluate((element) => {
+      const row = element.closest<HTMLElement>("[data-timeline-key]")
+      if (row) row.dataset.patchRow = "stable"
+    })
+
+    await timeline.send(
+      partUpdated(toolPart(second, "patch", "running", { patchText: "Update more files" }, { metadata: {} })),
+    )
+
+    const group = page.locator(`[data-timeline-part-ids="${first},${second}"]`)
+    await expect(group.locator("xpath=ancestor::*[@data-timeline-key]")).toHaveAttribute("data-patch-row", "stable")
+    await expect(group.locator('[data-slot="apply-patch-filename"]')).toHaveText(["first.ts"])
+    await expect(group.locator('[data-scope="apply-patch"] [data-type="update"] button')).toHaveAttribute(
+      "aria-expanded",
+      "true",
+    )
+
+    await timeline.send(
+      partUpdated(
+        toolPart(
+          second,
+          "patch",
+          "completed",
+          { patchText: "Update more files" },
+          {
+            metadata: {
+              files: [patchFile("src/first.ts", "modified"), patchFile("src/second.ts", "added")],
+            },
+          },
+        ),
+      ),
+    )
+
+    await expect(group.locator('[data-slot="apply-patch-filename"]')).toHaveText(["first.ts", "second.ts"])
+    await expect(group.locator('[data-scope="apply-patch"] [data-type="update"] button')).toHaveAttribute(
+      "aria-expanded",
+      "true",
+    )
+    await expect(group.locator('[data-scope="apply-patch"] [data-type="add"] button')).toHaveAttribute(
+      "aria-expanded",
+      "false",
+    )
+    await expect(page.locator(`[data-timeline-part-id="${first}"], [data-timeline-part-id="${second}"]`)).toHaveCount(0)
+  })
+
+  test("combines adjacent edit calls and repeated files into one group", async ({ page }) => {
+    const first = "prt_edit_first"
+    const second = "prt_edit_second"
+    await setupTimeline(page, {
+      messages: [
+        userMessage(),
+        assistantMessage([
+          toolPart(
+            first,
+            "edit",
+            "completed",
+            { path: "src/first.ts", oldString: "one", newString: "two" },
+            {
+              metadata: { files: [patchFile("src/first.ts", "modified")] },
+            },
+          ),
+          toolPart(
+            second,
+            "edit",
+            "completed",
+            { path: "src/first.ts", oldString: "two", newString: "three" },
+            {
+              metadata: { files: [patchFile("src/first.ts", "modified")] },
+            },
+          ),
+        ]),
+      ],
+      settings: { editToolPartsExpanded: true },
+    })
+
+    const group = page.locator(`[data-timeline-part-ids="${first},${second}"]`)
+    await expect(group.locator('[data-slot="basic-tool-tool-title"]')).toContainText("Edit")
+    await expect(group.getByText("1 file", { exact: true })).toBeVisible()
+    await expect(group.locator('[data-slot="apply-patch-filename"]')).toHaveText(["first.ts"])
+    await expect(group.locator('[data-scope="apply-patch"] button')).toHaveAttribute("aria-expanded", "true")
+  })
+
  test("projects gaps, dividers, assistant parts, and errors together", async ({ page }) => {
    const firstUser = userMessage(
      [
@@ -132,6 +250,43 @@ test.describe("session timeline projection", () => {
    expect(rows).toEqual(["AssistantPart", "TurnDivider", "AssistantPart"])
  })

+  test("renders aliased and long custom model notices", async ({ page }) => {
+    const shortName = "GPT-5.4 nano"
+    const longName = "Company Gateway Extra Long Context Model for Narrow Timeline Layouts"
+    await setupTimeline(page, {
+      viewport: { width: 420, height: 700 },
+      sessionMessages: [
+        {
+          id: "msg_model_fast_nano",
+          type: "model-switched",
+          time: { created: 1700000000000 },
+          model: { providerID: "company-gateway", id: "fast-nano", variant: "xhigh" },
+        },
+        {
+          id: "msg_model_long_context",
+          type: "model-switched",
+          time: { created: 1700000001000 },
+          model: { providerID: "company-gateway", id: "long-context" },
+        },
+        userMessage(),
+        assistantMessage(),
+      ],
+    })
+
+    const shortNotice = page.locator('[data-slot="session-timeline-notice"]').filter({ hasText: shortName })
+    const longNotice = page.locator('[data-slot="session-timeline-notice"]').filter({ hasText: longName })
+    await expect(shortNotice).toBeVisible()
+    await expect(shortNotice.getByText(`Switched to ${shortName}`, { exact: true })).toBeVisible()
+    await expect(shortNotice.locator('[data-slot="session-timeline-notice-variant"]')).toHaveText("xhigh")
+    await expect(page.getByText("fast-nano", { exact: true })).toHaveCount(0)
+    await expect(shortNotice.locator('[data-component="provider-icon"]')).toBeVisible()
+    await expect(longNotice).toBeVisible()
+    await expect(longNotice.locator('[data-component="provider-icon"]')).toBeVisible()
+    await expect(longNotice.locator('[data-slot="session-timeline-notice-variant"]')).toHaveCount(0)
+    await expect(longNotice.locator("[title]")).toHaveAttribute("title", `Switched to ${longName}`)
+    await expect.poll(() => longNotice.evaluate((element) => element.scrollWidth <= element.clientWidth)).toBe(true)
+  })
+
  test("renders user image, file attachment, file reference, and agent reference", async ({ page }) => {
    const text = "Use @explore with @src/a.ts and inspect the attachments"
    const parts: PartSeed<"user">[] = [
@@ -196,11 +351,7 @@ function patchPart(id: string) {
    { patchText: "Update the projected files" },
    {
      metadata: {
-        files: [
-          patchFile("src/a.ts", "modified"),
-          patchFile("src/b.ts", "added"),
-          patchFile("src/old.ts", "deleted"),
-        ],
+        files: [patchFile("src/a.ts", "modified")],
      },
    },
  )
@@ -98,12 +98,12 @@ test("labels completed searches with result counts", async ({ page }) => {

  const group = page.locator(`[data-timeline-part-ids="${glob},${grep}"]`)
  await group.locator('[data-slot="collapsible-trigger"]').click()
-  const rows = group.locator('[data-component="tool-trigger"]')
-  await expect(rows.nth(0)).toContainText("(1 match)")
-  await expect(rows.nth(1)).toContainText("(12 matches)")
+  const rows = group.locator('[data-component="context-tool-group-list"] [data-component="tool-trigger"]')
+  await expect(rows.filter({ hasText: "Glob" })).toContainText("(1 match)")
+  await expect(rows.filter({ hasText: "Grep" })).toContainText("(12 matches)")
 })

-test("labels V2 read tools from their path input", async ({ page }) => {
+test("labels read tools from their path input", async ({ page }) => {
  const id = "prt_read_path"
  await setupTimeline(page, {
    messages: [userMessage(), assistantMessage([toolPart(id, "read", "completed", { path: "src/a.ts" })])],
@@ -111,34 +111,37 @@ test("labels V2 read tools from their path input", async ({ page }) => {

  const group = page.locator(`[data-timeline-part-ids="${id}"]`)
  await group.locator('[data-slot="collapsible-trigger"]').click()
-  await expect(group.locator('[data-slot="basic-tool-tool-subtitle"]')).toHaveText("a.ts")
+  await expect(
+    group
+      .locator('[data-component="context-tool-group-list"] [data-component="tool-trigger"]')
+      .filter({ hasText: "Read" }),
+  ).toContainText("a.ts")
 })

-test("labels V2 skill tools from IDs and result metadata", async ({ page }) => {
+test("labels skill tools from IDs and result metadata", async ({ page }) => {
  const pending = "prt_skill_id"
  const completed = "prt_skill_name"
  await setupTimeline(page, {
    messages: [
      userMessage(),
      assistantMessage([
-        toolPart(pending, "skill", "running", { id: "sample-skill" }),
+        toolPart(pending, "skill", "running", { id: "frontend-design" }),
        toolPart(completed, "skill", "completed", { id: "opencode" }, { metadata: { name: "OpenCode" } }),
      ]),
    ],
  })

-  await expect(page.locator(`[data-timeline-part-id="${pending}"] [data-component="text-shimmer"]`)).toHaveAttribute(
-    "aria-label",
-    "sample-skill",
-  )
-  await expect(
-    page.locator(`[data-timeline-part-id="${completed}"] [data-component="text-shimmer"]`),
-  ).toHaveAttribute("aria-label", "OpenCode")
-  for (const id of [pending, completed]) {
+  for (const [id, name] of [
+    [pending, "frontend-design"],
+    [completed, "OpenCode"],
+  ] as const) {
    const skill = page.locator(`[data-timeline-part-id="${id}"]`)
-    await expect(skill.locator('[data-slot="skill-tool-label"]')).toHaveText("Skill")
-    await expect(skill.locator('[data-slot="skill-tool-separator"]')).toHaveText("·")
-    await expect(skill.locator('use[href="#opencode-v2-icon-post-skill"]')).toBeVisible()
+    const loaded = skill.locator('[data-component="tool-loaded-item"]')
+    await expect(loaded).toHaveAttribute("aria-label", `Loaded ${name} skill`)
+    await expect(loaded).toHaveCSS("line-height", "16px")
+    await expect(loaded.locator('[data-slot="tool-loaded-label"]')).toHaveText("Loaded")
+    await expect(loaded.locator('[data-slot="tool-loaded-kind"]')).toHaveText("skill")
+    await expect(loaded.locator('[data-component="text-shimmer"]')).toHaveAttribute("aria-label", name)
  }
 })

@@ -152,8 +155,7 @@ function errorInput(tool: string) {
  if (tool === "patch") return { patchText: "Update src/error.ts" }
  if (tool === "webfetch") return { url: "https://example.com" }
  if (tool === "websearch") return { query: "failure" }
-  if (tool === "subagent")
-    return { description: "Fail subagent", agent: "explore", prompt: "Inspect the failure." }
+  if (tool === "subagent") return { description: "Fail subagent", agent: "explore", prompt: "Inspect the failure." }
  if (tool === "skill") return { name: "failure" }
  return { target: "failure" }
 }
@@ -87,7 +87,7 @@ test("reconnects after a stream error", async ({ page }) => {
  expect((await timeline.transport.connections())[0]?.endedBy).toBe("error")
 })

-test("does not request replay when reconnecting the volatile V2 event stream", async ({ page }) => {
+test("does not request replay when reconnecting the volatile event stream", async ({ page }) => {
  const timeline = await setupTimeline(page, { eventRetry: 10 })
  const events = partUpdated(textPart("prt_transport_id", "event with id"))
  const first = (
--- a/Show More
+++ b/Show More