feat(browser): add experimental desktop browser

fix(app): keep pending steers after assistant work (#45435 )
refactor(core): use shared state for tool registry (#45414 )
2026-08-27 20:16:17 +00:00 · 2026-08-27 12:48:12 +10:00 · 2026-08-27 02:44:18 +00:00 · 2026-08-26 21:33:00 -05:00 · 2026-08-27 02:30:36 +00:00 · 2026-08-27 02:01:36 +00:00
535 changed files with 26921 additions and 7147 deletions
@@ -0,0 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Include both paths of renamed files in new snapshot change lists so undo restores the original file instead of only deleting the renamed file.
@@ -0,0 +1,10 @@
+---
+"@opencode-ai/core": patch
+---
+
+Make the experimental portable shell scanner authoritative, with no Tree-sitter
+fallback. Scan common Bash and PowerShell control flow, heredocs, functions,
+expressions, quoting, and substitutions natively. Preserve existing redirect and
+declaration permission matching, and make PowerShell saved approvals cover the
+original command spelling. Parser failures remain visible without changing the
+permission engine. The default Tree-sitter path is unchanged.
@@ -42,7 +42,7 @@ jobs:
      - name: Find affected packages
        id: packages
        env:
-          TURBO_SCM_BASE: ${{ github.event.pull_request.base.sha || github.event.before }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
          TURBO_SCM_HEAD: ${{ github.sha }}
        run: |
          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
@@ -119,7 +119,7 @@ jobs:
          GITHUB_ACTIONS=false bun turbo test --affected
        env:
          OPENCODE_EXPERIMENTAL_DISABLE_FILEWATCHER: ${{ runner.os == 'Windows' && 'true' || 'false' }}
-          TURBO_SCM_BASE: ${{ github.event.pull_request.base.sha || github.event.before }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
          TURBO_SCM_HEAD: ${{ github.sha }}

      - name: Verify published codemode package
@@ -137,7 +137,7 @@ jobs:
          fi
          bun turbo verify:package --affected --filter=@opencode-ai/sdk
        env:
-          TURBO_SCM_BASE: ${{ github.event.pull_request.base.sha || github.event.before }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
          TURBO_SCM_HEAD: ${{ github.sha }}

      - name: Verify compiled service lifecycle
@@ -237,9 +237,9 @@ jobs:
        working-directory: packages/app
        run: bunx playwright install chromium

-      - name: Run app e2e tests
+      - name: Run app e2e tests against production build
        if: env.E2E_ENABLED == 'true'
-        run: bun --cwd packages/app test:e2e:local
+        run: bun --cwd packages/app test:e2e:built
        env:
          CI: true
        timeout-minutes: 30
@@ -125,6 +125,7 @@
        "@effect/platform-node": "catalog:",
        "@opencode-ai/client": "workspace:*",
        "@opencode-ai/plugin": "workspace:*",
+        "@opencode-ai/pty": "0.1.12",
        "@opencode-ai/schema": "workspace:*",
        "@opencode-ai/server": "workspace:*",
        "@opencode-ai/tui": "workspace:*",
@@ -181,12 +182,14 @@
      "dependencies": {
        "@opencode-ai/protocol": "workspace:*",
        "@opencode-ai/schema": "workspace:*",
+        "ws": "8.21.0",
      },
      "devDependencies": {
        "@effect/platform-node": "catalog:",
        "@opencode-ai/httpapi-codegen": "workspace:*",
        "@tsconfig/bun": "catalog:",
        "@types/bun": "catalog:",
+        "@types/ws": "8.18.1",
        "@typescript/native-preview": "catalog:",
        "effect": "catalog:",
        "solid-js": "catalog:",
@@ -347,10 +350,8 @@
        "@ai-sdk/anthropic": "3.0.82",
        "@ai-sdk/azure": "3.0.88",
        "@ai-sdk/cohere": "3.0.27",
-        "@ai-sdk/deepinfra": "2.0.41",
        "@ai-sdk/gateway": "3.0.104",
        "@ai-sdk/google-vertex": "4.0.128",
-        "@ai-sdk/groq": "3.0.31",
        "@ai-sdk/mistral": "3.0.51",
        "@ai-sdk/openai-compatible": "2.0.41",
        "@ai-sdk/perplexity": "3.0.26",
@@ -365,6 +366,7 @@
        "@opencode-ai/ai": "workspace:*",
        "@opencode-ai/codemode": "workspace:*",
        "@opencode-ai/plugin": "workspace:*",
+        "@opencode-ai/pty": "0.1.12",
        "@opencode-ai/schema": "workspace:*",
        "@opencode-ai/util": "workspace:*",
        "@parcel/watcher": "2.5.1",
@@ -553,6 +555,20 @@
        "@typescript/native-preview": "catalog:",
      },
    },
+    "packages/latex": {
+      "name": "@opencode-ai/latex",
+      "version": "0.0.0",
+      "dependencies": {
+        "@opencode-ai/plugin": "workspace:*",
+        "@opentui/core": "catalog:",
+        "string-width": "catalog:",
+      },
+      "devDependencies": {
+        "@tsconfig/bun": "catalog:",
+        "@types/bun": "catalog:",
+        "@typescript/native-preview": "catalog:",
+      },
+    },
    "packages/merman": {
      "name": "@opencode-ai/merman",
      "version": "0.0.0",
@@ -731,6 +747,7 @@
      },
      "devDependencies": {
        "@happy-dom/global-registrator": "20.0.11",
+        "@playwright/test": "catalog:",
        "@tsconfig/node22": "catalog:",
        "@types/bun": "catalog:",
        "@types/luxon": "catalog:",
@@ -836,6 +853,7 @@
        "@opencode-ai/client": "workspace:*",
        "@opencode-ai/session-ui": "workspace:*",
        "@opencode-ai/ui": "workspace:*",
+        "@playwright/test": "catalog:",
        "@solidjs/meta": "catalog:",
        "@storybook/addon-a11y": "10.4.4",
        "@storybook/addon-docs": "10.4.4",
@@ -877,6 +895,7 @@
      "dependencies": {
        "@opencode-ai/client": "workspace:*",
        "@opencode-ai/core": "workspace:*",
+        "@opencode-ai/latex": "workspace:*",
        "@opencode-ai/merman": "workspace:*",
        "@opencode-ai/plugin": "workspace:*",
        "@opencode-ai/schema": "workspace:*",
@@ -986,7 +1005,6 @@
        "mime-types": "3.0.2",
        "minimatch": "10.2.5",
        "npm-package-arg": "13.0.2",
-        "pacote": "21.5.1",
        "resolve.exports": "catalog:",
      },
      "devDependencies": {
@@ -996,7 +1014,6 @@
        "@types/node": "catalog:",
        "@types/npm-package-arg": "6.1.4",
        "@types/npmcli__arborist": "6.3.3",
-        "@types/pacote": "11.1.8",
        "@typescript/native-preview": "catalog:",
      },
    },
@@ -1180,8 +1197,6 @@

    "@ai-sdk/deepgram": ["@ai-sdk/deepgram@2.0.52", "", { "dependencies": { "@ai-sdk/provider": "3.0.15", "@ai-sdk/provider-utils": "4.0.46" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-8pcrQvEQCbvrrQKnD6hclBbI0hUgSrgyADykRbabxv/g9vPurfMC6n23J7dD+KZ3EcCoW+qz3IUIfySJ58gBOg=="],

-    "@ai-sdk/deepinfra": ["@ai-sdk/deepinfra@2.0.41", "", { "dependencies": { "@ai-sdk/openai-compatible": "2.0.37", "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-y6RoOP7DGWmDSiSxrUSt5p18sbz+Ixe5lMVPmdE7x+Tr5rlrzvftyHhjWHfqlAtoYERZTGFbP6tPW1OfQcrb4A=="],
-
    "@ai-sdk/deepseek": ["@ai-sdk/deepseek@2.0.47", "", { "dependencies": { "@ai-sdk/provider": "3.0.14", "@ai-sdk/provider-utils": "4.0.38" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-MzcQ321JO8OY+TVLFI81A7cIIuoeLLxrLCDD+8C1E3Ro6UFyfMtRXo9bw9OhTMRSDMo6hgSDOo4Fekz8aJtQYQ=="],

    "@ai-sdk/elevenlabs": ["@ai-sdk/elevenlabs@2.0.52", "", { "dependencies": { "@ai-sdk/provider": "3.0.15", "@ai-sdk/provider-utils": "4.0.46" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-ZgkausouWvO9U4ZtowNJ093bSNYOvH8zqls3uLC3+oxzWvbbTZO8SOdmFk0+gGafsXFJvq2yUX9+rEeJPwOJLw=="],
@@ -2152,6 +2167,8 @@

    "@opencode-ai/httpapi-codegen": ["@opencode-ai/httpapi-codegen@workspace:packages/httpapi-codegen"],

+    "@opencode-ai/latex": ["@opencode-ai/latex@workspace:packages/latex"],
+
    "@opencode-ai/merman": ["@opencode-ai/merman@workspace:packages/merman"],

    "@opencode-ai/plugin": ["@opencode-ai/plugin@workspace:packages/plugin"],
@@ -2160,6 +2177,20 @@

    "@opencode-ai/protocol": ["@opencode-ai/protocol@workspace:packages/protocol"],

+    "@opencode-ai/pty": ["@opencode-ai/pty@0.1.12", "", { "optionalDependencies": { "@opencode-ai/pty-darwin-arm64": "0.1.12", "@opencode-ai/pty-darwin-x64": "0.1.12", "@opencode-ai/pty-linux-arm64-gnu": "0.1.12", "@opencode-ai/pty-linux-arm64-musl": "0.1.12", "@opencode-ai/pty-linux-x64-gnu": "0.1.12", "@opencode-ai/pty-linux-x64-musl": "0.1.12" }, "bin": { "opencode-pty": "bin/opencode-pty.js" } }, "sha512-dl4FyJUhTXThsWYY8txG/8/nwN7dE0M5Sic9r4L9f2pvtJnbR5zrCrPoiPIBIxZle1wVks1dhz4z/CfqLf5sCg=="],
+
+    "@opencode-ai/pty-darwin-arm64": ["@opencode-ai/pty-darwin-arm64@0.1.12", "", { "os": "darwin", "cpu": "arm64" }, "sha512-tMvoriq3VegVlj1uEglc6qE0M7VXy61nyf9Si7tTO7xa8JiyxuFJSXOZ1pGeErDu+pe24hvTyVOR+gkdew8w9g=="],
+
+    "@opencode-ai/pty-darwin-x64": ["@opencode-ai/pty-darwin-x64@0.1.12", "", { "os": "darwin", "cpu": "x64" }, "sha512-Sn5vMLL5giHOhx7J5H6zwDp4YjjXorY+QV0IEYY+SCT4wQfRBliokIyj23pRl6P2RK3u9bDLXJHDNMfDVZ2Rxg=="],
+
+    "@opencode-ai/pty-linux-arm64-gnu": ["@opencode-ai/pty-linux-arm64-gnu@0.1.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-HbnlKZy052l7G527wK0+05EXaUpZ4ykVAmNBEzqWCoi4TeQj2+Nr9kJ9trx9o1KrVcT4Ki58CCvN5QOls6Z0yQ=="],
+
+    "@opencode-ai/pty-linux-arm64-musl": ["@opencode-ai/pty-linux-arm64-musl@0.1.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-2nTN7ggu1h9XgjNcoQMYjP5sirfYnAskpdFCOqjokLqhytX/IMMmkRTQs+foaEaPz0dAIQD3DQplR2jZIgxp1w=="],
+
+    "@opencode-ai/pty-linux-x64-gnu": ["@opencode-ai/pty-linux-x64-gnu@0.1.12", "", { "os": "linux", "cpu": "x64" }, "sha512-FnD5ndnObTQKAoaVvxLKi5W+r3/+dsaMsobz6uK0B9hlmffXxY5CQ6HQyUU/h3aIKLWXxho5XYkA2b9yrp8/gA=="],
+
+    "@opencode-ai/pty-linux-x64-musl": ["@opencode-ai/pty-linux-x64-musl@0.1.12", "", { "os": "linux", "cpu": "x64" }, "sha512-prkrNu6uvjqoffxdGiDHSU5C0Y+kCSfv+lslu7dfRPgPKenVELNpRTAbOduyrWPac2vGt8j5NM61icJyodbJmA=="],
+
    "@opencode-ai/schema": ["@opencode-ai/schema@workspace:packages/schema"],

    "@opencode-ai/script": ["@opencode-ai/script@workspace:packages/script"],
@@ -5912,10 +5943,6 @@

    "@ai-sdk/deepgram/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.46", "", { "dependencies": { "@ai-sdk/provider": "3.0.15", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.8", "undici": "^6.28.0" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-tEtld97plCFiYevsJuOkGkeuhQndeMWFBVrJS4AjnbD5AqrNSXRCe0p+BZ3Cju/sxDeeZ9ym3q9YUV8fASA7aQ=="],

-    "@ai-sdk/deepinfra/@ai-sdk/openai-compatible": ["@ai-sdk/openai-compatible@2.0.37", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-+POSFVcgiu47BK64dhsI6OpcDC0/VAE2ZSaXdXGNNhpC/ava++uSRJYks0k2bpfY0wwCTgpAWZsXn/dG2Yppiw=="],
-
-    "@ai-sdk/deepinfra/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.21", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-MtFUYI1/8mgDvRmaBDjbLJPFFrMG777AvSgyIFQtZHIMzm88R/12vYBBpnk7pfiWLFE1DSZzY4WDYzGbKAcmiw=="],
-
    "@ai-sdk/deepseek/@ai-sdk/provider": ["@ai-sdk/provider@3.0.14", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-5X1k57JBJ4H7H1QjX7CnJYAB1I19r/trVZTMcSms7/kLNZ8RaU4Nt2agcwZzv82Hfx6Q7/TOLU7agAKeFfc8cA=="],

    "@ai-sdk/deepseek/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.38", "", { "dependencies": { "@ai-sdk/provider": "3.0.14", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.8" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-/HHGmtKllqjg1OLc023v9w9kK3laW7Z6TzfZukYQWCsGBbzB9p60zTvvpXFVcs44NZBVXL3viOa1HRKUbeee8g=="],
@@ -2,7 +2,7 @@
 exact = true
 # Only install newly resolved package versions published at least 3 days ago.
 minimumReleaseAge = 259200
-minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@brendonovich/vite-plugin-opencode", "@opencode-ai/sdk", "@opentui/core", "@opentui/core-darwin-arm64", "@opentui/core-darwin-x64", "@opentui/core-linux-arm64", "@opentui/core-linux-arm64-musl", "@opentui/core-linux-x64", "@opentui/core-linux-x64-musl", "@opentui/core-win32-arm64", "@opentui/core-win32-x64", "@opentui/keymap", "@opentui/solid", "opentui-spinner", "gitlab-ai-provider", "opencode-gitlab-auth", "@ff-labs/fff-node", "@ff-labs/fff-bun", "@ff-labs/fff-bin-darwin-arm64", "@ff-labs/fff-bin-darwin-x64", "@ff-labs/fff-bin-linux-arm64-gnu", "@ff-labs/fff-bin-linux-arm64-musl", "@ff-labs/fff-bin-linux-x64-gnu", "@ff-labs/fff-bin-linux-x64-musl", "@ff-labs/fff-bin-win32-arm64", "@ff-labs/fff-bin-win32-x64", "@pierre/diffs", "@pierre/theming", "app-builder-lib", "dmg-builder", "electron-builder", "electron-publish", "blume"]
+minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@brendonovich/vite-plugin-opencode", "@opencode-ai/sdk", "@opencode-ai/pty", "@opencode-ai/pty-darwin-arm64", "@opencode-ai/pty-darwin-x64", "@opencode-ai/pty-linux-arm64-gnu", "@opencode-ai/pty-linux-arm64-musl", "@opencode-ai/pty-linux-x64-gnu", "@opencode-ai/pty-linux-x64-musl", "@opentui/core", "@opentui/core-darwin-arm64", "@opentui/core-darwin-x64", "@opentui/core-linux-arm64", "@opentui/core-linux-arm64-musl", "@opentui/core-linux-x64", "@opentui/core-linux-x64-musl", "@opentui/core-win32-arm64", "@opentui/core-win32-x64", "@opentui/keymap", "@opentui/solid", "opentui-spinner", "gitlab-ai-provider", "opencode-gitlab-auth", "@ff-labs/fff-node", "@ff-labs/fff-bun", "@ff-labs/fff-bin-darwin-arm64", "@ff-labs/fff-bin-darwin-x64", "@ff-labs/fff-bin-linux-arm64-gnu", "@ff-labs/fff-bin-linux-arm64-musl", "@ff-labs/fff-bin-linux-x64-gnu", "@ff-labs/fff-bin-linux-x64-musl", "@ff-labs/fff-bin-win32-arm64", "@ff-labs/fff-bin-win32-x64", "@pierre/diffs", "@pierre/theming", "app-builder-lib", "dmg-builder", "electron-builder", "electron-publish", "blume"]

 [test]
 root = "./do-not-run-tests-from-root"
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-fOM/kGJJ1cipCHQIxioDZEB7NZykpSiqgwm7gIS6THI=",
-    "aarch64-linux": "sha256-XTY2C33HjsBMWO7VIiWc2MynjJrxbTLrOJ+6pM+afI0=",
-    "aarch64-darwin": "sha256-wX6+bC18djtPZ7A9ch+wryM7tDFfrAlT0xx0QTk6EJQ=",
-    "x86_64-darwin": "sha256-dcRRX4bYq5AmG4GcVmYq/M+06dlf4KJHn+clT2JY48g="
+    "x86_64-linux": "sha256-iYdVrLtyKmjlyypisF9SqzgyriWT90kSCh3crxw9AKU=",
+    "aarch64-linux": "sha256-BV2t4w5ujArbtSC/Qfm3gLzevQW9A6hMgOyPVp94g/o=",
+    "aarch64-darwin": "sha256-EwMq7zaxzzcsmH0Pjqu4ftGdcM8Lna8mvHgKzRcVI8g=",
+    "x86_64-darwin": "sha256-PokzxlkQy6JvHADF2ZMIIDI1u9ZjSNNedpmR9gvHS5c="
  }
 }
@@ -36,7 +36,12 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // Protocols whose wire format ignores inline cache markers (OpenAI's implicit
 // prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
 // whole policy pass for these — emitting hints would be harmless but pointless.
-const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "google-vertex-messages", "bedrock-converse", "openrouter"])
+const RESPECTS_INLINE_HINTS = new Set([
+  "anthropic-messages",
+  "google-vertex-messages",
+  "bedrock-converse",
+  "openrouter",
+])

 const makeHint = (ttlSeconds: number | undefined): CacheHint =>
  ttlSeconds !== undefined ? new CacheHint({ type: "ephemeral", ttlSeconds }) : new CacheHint({ type: "ephemeral" })
@@ -69,14 +69,22 @@ export interface OptionsInput {
  // SDK Metadata:2649 {user_id?: string | null}
  readonly metadata?: { readonly user_id?: string | null }
  // SDK MessageCreateParamsContainer:2596 ContainerParams|string
-  readonly container?: string | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
+  readonly container?:
+    | string
+    | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
  readonly inference_geo?: string | null
  readonly inferenceGeo?: string | null
  readonly cache_control?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
  readonly cacheControl?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
  // SDK OutputConfig:2684 {effort, format: JSONOutputFormat}
-  readonly output_config?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
-  readonly outputConfig?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
+  readonly output_config?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
+  readonly outputConfig?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
 }

 export type ProviderOptionsInput = OptionsInput
@@ -259,7 +267,11 @@ const AnthropicToolChoice = Schema.Union([
    type: Schema.Literals(["auto", "any", "none"]),
    disable_parallel_tool_use: Schema.optional(Schema.Boolean),
  }),
-  Schema.Struct({ type: Schema.tag("tool"), name: Schema.String, disable_parallel_tool_use: Schema.optional(Schema.Boolean) }),
+  Schema.Struct({
+    type: Schema.tag("tool"),
+    name: Schema.String,
+    disable_parallel_tool_use: Schema.optional(Schema.Boolean),
+  }),
 ])

 const AnthropicThinking = Schema.Union([
@@ -394,6 +406,7 @@ const AnthropicEvent = Schema.Struct({
 type AnthropicEvent = Schema.Schema.Type<typeof AnthropicEvent>

 interface ParserState {
+  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
  readonly reasoningSignatures: Readonly<Record<number, string>>
  readonly usage?: Usage
@@ -428,18 +441,18 @@ const cacheControl = (breakpoints: Cache.Breakpoints, cache: CacheHint | undefin
  return Cache.ttlBucket(cache.ttlSeconds) === "1h" ? EPHEMERAL_1H : EPHEMERAL_5M
 }

-const anthropicMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ anthropic: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })

-const signatureFromMetadata = (metadata: ProviderMetadata | undefined): string | undefined => {
-  const anthropic = metadata?.anthropic
-  if (!ProviderShared.isRecord(anthropic)) return undefined
-  return typeof anthropic.signature === "string" ? anthropic.signature : undefined
+const signatureFromMetadata = (metadata: ProviderMetadata | undefined, key: string): string | undefined => {
+  const provider = metadata?.[key]
+  if (!ProviderShared.isRecord(provider)) return undefined
+  return typeof provider.signature === "string" ? provider.signature : undefined
 }

-const redactedDataFromMetadata = (metadata: ProviderMetadata | undefined): string | undefined => {
-  const anthropic = metadata?.anthropic
-  if (!ProviderShared.isRecord(anthropic)) return undefined
-  return typeof anthropic.redactedData === "string" ? anthropic.redactedData : undefined
+const redactedDataFromMetadata = (metadata: ProviderMetadata | undefined, key: string): string | undefined => {
+  const provider = metadata?.[key]
+  if (!ProviderShared.isRecord(provider)) return undefined
+  return typeof provider.redactedData === "string" ? provider.redactedData : undefined
 }

 const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition, inputSchema: JsonSchema): AnthropicTool => ({
@@ -499,14 +512,21 @@ const serverToolResultType = (name: string): AnthropicServerToolResultType | und
  return undefined
 }

-const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (part: ToolResultPart) {
+const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (
+  part: ToolResultPart,
+  providerMetadataKey: string,
+) {
  const wireType = serverToolResultType(part.name)
  if (!wireType)
    return yield* invalid(`Anthropic Messages does not know how to round-trip server tool result for ${part.name}`)
  // Prefer the provider-owned replay payload; fall back to the result value for
  // histories constructed directly from provider events.
-  const payload = part.providerMetadata?.anthropic?.["result"] ?? part.result.value
-  return { type: wireType, tool_use_id: scrubToolCallID(part.id), content: payload } satisfies AnthropicServerToolResultBlock
+  const payload = part.providerMetadata?.[providerMetadataKey]?.["result"] ?? part.result.value
+  return {
+    type: wireType,
+    tool_use_id: scrubToolCallID(part.id),
+    content: payload,
+  } satisfies AnthropicServerToolResultBlock
 })

 const fileIdFromMetadata = (metadata: MediaPart["metadata"]): string | undefined => {
@@ -554,9 +574,7 @@ const documentContextFromMetadata = (metadata: MediaPart["metadata"]): string |
  return undefined
 }

-const citationsFromMetadata = (
-  metadata: MediaPart["metadata"],
-): AnthropicDocumentBlock["citations"] | undefined => {
+const citationsFromMetadata = (metadata: MediaPart["metadata"]): AnthropicDocumentBlock["citations"] | undefined => {
  if (!ProviderShared.isRecord(metadata)) return undefined
  const raw = ProviderShared.isRecord(metadata.anthropic)
    ? (metadata.anthropic.citations ?? metadata.citations)
@@ -706,8 +724,7 @@ const lowerToolResultContent = Effect.fnUntraced(function* (part: ToolResultPart
 })

 const requireThinkingSignature = (request: LLMRequest) => {
-  if (request.model.compatibility?.requireSignature !== undefined)
-    return request.model.compatibility.requireSignature
+  if (request.model.compatibility?.requireSignature !== undefined) return request.model.compatibility.requireSignature
  const provider = request.model.provider.toLowerCase()
  const model = request.model.id.toLowerCase()
  const baseURL = (request.model.route.endpoint.baseURL ?? "").toLowerCase()
@@ -791,6 +808,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
  breakpoints: Cache.Breakpoints,
 ) {
  const messages: AnthropicMessage[] = []
+  const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)

  for (const [index, message] of request.messages.entries()) {
    if (message.role === "system") {
@@ -836,8 +854,8 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
        if (part.type === "reasoning") {
          // A signature marks visible thinking; only signature-less parts carrying
          // redactedData round-trip as opaque redacted_thinking blocks.
-          const signature = part.encrypted ?? signatureFromMetadata(part.providerMetadata)
-          const redactedData = redactedDataFromMetadata(part.providerMetadata)
+          const signature = part.encrypted ?? signatureFromMetadata(part.providerMetadata, providerMetadataKey)
+          const redactedData = redactedDataFromMetadata(part.providerMetadata, providerMetadataKey)
          if (signature === undefined && redactedData !== undefined) {
            content.push({ type: "redacted_thinking", data: redactedData })
            continue
@@ -866,7 +884,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
          continue
        }
        if (part.type === "tool-result" && part.providerExecuted) {
-          content.push(yield* lowerServerToolResult(part))
+          content.push(yield* lowerServerToolResult(part, providerMetadataKey))
          continue
        }
        return yield* invalid(
@@ -900,21 +918,24 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (

 const resolveOptions = Effect.fn("AnthropicMessages.resolveOptions")(function* (request: LLMRequest) {
  const input = request.providerOptions as Record<string, unknown> | undefined
-  const rawServiceTier = (input as Record<string, unknown> | undefined)?.service_tier ?? (input as Record<string, unknown> | undefined)?.serviceTier
+  const rawServiceTier =
+    (input as Record<string, unknown> | undefined)?.service_tier ??
+    (input as Record<string, unknown> | undefined)?.serviceTier
  const service_tier =
    rawServiceTier === "auto" || rawServiceTier === "standard_only"
      ? (rawServiceTier as "auto" | "standard_only")
      : undefined
  const rawMetadata = (input as Record<string, unknown> | undefined)?.metadata
  const metadata =
-    ProviderShared.isRecord(rawMetadata) &&
-    (typeof rawMetadata.user_id === "string" || rawMetadata.user_id === null)
+    ProviderShared.isRecord(rawMetadata) && (typeof rawMetadata.user_id === "string" || rawMetadata.user_id === null)
      ? { user_id: rawMetadata.user_id as string | null }
      : undefined
  const container =
    typeof (input as Record<string, unknown> | undefined)?.container === "string" ||
    ProviderShared.isRecord((input as Record<string, unknown> | undefined)?.container)
-      ? ((input as Record<string, unknown>).container as string | { id?: string | null; skills?: ReadonlyArray<Record<string, unknown>> | null })
+      ? ((input as Record<string, unknown>).container as
+          | string
+          | { id?: string | null; skills?: ReadonlyArray<Record<string, unknown>> | null })
      : undefined
  const rawInferenceGeo =
    (input as Record<string, unknown> | undefined)?.inference_geo ??
@@ -965,8 +986,7 @@ const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function*
    input.display === "summarized" || input.display === "omitted"
      ? (input.display as "summarized" | "omitted")
      : undefined
-  if (input.type === "adaptive")
-    return { type: "adaptive" as const, ...(display === undefined ? {} : { display }) }
+  if (input.type === "adaptive") return { type: "adaptive" as const, ...(display === undefined ? {} : { display }) }
  if (input.type === "disabled") return { type: "disabled" as const }
  if (input.type !== "enabled") return undefined
  const budget =
@@ -1054,7 +1074,7 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // inclusive `inputTokens` the rest of the contract expects. Extended
 // thinking tokens are included in `output_tokens`; newer responses also
 // expose that subset through `output_tokens_details.thinking_tokens`.
-const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
+const mapUsage = (usage: AnthropicUsage | undefined, providerMetadataKey: string): Usage | undefined => {
  if (!usage) return undefined
  const nonCached = usage.input_tokens ?? undefined
  const cacheRead = usage.cache_read_input_tokens ?? undefined
@@ -1068,7 +1088,7 @@ const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
    cacheWriteInputTokens: cacheWrite,
    reasoningTokens: usage.output_tokens_details?.thinking_tokens,
    totalTokens: ProviderShared.totalTokens(inputTokens, usage.output_tokens, undefined),
-    providerMetadata: { anthropic: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
  })
 }

@@ -1077,7 +1097,7 @@ const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
 // field prefers `right` when defined, falls back to `left`. `inputTokens` is
 // recomputed from the merged breakdown so the inclusive total stays
 // consistent with `nonCached + cacheRead + cacheWrite`.
-const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
+const mergeUsage = (left: Usage | undefined, right: Usage | undefined, providerMetadataKey: string) => {
  if (!left) return right
  if (!right) return left
  const nonCachedInputTokens = right.nonCachedInputTokens ?? left.nonCachedInputTokens
@@ -1095,7 +1115,9 @@ const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
    reasoningTokens,
    totalTokens: ProviderShared.totalTokens(inputTokens, outputTokens, undefined),
    providerMetadata: {
-      anthropic: mergeJsonRecords(left.providerMetadata?.["anthropic"], right.providerMetadata?.["anthropic"]) ?? {},
+      [providerMetadataKey]:
+        mergeJsonRecords(left.providerMetadata?.[providerMetadataKey], right.providerMetadata?.[providerMetadataKey]) ??
+        {},
    },
  })
 }
@@ -1113,7 +1135,7 @@ const SERVER_TOOL_RESULT_NAMES: Record<AnthropicServerToolResultType, string> =

 const isServerToolResultType = (type: string): type is AnthropicServerToolResultType => type in SERVER_TOOL_RESULT_NAMES

-const serverToolResultEvent = (block: AnthropicStreamBlock): LLMEvent | undefined => {
+const serverToolResultEvent = (block: AnthropicStreamBlock, providerMetadataKey: string): LLMEvent | undefined => {
  if (!block.type || !isServerToolResultType(block.type)) return undefined
  const errorPayload =
    typeof block.content === "object" && block.content !== null && "type" in block.content
@@ -1127,7 +1149,7 @@ const serverToolResultEvent = (block: AnthropicStreamBlock): LLMEvent | undefine
    providerExecuted: true,
    // The complete payload is irreducible provider replay state: subsequent
    // stateless requests must round-trip the typed result block verbatim.
-    providerMetadata: anthropicMetadata({ blockType: block.type, result: block.content }),
+    providerMetadata: providerMetadata(providerMetadataKey, { blockType: block.type, result: block.content }),
  })
 }

@@ -1136,8 +1158,8 @@ type StepResult = readonly [ParserState, ReadonlyArray<LLMEvent>]
 const NO_EVENTS: StepResult["1"] = []

 const onMessageStart = (state: ParserState, event: AnthropicEvent): StepResult => {
-  const usage = mapUsage(event.message?.usage)
-  return [usage ? { ...state, usage: mergeUsage(state.usage, usage) } : state, NO_EVENTS]
+  const usage = mapUsage(event.message?.usage, state.providerMetadataKey)
+  return [usage ? { ...state, usage: mergeUsage(state.usage, usage, state.providerMetadataKey) } : state, NO_EVENTS]
 }

 const onContentBlockStart = (
@@ -1189,14 +1211,16 @@ const onContentBlockStart = (
  if (block.type === "thinking" && block.thinking !== undefined) {
    const events: LLMEvent[] = []
    const id = `reasoning-${event.index ?? 0}`
-    const providerMetadata =
-      block.signature === undefined ? undefined : anthropicMetadata({ signature: block.signature })
-    const lifecycle = Lifecycle.reasoningStart(state.lifecycle, events, id, providerMetadata)
+    const metadata =
+      block.signature === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { signature: block.signature })
+    const lifecycle = Lifecycle.reasoningStart(state.lifecycle, events, id, metadata)
    return [
      {
        ...state,
        lifecycle: block.thinking
-          ? Lifecycle.reasoningDelta(lifecycle, events, id, block.thinking, providerMetadata)
+          ? Lifecycle.reasoningDelta(lifecycle, events, id, block.thinking, metadata)
          : lifecycle,
        reasoningSignatures:
          event.index === undefined || block.signature === undefined
@@ -1219,14 +1243,14 @@ const onContentBlockStart = (
          state.lifecycle,
          events,
          `reasoning-${event.index ?? 0}`,
-          anthropicMetadata({ redactedData: block.data }),
+          providerMetadata(state.providerMetadataKey, { redactedData: block.data }),
        ),
      },
      events,
    ]
  }

-  const result = serverToolResultEvent(block)
+  const result = serverToolResultEvent(block, state.providerMetadataKey)
  if (!result) return [state, NO_EVENTS]
  const events: LLMEvent[] = []
  return [{ ...state, lifecycle: Lifecycle.stepStart(state.lifecycle, events) }, [...events, result]]
@@ -1306,7 +1330,7 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
        Lifecycle.textEnd(state.lifecycle, events, `text-${event.index}`),
        events,
        `reasoning-${event.index}`,
-        signature === undefined ? undefined : anthropicMetadata({ signature }),
+        signature === undefined ? undefined : providerMetadata(state.providerMetadataKey, { signature }),
      )
  events.push(...resultEvents)
  const reasoningSignatures = { ...state.reasoningSignatures }
@@ -1318,7 +1342,7 @@ const onMessageDelta = (
  state: ParserState,
  event: AnthropicEvent & { readonly delta?: AnthropicStreamDelta },
 ): StepResult => {
-  const usage = mergeUsage(state.usage, mapUsage(event.usage))
+  const usage = mergeUsage(state.usage, mapUsage(event.usage, state.providerMetadataKey), state.providerMetadataKey)
  return [
    {
      ...state,
@@ -1331,7 +1355,7 @@ const onMessageDelta = (
        providerMetadata:
          event.delta?.stop_sequence === null || event.delta?.stop_sequence === undefined
            ? undefined
-            : anthropicMetadata({ stopSequence: event.delta.stop_sequence }),
+            : providerMetadata(state.providerMetadataKey, { stopSequence: event.delta.stop_sequence }),
      },
    },
    NO_EVENTS,
@@ -1418,9 +1442,7 @@ const step = (state: ParserState, event: AnthropicEvent) => {
      if (event.index === undefined)
        return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic ${block.type} missing index`))
      if (!block.id)
-        return Effect.fail(
-          ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`),
-        )
+        return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`))
    }
    return Effect.succeed(onContentBlockStart(state, { ...event, content_block: block }))
  }
@@ -1459,7 +1481,8 @@ export const protocol = Protocol.make({
  },
  stream: {
    event: Protocol.jsonEvent(AnthropicEvent),
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
      reasoningSignatures: {},
      lifecycle: Lifecycle.initial(),
@@ -1473,10 +1496,9 @@ export const route = Route.make({
  provider: "anthropic",
  providerMetadataKey: "anthropic",
  protocol,
-  endpoint: Endpoint.path(
-    (input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH),
-    { baseURL: DEFAULT_BASE_URL },
-  ),
+  endpoint: Endpoint.path((input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH), {
+    baseURL: DEFAULT_BASE_URL,
+  }),
  auth: Auth.none,
  framing,
  headers: () => ({ "anthropic-version": "2023-06-01" }),
@@ -258,19 +258,21 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
    tool: (name) => ({ tool: { name } }) as const,
  })

-const bedrockMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ bedrock: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })

-const reasoningSignature = (part: ReasoningPart) => {
-  const bedrock = part.providerMetadata?.bedrock
+const reasoningSignature = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
  return (
    part.encrypted ??
-    (ProviderShared.isRecord(bedrock) && typeof bedrock.signature === "string" ? bedrock.signature : undefined)
+    (ProviderShared.isRecord(metadata) && typeof metadata.signature === "string" ? metadata.signature : undefined)
  )
 }

-const reasoningRedactedData = (part: ReasoningPart) => {
-  const bedrock = part.providerMetadata?.bedrock
-  return ProviderShared.isRecord(bedrock) && typeof bedrock.redactedData === "string" ? bedrock.redactedData : undefined
+const reasoningRedactedData = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
+  return ProviderShared.isRecord(metadata) && typeof metadata.redactedData === "string"
+    ? metadata.redactedData
+    : undefined
 }

 const lowerToolCall = (part: ToolCallPart): BedrockToolUseBlock => ({
@@ -318,6 +320,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
  breakpoints: BedrockCache.Breakpoints,
 ) {
  const messages: BedrockMessage[] = []
+  const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)

  for (const message of request.messages) {
    if (message.role === "system") {
@@ -365,8 +368,8 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
          continue
        }
        if (part.type === "reasoning") {
-          const signature = reasoningSignature(part)
-          const redactedData = reasoningRedactedData(part)
+          const signature = reasoningSignature(part, providerMetadataKey)
+          const redactedData = reasoningRedactedData(part, providerMetadataKey)
          if (signature === undefined && redactedData !== undefined) {
            content.push({ reasoningContent: { redactedContent: redactedData } })
            continue
@@ -466,7 +469,7 @@ const mapFinishReason = (reason: string): FinishReason => {

 // AWS reports inputTokens separately from cache reads and writes.
 // Bedrock does not break reasoning out of outputTokens for current models.
-const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
+const mapUsage = (usage: BedrockUsageSchema | undefined, providerMetadataKey: string): Usage | undefined => {
  if (!usage) return undefined
  const inputTokens = ProviderShared.sumTokens(
    usage.inputTokens,
@@ -480,11 +483,12 @@ const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
    cacheReadInputTokens: usage.cacheReadInputTokens,
    cacheWriteInputTokens: usage.cacheWriteInputTokens,
    totalTokens: ProviderShared.totalTokens(inputTokens, usage.outputTokens, usage.totalTokens),
-    providerMetadata: { bedrock: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
  })
 }

 interface ParserState {
+  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
  // Bedrock splits the finish into `messageStop` (carries `stopReason`) and
  // `metadata` (carries usage). Hold the terminal event in state so `onHalt`
@@ -541,20 +545,14 @@ const step = (state: ParserState, event: BedrockEvent) =>
      const reasoning = event.contentBlockDelta.delta.reasoningContent
      const events: LLMEvent[] = []
      const redactedData = reasoning.redactedContent ?? reasoning.data
-      const providerMetadata = reasoning.signature
-        ? bedrockMetadata({ signature: reasoning.signature })
+      const metadata = reasoning.signature
+        ? providerMetadata(state.providerMetadataKey, { signature: reasoning.signature })
        : redactedData !== undefined
-          ? bedrockMetadata({ redactedData })
+          ? providerMetadata(state.providerMetadataKey, { redactedData })
          : undefined
      const lifecycle =
-        reasoning.text !== undefined || providerMetadata !== undefined
-          ? Lifecycle.reasoningDelta(
-              state.lifecycle,
-              events,
-              `reasoning-${index}`,
-              reasoning.text ?? "",
-              providerMetadata,
-            )
+        reasoning.text !== undefined || metadata !== undefined
+          ? Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text ?? "", metadata)
          : state.lifecycle
      return [
        {
@@ -596,7 +594,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
            events,
            `reasoning-${index}`,
            state.reasoningSignatures[index]
-              ? bedrockMetadata({ signature: state.reasoningSignatures[index] })
+              ? providerMetadata(state.providerMetadataKey, { signature: state.reasoningSignatures[index] })
              : undefined,
          )
      events.push(...resultEvents)
@@ -633,7 +631,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
    }

    if (event.metadata) {
-      const usage = mapUsage(event.metadata.usage) ?? state.pendingFinish?.usage
+      const usage = mapUsage(event.metadata.usage, state.providerMetadataKey) ?? state.pendingFinish?.usage
      return [
        {
          ...state,
@@ -652,7 +650,9 @@ const step = (state: ParserState, event: BedrockEvent) =>
        method: "stream",
        reason: classifyProviderFailure({
          message:
-            event.exception.details.message ?? event.exception.details.originalMessage ?? "Bedrock Converse stream error",
+            event.exception.details.message ??
+            event.exception.details.originalMessage ??
+            "Bedrock Converse stream error",
          code: event.exception.type,
        }),
      })
@@ -696,7 +696,8 @@ export const protocol = Protocol.make({
  },
  stream: {
    event: BedrockEvent,
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
      pendingFinish: undefined,
      hasToolCalls: false,
@@ -82,7 +82,9 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
        "Failed to parse Bedrock Converse event-stream payload",
      )) as Record<string, unknown>
      delete parsed.p
-      out.push(messageType === "exception" ? { exception: { type: eventType, details: parsed } } : { [eventType]: parsed })
+      out.push(
+        messageType === "exception" ? { exception: { type: eventType, details: parsed } } : { [eventType]: parsed },
+      )
    }
    return [cursor, out] as const
  })
@@ -229,6 +229,7 @@ type GeminiEvent = Schema.Schema.Type<typeof GeminiEvent>

 interface ParserState {
  readonly route: string
+  readonly providerMetadataKey: string
  readonly finishReason?: string
  readonly hasToolCalls: boolean
  readonly promptFeedback?: GeminiPromptFeedback
@@ -285,22 +286,23 @@ const lowerUserPart = Effect.fn("Gemini.lowerUserPart")(function* (part: TextPar
  return { inlineData: { mimeType: media.mime, data: media.base64 } }
 })

-const googleMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ google: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })

-const thoughtSignature = (providerMetadata: ProviderMetadata | undefined) => {
-  const google = providerMetadata?.google
-  return ProviderShared.isRecord(google) && typeof google.thoughtSignature === "string"
-    ? google.thoughtSignature
+const thoughtSignature = (metadata: ProviderMetadata | undefined, key: string) => {
+  const value = metadata?.[key]
+  return ProviderShared.isRecord(value) && typeof value.thoughtSignature === "string"
+    ? value.thoughtSignature
    : undefined
 }

-const lowerToolCall = (part: ToolCallPart, omitIds: boolean) => ({
+const lowerToolCall = (part: ToolCallPart, omitIds: boolean, metadataKey: string) => ({
  functionCall: { ...(omitIds ? {} : { id: part.id }), name: part.name, args: part.input },
-  thoughtSignature: thoughtSignature(part.providerMetadata),
+  thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
 })

 const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMRequest) {
  const contents: GeminiContent[] = []
+  const metadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
  const omitCallIds = omitsFunctionCallIds(request.model.id)
  const legacyToolMedia = routesLegacyToolMedia(request.model.id)
  let pendingMedia: GeminiInlineDataPart[] | undefined
@@ -342,15 +344,19 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
        if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call"]))
          return yield* ProviderShared.unsupportedContent("Gemini", "assistant", ["text", "reasoning", "tool-call"])
        if (part.type === "text") {
-          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey) })
          continue
        }
        if (part.type === "reasoning") {
-          parts.push({ text: part.text, thought: true, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({
+            text: part.text,
+            thought: true,
+            thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
+          })
          continue
        }
        if (part.type === "tool-call") {
-          const lowered = lowerToolCall(part, omitCallIds)
+          const lowered = lowerToolCall(part, omitCallIds, metadataKey)
          const signature = lowered.thoughtSignature
          parts.push({
            ...lowered,
@@ -498,7 +504,7 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
 // `cachedContentTokenCount` subset. `candidatesTokenCount` is *exclusive*
 // of `thoughtsTokenCount` — visible-only, not a total — so we sum the two
 // to produce the inclusive `outputTokens` the rest of the contract expects.
-const mapUsage = (usage: GeminiUsage | undefined) => {
+const mapUsage = (usage: GeminiUsage | undefined, metadataKey: string) => {
  if (!usage) return undefined
  // Explicit provider nulls decode as `null`; normalize to `undefined` so the
  // token arithmetic below treats them like absent counts.
@@ -519,7 +525,7 @@ const mapUsage = (usage: GeminiUsage | undefined) => {
    cacheReadInputTokens: cached,
    reasoningTokens: thoughts,
    totalTokens: ProviderShared.totalTokens(promptTokens, outputTokens, usage.totalTokenCount ?? undefined),
-    providerMetadata: { google: usage },
+    providerMetadata: providerMetadata(metadataKey, usage),
  })
 }

@@ -567,10 +573,15 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
      lifecycle,
      events,
      "reasoning-0",
-      googleMetadata({ thoughtSignature: state.reasoningSignature }),
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.reasoningSignature }),
    )
  if (state.textSignature !== undefined)
-    lifecycle = Lifecycle.textEnd(lifecycle, events, "text-0", googleMetadata({ thoughtSignature: state.textSignature }))
+    lifecycle = Lifecycle.textEnd(
+      lifecycle,
+      events,
+      "text-0",
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.textSignature }),
+    )
  Lifecycle.finish(lifecycle, events, {
    reason: {
      normalized:
@@ -579,7 +590,9 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
    },
    usage: state.usage,
    providerMetadata:
-      state.promptFeedback === undefined ? undefined : googleMetadata({ promptFeedback: state.promptFeedback }),
+      state.promptFeedback === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { promptFeedback: state.promptFeedback }),
  })
  return events
 }
@@ -588,7 +601,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
  const nextState = {
    ...state,
    promptFeedback: event.promptFeedback ?? state.promptFeedback,
-    usage: event.usageMetadata ? (mapUsage(event.usageMetadata) ?? state.usage) : state.usage,
+    usage: event.usageMetadata
+      ? (mapUsage(event.usageMetadata, state.providerMetadataKey) ?? state.usage)
+      : state.usage,
  }
  const candidate = event.candidates?.[0]
  if (!candidate?.content)
@@ -632,7 +647,7 @@ const step = (state: ParserState, event: GeminiEvent) => {
          events,
          "reasoning-0",
          part.text,
-          signature ? googleMetadata({ thoughtSignature: signature }) : undefined,
+          signature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: signature }) : undefined,
        )
        continue
      }
@@ -640,14 +655,16 @@ const step = (state: ParserState, event: GeminiEvent) => {
        lifecycle,
        events,
        "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
      )
      lifecycle = Lifecycle.textDelta(
        lifecycle,
        events,
        "text-0",
        part.text,
-        textSignature ? googleMetadata({ thoughtSignature: textSignature }) : undefined,
+        textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
      )
      textSignature = undefined
      continue
@@ -667,7 +684,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
        lifecycle,
        events,
        "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
      )
      lifecycle = Lifecycle.stepStart(lifecycle, events)
      events.push(
@@ -675,8 +694,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
          id,
          name: part.functionCall.name,
          input,
-          providerMetadata:
-            part.thoughtSignature ? googleMetadata({ thoughtSignature: part.thoughtSignature }) : undefined,
+          providerMetadata: part.thoughtSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: part.thoughtSignature })
+            : undefined,
        }),
      )
      hasToolCalls = true
@@ -714,6 +734,7 @@ export const protocol = Protocol.make({
    event: Protocol.jsonEvent(GeminiEvent),
    initial: (request) => ({
      route: `${request.model.provider}/${request.model.route.id}`,
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      hasToolCalls: false,
      lifecycle: Lifecycle.initial(),
    }),
@@ -196,7 +196,7 @@ type LoweredInputItem =
 // multiple streamed summary parts into the same item before flushing.
 type OpenResponsesReasoningInput = {
  type: "reasoning"
-  id: string
+  id?: string
  summary: Array<{ type: "summary_text"; text: string }>
  encrypted_content?: string | null
 }
@@ -346,7 +346,8 @@ export const Event = Schema.StructWithRest(
    item_id: Schema.optional(Schema.String),
    output_index: Schema.optional(Schema.Number),
    summary_index: Schema.optional(Schema.Number),
-    item: Schema.optional(StreamItem),
+    // OutputItemAdded/Done permit a null item in the Open Responses OpenAPI schema.
+    item: optionalNull(StreamItem),
    response: Schema.optional(
      Schema.StructWithRest(
        Schema.Struct({
@@ -372,9 +373,6 @@ export const Event = Schema.StructWithRest(
 )
 export type Event = Schema.Schema.Type<typeof Event>

-// Which lowered input item a persisted item id is about to be attached to.
-export type ItemKind = "message" | "reasoning" | "function-call" | "hosted-tool"
-
 export interface Extension {
  readonly id: string
  readonly name: string
@@ -384,10 +382,6 @@ export interface Extension {
    readonly request: LLMRequest
  }) => MediaInput | undefined
  readonly lowerHostedToolItem?: (item: unknown) => ExtendedHostedToolItem | 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 }
@@ -447,53 +441,37 @@ 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}$/
+// Server-issued item ids need a nonempty prefix and suffix, but the prefix is
+// provider-defined and does not necessarily identify the item's semantic type.
 const itemID = (providerMetadata: ProviderMetadata | undefined, providerMetadataKey: string) => {
  const metadata = providerMetadata?.[providerMetadataKey]
-  return ProviderShared.isRecord(metadata) &&
-    typeof metadata.itemId === "string" &&
-    ITEM_ID_PATTERN.test(metadata.itemId)
-    ? metadata.itemId
-    : undefined
+  if (!ProviderShared.isRecord(metadata) || typeof metadata.itemId !== "string") return undefined
+  const separator = metadata.itemId.indexOf("_")
+  return separator > 0 && separator < metadata.itemId.length - 1 ? metadata.itemId : undefined
 }

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

-const lowerReasoning = (
-  part: ReasoningPart,
-  providerMetadataKey: string,
-  extension: Extension,
-): OpenResponsesReasoningInput | undefined => {
+const lowerReasoning = (part: ReasoningPart, providerMetadataKey: string): OpenResponsesReasoningInput | undefined => {
  const metadata = part.providerMetadata?.[providerMetadataKey]
+  if (!ProviderShared.isRecord(metadata)) return undefined
  const id = itemID(part.providerMetadata, providerMetadataKey)
-  if (!ProviderShared.isRecord(metadata) || !acceptsItemID(extension, "reasoning", id)) return undefined
  const encryptedContent =
    typeof metadata.reasoningEncryptedContent === "string" || metadata.reasoningEncryptedContent === null
      ? metadata.reasoningEncryptedContent
      : undefined
  return {
    type: "reasoning",
-    id,
+    ...(id === undefined ? {} : { id }),
    summary: part.text.length > 0 ? [{ type: "summary_text", text: part.text }] : [],
    encrypted_content: encryptedContent,
  }
@@ -614,8 +592,7 @@ 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 rawID = itemID(part.providerMetadata, providerMetadataKey)
-          const id = acceptsItemID(extension, "message", rawID) ? rawID : undefined
+          const id = itemID(part.providerMetadata, providerMetadataKey)
          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)
@@ -640,23 +617,23 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        }
        if (part.type === "reasoning") {
          flushText()
-          const reasoning = lowerReasoning(part, providerMetadataKey, extension)
+          const reasoning = lowerReasoning(part, providerMetadataKey)
          if (!reasoning) continue
-          const existing = reasoningItems[reasoning.id]
+          const existing = reasoning.id === undefined ? undefined : reasoningItems[reasoning.id]
          if (existing) {
            existing.summary.push(...reasoning.summary)
            if (typeof reasoning.encrypted_content === "string")
              existing.encrypted_content = reasoning.encrypted_content
            continue
          }
-          reasoningItems[reasoning.id] = reasoning
+          if (reasoning.id !== undefined) reasoningItems[reasoning.id] = reasoning
          input.push(reasoning)
          continue
        }
        if (part.type === "tool-call") {
          flushText()
          if (part.providerExecuted === true) continue
-          input.push(lowerToolCall(part, providerMetadataKey, extension))
+          input.push(lowerToolCall(part, providerMetadataKey))
          continue
        }
        if (part.type === "tool-result" && part.providerExecuted === true) {
@@ -668,7 +645,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
              : Schema.is(HostedToolItem)(part.result.value)
                ? part.result.value
                : extension.lowerHostedToolItem?.(part.result.value)
-          if (acceptsItemID(extension, "hosted-tool", id) && hosted?.id === id) {
+          if (id !== undefined && hosted?.id === id) {
            if (!hostedToolItems.has(id)) {
              input.push(hosted)
              hostedToolItems.add(id)
@@ -836,7 +813,7 @@ export const providerMetadata = (state: ParserState, metadata: Record<string, un
 })

 const isReasoningItem = (item: StreamItem): item is StreamItem & { type: "reasoning"; id: string } =>
-  item.type === "reasoning" && typeof item.id === "string" && item.id.length > 0
+  item.type === "reasoning" && typeof item.id === "string"

 export type StepResult = readonly [ParserState, ReadonlyArray<LLMEvent>]

@@ -915,7 +892,7 @@ 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 !== undefined) {
    const phase = messagePhase(item.phase)
    return [
      {
@@ -946,7 +923,7 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
  }
  if (item?.type !== "function_call" || !item.call_id) return [state, NO_EVENTS]
  const id = item.id ?? item.call_id
-  const metadata = item.id ? providerMetadata(state, { itemId: item.id }) : undefined
+  const metadata = item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined
  const events: LLMEvent[] = []
  const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
  return [
@@ -965,7 +942,7 @@ 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]
+  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
  const item = state.reasoningItems[event.item_id]
  if (!item) return [state, NO_EVENTS]
  if (event.summary_index === 0) return [state, NO_EVENTS]
@@ -1012,7 +989,7 @@ const onReasoningSummaryPartAdded = (state: ParserState, event: Event): StepResu
 }

 const onReasoningSummaryPartDone = (state: ParserState, event: Event): StepResult => {
-  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
+  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
  const item = state.reasoningItems[event.item_id]
  if (!item) return [state, NO_EVENTS]
  return [
@@ -1037,7 +1014,7 @@ const onFunctionCallArgumentsDelta = Effect.fn("OpenResponses.onFunctionCallArgu
  state: ParserState,
  event: Event,
 ) {
-  if (!event.item_id) return [state, NO_EVENTS] satisfies StepResult
+  if (event.item_id === undefined) return [state, NO_EVENTS] satisfies StepResult
  const tool = state.tools[event.item_id]
  if (!tool) return [state, NO_EVENTS] satisfies StepResult
  const final = event.type === "response.function_call_arguments.done" ? event.arguments : undefined
@@ -1068,7 +1045,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
  const item = event.item
  if (!item) return [state, NO_EVENTS] satisfies StepResult

-  if (item.type === "message" && item.id) {
+  if (item.type === "message" && item.id !== undefined) {
    const itemPhase = messagePhase(item.phase)
    const phase = itemPhase === undefined ? state.messagePhases[item.id] : itemPhase
    const events: LLMEvent[] = []
@@ -1099,7 +1076,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
      : ToolStream.start(state.tools, id, {
          id: item.call_id,
          name: item.name,
-          providerMetadata: item.id ? providerMetadata(state, { itemId: item.id }) : undefined,
+          providerMetadata: item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined,
        })
    const result =
      item.arguments === undefined
@@ -1160,7 +1137,7 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
          ([current, events], item) => {
            const id = item.id ?? (item.type === "function_call" ? item.call_id : undefined)
            if (
-              !id ||
+              id === undefined ||
              ((item.type !== "function_call" || !current.tools[id]) &&
                (item.type !== "reasoning" || !current.reasoningItems[id]))
            )
@@ -1243,12 +1220,13 @@ export const providerFailure = (id: string, event: Event, fallback: string) => {
 const providerError = (state: ParserState, event: Event, fallback: string) => providerFailure(state.id, event, fallback)

 export const step = (state: ParserState, input: Event) => {
+  // The OpenAPI requires string IDs but imposes no minLength; empty is not missing.
  const event =
-    input.item_id && outputItemID(state, input) !== input.item_id
+    input.item_id !== undefined && outputItemID(state, input) !== input.item_id
      ? { ...input, item_id: outputItemID(state, input) }
      : input
  if (event.type === "response.output_text.delta" || event.type === "response.output_text.done") {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(
      event.type === "response.output_text.delta"
        ? onOutputTextDelta(state, event, event.item_id)
@@ -1257,7 +1235,7 @@ export const step = (state: ParserState, input: Event) => {
  }
  if (event.type === "response.refusal.delta" || event.type === "response.refusal.done") {
    const value = event.type === "response.refusal.delta" ? event.delta : event.refusal
-    if (!event.item_id || typeof value !== "string")
+    if (event.item_id === undefined || typeof value !== "string")
      return ProviderShared.eventError(state.id, `${event.type} is malformed`)
    return Effect.succeed(
      event.type === "response.refusal.delta"
@@ -1266,7 +1244,7 @@ export const step = (state: ParserState, input: Event) => {
    )
  }
  if (event.type === "response.reasoning.delta" || event.type === "response.reasoning_summary_text.delta") {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(onReasoningDelta(state, event, event.item_id))
  }
  if (
@@ -1274,24 +1252,24 @@ export const step = (state: ParserState, input: Event) => {
    event.type === "response.reasoning_summary_text.done" ||
    event.type === "response.reasoning_text.done"
  ) {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(onReasoningDone(state, event, event.item_id))
  }
  if (event.type === "response.reasoning_summary_part.added")
-    return event.item_id
+    return event.item_id !== undefined
      ? Effect.succeed(onReasoningSummaryPartAdded(state, event))
      : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
  if (event.type === "response.reasoning_summary_part.done")
-    return event.item_id
+    return event.item_id !== undefined
      ? Effect.succeed(onReasoningSummaryPartDone(state, event))
      : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
  if (event.type === "response.output_item.added") {
-    if (event.item?.type === "message" && !event.item.id)
+    if (event.item?.type === "message" && event.item.id === undefined)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
    const id = event.item?.id ?? (event.item?.type === "function_call" ? event.item.call_id : undefined)
    return Effect.succeed(
      onOutputItemAdded(
-        event.output_index !== undefined && id
+        event.output_index !== undefined && id !== undefined
          ? { ...state, outputItems: { ...state.outputItems, [event.output_index]: id } }
          : state,
        event,
@@ -1299,11 +1277,11 @@ export const step = (state: ParserState, input: Event) => {
    )
  }
  if (event.type === "response.function_call_arguments.delta" || event.type === "response.function_call_arguments.done")
-    return event.item_id
+    return event.item_id !== undefined
      ? onFunctionCallArgumentsDelta(state, event)
      : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
  if (event.type === "response.output_item.done") {
-    if (event.item?.type === "message" && !event.item.id)
+    if (event.item?.type === "message" && event.item.id === undefined)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
    return onOutputItemDone(state, event)
  }
@@ -253,6 +253,7 @@ interface PendingToolDelta {
 }

 export interface ParserState {
+  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
  readonly pendingTools: Partial<Record<number, PendingToolDelta>>
  readonly toolCallEvents: ReadonlyArray<LLMEvent>
@@ -324,17 +325,18 @@ const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart
 const openAICompatibleReasoningContent = (native: unknown) =>
  isRecord(native) && typeof native.reasoning_content === "string" ? native.reasoning_content : undefined

-const reasoningField = (part: ReasoningPart) => {
-  const field = part.providerMetadata?.openai?.reasoningField
+const reasoningField = (part: ReasoningPart, providerMetadataKey: string) => {
+  const field = part.providerMetadata?.[providerMetadataKey]?.reasoningField
  return typeof field === "string" ? field : undefined
 }

-const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown) => {
+const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown, providerMetadataKey: string) => {
  const observed = parts.flatMap((part) => {
-    const details = part.providerMetadata?.openai?.reasoningDetails
+    const details = part.providerMetadata?.[providerMetadataKey]?.reasoningDetails
    return Array.isArray(details) ? details : []
  })
-  if (parts.some((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))) return observed
+  if (parts.some((part) => Array.isArray(part.providerMetadata?.[providerMetadataKey]?.reasoningDetails)))
+    return observed
  if (isRecord(native) && Array.isArray(native.reasoning_details)) return native.reasoning_details
 }

@@ -366,7 +368,7 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
  message: OpenAIChatRequestMessage,
  configuredField: string | undefined,
  requireReasoning: boolean,
-  options: LoweringOptions,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
  const content: TextPart[] = []
  const reasoning: ReasoningPart[] = []
@@ -388,10 +390,14 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
    }
  }
  const text = reasoning.map((part) => part.text).join("")
-  const details = reasoningDetails(reasoning, message.native?.openaiCompatible)
-  const observedField = reasoning.map(reasoningField).find((value) => value !== undefined)
+  const details = reasoningDetails(reasoning, message.native?.openaiCompatible, options.providerMetadataKey)
+  const observedField = reasoning
+    .map((part) => reasoningField(part, options.providerMetadataKey))
+    .find((value) => value !== undefined)
  const nativeReasoning = openAICompatibleReasoningContent(message.native?.openaiCompatible)
-  const fullyStructured = reasoning.every((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))
+  const fullyStructured = reasoning.every((part) =>
+    Array.isArray(part.providerMetadata?.[options.providerMetadataKey]?.reasoningDetails),
+  )
  const field = (() => {
    if (configuredField !== undefined && (requireReasoning || reasoning.length > 0 || nativeReasoning !== undefined))
      return configuredField
@@ -459,7 +465,7 @@ const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
  message: OpenAIChatRequestMessage,
  reasoningField: string | undefined,
  requireReasoning: boolean,
-  options: LoweringOptions,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
  if (message.role === "user") return [yield* lowerUserMessage(message, options)]
  if (message.role === "assistant")
@@ -495,8 +501,13 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
  const mistral = ["mistral", "devstral", "codestral", "pixtral", "mixtral"].some((family) => modelID.includes(family))
  const lowering = {
    ...options,
+    providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
    toolCallID: (id: string) => {
-      if (mistral) return id.replace(/[^a-zA-Z0-9]/g, "").slice(0, 9).padEnd(9, "0")
+      if (mistral)
+        return id
+          .replace(/[^a-zA-Z0-9]/g, "")
+          .slice(0, 9)
+          .padEnd(9, "0")
      if (modelID.includes("claude")) return id.replace(/[^a-zA-Z0-9_-]/g, "_")
      if (request.model.provider === "openai" || request.model.provider === "azure" || modelID.startsWith("openai/"))
        return id.slice(0, 40)
@@ -505,7 +516,8 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
  }
  const requireAssistantAfterTool = request.model.compatibility?.requireAssistantAfterTool ?? mistral
  const bridgeTools = () => {
-    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool") messages.push({ role: "assistant", content: "Done." })
+    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool")
+      messages.push({ role: "assistant", content: "Done." })
  }
  const pendingImages: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
  const flushImages = () => {
@@ -557,7 +569,10 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
        )
      continue
    }
-    if (message.role === "assistant" && message.content.every((part) => part.type === "text" && part.text.trim() === ""))
+    if (
+      message.role === "assistant" &&
+      message.content.every((part) => part.type === "text" && part.text.trim() === "")
+    )
      continue
    if (message.role === "tool") {
      const lowered = yield* lowerToolMessages(message, lowering)
@@ -588,7 +603,10 @@ const hasToolHistory = (messages: ReadonlyArray<LLMRequest["messages"][number]>)
 // models.dev provider naming: DeepSeek, Moonshot AI, Together AI, ZAI
 // (Zhipu + Coding Plan variants), Nvidia, Cerebras, Chutes, etc. still
 // require `max_tokens`.
-const detectMaxTokensField = (provider: string, baseURL: string | undefined): "max_tokens" | "max_completion_tokens" => {
+const detectMaxTokensField = (
+  provider: string,
+  baseURL: string | undefined,
+): "max_tokens" | "max_completion_tokens" => {
  const p = provider.toLowerCase()
  const url = (baseURL ?? "").toLowerCase()
  if (
@@ -638,7 +656,8 @@ const detectSupportsStore = (provider: string, baseURL: string | undefined): boo
  const isChutes = p === "chutes" || url.includes("chutes.ai")
  const isCloudflareWorkersAI = p === "cloudflare-workers-ai" || url.includes("api.cloudflare.com")
  const isCloudflareAiGateway = p === "cloudflare-ai-gateway" || url.includes("gateway.ai.cloudflare.com")
-  const isVercelAiGateway = p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
+  const isVercelAiGateway =
+    p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
  const isAntLing = p === "ant-ling" || url.includes("api.ant-ling.com")
  const isOpencode = p === "opencode" || url.includes("opencode.ai")
  const isNonStandard =
@@ -670,11 +689,7 @@ const detectSupportsStrictMode = (provider: string, baseURL: string | undefined)
  return !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia
 }

-const detectZaiToolStream = (
-  provider: string,
-  baseURL: string | undefined,
-  modelID: string,
-): boolean => {
+const detectZaiToolStream = (provider: string, baseURL: string | undefined, modelID: string): boolean => {
  const p = provider.toLowerCase()
  const url = (baseURL ?? "").toLowerCase()
  const isZai =
@@ -724,10 +739,10 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
  const supportsStore = request.model.compatibility?.supportsStore ?? detectSupportsStore(provider, baseURL)
  const supportsUsageInStreaming =
    request.model.compatibility?.supportsUsageInStreaming ?? detectSupportsUsageInStreaming()
-  const supportsStrictMode = request.model.compatibility?.supportsStrictMode ?? detectSupportsStrictMode(provider, baseURL)
+  const supportsStrictMode =
+    request.model.compatibility?.supportsStrictMode ?? detectSupportsStrictMode(provider, baseURL)
  const zaiToolStream =
-    request.model.compatibility?.zaiToolStream ??
-    detectZaiToolStream(provider, baseURL, request.model.id)
+    request.model.compatibility?.zaiToolStream ?? detectZaiToolStream(provider, baseURL, request.model.id)
  const hasHistory = hasToolHistory(request.messages)
  const hasActiveTools = request.tools.length > 0
  return {
@@ -812,15 +827,14 @@ const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event
 // Providers differ on cache-hit location: OpenAI uses
 // `prompt_tokens_details.cached_tokens`, DeepSeek uses
 // `prompt_cache_hit_tokens`, and Zai uses top-level `cached_tokens`.
-const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
+const mapUsage = (usage: OpenAIChatEvent["usage"], providerMetadataKey: string): Usage | undefined => {
  if (!usage) return undefined
  const input = usage.prompt_tokens ?? undefined
  const output = usage.completion_tokens ?? undefined
-  const cached =
-    (usage.prompt_tokens_details?.cached_tokens ??
-      (usage as { prompt_cache_hit_tokens?: number | null }).prompt_cache_hit_tokens ??
-      (usage as { cached_tokens?: number | null }).cached_tokens ??
-      undefined) as number | undefined
+  const cached = (usage.prompt_tokens_details?.cached_tokens ??
+    (usage as { prompt_cache_hit_tokens?: number | null }).prompt_cache_hit_tokens ??
+    (usage as { cached_tokens?: number | null }).cached_tokens ??
+    undefined) as number | undefined
  const cacheWrite = usage.prompt_tokens_details?.cache_write_tokens ?? undefined
  const reasoning = usage.completion_tokens_details?.reasoning_tokens ?? undefined
  const nonCached = ProviderShared.subtractTokens(input, ProviderShared.sumTokens(cached, cacheWrite))
@@ -832,7 +846,7 @@ const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
    cacheWriteInputTokens: cacheWrite,
    reasoningTokens: reasoning,
    totalTokens: ProviderShared.totalTokens(input, output, usage.total_tokens ?? undefined),
-    providerMetadata: { openai: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
  })
 }

@@ -906,8 +920,12 @@ const conflictingReasoningTextDetails = (previous: Record<string, unknown>, curr
 const conflictingDetailValue = (previous: unknown, current: unknown) =>
  previous !== undefined && previous !== null && current !== undefined && current !== null && previous !== current

-const reasoningMetadata = (field: ParserState["reasoningField"], details?: ReadonlyArray<unknown>) => ({
-  openai: {
+const reasoningMetadata = (
+  providerMetadataKey: string,
+  field: ParserState["reasoningField"],
+  details?: ReadonlyArray<unknown>,
+) => ({
+  [providerMetadataKey]: {
    ...(field ? { reasoningField: field } : {}),
    ...(details ? { reasoningDetails: details } : {}),
  },
@@ -934,15 +952,17 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
    // Moonshot (and a few other OpenAI-compatible providers) attach usage to
    // `choice.usage` instead of the top-level `usage` field.
    const choiceUsage = (choice as unknown as { usage?: OpenAIChatEvent["usage"] })?.usage
-    const usage = mapUsage(event.usage) ?? (choiceUsage ? mapUsage(choiceUsage) : undefined) ?? state.usage
+    const usage =
+      mapUsage(event.usage, state.providerMetadataKey) ??
+      (choiceUsage ? mapUsage(choiceUsage, state.providerMetadataKey) : undefined) ??
+      state.usage
    const rawFinishReason = choice?.finish_reason
-    const finishReason =
-      rawFinishReason
-        ? {
-            normalized: yield* mapFinishReason(event, rawFinishReason),
-            raw: choice?.native_finish_reason ?? rawFinishReason,
-          }
-        : state.finishReason
+    const finishReason = rawFinishReason
+      ? {
+          normalized: yield* mapFinishReason(event, rawFinishReason),
+          raw: choice?.native_finish_reason ?? rawFinishReason,
+        }
+      : state.finishReason
    const delta = choice?.delta
    const toolDeltas = delta?.tool_calls ?? []
    let tools = state.tools
@@ -973,7 +993,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
    const detailDelta = Array.isArray(delta?.reasoning_details) ? delta.reasoning_details : undefined
    if (detailDelta !== undefined) appendReasoningDetails(state.reasoningDetails, detailDelta)
    const reasoningDetailsObserved = state.reasoningDetailsObserved || detailDelta !== undefined
-    const deltaMetadata = reasoningMetadata(reasoningField)
+    const deltaMetadata = reasoningMetadata(state.providerMetadataKey, reasoningField)
    const text = detailDelta?.length ? (detailText(detailDelta) ?? reasoning?.text) : reasoning?.text
    if (text !== undefined) lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", text, deltaMetadata)
    else if (
@@ -989,7 +1009,11 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
        lifecycle,
        events,
        "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
+        reasoningMetadata(
+          state.providerMetadataKey,
+          reasoningField,
+          reasoningDetailsObserved ? state.reasoningDetails : undefined,
+        ),
      )
      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)
    }
@@ -999,7 +1023,11 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
        lifecycle,
        events,
        "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
+        reasoningMetadata(
+          state.providerMetadataKey,
+          reasoningField,
+          reasoningDetailsObserved ? state.reasoningDetails : undefined,
+        ),
      )
      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)
    }
@@ -1060,6 +1088,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>

    return [
      {
+        providerMetadataKey: state.providerMetadataKey,
        tools: finished?.tools ?? tools,
        pendingTools,
        toolCallEvents: finished?.events ?? state.toolCallEvents,
@@ -1103,12 +1132,18 @@ const finishEvents = Effect.fn("OpenAIChat.finishEvents")(function* (state: Pars
      }
    : { normalized: hasToolCalls ? ("tool-calls" as const) : ("stop" as const) }
  const metadata = reasoningMetadata(
+    state.providerMetadataKey,
    state.reasoningField,
    state.reasoningDetailsObserved ? state.reasoningDetails : undefined,
  )
  const started =
    state.reasoningDetailsObserved && !state.reasoningEmitted
-      ? Lifecycle.reasoningStart(state.lifecycle, events, "reasoning-0", reasoningMetadata(state.reasoningField))
+      ? Lifecycle.reasoningStart(
+          state.lifecycle,
+          events,
+          "reasoning-0",
+          reasoningMetadata(state.providerMetadataKey, state.reasoningField),
+        )
      : state.lifecycle
  const ended = Lifecycle.reasoningEnd(started, events, "reasoning-0", metadata)
  const lifecycle = toolCallEvents.length ? Lifecycle.stepStart(ended, events) : ended
@@ -1135,6 +1170,7 @@ export const protocol = Protocol.make({
  stream: {
    event: Protocol.jsonEvent(OpenAIChatEvent),
    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
      pendingTools: {},
      toolCallEvents: [],
@@ -86,28 +86,10 @@ const OpenAIResponsesBody = Schema.Struct({
 })
 export type OpenAIResponsesBody = Schema.Schema.Type<typeof OpenAIResponsesBody>

-// Replayed items are paired with stored server state by id, so a foreign or
-// synthetic token can fail request validation even when `call_id` pairing is
-// intact. Only resend ids in each item kind's own grammar; hosted tool
-// items 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 items keep generic validation.
-  "hosted-tool": [],
-}
-
 const extension = {
  id: ADAPTER,
  name: NAME,
  lowerHostedToolItem: (item: unknown) => (Schema.is(OpenAIResponsesHostedToolItem)(item) ? item : undefined),
-  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

 const nativeImageToolInput = (tool: ToolDefinition) => {
@@ -202,7 +184,7 @@ const HOSTED_TOOLS = {

 const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
  if (event.type === "response.reasoning_text.delta")
-    return event.item_id
+    return event.item_id !== undefined
      ? Effect.succeed(
          OpenResponses.onReasoningDelta(state, event, OpenResponses.outputItemID(state, event) ?? event.item_id),
        )
@@ -29,10 +29,9 @@ export type ResponseIncludable = (typeof ResponseIncludables)[number] | (string

 export const ServiceTiers = ["auto", "default", "flex", "priority"] as const
 export type ServiceTier = (typeof ServiceTiers)[number] | (string & {})
-export const ServiceTier = Schema.declare<ServiceTier>(
-  (value): value is ServiceTier => typeof value === "string",
-  { title: "ServiceTier" },
-)
+export const ServiceTier = Schema.declare<ServiceTier>((value): value is ServiceTier => typeof value === "string", {
+  title: "ServiceTier",
+})

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

 export * as ResponsesHostedTools from "./responses-hosted-tools.js"
@@ -40,15 +40,16 @@ const patterns = [
  /model_context_window_exceeded/i,
  /too many tokens/i,
  /token limit exceeded/i,
+  /request_too_large/i,
 ]

-const payloadPatterns = [/request_too_large/i, /request entity too large/i, /payload too large/i, /request too large/i]
+const payloadPatterns = [/request entity too large/i, /payload too large/i, /request too large/i]

 const exclusions = [/^(throttling error|service unavailable):/i, /rate limit/i, /too many requests/i]

 export const isContextOverflow = (message: string) =>
  !exclusions.some((pattern) => pattern.test(message)) &&
-  (patterns.some((pattern) => pattern.test(message)) || /^400\s*(status code)?\s*\(no body\)/i.test(message))
+  (patterns.some((pattern) => pattern.test(message)) || /^4(?:00|13)\s*(status code)?\s*\(no body\)/i.test(message))

 export const isPayloadTooLarge = (message: string) => payloadPatterns.some((pattern) => pattern.test(message))

@@ -106,6 +107,7 @@ export function classifyProviderFailure(input: ProviderFailure): AIError["reason
    clientScoped &&
    (codes.includes("context_length_exceeded") ||
      codes.includes("model_context_window_exceeded") ||
+      codes.includes("request_too_large") ||
      isContextOverflow(text))
  )
    return new InvalidRequestReason({ ...common, classification: "context-overflow" })
@@ -1,5 +1,5 @@
 import { Auth } from "../route/auth.js"
-import type { Route as RouteDef, RouteDefaultsInput } from "../route/client.js"
+import { Route, type RouteDefaultsInput } from "../route/client.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { OpenAIChat } from "../protocols/openai-chat.js"
 import { OpenAIResponses } from "../protocols/openai-responses.js"
@@ -23,22 +23,30 @@ export interface Settings extends ProviderPackage.Settings {
  readonly baseURL?: string
  readonly credentials?: Credentials
  readonly region?: string
+  readonly topP?: number
  readonly providerOptions?: OpenAIProviderOptionsInput
 }

-const responsesRoute = OpenAIResponses.route.with({
+const responsesRoute = Route.make({
  id: "bedrock-mantle-responses",
  provider: id,
+  providerMetadataKey: "mantle",
+  protocol: OpenAIResponses.protocol,
+  endpoint: OpenAIResponses.route.endpoint,
+  auth: OpenAIResponses.route.auth,
+  transport: OpenAIResponses.httpTransport,
+  defaults: OpenAIResponses.route.defaults,
 })

 const chatRoute = OpenAIChat.route.with({
  id: "bedrock-mantle-chat",
  provider: id,
+  providerMetadataKey: "mantle",
 })

 export const routes = [responsesRoute, chatRoute]

-const configuredRoute = <Body, Prepared>(route: RouteDef<Body, Prepared>, input: Config) => {
+const configuredRoute = <Body, Prepared>(route: Route<Body, Prepared>, input: Config) => {
  const region = input.region ?? input.credentials?.region ?? "us-east-1"
  const credentials = input.credentials === undefined ? undefined : { ...input.credentials, region }
  return route.with({
@@ -70,7 +78,7 @@ export const configure = (input: Config = {}) => {

  return {
    id,
-    model: chat,
+    model: responses,
    chat,
    responses,
    configure,
@@ -88,6 +96,7 @@ const config = (settings: Settings): Config => {
    apiKey: settings.auth === "sigv4" ? undefined : settings.apiKey,
    baseURL: settings.baseURL,
    credentials: settings.credentials,
+    generation: settings.topP === undefined ? undefined : { topP: settings.topP },
    headers: settings.headers === undefined ? undefined : { ...settings.headers },
    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
    providerOptions: settings.providerOptions,
@@ -103,4 +112,4 @@ export const responsesModel: ProviderPackage.Definition<Settings, OpenAIProvider
  modelID,
  settings,
 ) => configure(config(settings)).responses(modelID)
-export const model = chatModel
+export const model = responsesModel
@@ -35,6 +35,7 @@ const configuredRoute = (input: Config) => {
  return BedrockConverse.route.with({
    ...rest,
    provider: id,
+    providerMetadataKey: "bedrock",
    endpoint: { baseURL: baseURL ?? bedrockBaseURL(resolvedRegion) },
    auth: apiKey === undefined ? BedrockConverse.sigV4Auth(credentials) : Auth.bearer(apiKey),
  })
@@ -1,2 +1,2 @@
-export { chatModel as model } from "../amazon-bedrock-mantle.js"
+export { responsesModel as model } from "../amazon-bedrock-mantle.js"
 export type { Settings } from "../amazon-bedrock-mantle.js"
@@ -0,0 +1,61 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("deepinfra")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "deepinfra-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.deepinfra.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const root = baseURL?.replace(/\/+$/, "")
+  const configured = route.with({
+    ...defaults,
+    endpoint: {
+      baseURL: root === undefined ? profiles.deepinfra.baseURL : root.endsWith("/openai") ? root : `${root}/openai`,
+    },
+    auth: AuthOptions.bearer(input, "DEEPINFRA_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning_content", supportsStore: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
@@ -27,6 +27,7 @@ export interface Settings extends ProviderPackage.Settings {
 const route = OpenAICompatibleChat.route.with({
  id: "google-vertex-chat",
  provider: id,
+  providerMetadataKey: "vertex",
 })

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

@@ -68,7 +68,7 @@ const protocol = {
 const route = Route.make({
  id: "google-vertex-gemini",
  provider: id,
-  providerMetadataKey: "google",
+  providerMetadataKey: "vertex",
  protocol,
  endpoint: Endpoint.path(({ request }) => {
    const model = String(request.model.id)
@@ -0,0 +1,116 @@
+import { Effect, Schema } from "effect"
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAIChat } from "../protocols/openai-chat.js"
+import { ProviderShared } from "../protocols/shared.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { Route, type RouteDefaultsInput } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Framing } from "../route/framing.js"
+import { Protocol } from "../route/protocol.js"
+import { ProviderID, type ModelID, type LLMRequest } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("groq")
+
+export type ProviderOptions = Pick<OpenAIProviderOptionsInput, "reasoningEffort"> & {
+  /** Controls visible reasoning on GPT-OSS; other models always use parsed reasoning. */
+  readonly includeReasoning?: boolean
+  readonly parallelToolCalls?: boolean
+  readonly serviceTier?: "on_demand" | "flex" | "auto" | "performance" | (string & {})
+  readonly user?: string
+}
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: ProviderOptions
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: ProviderOptions
+}
+
+const Options = Schema.Struct({
+  includeReasoning: Schema.optional(Schema.Boolean),
+  parallelToolCalls: Schema.optional(Schema.Boolean),
+  serviceTier: Schema.optional(Schema.String),
+  user: Schema.optional(Schema.String),
+})
+
+export const protocol = Protocol.make({
+  id: "groq-chat",
+  body: {
+    schema: Schema.Struct({
+      ...OpenAIChat.bodyFields,
+      reasoning_format: Schema.optional(Schema.Literal("parsed")),
+      include_reasoning: Schema.optional(Schema.Boolean),
+      parallel_tool_calls: Schema.optional(Schema.Boolean),
+      service_tier: Schema.optional(Schema.String),
+      user: Schema.optional(Schema.String),
+    }),
+    from: Effect.fn("Groq.fromRequest")(function* (request: LLMRequest) {
+      const options = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(
+        request.providerOptions ?? {},
+      )
+      const gptOSS = request.model.id.startsWith("openai/gpt-oss-")
+      return {
+        ...(yield* OpenAIChat.fromRequest(request)),
+        reasoning_format: gptOSS ? undefined : ("parsed" as const),
+        include_reasoning: gptOSS ? options.includeReasoning : undefined,
+        parallel_tool_calls: options.parallelToolCalls,
+        service_tier: options.serviceTier,
+        user: options.user,
+      }
+    }),
+  },
+  stream: OpenAIChat.protocol.stream,
+})
+
+export const route = Route.make({
+  id: "groq-chat",
+  provider: id,
+  providerMetadataKey: "openai",
+  protocol,
+  endpoint: Endpoint.path("/chat/completions", { baseURL: profiles.groq.baseURL }),
+  framing: Framing.sse,
+})
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.groq.baseURL },
+    auth: AuthOptions.bearer(input, "GROQ_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<ProviderOptions>({
+        id: modelID,
+        compatibility: {
+          maxTokensField: "max_completion_tokens",
+          reasoningField: "reasoning",
+          requireReasoning: false,
+          supportsStore: false,
+          supportsStrictMode: false,
+        },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, ProviderOptions>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
+
+export * as Groq from "./groq.js"
@@ -6,11 +6,13 @@ export * as Azure from "./azure.js"
 export * as Cerebras from "./cerebras.js"
 export * as Cloudflare from "./cloudflare.js"
 export { CloudflareAIGateway, CloudflareWorkersAI } from "./cloudflare.js"
+export * as DeepInfra from "./deepinfra.js"
 export * as Google from "./google.js"
 export * as GoogleVertex from "./google-vertex.js"
 export * as GoogleVertexChat from "./google-vertex-chat.js"
 export * as GoogleVertexMessages from "./google-vertex-messages.js"
 export * as GoogleVertexResponses from "./google-vertex-responses.js"
+export * as Groq from "./groq.js"
 export * as OpenAI from "./openai.js"
 export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
@@ -164,6 +164,7 @@ const bodyOptions = (input: unknown) => {
 export const route = Route.make({
  id: ADAPTER,
  provider: profile.provider,
+  providerMetadataKey: "openrouter",
  protocol,
  endpoint: Endpoint.path("/chat/completions", { baseURL: profile.baseURL }),
  framing: Framing.sse,
@@ -89,6 +89,7 @@ export interface RouteDefaultsInput {
 export interface RoutePatch<Body, Prepared> extends RouteDefaultsInput {
  readonly id?: string
  readonly provider?: string | ProviderID
+  readonly providerMetadataKey?: string
  readonly auth?: Auth.Definition
  readonly transport?: Transport<Body, Prepared, unknown>
  readonly endpoint?: EndpointPatch<Body>
@@ -289,11 +290,16 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
      defaults: routeInput.defaults ?? {},
      body: protocol.body,
      with: (patch: RoutePatch<Body, Prepared>) => {
-        const { id, provider, auth, transport, endpoint, ...defaults } = patch
+        const { id, provider, providerMetadataKey, auth, transport, endpoint, ...defaults } = patch
        return build({
          ...routeInput,
          id: id ?? routeInput.id,
          provider: provider ?? routeInput.provider,
+          providerMetadataKey:
+            providerMetadataKey ??
+            (provider !== undefined && String(provider) !== String(routeInput.provider)
+              ? String(provider)
+              : routeInput.providerMetadataKey),
          auth: auth ?? routeInput.auth,
          endpoint: endpoint ? Endpoint.merge(routeInput.endpoint, endpoint) : routeInput.endpoint,
          transport: (transport as Transport<Body, Prepared, Frame> | undefined) ?? routeInput.transport,
@@ -339,9 +345,7 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
                return onHalt
                  ? parsed.pipe(
                      Stream.concat(
-                        Stream.suspend(() =>
-                          Stream.unwrap(onHalt(state).pipe(Effect.map(Stream.fromIterable))),
-                        ),
+                        Stream.suspend(() => Stream.unwrap(onHalt(state).pipe(Effect.map(Stream.fromIterable)))),
                      ),
                    )
                  : parsed
@@ -40,17 +40,6 @@ const headerDetails = (headers: Headers.Headers) =>
 const normalizedHeaders = (headers: Headers.Headers) =>
  Object.fromEntries(Object.entries(headers).map(([key, value]) => [key.toLowerCase(), value]))

-const requestId = (headers: Record<string, string>) => {
-  return (
-    headers["x-request-id"] ??
-    headers["request-id"] ??
-    headers["x-amzn-requestid"] ??
-    headers["x-amz-request-id"] ??
-    headers["x-goog-request-id"] ??
-    headers["cf-ray"]
-  )
-}
-
 const retryAfterMs = (headers: Record<string, string>) => {
  const millis = Number(headers["retry-after-ms"])
  if (Number.isFinite(millis)) return Math.max(0, millis)
@@ -147,14 +136,12 @@ const responseHttp = (input: {
  readonly request: HttpClientRequest.HttpClientRequest
  readonly response: HttpClientResponse.HttpClientResponse
  readonly body: ReturnType<typeof responseBody>
-  readonly requestId?: string | undefined
  readonly rateLimit?: HttpRateLimitDetails | undefined
 }) =>
  new HttpContext({
    request: requestDetails(input.request),
    response: responseDetails(input.response),
    ...input.body,
-    requestId: input.requestId,
    rateLimit: input.rateLimit,
  })

@@ -179,7 +166,6 @@ const statusError =
            request,
            response,
            body: details,
-            requestId: requestId(headers),
            rateLimit,
          }),
        }),
@@ -216,7 +202,6 @@ export const classifyHttpFailure = (input: {
          ? undefined
          : new HttpResponseDetails({ status: input.status, headers: headerDetails(Headers.fromInput(headers)) }),
      ...details,
-      requestId: requestId(headers),
      rateLimit,
    }),
  })
@@ -29,7 +29,6 @@ export class HttpContext extends Schema.Class<HttpContext>("AI.HttpContext")({
  response: Schema.optional(HttpResponseDetails),
  body: Schema.optional(Schema.String),
  bodyTruncated: Schema.optional(Schema.Boolean),
-  requestId: Schema.optional(Schema.String),
  rateLimit: Schema.optional(HttpRateLimitDetails),
 }) {}

@@ -211,6 +211,28 @@ describe("RequestExecutor", () => {
    }).pipe(Effect.provide(responsesLayer([new Response("request too large", { status: 413 })]))),
  )

+  it.effect("classifies Anthropic request_too_large as context overflow", () =>
+    Effect.gen(function* () {
+      const executor = yield* RequestExecutor.Service
+      const error = yield* executor.execute(request).pipe(Effect.flip)
+
+      expectAIError(error)
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidRequest",
+        classification: "context-overflow",
+        http: { response: { status: 413 } },
+      })
+    }).pipe(
+      Effect.provide(
+        responsesLayer([
+          new Response('{"error":{"type":"request_too_large","message":"Request exceeds the maximum size"}}', {
+            status: 413,
+          }),
+        ]),
+      ),
+    ),
+  )
+
  it.effect("does not classify ordinary invalid requests as context overflow", () =>
    Effect.gen(function* () {
      const executor = yield* RequestExecutor.Service
@@ -305,7 +327,6 @@ describe("RequestExecutor", () => {
          retryAfterMs: 0,
          rateLimit: { retryAfterMs: 0 },
          http: {
-            requestId: "req_123",
            request: {
              method: "POST",
              url: "https://provider.test/v1/chat?api_key=secret&key=secret&debug=1",
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/chat-streams-text",
    "recordedAt": "2026-08-23T17:21:53.198Z"
  },
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/responses-calls-a-tool",
    "recordedAt": "2026-08-23T17:21:55.170Z"
  },
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/responses-continues-after-a-tool-result",
    "recordedAt": "2026-08-23T17:21:56.397Z"
  },
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/responses-streams-text",
    "recordedAt": "2026-08-23T17:21:54.158Z"
  },
@@ -2,11 +2,7 @@
  "version": 1,
  "metadata": {
    "model": "openai.gpt-oss-120b",
-    "tags": [
-      "prefix:bedrock-mantle",
-      "provider:amazon-bedrock",
-      "protocol:openai-responses"
-    ],
+    "tags": ["prefix:bedrock-mantle", "provider:amazon-bedrock", "protocol:openai-responses"],
    "name": "bedrock-mantle/streams-text",
    "recordedAt": "2026-08-25T03:29:02.968Z"
  },
@@ -7,13 +7,7 @@
    "route": "cloudflare-workers-ai",
    "transport": "http",
    "model": "@cf/openai/gpt-oss-20b",
-    "tags": [
-      "prefix:cloudflare-workers-ai",
-      "provider:cloudflare-workers-ai",
-      "tool",
-      "tool-call",
-      "golden"
-    ]
+    "tags": ["prefix:cloudflare-workers-ai", "provider:cloudflare-workers-ai", "tool", "tool-call", "golden"]
  },
  "interactions": [
    {
@@ -35,4 +29,4 @@
      }
    }
  ]
-}
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "text", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-text",
+    "recordedAt": "2026-08-26T00:34:03.019Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":40,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Hello\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"!\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":3.46e-6,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-call", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-call",
+    "recordedAt": "2026-08-26T00:34:04.173Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_SMfBjXa8eCmHLyjfeyARxe3a\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":0.000027480000000000005,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,50 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-loop", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-loop",
+    "recordedAt": "2026-08-26T00:34:05.656Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":0.000028780000000000002,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}],\"reasoning_content\":\"\"},{\"role\":\"tool\",\"tool_call_id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Paris\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" is\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" sunny\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\".\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":0.0000292,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -1,11 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
    "name": "google-vertex/calls-a-tool",
    "recordedAt": "2026-08-23T17:21:51.036Z"
  },
@@ -1,11 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
    "name": "google-vertex/continues-after-a-tool-result",
    "recordedAt": "2026-08-23T17:21:51.853Z"
  },
@@ -1,11 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
    "name": "google-vertex/streams-text",
    "recordedAt": "2026-08-23T17:21:50.112Z"
  },
@@ -0,0 +1,29 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "openai/gpt-oss-20b",
+    "tags": ["prefix:groq-chat", "provider:groq", "protocol:groq-chat", "text", "usage"],
+    "name": "groq-chat/streams-text-with-usage",
+    "recordedAt": "2026-08-26T14:40:09.833Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.groq.com/openai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"openai/gpt-oss-20b\",\"messages\":[{\"role\":\"user\",\"content\":\"Reply with exactly one word: hello\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"reasoning_effort\":\"low\",\"max_completion_tokens\":512,\"include_reasoning\":false,\"service_tier\":\"on_demand\",\"user\":\"recorded-test\"}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"logprobs\":null,\"finish_reason\":null}],\"x_groq\":{\"id\":\"req_01m0z87915eep9bpf10gg7331e\",\"seed\":94036161}}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"hello\"},\"logprobs\":null,\"finish_reason\":null}]}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"x_groq\":{\"id\":\"req_01m0z87915eep9bpf10gg7331e\",\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}}},\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}}}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[],\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}},\"service_tier\":\"on_demand\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -1,12 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:openai-responses-cache",
-      "provider:openai",
-      "protocol:openai-responses",
-      "cache"
-    ],
+    "tags": ["prefix:openai-responses-cache", "provider:openai", "protocol:openai-responses", "cache"],
    "name": "openai-responses-cache/reports-cached-tokens-on-identical-second-call",
    "recordedAt": "2026-08-25T03:29:25.124Z"
  },
@@ -5,14 +5,7 @@
    "route": "openai-responses",
    "transport": "http",
    "model": "gpt-5.5",
-    "tags": [
-      "prefix:openai-responses",
-      "provider:openai",
-      "flagship",
-      "tool",
-      "tool-loop",
-      "golden"
-    ],
+    "tags": ["prefix:openai-responses", "provider:openai", "flagship", "tool", "tool-loop", "golden"],
    "name": "openai-responses/openai-responses-gpt-5-5-tool-loop",
    "recordedAt": "2026-08-20T06:30:22.262Z"
  },
@@ -1,14 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:openai",
-      "protocol:openai-responses",
-      "tool",
-      "tool-result"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:openai", "protocol:openai-responses", "tool", "tool-result"],
    "name": "pdf/openai-tool-result",
    "recordedAt": "2026-08-25T03:29:08.297Z"
  },
@@ -1,13 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:openai",
-      "protocol:openai-responses",
-      "user-input"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:openai", "protocol:openai-responses", "user-input"],
    "name": "pdf/openai-user-input",
    "recordedAt": "2026-08-25T03:29:05.645Z"
  },
@@ -1,14 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:xai",
-      "protocol:xai-responses",
-      "tool",
-      "tool-result"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:xai", "protocol:xai-responses", "tool", "tool-result"],
    "name": "pdf/xai-tool-result",
    "recordedAt": "2026-08-25T03:29:11.774Z"
  },
@@ -1,13 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:xai",
-      "protocol:xai-responses",
-      "user-input"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:xai", "protocol:xai-responses", "user-input"],
    "name": "pdf/xai-user-input",
    "recordedAt": "2026-08-25T03:29:10.612Z"
  },
@@ -52,4 +52,4 @@
      }
    }
  ]
-}
+}
@@ -2,12 +2,7 @@
  "version": 1,
  "metadata": {
    "model": "anthropic/claude-sonnet-4.6",
-    "tags": [
-      "prefix:openai-compatible-chat",
-      "provider:vercel-ai-gateway",
-      "protocol:openai-chat",
-      "reasoning"
-    ],
+    "tags": ["prefix:openai-compatible-chat", "provider:vercel-ai-gateway", "protocol:openai-chat", "reasoning"],
    "name": "vercel-ai-gateway-reasoning",
    "recordedAt": "2026-07-18T11:28:42.077Z"
  },
@@ -31,4 +26,4 @@
      }
    }
  ]
-}
+}
@@ -18,13 +18,19 @@ describe("provider error classification", () => {
    expect(messages.every(isContextOverflow)).toBe(true)
  })

-  test("classifies request size failures separately from context overflow", () => {
-    const failures = [
-      classifyProviderFailure({ message: "request too large", status: 413 }),
+  test("classifies Anthropic request_too_large as recoverable overflow", () => {
+    expect(
      classifyProviderFailure({
        message: '{"error":{"type":"request_too_large","message":"Request exceeds the maximum size"}}',
        status: 400,
      }),
+    ).toMatchObject({ _tag: "InvalidRequest", classification: "context-overflow" })
+    expect(isContextOverflow("413 status code (no body)")).toBe(true)
+  })
+
+  test("classifies generic request size failures separately from context overflow", () => {
+    const failures = [
+      classifyProviderFailure({ message: "request too large", status: 413 }),
      classifyProviderFailure({ message: "upstream request entity too large", status: 502 }),
    ]

@@ -33,7 +39,6 @@ describe("provider error classification", () => {
        expect.objectContaining({ _tag: "InvalidRequest", classification: "payload-too-large" }),
      ),
    )
-    expect(isContextOverflow("413 status code (no body)")).toBe(false)
  })

  test("does not classify rate limits as context overflow", () => {
@@ -84,9 +89,7 @@ describe("provider error classification", () => {

  test("classifies network error text as provider internal", () => {
    expect(
-      ["network error", "network-error", "network_error"].map(
-        (message) => classifyProviderFailure({ message })._tag,
-      ),
+      ["network error", "network-error", "network_error"].map((message) => classifyProviderFailure({ message })._tag),
    ).toEqual(["ProviderInternal", "ProviderInternal", "ProviderInternal"])
  })

@@ -28,6 +28,8 @@ describe("provider package entrypoints", () => {
      import("@opencode-ai/ai/providers/amazon-bedrock/mantle/responses"),
      import("@opencode-ai/ai/providers/togetherai"),
      import("@opencode-ai/ai/providers/cerebras"),
+      import("@opencode-ai/ai/providers/deepinfra"),
+      import("@opencode-ai/ai/providers/groq"),
    ])

    for (const module of modules) expect(module.model).toBeFunction()
@@ -37,6 +39,24 @@ describe("provider package entrypoints", () => {
    expect(modules[19].model).toBe(modules[20].model)
  })

+  test("maps DeepInfra package settings onto its native executable model", async () => {
+    const DeepInfra = await import("@opencode-ai/ai/providers/deepinfra")
+    const settings = {
+      apiKey: "fixture",
+      baseURL: "https://provider.example.test/v1/",
+      headers: { "x-application": "opencode" },
+      body: { service_tier: "priority" },
+      providerOptions: { reasoningEffort: "high" as const },
+    }
+    const deepinfra = DeepInfra.model("google/gemma-3-27b-it", settings)
+
+    expect(deepinfra.route.id).toBe("deepinfra-chat")
+    expect(deepinfra.route.endpoint.baseURL).toBe("https://provider.example.test/v1/openai")
+    expect(deepinfra.route.defaults.providerOptions).toEqual(settings.providerOptions)
+    expect(deepinfra.route.defaults.headers).toEqual(settings.headers)
+    expect(deepinfra.route.defaults.http?.body).toEqual(settings.body)
+  })
+
  test("maps OpenRouter and xAI package settings onto executable models", async () => {
    const OpenRouter = await import("@opencode-ai/ai/providers/openrouter")
    const XAI = await import("@opencode-ai/ai/providers/xai")
@@ -2,7 +2,7 @@ import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { CacheHint, LLM, AIError, LLMRequest, Message, ToolCallPart, ToolDefinition, Usage } from "../../src/index.js"
-import { Auth, LLMClient } from "../../src/route.js"
+import { Auth, Endpoint, LLMClient, Route } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
 import * as AnthropicMessages from "../../src/protocols/anthropic-messages.js"
 import { GoogleVertexMessages } from "../../src/providers.js"
@@ -810,6 +810,99 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("round-trips compatible provider metadata in its own namespace", () =>
+    Effect.gen(function* () {
+      const compatible = Route.make({
+        id: "custom-anthropic-messages",
+        provider: "custom-anthropic",
+        protocol: AnthropicMessages.protocol,
+        endpoint: Endpoint.path("/messages", { baseURL: "https://compatible.test/v1" }),
+        auth: Auth.header("x-api-key", "test"),
+        framing: AnthropicMessages.framing,
+      }).model({ id: "custom-model" })
+      const result = [
+        {
+          type: "web_search_result",
+          url: "https://example.com",
+          citations: [{ type: "web_search_result_location", cited_text: "Example" }],
+        },
+      ]
+      const response = yield* LLMClient.generate(LLM.request({ model: compatible, prompt: "Search." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5, custom_start: true } } },
+              { type: "content_block_start", index: 0, content_block: { type: "thinking", thinking: "Thinking." } },
+              { type: "content_block_delta", index: 0, delta: { type: "signature_delta", signature: "custom_sig" } },
+              { type: "content_block_stop", index: 0 },
+              {
+                type: "content_block_start",
+                index: 1,
+                content_block: { type: "redacted_thinking", data: "custom_redacted" },
+              },
+              { type: "content_block_stop", index: 1 },
+              {
+                type: "content_block_start",
+                index: 2,
+                content_block: {
+                  type: "server_tool_use",
+                  id: "custom_tool",
+                  name: "web_search",
+                  input: { query: "example" },
+                },
+              },
+              { type: "content_block_stop", index: 2 },
+              {
+                type: "content_block_start",
+                index: 3,
+                content_block: { type: "web_search_tool_result", tool_use_id: "custom_tool", content: result },
+              },
+              { type: "content_block_stop", index: 3 },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "end_turn", stop_sequence: "custom_stop" },
+                usage: { output_tokens: 2, custom_terminal: true },
+              },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toMatchObject([
+        { type: "reasoning", text: "Thinking.", providerMetadata: { "custom-anthropic": { signature: "custom_sig" } } },
+        { type: "reasoning", text: "", providerMetadata: { "custom-anthropic": { redactedData: "custom_redacted" } } },
+        { type: "tool-call", id: "custom_tool", providerExecuted: true },
+        {
+          type: "tool-result",
+          providerExecuted: true,
+          providerMetadata: { "custom-anthropic": { blockType: "web_search_tool_result", result } },
+        },
+      ])
+      expect(response.usage?.providerMetadata).toEqual({
+        "custom-anthropic": { input_tokens: 5, custom_start: true, output_tokens: 2, custom_terminal: true },
+      })
+      expect(response.events.at(-1)).toMatchObject({
+        providerMetadata: { "custom-anthropic": { stopSequence: "custom_stop" } },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model: compatible, messages: [response.message], cache: "none" }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { type: "thinking", thinking: "Thinking.", signature: "custom_sig" },
+            { type: "redacted_thinking", data: "custom_redacted" },
+            { type: "server_tool_use", id: "custom_tool", name: "web_search", input: { query: "example" } },
+            { type: "web_search_tool_result", tool_use_id: "custom_tool", content: result },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("parses text, reasoning, and usage stream fixtures", () =>
    Effect.gen(function* () {
      const body = sseEvents(
@@ -569,6 +569,57 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("round-trips reassigned provider reasoning and usage metadata in its own namespace", () =>
+    Effect.gen(function* () {
+      const compatible = model.route.with({ provider: "custom-bedrock" }).model({ id: model.id })
+      const redactedData = "cmVkYWN0ZWQtdGhpbmtpbmc="
+      const response = yield* LLMClient.generate(LLMRequest.update(baseRequest, { model: compatible })).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["messageStart", { role: "assistant" }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { text: "Let me think." } } }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { signature: "custom_sig" } } }],
+              ["contentBlockStop", { contentBlockIndex: 0 }],
+              [
+                "contentBlockDelta",
+                { contentBlockIndex: 1, delta: { reasoningContent: { redactedContent: redactedData } } },
+              ],
+              ["contentBlockStop", { contentBlockIndex: 1 }],
+              ["messageStop", { stopReason: "end_turn" }],
+              ["metadata", { usage: { inputTokens: 5, outputTokens: 2, totalTokens: 7 } }],
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Let me think.",
+          providerMetadata: { "custom-bedrock": { signature: "custom_sig" } },
+        },
+        { type: "reasoning", text: "", providerMetadata: { "custom-bedrock": { redactedData } } },
+      ])
+      expect(response.usage?.providerMetadata).toEqual({
+        "custom-bedrock": { inputTokens: 5, outputTokens: 2, totalTokens: 7 },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model: compatible, messages: [response.message], cache: "none" }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { reasoningContent: { reasoningText: { text: "Let me think.", signature: "custom_sig" } } },
+            { reasoningContent: { redactedContent: redactedData } },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("preserves reasoning signatures when contentBlockStop is missing", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(baseRequest).pipe(
@@ -1,11 +1,13 @@
 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 } from "../../src/index.js"
 import { AmazonBedrockMantle } from "../../src/providers.js"
+import { model } from "../../src/providers/amazon-bedrock/mantle.js"
+import { OpenAIResponses } from "../../src/protocols/openai-responses.js"
 import { compileRequest, LLMClient } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
-import { dynamicResponse } from "../lib/http.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
 import { sseEvents } from "../lib/sse.js"
 import { recordedTests } from "../recorded-test.js"

@@ -16,12 +18,16 @@ const credentials = {
 }

 describe("Amazon Bedrock Mantle provider", () => {
-  it.effect("uses Chat by default and exposes Responses", () =>
+  it.effect("uses Responses by default and exposes Chat explicitly", () =>
    Effect.gen(function* () {
      const provider = AmazonBedrockMantle.configure({ credentials })
-      const chat = yield* compileRequest(LLM.request({ model: provider.model("openai.gpt-oss-120b"), prompt: "Hi" }))
+      expect(provider.model).toBe(provider.responses)
+      expect(AmazonBedrockMantle.model).toBe(AmazonBedrockMantle.responsesModel)
+      expect(model).toBe(AmazonBedrockMantle.responsesModel)
+      expect(provider.model("openai.gpt-oss-120b").route.transport).toBe(OpenAIResponses.httpTransport)
+      const chat = yield* compileRequest(LLM.request({ model: provider.chat("openai.gpt-oss-120b"), prompt: "Hi" }))
      const responses = yield* compileRequest(
-        LLM.request({ model: provider.responses("openai.gpt-oss-120b"), prompt: "Hi" }),
+        LLM.request({ model: provider.model("openai.gpt-oss-120b"), prompt: "Hi" }),
      )

      expect(chat).toMatchObject({
@@ -34,6 +40,23 @@ describe("Amazon Bedrock Mantle provider", () => {
        protocol: "openai-responses",
        body: { model: "openai.gpt-oss-120b", store: false },
      })
+      expect(provider.model("openai.gpt-oss-120b").route.providerMetadataKey).toBe("mantle")
+      expect(provider.chat("openai.gpt-oss-120b").route.providerMetadataKey).toBe("mantle")
+    }),
+  )
+
+  it.effect("preserves configured top-p generation defaults for Chat and Responses", () =>
+    Effect.gen(function* () {
+      const settings = { apiKey: "test-key", topP: 0.8 }
+      const chat = yield* compileRequest(
+        LLM.request({ model: AmazonBedrockMantle.chatModel("openai.gpt-oss-safeguard-20b", settings), prompt: "Hi" }),
+      )
+      const responses = yield* compileRequest(
+        LLM.request({ model: AmazonBedrockMantle.responsesModel("openai.gpt-oss-120b", settings), prompt: "Hi" }),
+      )
+
+      expect(chat.body.top_p).toBe(0.8)
+      expect(responses.body.top_p).toBe(0.8)
    }),
  )

@@ -83,6 +106,42 @@ describe("Amazon Bedrock Mantle provider", () => {
      expect(seen).toEqual([{ url: "https://mantle.test/v1/chat/completions", authorization: "Bearer test-key" }])
    }),
  )
+
+  it.effect("replays reasoning with Mantle's message-prefixed item ids", () =>
+    Effect.gen(function* () {
+      const model = AmazonBedrockMantle.configure({ apiKey: "test-key" }).responses("openai.gpt-oss-120b")
+      const item = { type: "reasoning", id: "msg_95d4d0af4350432a", encrypted_content: "mantle-state" }
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", item },
+              { type: "response.reasoning_summary_text.delta", item_id: item.id, delta: "Considering." },
+              { type: "response.output_item.done", item },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model, messages: [response.message, Message.user("Continue.")] }),
+      )
+
+      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+        mantle: { itemId: "msg_95d4d0af4350432a", reasoningEncryptedContent: "mantle-state" },
+      })
+      expect(prepared.body.input).toEqual([
+        {
+          type: "reasoning",
+          id: "msg_95d4d0af4350432a",
+          summary: [{ type: "summary_text", text: "Considering." }],
+          encrypted_content: "mantle-state",
+        },
+        { role: "user", content: [{ type: "input_text", text: "Continue." }] },
+      ])
+    }),
+  )
 })

 const recorded = recordedTests({
@@ -126,7 +126,7 @@ describe("Cloudflare", () => {
      expect(response.reasoning).toBe("Thinking")
      expect(response.events.filter(LLMEvent.is.reasoningDelta)).toHaveLength(2)
      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
-        openai: { reasoningField: "reasoning", reasoningDetails: merged },
+        "cloudflare-ai-gateway": { reasoningField: "reasoning", reasoningDetails: merged },
      })

      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
@@ -515,7 +515,10 @@ describe("Gemini route", () => {
        {
          role: "model",
          parts: [
-            { functionCall: { id: "call_image", name: "read", args: { path: "pixel.png" } }, thoughtSignature: "sig_1" },
+            {
+              functionCall: { id: "call_image", name: "read", args: { path: "pixel.png" } },
+              thoughtSignature: "sig_1",
+            },
          ],
        },
        {
@@ -606,10 +609,7 @@ describe("Gemini route", () => {
      expect(prepared.body.contents).toEqual([
        {
          role: "model",
-          parts: [
-            { functionCall: { name: "shot", args: {} } },
-            { functionCall: { name: "shot", args: {} } },
-          ],
+          parts: [{ functionCall: { name: "shot", args: {} } }, { functionCall: { name: "shot", args: {} } }],
        },
        {
          role: "user",
@@ -1071,7 +1071,9 @@ describe("Gemini route", () => {
      const prepared = yield* compileRequest(
        LLM.request({
          model,
-          messages: [Message.assistant([{ type: "text", text: "All done.", providerMetadata: delta?.providerMetadata }])],
+          messages: [
+            Message.assistant([{ type: "text", text: "All done.", providerMetadata: delta?.providerMetadata }]),
+          ],
        }),
      )
      expect(prepared.body.contents).toEqual([
@@ -1572,9 +1574,7 @@ describe("Gemini route", () => {
              { candidates: [{ content: { role: "model", parts: null } }] },
              { candidates: [{ content: null, finishReason: null }] },
              {
-                candidates: [
-                  { content: { role: "model", parts: [{ text: "Hello" }] }, finishReason: "STOP" as const },
-                ],
+                candidates: [{ content: { role: "model", parts: [{ text: "Hello" }] }, finishReason: "STOP" as const }],
              },
            ),
          ),
@@ -1,6 +1,6 @@
 import * as Anthropic from "../../src/providers/anthropic.js"
 import * as AnthropicCompatible from "../../src/providers/anthropic-compatible.js"
-import { Cerebras, TogetherAI } from "../../src/providers/index.js"
+import { Cerebras, DeepInfra, TogetherAI } from "../../src/providers/index.js"
 import { CloudflareAIGateway, CloudflareWorkersAI } from "../../src/providers/cloudflare.js"
 import * as Google from "../../src/providers/google.js"
 import * as OpenAI from "../../src/providers/openai.js"
@@ -55,6 +55,9 @@ const cerebras = Cerebras.configure({ apiKey: process.env.CEREBRAS_API_KEY ?? "f
 const groq = OpenAICompatible.groq
  .configure({ apiKey: process.env.GROQ_API_KEY ?? "fixture" })
  .model("llama-3.3-70b-versatile")
+const deepInfra = DeepInfra.configure({ apiKey: process.env.DEEPINFRA_API_KEY ?? "fixture" }).model(
+  "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+)
 const openRouter = OpenRouter.configure({ apiKey: process.env.OPENROUTER_API_KEY ?? "fixture" })
 const openrouter = openRouter.model("openai/gpt-4o-mini")
 const openrouterGpt55 = openRouter.model("openai/gpt-5.5")
@@ -222,6 +225,13 @@ describeRecordedGoldenScenarios([
    requires: ["GROQ_API_KEY"],
    scenarios: ["text", "tool-call", { id: "tool-loop", timeout: 30_000 }],
  },
+  {
+    name: "DeepInfra Llama 3.3 70B",
+    prefix: "deepinfra-chat",
+    model: deepInfra,
+    requires: ["DEEPINFRA_API_KEY"],
+    scenarios: ["text", "tool-call", { id: "tool-loop", timeout: 30_000 }],
+  },
  {
    name: "OpenRouter gpt-4o-mini",
    prefix: "openai-compatible-chat",
@@ -26,9 +26,7 @@ const recorded = recordedTests({
 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" }),
-      )
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Reply with exactly one word: hello" }))

      expect(response.text.toLowerCase()).toContain("hello")
    }),
@@ -6,7 +6,7 @@ import { GoogleVertex, GoogleVertexChat, GoogleVertexMessages, GoogleVertexRespo
 import { LLMClient } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
-import { dynamicResponse } from "../lib/http.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
 import { deltaChunk, finishChunk } from "../lib/openai-chunks.js"
 import { sseEvents } from "../lib/sse.js"

@@ -89,7 +89,7 @@ describe("Google Vertex providers", () => {
                id: "call_1",
                name: "lookup",
                input: { query: "weather" },
-                providerMetadata: { google: { functionCallId: "provider_call_1" } },
+                providerMetadata: { vertex: { functionCallId: "provider_call_1" } },
              }),
            ]),
            Message.tool({
@@ -97,7 +97,7 @@ describe("Google Vertex providers", () => {
              name: "lookup",
              result: "sunny",
              resultType: "text",
-              providerMetadata: { google: { functionCallId: "provider_call_1" } },
+              providerMetadata: { vertex: { functionCallId: "provider_call_1" } },
            }),
          ],
        }),
@@ -122,6 +122,91 @@ describe("Google Vertex providers", () => {
    }),
  )

+  it.effect("round-trips Vertex Gemini metadata through signed content, tool calls, and usage", () =>
+    Effect.gen(function* () {
+      const model = GoogleVertex.configure({
+        accessToken: "vertex-token",
+        project: "vertex-project",
+      }).model("gemini-3.5-flash")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Check the weather." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents({
+              candidates: [
+                {
+                  content: {
+                    role: "model",
+                    parts: [
+                      { text: "Thinking.", thought: true, thoughtSignature: "reasoning_sig" },
+                      { text: "Checking.", thoughtSignature: "text_sig" },
+                      {
+                        functionCall: { id: "provider_call_1", name: "lookup", args: { query: "weather" } },
+                        thoughtSignature: "tool_sig",
+                      },
+                    ],
+                  },
+                  finishReason: "STOP",
+                },
+              ],
+              promptFeedback: { blockReasonMessage: "Reviewed" },
+              usageMetadata: { promptTokenCount: 5, candidatesTokenCount: 2, thoughtsTokenCount: 1 },
+            }),
+          ),
+        ),
+      )
+      const reasoning = response.events.find((event) => event.type === "reasoning-end")
+      const text = response.events.find((event) => event.type === "text-delta")
+      const toolCall = response.toolCalls[0]
+
+      expect(reasoning?.providerMetadata).toEqual({ vertex: { thoughtSignature: "reasoning_sig" } })
+      expect(text?.providerMetadata).toEqual({ vertex: { thoughtSignature: "text_sig" } })
+      expect(toolCall).toMatchObject({
+        id: "provider_call_1",
+        providerMetadata: { vertex: { thoughtSignature: "tool_sig" } },
+      })
+      expect(response.usage?.providerMetadata).toEqual({
+        vertex: { promptTokenCount: 5, candidatesTokenCount: 2, thoughtsTokenCount: 1 },
+      })
+      expect(response.events.at(-1)?.providerMetadata).toEqual({
+        vertex: { promptFeedback: { blockReasonMessage: "Reviewed" } },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              { type: "reasoning", text: "Thinking.", providerMetadata: reasoning?.providerMetadata },
+              { type: "text", text: "Checking.", providerMetadata: text?.providerMetadata },
+              ToolCallPart.make({
+                id: toolCall.id,
+                name: toolCall.name,
+                input: toolCall.input,
+                providerMetadata: toolCall.providerMetadata,
+              }),
+            ]),
+            Message.tool({ id: toolCall.id, name: toolCall.name, result: "sunny", resultType: "text" }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          parts: [
+            { text: "Thinking.", thought: true, thoughtSignature: "reasoning_sig" },
+            { text: "Checking.", thoughtSignature: "text_sig" },
+            { functionCall: { name: "lookup", args: { query: "weather" } }, thoughtSignature: "tool_sig" },
+          ],
+        },
+        {
+          role: "user",
+          parts: [{ functionResponse: { 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({
@@ -0,0 +1,185 @@
+import { configure } from "@opencode-ai/ai/providers/groq"
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMEvent, LLMRequest, LLMResponse, Message, ToolChoice, ToolDefinition } from "../../src/index.js"
+import { LLMClient } from "../../src/route.js"
+import { compileRequest } from "../../src/route/client.js"
+import { recordedTests } from "../recorded-test.js"
+
+const apiKey = process.env.GROQ_API_KEY ?? "fixture"
+const recorded = recordedTests({
+  prefix: "groq-chat",
+  provider: "groq",
+  protocol: "groq-chat",
+  requires: ["GROQ_API_KEY"],
+})
+
+const weather = ToolDefinition.make({
+  name: "lookup_weather",
+  description: "Look up the current weather for a city",
+  inputSchema: {
+    type: "object",
+    properties: { city: { type: "string", enum: ["Paris", "London"] } },
+    required: ["city"],
+    additionalProperties: false,
+  },
+})
+
+describe("Groq recorded", () => {
+  recorded.effect.with(
+    "streams text with usage",
+    { tags: ["text", "usage"], metadata: { model: "openai/gpt-oss-20b" } },
+    () =>
+      Effect.gen(function* () {
+        const request = LLM.request({
+          model: configure({
+            apiKey,
+            providerOptions: {
+              includeReasoning: false,
+              reasoningEffort: "low",
+              serviceTier: "on_demand",
+              user: "recorded-test",
+            },
+          }).model("openai/gpt-oss-20b"),
+          prompt: "Reply with exactly one word: hello",
+          generation: { maxTokens: 512 },
+        })
+        const compiled = yield* compileRequest(request)
+        expect(compiled.body).toMatchObject({
+          max_completion_tokens: 512,
+          stream_options: { include_usage: true },
+          include_reasoning: false,
+          service_tier: "on_demand",
+          user: "recorded-test",
+        })
+        expect(compiled.body.max_tokens).toBeUndefined()
+        expect(compiled.body.store).toBeUndefined()
+        expect(compiled.body.reasoning_format).toBeUndefined()
+
+        const response = yield* LLMClient.generate(request)
+        expect(response.text.toLowerCase().trim()).toBe("hello")
+        expect(response.reasoning).toBe("")
+        expect(response.events.some(LLMEvent.is.textDelta)).toBe(true)
+        expectUsage(response)
+      }),
+    60_000,
+  )
+
+  for (const item of [
+    {
+      name: "continues Qwen parallel tool calls",
+      model: configure({ apiKey, providerOptions: { parallelToolCalls: true, reasoningEffort: "none" } }).model(
+        "qwen/qwen3.6-27b",
+      ),
+      cities: ["Paris", "London"],
+      reasoning: false,
+    },
+    {
+      name: "replays GPT OSS reasoning through a tool loop",
+      model: configure({ apiKey, providerOptions: { includeReasoning: true, reasoningEffort: "low" } }).model(
+        "openai/gpt-oss-20b",
+      ),
+      cities: ["Paris"],
+      reasoning: true,
+    },
+  ]) {
+    recorded.effect.with(
+      item.name,
+      {
+        tags: ["tool", "tool-loop", "usage", item.reasoning ? "reasoning" : "parallel"],
+        metadata: { model: item.model.id },
+      },
+      () =>
+        Effect.gen(function* () {
+          const request = LLM.request({
+            model: item.model,
+            prompt: `Look up the current weather in ${item.cities.join(" and ")}. Call lookup_weather once for each city in the same response before answering. After receiving all results, report each city's weather in one short sentence.`,
+            tools: [weather],
+            toolChoice: "required",
+            generation: { maxTokens: 1536 },
+          })
+          const compiled = yield* compileRequest(request)
+          expect(compiled.body.stream_options).toEqual({ include_usage: true })
+          expect(compiled.body.store).toBeUndefined()
+          expect(compiled.body.reasoning_format).toBe(item.reasoning ? undefined : "parsed")
+          expect(compiled.body.tools[0].function.strict).toBeUndefined()
+          if (!item.reasoning) expect(compiled.body.parallel_tool_calls).toBe(true)
+
+          const first = yield* LLMClient.generate(request)
+          expect(first.finishReason.normalized).toBe("tool-calls")
+          expect(first.toolCalls).toHaveLength(item.cities.length)
+          expect(new Set(first.toolCalls.map((call) => call.id)).size).toBe(item.cities.length)
+          expect(first.toolCalls.map((call) => call.input)).toEqual(
+            expect.arrayContaining(item.cities.map((city) => ({ city }))),
+          )
+          expect(first.toolCalls.every((call) => call.name === "lookup_weather")).toBe(true)
+          expectUsage(first)
+          if (item.reasoning) {
+            expect(first.reasoning.length).toBeGreaterThan(0)
+            expect(first.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
+          }
+
+          const followUp = LLMRequest.update(request, {
+            toolChoice: ToolChoice.make("none"),
+            messages: [
+              ...request.messages,
+              first.message,
+              ...first.toolCalls.map((call) =>
+                Message.tool({ id: call.id, name: call.name, result: { condition: "sunny", temperature: "18C" } }),
+              ),
+            ],
+          })
+          const replay = yield* compileRequest(followUp)
+          if (item.reasoning) {
+            expect(replay.body.messages).toEqual(
+              expect.arrayContaining([expect.objectContaining({ role: "assistant", reasoning: first.reasoning })]),
+            )
+          }
+          expect(replay.body.reasoning_format).toBe(item.reasoning ? undefined : "parsed")
+
+          const second = yield* LLMClient.generate(followUp)
+          expect(second.finishReason.normalized).toBe("stop")
+          expect(second.toolCalls).toHaveLength(0)
+          expect(second.text.toLowerCase()).toContain("sunny")
+          item.cities.forEach((city) => expect(second.text).toContain(city))
+          expectUsage(second)
+        }),
+      60_000,
+    )
+  }
+
+  recorded.effect.with(
+    "streams Qwen parsed reasoning",
+    { tags: ["reasoning", "usage"], metadata: { model: "qwen/qwen3.6-27b" } },
+    () =>
+      Effect.gen(function* () {
+        const request = LLM.request({
+          model: configure({
+            apiKey,
+            providerOptions: { reasoningEffort: "default" },
+          }).model("qwen/qwen3.6-27b"),
+          prompt:
+            "What is 173 multiplied by 219? Think through the arithmetic, then reply with only the final integer.",
+          generation: { maxTokens: 2048 },
+        })
+        const compiled = yield* compileRequest(request)
+        expect(compiled.body).toMatchObject({ reasoning_format: "parsed", reasoning_effort: "default" })
+        expect(compiled.body.include_reasoning).toBeUndefined()
+
+        const response = yield* LLMClient.generate(request)
+        expect(response.text.replaceAll(",", "").trim()).toBe("37887")
+        expect(response.text).not.toContain("<think>")
+        expect(response.reasoning.length).toBeGreaterThan(0)
+        expect(response.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
+        expectUsage(response)
+      }),
+    60_000,
+  )
+})
+
+function expectUsage(response: LLMResponse) {
+  expect(response.usage).toBeDefined()
+  expect(response.usage?.inputTokens).toBeGreaterThan(0)
+  expect(response.usage?.outputTokens).toBeGreaterThan(0)
+  expect(response.events.filter(LLMEvent.is.finish)).toHaveLength(1)
+}
@@ -0,0 +1,112 @@
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { LanguageModel, LLM, Message } from "../../src/index.js"
+import { OpenAIChat } from "../../src/protocols/openai-chat.js"
+import { Groq } from "../../src/providers/groq.js"
+import { compileRequest } from "../../src/route/client.js"
+import { it } from "../lib/effect.js"
+import { weatherTool } from "../recorded-scenarios.js"
+
+it.effect("Groq reuses Chat streaming and defaults to parsed reasoning", () =>
+  Effect.gen(function* () {
+    expect(Groq.protocol.stream).toBe(OpenAIChat.protocol.stream)
+    const model = Groq.configure({ apiKey: "fixture" }).model("llama-3.3-70b-versatile")
+    expect(model.route.endpoint.baseURL).toBe("https://api.groq.com/openai/v1")
+    const compiled = yield* compileRequest(
+      LLM.request({ model, prompt: "Hello", tools: [weatherTool], generation: { maxTokens: 64 } }),
+    )
+    expect(compiled.body).toMatchObject({
+      max_completion_tokens: 64,
+      stream_options: { include_usage: true },
+      reasoning_format: "parsed",
+    })
+    for (const key of ["store", "max_tokens", "include_reasoning", "parallel_tool_calls", "service_tier", "user"])
+      expect(compiled.body[key]).toBeUndefined()
+    expect(compiled.body.tools?.[0]?.function).not.toHaveProperty("strict")
+  }),
+)
+
+it.effect("Groq lowers its own options for custom catalog identities and endpoints", () =>
+  Effect.gen(function* () {
+    const model = LanguageModel.update(
+      Groq.model("qwen/qwen3.6-27b", {
+        apiKey: "fixture",
+        baseURL: "https://gateway.example/v1",
+        headers: { "x-client": "test" },
+        body: { custom: "value" },
+        providerOptions: {
+          reasoningEffort: "default",
+          parallelToolCalls: true,
+          serviceTier: "flex",
+          user: "test-user",
+        },
+      }),
+      { provider: "custom-groq" },
+    )
+    const compiled = yield* compileRequest(
+      LLM.request({ model, prompt: "Hello", providerOptions: { parallelToolCalls: false, includeReasoning: false } }),
+    )
+    expect(model.route.endpoint.baseURL).toBe("https://gateway.example/v1")
+    expect(model.route.defaults.headers).toEqual({ "x-client": "test" })
+    expect(model.route.defaults.http?.body).toEqual({ custom: "value" })
+    expect(compiled.body).toMatchObject({
+      reasoning_effort: "default",
+      reasoning_format: "parsed",
+      parallel_tool_calls: false,
+      service_tier: "flex",
+      user: "test-user",
+    })
+    expect(compiled.body.include_reasoning).toBeUndefined()
+    for (const key of ["reasoningFormat", "reasoningEffort", "parallelToolCalls", "serviceTier"])
+      expect(compiled.body).not.toHaveProperty(key)
+  }),
+)
+
+it.effect("Groq replays reasoning only when present and preserves explicit reasoning exclusion", () =>
+  Effect.gen(function* () {
+    const compiled = yield* compileRequest(
+      LLM.request({
+        model: Groq.configure({ apiKey: "fixture" }).model("openai/gpt-oss-20b"),
+        messages: [
+          Message.user("Think"),
+          Message.assistant([
+            { type: "reasoning", text: "Thinking" },
+            { type: "text", text: "Answer" },
+          ]),
+          Message.user("Again"),
+          Message.assistant("Answer only"),
+          Message.user("Continue"),
+        ],
+        providerOptions: { reasoningEffort: "low", includeReasoning: false },
+      }),
+    )
+    expect(compiled.body).toMatchObject({ reasoning_effort: "low", include_reasoning: false })
+    expect(compiled.body.reasoning_format).toBeUndefined()
+    expect(compiled.body.messages[1]).toMatchObject({ reasoning: "Thinking", content: "Answer" })
+    expect(compiled.body.messages[1]).not.toHaveProperty("reasoning_content")
+    expect(compiled.body.messages[3]).not.toHaveProperty("reasoning")
+  }),
+)
+
+it.effect("Groq omits reasoning_format for the GPT-OSS family by default", () =>
+  Effect.gen(function* () {
+    for (const id of ["openai/gpt-oss-20b", "openai/gpt-oss-120b", "openai/gpt-oss-safeguard-20b"]) {
+      const compiled = yield* compileRequest(
+        LLM.request({ model: Groq.configure({ apiKey: "fixture" }).model(id), prompt: "Hello" }),
+      )
+      expect(compiled.body.reasoning_format).toBeUndefined()
+      expect(compiled.body.include_reasoning).toBeUndefined()
+    }
+  }),
+)
+
+it.effect("Groq validates option types", () =>
+  Effect.gen(function* () {
+    for (const providerOptions of [{ includeReasoning: "false" }, { parallelToolCalls: "false" }]) {
+      const error = yield* compileRequest(
+        LLM.request({ model: Groq.configure({ apiKey: "fixture" }).model("qwen"), prompt: "Hello", providerOptions }),
+      ).pipe(Effect.flip)
+      expect(error.reason._tag).toBe("InvalidRequest")
+    }
+  }),
+)
@@ -2,13 +2,80 @@ import { describe, expect } from "bun:test"
 import { ConfigProvider, Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, Message, ToolDefinition } from "../../src/index.js"
-import { Cerebras, TogetherAI } from "../../src/providers/index.js"
+import {
+  AmazonBedrock,
+  AmazonBedrockMantle,
+  Anthropic,
+  AnthropicCompatible,
+  Azure,
+  Cerebras,
+  CloudflareAIGateway,
+  CloudflareWorkersAI,
+  DeepInfra,
+  Google,
+  GoogleVertex,
+  GoogleVertexChat,
+  GoogleVertexMessages,
+  GoogleVertexResponses,
+  Groq,
+  OpenAI,
+  OpenAICompatible,
+  OpenAICompatibleResponses,
+  OpenRouter,
+  TogetherAI,
+  XAI,
+} from "../../src/providers/index.js"
 import { compileRequest } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
 import { dynamicResponse } from "../lib/http.js"
 import { sseEvents } from "../lib/sse.js"

 describe("native OpenAI-compatible providers", () => {
+  it.effect("assigns provider-owned metadata namespaces across native routes", () =>
+    Effect.gen(function* () {
+      const vertex = { project: "project", accessToken: "token" }
+      const providers = [
+        [OpenAI.configure({ apiKey: "test" }).chat("model"), "openai"],
+        [OpenAI.configure({ apiKey: "test" }).responses("model"), "openai"],
+        [Azure.configure({ resourceName: "resource", apiKey: "test" }).chat("model"), "azure"],
+        [Azure.configure({ resourceName: "resource", apiKey: "test" }).responses("model"), "azure"],
+        [AmazonBedrock.configure({ apiKey: "test" }).model("model"), "bedrock"],
+        [AmazonBedrockMantle.configure({ apiKey: "test" }).chat("model"), "mantle"],
+        [AmazonBedrockMantle.configure({ apiKey: "test" }).responses("model"), "mantle"],
+        [Google.configure({ apiKey: "test" }).model("model"), "google"],
+        [GoogleVertex.configure(vertex).model("model"), "vertex"],
+        [GoogleVertexChat.configure(vertex).model("model"), "vertex"],
+        [GoogleVertexResponses.configure(vertex).model("model"), "vertex"],
+        [GoogleVertexMessages.configure(vertex).model("model"), "anthropic"],
+        [Anthropic.configure({ apiKey: "test" }).model("model"), "anthropic"],
+        [
+          AnthropicCompatible.configure({ baseURL: "https://example.test/v1", provider: "minimax" }).model("model"),
+          "minimax",
+        ],
+        [
+          OpenAICompatible.configure({ baseURL: "https://example.test/v1", provider: "custom" }).model("model"),
+          "custom",
+        ],
+        [
+          OpenAICompatibleResponses.configure({ baseURL: "https://example.test/v1", provider: "custom" }).model(
+            "model",
+          ),
+          "custom",
+        ],
+        [Cerebras.configure({ apiKey: "test" }).model("model"), "cerebras"],
+        [DeepInfra.configure({ apiKey: "test" }).model("model"), "deepinfra"],
+        [TogetherAI.configure({ apiKey: "test" }).model("model"), "togetherai"],
+        [CloudflareAIGateway.configure({ accountId: "account" }).model("model"), "cloudflare-ai-gateway"],
+        [CloudflareWorkersAI.configure({ accountId: "account" }).model("model"), "cloudflare-workers-ai"],
+        [OpenRouter.configure({ apiKey: "test" }).model("model"), "openrouter"],
+        [XAI.configure({ apiKey: "test" }).chat("model"), "xai"],
+        [XAI.configure({ apiKey: "test" }).responses("model"), "xai"],
+      ] as const
+
+      for (const [model, key] of providers) expect(model.route.providerMetadataKey).toBe(key)
+    }),
+  )
+
  it.effect("preserves native Together AI and Cerebras provider and route identities", () =>
    Effect.gen(function* () {
      const together = TogetherAI.configure({ apiKey: "fixture" }).model("meta-llama/Llama-3.3-70B")
@@ -29,6 +96,18 @@ describe("native OpenAI-compatible providers", () => {
    }),
  )

+  it.effect("preserves native DeepInfra provider and route identity", () =>
+    Effect.gen(function* () {
+      const deepinfra = DeepInfra.configure({ apiKey: "fixture" }).model("google/gemma-3-27b-it")
+      expect(deepinfra).toMatchObject({
+        provider: "deepinfra",
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning_content", supportsStore: false },
+        route: { id: "deepinfra-chat", protocol: "openai-chat" },
+      })
+      expect(deepinfra.route.endpoint.baseURL).toBe("https://api.deepinfra.com/v1/openai")
+    }),
+  )
+
  it.effect("applies native provider request defaults even with a custom gateway URL", () =>
    Effect.gen(function* () {
      const together = yield* compileRequest(
@@ -82,6 +161,21 @@ describe("native OpenAI-compatible providers", () => {
    }),
  )

+  it.effect("normalizes DeepInfra API roots without duplicating the OpenAI path", () =>
+    Effect.gen(function* () {
+      for (const baseURL of [
+        "https://gateway.example/v1",
+        "https://gateway.example/v1/",
+        "https://gateway.example/v1/openai",
+        "https://gateway.example/v1/openai/",
+      ]) {
+        expect(DeepInfra.configure({ apiKey: "fixture", baseURL }).model("gemma").route.endpoint.baseURL).toBe(
+          "https://gateway.example/v1/openai",
+        )
+      }
+    }),
+  )
+
  it.effect("maps package settings onto native executable models", () =>
    Effect.gen(function* () {
      for (const native of [TogetherAI, Cerebras]) {
@@ -122,6 +216,18 @@ describe("native OpenAI-compatible providers", () => {
          token: "cerebras-secret",
          url: "https://api.cerebras.ai/v1/chat/completions",
        },
+        {
+          model: DeepInfra.configure().model("gemma"),
+          env: { DEEPINFRA_API_KEY: "deepinfra-secret" },
+          token: "deepinfra-secret",
+          url: "https://api.deepinfra.com/v1/openai/chat/completions",
+        },
+        {
+          model: Groq.configure().model("llama"),
+          env: { GROQ_API_KEY: "groq-secret" },
+          token: "groq-secret",
+          url: "https://api.groq.com/openai/v1/chat/completions",
+        },
      ]

      yield* Effect.forEach(scenarios, (scenario) =>
@@ -68,11 +68,13 @@ for (const item of cases) {
          expect(response.text.replaceAll(",", "").trim()).toBe("37887")
          expect(response.reasoning.length).toBeGreaterThan(0)
          expect(response.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
-          const metadata = response.message.content.find((part) => part.type === "reasoning")?.providerMetadata
-          expect(metadata?.openai?.reasoningField).toBe(item.structured ? "reasoning" : "reasoning_content")
-          expect(Array.isArray(metadata?.openai?.reasoningDetails)).toBe(item.structured)
+          const metadata = response.message.content.find((part) => part.type === "reasoning")?.providerMetadata?.[
+            item.model.route.providerMetadataKey ?? String(item.model.provider)
+          ]
+          expect(metadata?.reasoningField).toBe(item.structured ? "reasoning" : "reasoning_content")
+          expect(Array.isArray(metadata?.reasoningDetails)).toBe(item.structured)
          if (!item.structured) return
-          const details = metadata?.openai?.reasoningDetails
+          const details = metadata?.reasoningDetails
          if (!Array.isArray(details)) return
          expect(
            details.some(
@@ -126,7 +128,11 @@ for (const item of cases) {
          ).toMatch(/^Paris is sunny\.?$/)
          const details = events
            .filter(LLMEvent.is.reasoningEnd)
-            .map((event) => event.providerMetadata?.openai?.reasoningDetails)
+            .map(
+              (event) =>
+                event.providerMetadata?.[item.model.route.providerMetadataKey ?? String(item.model.provider)]
+                  ?.reasoningDetails,
+            )
            .find(Array.isArray)
          expect(Array.isArray(details)).toBe(item.structured)
          if (!item.structured || !Array.isArray(details)) return
@@ -903,6 +903,70 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("uses the configured provider metadata namespace for reasoning and usage", () =>
+    Effect.gen(function* () {
+      const selected = LanguageModel.update(model, {
+        route: { ...model.route, providerMetadataKey: "vendor" },
+      })
+      const details = [{ type: "reasoning.text", text: "thinking", signature: "signed" }]
+      const response = yield* LLMClient.generate(LLMRequest.update(request, { model: selected })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { choices: [{ delta: { reasoning: "thinking", reasoning_details: details } }] },
+              deltaChunk({ content: "Hello" }),
+              deltaChunk({}, "stop"),
+              usageChunk({ prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 }),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+        vendor: { reasoningField: "reasoning", reasoningDetails: details },
+      })
+      expect(response.usage?.providerMetadata).toEqual({
+        vendor: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+      })
+
+      const replay = yield* compileRequest(LLM.request({ model: selected, messages: [response.message] }))
+      expect(replay.body.messages).toEqual([
+        { role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: details },
+      ])
+    }),
+  )
+
+  it.effect("falls back to the selected provider for the metadata namespace", () =>
+    Effect.gen(function* () {
+      const compatible = model.route.with({ provider: "deepseek" }).model({ id: "deepseek-chat" })
+      const selected = LanguageModel.update(compatible, {
+        route: { ...compatible.route, providerMetadataKey: undefined },
+      })
+      const response = yield* LLMClient.generate(LLMRequest.update(request, { model: selected })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              deltaChunk({ reasoning_content: "thinking" }),
+              deltaChunk({ content: "Hello" }),
+              deltaChunk({}, "stop"),
+              usageChunk({ prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 }),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+        deepseek: { reasoningField: "reasoning_content" },
+      })
+      expect(response.usage?.providerMetadata).toEqual({
+        deepseek: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+      })
+
+      const replay = yield* compileRequest(LLM.request({ model: selected, messages: [response.message] }))
+      expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_content: "thinking" }])
+    }),
+  )
+
  it.effect("parses and replays a configured custom reasoning field", () =>
    Effect.gen(function* () {
      const custom = LanguageModel.update(model, { compatibility: { reasoningField: "vendor_reasoning" } })
@@ -437,7 +437,7 @@ describe("OpenAI-compatible Chat route", () => {
        outputTokens: undefined,
        totalTokens: undefined,
        providerMetadata: {
-          openai: {
+          deepseek: {
            prompt_tokens: null,
            completion_tokens: null,
            total_tokens: null,
@@ -195,15 +195,19 @@ describe("Open Responses-compatible route", () => {
          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: "Kept.", providerMetadata: { "openai-compatible": { itemId: "history_1" } } },
              {
                type: "text",
-                text: "Dropped.",
-                providerMetadata: { openresponses: { itemId: `m${"a".repeat(64)}` } },
+                text: "Long.",
+                providerMetadata: { "openai-compatible": { itemId: `history_${"a".repeat(64)}` } },
              },
+              {
+                type: "text",
+                text: "Opaque.",
+                providerMetadata: { "openai-compatible": { itemId: "provider_value/with+symbols" } },
+              },
+              { type: "text", text: "No suffix.", providerMetadata: { "openai-compatible": { itemId: "msg_" } } },
+              { type: "text", text: "No prefix.", providerMetadata: { "openai-compatible": { itemId: "_item" } } },
            ]),
          ],
        }),
@@ -218,8 +222,23 @@ describe("Open Responses-compatible route", () => {
        },
        {
          type: "message",
+          id: `history_${"a".repeat(64)}`,
          role: "assistant",
-          content: [{ type: "output_text", text: "Dropped." }],
+          content: [{ type: "output_text", text: "Long." }],
+        },
+        {
+          type: "message",
+          id: "provider_value/with+symbols",
+          role: "assistant",
+          content: [{ type: "output_text", text: "Opaque." }],
+        },
+        {
+          type: "message",
+          role: "assistant",
+          content: [
+            { type: "output_text", text: "No suffix." },
+            { type: "output_text", text: "No prefix." },
+          ],
        },
      ])
    }),
@@ -248,7 +267,7 @@ describe("Open Responses-compatible route", () => {
              name: item.type,
              result: { type: "json", value: item },
              providerExecuted: true,
-              providerMetadata: { openresponses: { itemId: item.id } },
+              providerMetadata: { example: { itemId: item.id } },
            }),
          ),
        }),
@@ -283,11 +302,296 @@ describe("Open Responses-compatible route", () => {
      )

      expect(response.message.content).toEqual([
-        { type: "text", text: "Indexed", providerMetadata: { openresponses: { itemId: "msg_1" } } },
+        { type: "text", text: "Indexed", providerMetadata: { "openai-compatible": { itemId: "msg_1" } } },
      ])
    }),
  )

+  describe("stream validation", () => {
+    const request = LLM.request({
+      model: configure({ apiKey: "test-key", baseURL: "https://responses.example.test/v1" }).model("example-model"),
+      prompt: "Respond.",
+    })
+
+    const fixtures = [
+      {
+        item: { type: "message" },
+        events: [
+          { type: "response.output_text.delta", delta: "Preserved" },
+          { type: "response.output_text.done", text: "Preserved" },
+          { type: "response.refusal.delta", delta: "Preserved" },
+          { type: "response.refusal.done", refusal: "Preserved" },
+        ],
+      },
+      {
+        item: { type: "reasoning", encrypted_content: "encrypted-state" },
+        events: [
+          { type: "response.reasoning.delta", delta: "Preserved" },
+          { type: "response.reasoning.done", text: "Preserved" },
+          { type: "response.reasoning_summary_text.delta", delta: "Preserved" },
+          { type: "response.reasoning_summary_text.done", text: "Preserved" },
+          { type: "response.reasoning_text.done", text: "Preserved" },
+        ],
+      },
+      {
+        item: { type: "function_call", call_id: "call_1", name: "lookup" },
+        events: [
+          { type: "response.function_call_arguments.delta", delta: '{"query":"Preserved"}' },
+          { type: "response.function_call_arguments.done", arguments: '{"query":"Preserved"}' },
+        ],
+      },
+    ]
+
+    const routings = [
+      { name: "empty item and event IDs", id: "", item_id: "" },
+      { name: "empty event ID with registered index", id: "item_1", item_id: "", output_index: 2 },
+      { name: "empty stored ID with registered index", id: "", item_id: "wrong_item", output_index: 2 },
+      { name: "empty item and event IDs with registered index", id: "", item_id: "", output_index: 2 },
+    ]
+
+    fixtures.forEach((fixture) => {
+      fixture.events.forEach((event) => {
+        routings.forEach((routing) => {
+          it.effect(`${event.type} preserves content with ${routing.name}`, () =>
+            Effect.gen(function* () {
+              const item = { ...fixture.item, id: routing.id }
+              const response = yield* LLMClient.generate(request).pipe(
+                Effect.provide(
+                  fixedResponse(
+                    sseEvents(
+                      { type: "response.output_item.added", output_index: routing.output_index, item },
+                      { ...event, item_id: routing.item_id, output_index: routing.output_index },
+                      { type: "response.output_item.done", output_index: routing.output_index, item },
+                      { type: "response.completed", response: { id: "resp_1" } },
+                    ),
+                  ),
+                ),
+              )
+
+              const metadata = { "openai-compatible": { itemId: routing.id } }
+              if (fixture.item.type === "function_call") {
+                expect(response.toolCalls).toEqual([
+                  expect.objectContaining({
+                    id: "call_1",
+                    name: "lookup",
+                    input: { query: "Preserved" },
+                    providerMetadata: metadata,
+                  }),
+                ])
+                return
+              }
+              if (fixture.item.type === "reasoning") {
+                expect(response.message.content).toEqual([
+                  {
+                    type: "reasoning",
+                    text: "Preserved",
+                    providerMetadata: {
+                      "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "encrypted-state" },
+                    },
+                  },
+                ])
+                expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
+                return
+              }
+              expect(response.message.content).toEqual([
+                { type: "text", text: "Preserved", providerMetadata: metadata },
+              ])
+              expect(response.events.filter(LLMEvent.is.textEnd)).toEqual([
+                expect.objectContaining({ id: routing.id, providerMetadata: metadata }),
+              ])
+            }),
+          )
+        })
+      })
+    })
+
+    routings.forEach((routing) => {
+      it.effect(`preserves reasoning summary boundaries and terminal metadata with ${routing.name}`, () =>
+        Effect.gen(function* () {
+          const address = { item_id: routing.item_id, output_index: routing.output_index }
+          const response = yield* LLMClient.generate(request).pipe(
+            Effect.provide(
+              fixedResponse(
+                sseEvents(
+                  {
+                    type: "response.output_item.added",
+                    output_index: routing.output_index,
+                    item: { type: "reasoning", id: routing.id },
+                  },
+                  { type: "response.reasoning_summary_part.added", ...address, summary_index: 0 },
+                  { type: "response.reasoning_summary_text.delta", ...address, summary_index: 0, delta: "First." },
+                  { type: "response.reasoning_summary_text.done", ...address, summary_index: 0, text: "First." },
+                  { type: "response.reasoning_summary_part.done", ...address, summary_index: 0 },
+                  { type: "response.reasoning_summary_part.added", ...address, summary_index: 1 },
+                  { type: "response.reasoning_summary_text.done", ...address, summary_index: 1, text: "Second." },
+                  { type: "response.reasoning_summary_part.done", ...address, summary_index: 1 },
+                  {
+                    type: "response.completed",
+                    response: { output: [{ type: "reasoning", id: routing.id, encrypted_content: "final-state" }] },
+                  },
+                ),
+              ),
+            ),
+          )
+
+          expect(response.message.content).toEqual([
+            {
+              type: "reasoning",
+              text: "First.",
+              providerMetadata: { "openai-compatible": { itemId: routing.id } },
+            },
+            {
+              type: "reasoning",
+              text: "Second.",
+              providerMetadata: {
+                "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "final-state" },
+              },
+            },
+          ])
+          expect(response.events.filter(LLMEvent.is.reasoningEnd)).toEqual([
+            expect.objectContaining({
+              id: `${routing.id}:0`,
+              providerMetadata: { "openai-compatible": { itemId: routing.id } },
+            }),
+            expect.objectContaining({
+              id: `${routing.id}:1`,
+              providerMetadata: {
+                "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "final-state" },
+              },
+            }),
+          ])
+        }),
+      )
+    })
+
+    it.effect("reconciles pending empty-ID function arguments from completed output", () =>
+      Effect.gen(function* () {
+        const item = { type: "function_call", id: "", call_id: "call_1", name: "lookup" }
+        const response = yield* LLMClient.generate(request).pipe(
+          Effect.provide(
+            fixedResponse(
+              sseEvents(
+                { type: "response.output_item.added", item },
+                { type: "response.function_call_arguments.delta", item_id: "", delta: '{"query":"partial' },
+                {
+                  type: "response.completed",
+                  response: { output: [{ ...item, arguments: '{"query":"complete"}' }] },
+                },
+              ),
+            ),
+          ),
+        )
+
+        expect(response.toolCalls).toEqual([
+          expect.objectContaining({
+            id: "call_1",
+            name: "lookup",
+            input: { query: "complete" },
+            providerMetadata: { "openai-compatible": { itemId: "" } },
+          }),
+        ])
+      }),
+    )
+
+    it.effect("treats null output items as no-ops without disturbing registered items", () =>
+      Effect.gen(function* () {
+        const response = yield* LLMClient.generate(request).pipe(
+          Effect.provide(
+            fixedResponse(
+              sseEvents(
+                { type: "response.output_item.added", output_index: 0, item: null },
+                { type: "response.output_item.done", output_index: 0, item: null },
+                { type: "response.output_item.added", output_index: 0, item: { type: "message", id: "msg_1" } },
+                { type: "response.output_text.delta", output_index: 0, item_id: "wrong_item", delta: "Before " },
+                { type: "response.output_item.added", output_index: 0, item: null },
+                { type: "response.output_item.done", output_index: 0, item: null },
+                { type: "response.output_text.delta", output_index: 0, item_id: "wrong_item", delta: "after" },
+                { type: "response.output_item.done", output_index: 0, item: { type: "message", id: "msg_1" } },
+                { type: "response.completed", response: { id: "resp_1" } },
+              ),
+            ),
+          ),
+        )
+
+        expect(response.message.content).toEqual([
+          { type: "text", text: "Before after", providerMetadata: { "openai-compatible": { itemId: "msg_1" } } },
+        ])
+        expect(response.events.map((event) => event.type)).toEqual([
+          "step-start",
+          "text-start",
+          "text-delta",
+          "text-delta",
+          "text-end",
+          "step-finish",
+          "finish",
+        ])
+      }),
+    )
+
+    it.effect("rejects missing, null, and non-string event IDs even with a registered output index", () =>
+      Effect.gen(function* () {
+        yield* Effect.forEach(
+          [
+            ...fixtures.flatMap((fixture) => fixture.events.map((event) => ({ item: fixture.item, event }))),
+            ...["response.reasoning_summary_part.added", "response.reasoning_summary_part.done"].map((type) => ({
+              item: { type: "reasoning" },
+              event: { type, summary_index: 0 },
+            })),
+          ],
+          (fixture) =>
+            Effect.forEach([undefined, null, 0, false, {}, []], (item_id) =>
+              Effect.gen(function* () {
+                const error = yield* LLMClient.generate(request).pipe(
+                  Effect.provide(
+                    fixedResponse(
+                      sseEvents(
+                        {
+                          type: "response.output_item.added",
+                          output_index: 0,
+                          item: { ...fixture.item, id: "item_1" },
+                        },
+                        { ...fixture.event, output_index: 0, item_id },
+                        { type: "response.completed", response: { id: "resp_1" } },
+                      ),
+                    ),
+                  ),
+                  Effect.flip,
+                )
+                expect(error.reason._tag).toBe("InvalidProviderOutput")
+              }),
+            ),
+        )
+      }),
+    )
+
+    it.effect("keeps malformed output item IDs invalid", () =>
+      Effect.gen(function* () {
+        yield* Effect.forEach(["response.output_item.added", "response.output_item.done"], (type) =>
+          Effect.forEach(fixtures, (fixture) =>
+            Effect.forEach(
+              fixture.item.type === "message" ? [undefined, null, 0, false, {}, []] : [null, 0, false, {}, []],
+              (id) =>
+                Effect.gen(function* () {
+                  const error = yield* LLMClient.generate(request).pipe(
+                    Effect.provide(
+                      fixedResponse(
+                        sseEvents(
+                          { type, item: { ...fixture.item, id } },
+                          { type: "response.completed", response: { id: "resp_1" } },
+                        ),
+                      ),
+                    ),
+                    Effect.flip,
+                  )
+                  expect(error.reason._tag).toBe("InvalidProviderOutput")
+                }),
+            ),
+          ),
+        )
+      }),
+    )
+  })
+
  it.effect("streams function calls without optional item ids through the shared baseline", () =>
    Effect.gen(function* () {
      const model = configure({
@@ -362,7 +666,7 @@ describe("Open Responses-compatible route", () => {

      expect(response.events.find(LLMEvent.is.toolCall)).toMatchObject({
        input: { query: "complete" },
-        providerMetadata: { openresponses: { itemId: "item_1" } },
+        providerMetadata: { example: { itemId: "item_1" } },
      })
    }),
  )
@@ -394,7 +698,7 @@ describe("Open Responses-compatible route", () => {
      )

      expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
-        providerMetadata: { openresponses: { itemId: "rs_raw", reasoningEncryptedContent: "raw-state" } },
+        providerMetadata: { "openai-compatible": { itemId: "rs_raw", reasoningEncryptedContent: "raw-state" } },
      })
    }),
  )
@@ -443,7 +747,7 @@ describe("Open Responses-compatible route", () => {
            Message.assistant({
              type: "text",
              text: "Unclassified.",
-              providerMetadata: { openresponses: { phase: null } },
+              providerMetadata: { "openai-compatible": { phase: null } },
            }),
          ],
        }),
@@ -502,7 +806,7 @@ describe("Open Responses-compatible route", () => {
        {
          type: "text",
          text: "I can't help with that.",
-          providerMetadata: { openresponses: { itemId: "msg_refusal" } },
+          providerMetadata: { example: { itemId: "msg_refusal" } },
        },
      ])

@@ -591,7 +895,7 @@ describe("Open Responses-compatible route", () => {

      expect(response.toolCalls).toEqual([])
      expect(response.events.find(LLMEvent.is.finish)).toMatchObject({
-        providerMetadata: { openresponses: { responseId: "resp_1" } },
+        providerMetadata: { example: { responseId: "resp_1" } },
      })
    }),
  )
@@ -2229,6 +2229,35 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("accepts empty IDs for native reasoning text deltas", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 1, item: { type: "reasoning", id: "" } },
+              { type: "response.reasoning_text.delta", output_index: 1, item_id: "", delta: "Raw" },
+              {
+                type: "response.output_item.done",
+                output_index: 1,
+                item: { type: "reasoning", id: "", encrypted_content: "state" },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Raw",
+          providerMetadata: { openai: { itemId: "", reasoningEncryptedContent: "state" } },
+        },
+      ])
+    }),
+  )
+
  it.effect("falls back to item ids when an output index was not registered", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
@@ -3150,38 +3179,50 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("drops replayed item ids outside the server's grammar", () =>
+  it.effect("preserves provider-issued item ids and removes malformed ids without dropping items", () =>
    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${"a".repeat(64)}` } },
+                providerMetadata: { openai: { itemId: `message_${"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" } },
              },
+              {
+                type: "reasoning",
+                text: "Missing suffix.",
+                providerMetadata: { openai: { itemId: "rs_", reasoningEncryptedContent: "another-state" } },
+              },
+              {
+                type: "reasoning",
+                text: "No prefix separator.",
+                providerMetadata: { openai: { itemId: "550e8400-e29b-41d4-a716-446655440000" } },
+              },
              ToolCallPart.make({
                id: "call_1",
                name: "lookup",
                input: { query: "weather" },
                providerMetadata: { openai: { itemId: "toolu_01A" } },
              }),
+              ToolCallPart.make({
+                id: "call_2",
+                name: "lookup",
+                input: { query: "news" },
+                providerMetadata: { openai: { itemId: "fc_" } },
+              }),
            ]),
          ],
        }),
@@ -3190,18 +3231,44 @@ describe("OpenAI Responses route", () => {
      expect(prepared.body.input).toEqual([
        {
          type: "message",
+          id: "history_1",
          role: "assistant",
-          content: [
-            { type: "output_text", text: "Hello" },
-            { type: "output_text", text: "World" },
-          ],
+          content: [{ type: "output_text", text: "Hello" }],
+        },
+        {
+          type: "message",
+          id: `message_${"a".repeat(64)}`,
+          role: "assistant",
+          content: [{ type: "output_text", text: "World" }],
+        },
+        {
+          type: "reasoning",
+          id: "thinking_1",
+          summary: [{ type: "summary_text", text: "Checked the diff." }],
+          encrypted_content: "encrypted-state",
+        },
+        {
+          type: "reasoning",
+          summary: [{ type: "summary_text", text: "Missing suffix." }],
+          encrypted_content: "another-state",
+        },
+        {
+          type: "reasoning",
+          summary: [{ type: "summary_text", text: "No prefix separator." }],
        },
        {
          type: "function_call",
+          id: "toolu_01A",
          call_id: "call_1",
          name: "lookup",
          arguments: '{"query":"weather"}',
        },
+        {
+          type: "function_call",
+          call_id: "call_2",
+          name: "lookup",
+          arguments: '{"query":"news"}',
+        },
      ])
    }),
  )
@@ -295,7 +295,7 @@ describe("OpenRouter", () => {
              {
                type: "reasoning",
                text: "Thinking",
-                providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
+                providerMetadata: { openrouter: { reasoningField: "reasoning", reasoningDetails: details } },
              },
            ]),
          ],
@@ -328,7 +328,7 @@ describe("OpenRouter", () => {
            Message.assistant({
              type: "reasoning",
              text: "Thinking",
-              providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
+              providerMetadata: { openrouter: { reasoningField: "reasoning", reasoningDetails: details } },
            }),
          ],
        }),
@@ -354,7 +354,7 @@ describe("OpenRouter", () => {
            Message.assistant({
              type: "reasoning",
              text: "AB",
-              providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
+              providerMetadata: { openrouter: { reasoningField: "reasoning", reasoningDetails: details } },
            }),
          ],
        }),
@@ -80,7 +80,9 @@ describe("Z.ai Images", () => {

  it.effect("sanitizes unpaired surrogates in outbound image requests", () =>
    Image.generate({
-      model: ZAI.configure({ apiKey: "test", http: { body: { metadata: { source: "default\uDC00" } } } }).image("model"),
+      model: ZAI.configure({ apiKey: "test", http: { body: { metadata: { source: "default\uDC00" } } } }).image(
+        "model",
+      ),
      prompt: "A red circle \uD800 on a white background \u{1F600}",
    }).pipe(
      Effect.provide(
@@ -40,4 +40,13 @@ describe("Route.with", () => {
      "x-patch": "patch",
    })
  })
+
+  test("assigns metadata ownership to a replacement provider and preserves explicit overrides", () => {
+    const route = OpenAIChat.route.with({ provider: "azure" })
+    const overridden = route.with({ providerMetadataKey: "custom-azure" }).with({ headers: { "x-test": "value" } })
+
+    expect(route.providerMetadataKey).toBe("azure")
+    expect(overridden.providerMetadataKey).toBe("custom-azure")
+    expect(overridden.defaults).not.toHaveProperty("providerMetadataKey")
+  })
 })
@@ -1,3 +1,5 @@
 src/assets/theme.css
 e2e/test-results
 e2e/playwright-report
+component-tests/test-results
+component-tests/playwright-report
@@ -31,7 +31,7 @@ Your app is ready to be deployed!

 ## E2E Testing

-Playwright starts the Vite dev server automatically via `webServer`, and UI tests expect an opencode backend at `localhost:4096` by default.
+Locally, Playwright starts the Vite dev server automatically via `webServer`, or reuses one already running at the configured address. The browser suite uses isolated API fixtures rather than a live opencode backend.

 ```bash
 bunx playwright install chromium
@@ -39,11 +39,35 @@ bun run test:e2e:local
 bun run test:e2e:local -- --grep "settings"
 ```

+CI builds the app once and runs the same suite against Vite preview, serving production assets from `dist`. Managed built runs never reuse an existing server, so a running dev server cannot silently replace the production build. To run this mode locally:
+
+```bash
+bun run test:e2e:built
+bun run test:e2e:built -- --grep "settings"
+```
+
+To test an already-running dev server without starting or building a server:
+
+```bash
+PLAYWRIGHT_BASE_URL=http://127.0.0.1:4444 bun run test:e2e
+```
+
+For an already-running production build, also set `PLAYWRIGHT_BUILD=1` so the fixture API uses the app's origin:
+
+```bash
+PLAYWRIGHT_BUILD=1 PLAYWRIGHT_BASE_URL=http://127.0.0.1:4444 bun run test:e2e
+```
+
+External targets must use HTTP because fixture URLs use HTTP. `PLAYWRIGHT_BASE_URL` skips server startup and building in either mode.
+
+Compiled CLI startup and service lifecycle coverage runs separately in CI via `packages/cli/script/service-smoke.ts`.
+
 Environment options:

- `PLAYWRIGHT_SERVER_HOST` / `PLAYWRIGHT_SERVER_PORT` (backend address, default: `localhost:4096`)
- `PLAYWRIGHT_PORT` (Vite dev server port, default: `3000`)
- `PLAYWRIGHT_BASE_URL` (override base URL, default: `http://localhost:<PLAYWRIGHT_PORT>`)
+- `PLAYWRIGHT_BUILD=1` (build and preview locally; always enabled when `CI` is set)
+- `PLAYWRIGHT_SERVER_HOST` / `PLAYWRIGHT_SERVER_PORT` (dev fixture API address, default: `127.0.0.1:4096`; built runs use the app's origin, matching production)
+- `PLAYWRIGHT_PORT` (managed dev or preview server port, default: `3000`)
+- `PLAYWRIGHT_BASE_URL` (use an externally managed app instead of starting a server; otherwise defaults to `http://127.0.0.1:<PLAYWRIGHT_PORT>`)

 ## Deployment

@@ -0,0 +1,26 @@
+import { expect, story } from "../../storybook/playwright/story"
+
+// Moved from packages/app/e2e/regression/prompt-thinking-level.spec.ts
+story("shows the thinking level control while relevant", async ({ mount, page }) => {
+  const component = await mount("opencode-composer-flow--model-and-variant")
+  const composer = component.locator('[data-component="composer"]')
+  const input = composer.locator('[data-component="composer-editor"]')
+  const control = composer.getByRole("button", { name: "Choose model variant" })
+
+  await page.mouse.move(0, 0)
+  await page.evaluate(() => (document.activeElement as HTMLElement | null)?.blur())
+  await expect(control).toBeVisible()
+
+  await control.click()
+  const high = page.getByRole("menuitemradio", { name: "high" })
+  await expect(high).toBeVisible()
+  await page.mouse.move(0, 0)
+  await expect(control).toBeVisible()
+  await expect(high).toBeVisible()
+  await high.click()
+
+  await input.focus()
+  await expect(control).toBeVisible()
+  await input.blur()
+  await expect(control).toBeVisible()
+})
@@ -44,6 +44,7 @@ The suite contains:
 - single-session tab close timing through stable home restoration
 - cached session repaint and mutation tracing
 - streaming timeline throughput, RAF-gap, long-task, geometry, and remount diagnostics
+- retained renderer heap with a large model catalog across repeated session navigation

 All benchmarks import the shared `benchmark` fixture. Pages created through Playwright's `page` fixture automatically capture main-frame navigation history and emit a Chrome trace when `OPENCODE_PERFORMANCE_TRACE_DIR` is set. Benchmarks that need isolated browser contexts use `withBenchmarkPage`, which owns the context and the same diagnostics lifecycle.

@@ -78,6 +79,17 @@ Benchmarks do not assert machine-dependent performance budgets. Streaming proces

 Committed smoke and regression tests continue to own correctness coverage for pagination, tab paint, context resize, collapse state, and composer spacing.

+## Retained renderer memory
+
+Run the catalog workload against the production app bundle:
+
+```sh
+bunx playwright test --config e2e/performance/playwright.config.ts \
+  timeline/provider-memory-benchmark.spec.ts --repeat-each=3
+```
+
+`PROVIDER_MEMORY_MODELS` defaults to 1,200 and `PROVIDER_MEMORY_SWITCHES` defaults to 10. Each sample records Chromium's `Runtime.getHeapUsage` and `Memory.getDOMCounters` after an explicit garbage collection. This measures retained state, not allocation peaks or normal GC timing. It does not include worker heaps, the Electron main/GPU processes, or the OpenCode server, and must not be reported as total desktop RAM. Use identical model counts and navigation sequences for before/after comparisons.
+
 ## Chrome traces

 Set `OPENCODE_PERFORMANCE_TRACE_DIR` to emit a standard Chrome DevTools trace for every benchmark page automatically:
@@ -1,7 +1,7 @@
 import config from "../../playwright.config"

 const port = Number(process.env.PLAYWRIGHT_PORT ?? 3000)
-process.env.PLAYWRIGHT_SERVER_PORT = String(port)
+process.env.PLAYWRIGHT_SERVER_PORT = new URL(process.env.PLAYWRIGHT_BASE_URL ?? `http://127.0.0.1:${port}`).port || "80"
 process.env.OPENCODE_PERFORMANCE_RUN_ID ??= `${new Date().toISOString().replace(/[:.]/g, "-")}-${process.pid}`

 export default {
@@ -12,9 +12,11 @@ export default {
  fullyParallel: false,
  workers: 1,
  reporter: [["html", { outputFolder: "../playwright-report/performance", open: "never" }], ["line"]],
-  webServer: {
-    ...config.webServer,
-    command: `bun run build && bun run serve -- --host 0.0.0.0 --port ${port} --strictPort`,
-    reuseExistingServer: false,
-  },
+  webServer: config.webServer
+    ? {
+        ...config.webServer,
+        command: `bun run build && bun run serve -- --host 0.0.0.0 --port ${port} --strictPort`,
+        reuseExistingServer: false,
+      }
+    : undefined,
 }
@@ -0,0 +1,98 @@
+import { benchmark, expect } from "../benchmark"
+import { mockOpenCodeServer } from "../../utils/mock-server"
+import { expectSessionTitle } from "../../utils/waits"
+import { fixture, pageMessages } from "./session-timeline-stress.fixture"
+import { installStressSessionTabs, installTimelineSettings, stressSessionHref } from "./timeline-test-helpers"
+import { waitForStableTimeline } from "./session-tab-switch-probe"
+import type { CatalogUpdated } from "@opencode-ai/client/promise"
+
+benchmark("measures retained renderer memory with a large model catalog", async ({ page, report }) => {
+  benchmark.setTimeout(120_000)
+  const count = Number(process.env.PROVIDER_MEMORY_MODELS ?? 1200)
+  const switches = Number(process.env.PROVIDER_MEMORY_SWITCHES ?? 10)
+  const provider = fixture.provider.all[0]
+  const selected = { ...provider.models["claude-opus-4-6"] }
+  await mockOpenCodeServer(page, {
+    directory: fixture.directory,
+    project: fixture.project,
+    sessions: fixture.sessions,
+    pageMessages,
+    provider: {
+      ...fixture.provider,
+      all: [
+        {
+          ...provider,
+          models: {
+            [selected.id]: selected,
+            ...Object.fromEntries(
+              Array.from({ length: count - 1 }, (_, index) => {
+                const id = `catalog-model-${index}`
+                return [
+                  id,
+                  {
+                    id,
+                    name: `Catalog model ${index}`,
+                    cost: { input: 1, output: 2 },
+                    limit: { context: 200_000, output: 8192 },
+                    variants: { high: { reasoningEffort: "high" } },
+                  },
+                ]
+              }),
+            ),
+          },
+        },
+      ],
+    },
+  })
+  await installTimelineSettings(page)
+  await installStressSessionTabs(page)
+  await page.goto(stressSessionHref(fixture.sourceID))
+  await expectSessionTitle(page, fixture.expected.sourceTitle)
+  await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
+  await expect(page.locator('[data-action="composer-model"]')).toContainText("Claude Opus 4.6")
+  const cdp = await page.context().newCDPSession(page)
+  const samples = []
+  for (let index = 0; index <= switches; index++) {
+    if (index > 0) {
+      const target = index % 2 === 1
+      const id = target ? fixture.targetID : fixture.sourceID
+      await page.locator(`[data-slot="titlebar-tabs"] a[href="${stressSessionHref(id)}"]`).click()
+      await expectSessionTitle(page, target ? fixture.expected.targetTitle : fixture.expected.sourceTitle)
+      await waitForStableTimeline(
+        page,
+        (target ? fixture.expected.targetMessageIDs : fixture.expected.sourceMessageIDs).at(-1)!,
+      )
+      await expect(page.locator('[data-action="composer-model"]')).toContainText("Claude Opus 4.6")
+    }
+    // GC is an explicit retained-heap measurement, not an application optimization or readiness wait.
+    await cdp.send("HeapProfiler.collectGarbage")
+    samples.push({
+      switches: index,
+      heap: await cdp.send("Runtime.getHeapUsage"),
+      dom: await cdp.send("Memory.getDOMCounters"),
+    })
+  }
+  expect(samples).toHaveLength(switches + 1)
+  expect(samples.every((sample) => sample.heap.usedSize > 0)).toBe(true)
+  selected.name = "Updated catalog model"
+  await page.evaluate(
+    (event) => {
+      const host = window as Window & { __mockServerStream?: { push: (events: CatalogUpdated[]) => void } }
+      if (!host.__mockServerStream) throw new Error("Missing fixture event stream")
+      host.__mockServerStream.push([event])
+    },
+    {
+      id: "evt_catalog_refresh",
+      created: Date.now(),
+      type: "catalog.updated",
+      location: { directory: fixture.directory },
+      data: {},
+    } satisfies CatalogUpdated,
+  )
+  await expect(page.locator('[data-action="composer-model"]')).toContainText(selected.name)
+  report(
+    { samples },
+    { models: count, switches, gc: "explicit", scope: "renderer main isolate; not total desktop RAM" },
+  )
+  await cdp.detach()
+})
@@ -3,7 +3,7 @@ import { base64Encode } from "@opencode-ai/util/encode"
 import { currentSession } from "../utils/mock-server"
 import { installSseTransport } from "../utils/sse-transport"

-const serverA = "http://127.0.0.1:4096"
+const serverA = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const serverB = "http://127.0.0.1:4097"
 const sessionA = session("ses_server_a", "C:/server-a", "Server A session")
 const sessionB = session("ses_server_b", "/home/server-b", "Server B session")
@@ -12,17 +12,17 @@ test("closing the active server's last tab opens the remaining server tab", asyn
  const requests: string[] = []
  await mockServers(page, requests)
  await page.addInitScript(
-    ({ serverB, sessionA, sessionB }) => {
+    ({ serverA, serverB, sessionA, sessionB }) => {
      localStorage.setItem("opencode.global.dat:server", JSON.stringify({ list: [serverB] }))
      localStorage.setItem(
        "opencode.window.browser.dat:tabs",
        JSON.stringify([
-          { type: "session", server: "http://127.0.0.1:4096", sessionId: sessionA },
+          { type: "session", server: serverA, sessionId: sessionA },
          { type: "session", server: serverB, sessionId: sessionB },
        ]),
      )
    },
-    { serverB, sessionA: sessionA.id, sessionB: sessionB.id },
+    { serverA, serverB, sessionA: sessionA.id, sessionB: sessionB.id },
  )

  const hrefA = `/server/${base64Encode(serverA)}/session/${sessionA.id}`
@@ -55,7 +55,7 @@ function session(id: string, directory: string, title: string) {
 async function mockServers(page: Page, requests: string[]) {
  await installSseTransport(page, { server: serverA })
  await installSseTransport(page, { server: serverB })
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
    const url = new URL(route.request().url())
    if (url.origin !== serverA && url.origin !== serverB) return route.fallback()
    requests.push(url.toString())
@@ -0,0 +1,103 @@
+import { expect, test } from "@playwright/test"
+import { base64Encode } from "@opencode-ai/util/encode"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+const directory = "C:/Projects/extensions-demo"
+const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
+const session = {
+  id: "ses_project_extensions",
+  title: "Existing session",
+  directory,
+  projectID: "proj_extensions_demo",
+  time: { created: 1700000000000, updated: 1700000000000 },
+}
+
+test.use({ viewport: { width: 1440, height: 1000 }, colorScheme: "dark" })
+
+test("project Extensions stays inside settings while plugins load", async ({ page }) => {
+  await mockOpenCodeServer(page, {
+    directory,
+    project: {
+      id: session.projectID,
+      canonical: directory,
+      name: "Extensions demo",
+      vcs: "git",
+      time: session.time,
+      sandboxes: [],
+    },
+    provider: { all: [], connected: [], default: {} },
+    sessions: [session],
+    pageMessages: () => ({ items: [] }),
+  })
+  await page.addInitScript(
+    ({ server, sessionID, directory }) => {
+      localStorage.setItem(
+        "opencode.global.dat:server",
+        JSON.stringify({ projects: { local: [{ worktree: directory, expanded: true }] } }),
+      )
+      localStorage.setItem(
+        "opencode.window.browser.dat:tabs",
+        JSON.stringify([{ type: "session", server, sessionId: sessionID }]),
+      )
+    },
+    { server, sessionID: session.id, directory },
+  )
+  const href = `/server/${base64Encode(server)}/session/${session.id}`
+  await page.goto(href)
+  await expect(page.getByRole("heading", { name: session.title, exact: true })).toBeVisible()
+  await page.keyboard.press("Control+,")
+  const settings = page.getByTestId("settings-screen")
+  await settings.getByRole("tab", { name: "Projects", exact: true }).click()
+  await settings.getByText("Extensions demo", { exact: true }).click()
+  const dialog = page.getByRole("dialog")
+  await expect(dialog.getByRole("textbox", { name: "Name", exact: true })).toBeFocused()
+
+  const globalPlugins = Promise.withResolvers<void>()
+  const projectPlugins = Promise.withResolvers<void>()
+  await page.route(
+    (url) => url.pathname === "/api/plugin",
+    async (route) => {
+      const project = new URL(route.request().url()).searchParams.get("location[directory]")
+      await (project ? projectPlugins : globalPlugins).promise
+      await route.fulfill({
+        json: {
+          location: project ? { directory: project } : {},
+          data: (project ? ["shared-plugin", "project-plugin"] : ["shared-plugin"]).map((id) => ({
+            id,
+            source: { type: "package", package: id },
+            status: "active",
+            tui: false,
+          })),
+        },
+      })
+    },
+  )
+  const requested = page.waitForRequest((request) => {
+    const url = new URL(request.url())
+    return url.pathname === "/api/plugin" && url.searchParams.get("location[directory]") === directory
+  })
+  await dialog.getByRole("tab", { name: "Extensions", exact: true }).click()
+  await requested
+  await expect(page).toHaveURL(href)
+  await expect(dialog.getByRole("heading", { name: "Extensions", exact: true })).toBeVisible()
+  await expect(settings).toBeVisible()
+  await expect(page.getByRole("heading", { name: session.title, exact: true, includeHidden: true })).toBeHidden()
+  await dialog.getByRole("tab", { name: "Plugins", exact: true }).click()
+  await expect(dialog.getByRole("tab", { name: "Plugins", exact: true })).toHaveAttribute("aria-selected", "true")
+
+  globalPlugins.resolve()
+  await dialog.getByRole("tab", { name: "Scripts", exact: true }).click()
+  await expect(dialog.getByRole("heading", { name: "Scripts", exact: true })).toBeVisible()
+  await dialog.getByRole("tab", { name: "Extensions", exact: true }).click()
+  projectPlugins.resolve()
+  await dialog.getByRole("tab", { name: "Plugins", exact: true }).click()
+  await expect(dialog.getByText("project-plugin", { exact: true })).toBeVisible()
+  await dialog.getByRole("button", { name: "Shared with all projects 1", exact: true }).click()
+  await expect(dialog.getByText("shared-plugin", { exact: true })).toBeVisible()
+  await expect(page).toHaveURL(href)
+
+  await page.keyboard.press("Escape")
+  await expect(dialog).toBeHidden()
+  await expect(settings.getByRole("tab", { name: "Projects", exact: true })).toHaveAttribute("aria-selected", "true")
+  await expect(page.getByRole("heading", { name: session.title, exact: true, includeHidden: true })).toBeHidden()
+})
@@ -3,7 +3,7 @@ import { expect, test, type Page, type Route } from "@playwright/test"
 import { installSseTransport } from "../utils/sse-transport"
 import { currentSession } from "../utils/mock-server"

-const serverA = `http://127.0.0.1:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
+const serverA = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const serverB = "http://127.0.0.1:4097"
 const directoryA = "C:/server-a"
 const directoryB = "/home/server-b"
@@ -24,11 +24,17 @@ test("session settings use the remote server context", async ({ page }) => {
  await configureServers(page)

  await page.goto(`/server/${base64Encode(serverB)}/session/${sessionB.id}`)
-  await expect(page.getByRole("heading", { name: sessionB.title, exact: true })).toBeVisible()
+  const sessionHeading = page.getByRole("heading", { name: sessionB.title, exact: true, includeHidden: true })
+  await expect(sessionHeading).toBeVisible()
  await page.keyboard.press("Control+,")

-  const dialog = page.locator(".settings-dialog")
-  const autoAccept = dialog.locator('[data-action="settings-auto-accept-permissions"]')
+  const settings = page.getByTestId("settings-screen")
+  await expect(settings).toBeVisible()
+  await expect(page.getByRole("dialog")).toHaveCount(0)
+  await expect(settings.getByRole("tablist")).toHaveCSS("width", "328px")
+  await expect(sessionHeading).toBeAttached()
+  await expect(sessionHeading).toBeHidden()
+  const autoAccept = settings.locator('[data-action="settings-auto-accept-permissions"]')
  const input = autoAccept.getByRole("switch")
  await expect(autoAccept).toBeVisible()
  await expect(input).toBeEnabled()
@@ -55,9 +61,12 @@ test("session settings use the remote server context", async ({ page }) => {
      },
    ])

-  await dialog.getByRole("tab", { name: "Models" }).click()
-  await expect(dialog.getByRole("switch", { name: "Server B Model" })).toBeEnabled()
-  await expect(dialog.getByRole("switch", { name: "Server A Model" })).toHaveCount(0)
+  await settings.getByRole("tab", { name: "Models" }).click()
+  await expect(settings.getByRole("switch", { name: "Server B Model" })).toBeEnabled()
+  await expect(settings.getByRole("switch", { name: "Server A Model" })).toHaveCount(0)
+  await settings.getByRole("button", { name: "Back to app" }).click()
+  await expect(settings).toBeHidden()
+  await expect(sessionHeading).toBeVisible()
 })

 test("auto-accept responds for an unfocused server session", async ({ page }) => {
@@ -78,7 +87,7 @@ 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-dialog").locator('[data-action="settings-auto-accept-permissions"]')
+  const autoAccept = page.getByTestId("settings-screen").locator('[data-action="settings-auto-accept-permissions"]')
  await autoAccept.locator('[data-slot="switch-control"]').click()
  await expect(autoAccept.getByRole("switch")).toBeChecked()
  await expect
@@ -178,7 +187,7 @@ test("auto-accept sweeps again after a reconnect", async ({ page }) => {
  const first = await transport.waitForConnection()

  await page.keyboard.press("Control+,")
-  const autoAccept = page.locator(".settings-dialog").locator('[data-action="settings-auto-accept-permissions"]')
+  const autoAccept = page.getByTestId("settings-screen").locator('[data-action="settings-auto-accept-permissions"]')
  await autoAccept.locator('[data-slot="switch-control"]').click()
  await expect(autoAccept.getByRole("switch")).toBeChecked()
  await expect
@@ -234,7 +243,7 @@ test("auto-accept approves a request discovered by opening a session", async ({
  await expect(page.getByRole("heading", { name: sessionA.title, exact: true })).toBeVisible()

  await page.keyboard.press("Control+,")
-  const autoAccept = page.locator(".settings-dialog").locator('[data-action="settings-auto-accept-permissions"]')
+  const autoAccept = page.getByTestId("settings-screen").locator('[data-action="settings-auto-accept-permissions"]')
  await autoAccept.locator('[data-slot="switch-control"]').click()
  await expect(autoAccept.getByRole("switch")).toBeChecked()

@@ -299,7 +308,7 @@ async function mockServers(
  permissionResponses: PermissionResponse[] = [],
  options: MockServerOptions = {},
 ) {
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
    const url = new URL(route.request().url())
    if (url.origin !== serverA && url.origin !== serverB) return route.fallback()
    const remote = url.origin === serverB
@@ -2,7 +2,7 @@ import { expect, test, type Page, type Route } from "@playwright/test"
 import { base64Encode } from "@opencode-ai/util/encode"
 import { currentSession } from "../utils/mock-server"

-const serverA = "http://127.0.0.1:4096"
+const serverA = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const serverB = "http://127.0.0.1:4097"
 const sessionA = session("ses_server_a", "C:/server-a", "Server A session")
 const sessionB = session("ses_server_b", "/home/server-b", "Server B session")
@@ -51,7 +51,7 @@ function session(id: string, directory: string, title: string) {
 }

 async function mockServers(page: Page) {
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
    const url = new URL(route.request().url())
    if (url.origin !== serverA && url.origin !== serverB) return route.fallback()
    const current = url.origin === serverA ? sessionA : sessionB
@@ -12,54 +12,6 @@ test.beforeEach(async ({ page }) => {
  await openReview(page)
 })

-test("opens the comment editor when code is clicked", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const line = review.getByText("export const value = 'after'", { exact: true })
-  await expectAppVisible(line)
-  await line.click()
-
-  await expect(review.getByRole("textbox")).toBeVisible()
-  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on line 2")
-})
-
-test("opens the comment editor when a line number is clicked", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const lineNumber = review.locator('[data-column-number="1"]').last()
-  await expectAppVisible(lineNumber)
-  await lineNumber.click()
-
-  await expect(review.getByRole("textbox")).toBeVisible()
-  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on line 1")
-})
-
-test("opens the comment editor for a line number range", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const start = review.locator('[data-column-number="1"]').last()
-  const end = review.locator('[data-column-number="3"]').last()
-  await expectAppVisible(start)
-  await expectAppVisible(end)
-
-  await start.dragTo(end)
-
-  await expect(review.getByRole("textbox")).toBeVisible()
-  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on lines 1-3")
-})
-
-test("shows a comment button when a diff line is hovered", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const line = review.getByText("export const first = 1", { exact: true })
-  await expectAppVisible(line)
-
-  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")
-})
-
 test("stages a submitted line comment in the prompt context", async ({ page }) => {
  page.on("request", (request) => {
    expect.soft(request.method(), `unexpected ${request.method()} ${new URL(request.url()).pathname}`).toBe("GET")
@@ -2,7 +2,7 @@ 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 }) => {
+test("server dialog keeps focus above fullscreen settings", async ({ page }) => {
  await page.addInitScript((server) => {
    localStorage.setItem("opencode.global.dat:server", JSON.stringify({ list: [server] }))
  }, server)
@@ -24,8 +24,9 @@ test("nested server dialog keeps focus inside the top layer", async ({ page }) =

  await page.goto("/")
  await page.keyboard.press("Control+,")
-  const settings = page.locator(".settings-dialog")
+  const settings = page.getByTestId("settings-screen")
  await expect(settings).toBeVisible()
+  await expect(page.getByRole("dialog")).toHaveCount(0)
  await settings.getByRole("tab", { name: "Servers" }).click()
  await settings.getByRole("button", { name: "Add server" }).click()

@@ -41,6 +42,9 @@ test("nested server dialog keeps focus inside the top layer", async ({ page }) =
  await expect(password).toBeFocused()
  await password.fill("secret")
  await expect(password).toHaveValue("secret")
+  await page.keyboard.press("Escape")
+  await expect(editor).toBeHidden()
+  await expect(settings).toBeVisible()
 })

 function json(route: Route, body: unknown, status = 200) {
@@ -1,5 +1,5 @@
 import { expect, test, type Page } from "@playwright/test"
-import type { OpenCodeEvent } from "@opencode-ai/client/promise"
+import type { OpenCodeEvent, SessionMessageInfo } from "@opencode-ai/client/promise"
 import { base64Encode } from "@opencode-ai/util/encode"
 import { mockOpenCodeServer } from "../utils/mock-server"
 import { expectAppVisible } from "../utils/waits"
@@ -18,7 +18,7 @@ type InboxRow = {
  delivery: "steer" | "queue"
 }

-function createQueueMock(seed: string[]) {
+function createQueueMock(seed: string[], messages: SessionMessageInfo[] = []) {
  const rows: InboxRow[] = seed.map((text, index) => ({
    id: `inb_seed_${index + 1}`,
    sessionID,
@@ -32,13 +32,16 @@ function createQueueMock(seed: string[]) {
  const changes: { inboxID: string; action: "cancel" | "steer" }[] = []
  const log: string[] = []
  let sequence = 0
-  const emit = (type: OpenCodeEvent["type"], data: OpenCodeEvent["data"]) => {
+  const emit = <Type extends OpenCodeEvent["type"]>(
+    type: Type,
+    data: Extract<OpenCodeEvent, { type: Type }>["data"],
+  ) => {
    sequence += 1
    events.push({
      id: `evt_queue_${sequence}`,
      type,
      created: Date.now(),
-      durable: { aggregateID: sessionID, seq: sequence, version: 1 },
+      durable: { aggregateID: sessionID, seq: sequence, version: type === "session.tool.success" ? 2 : 1 },
      data,
    } as OpenCodeEvent)
  }
@@ -47,6 +50,8 @@ function createQueueMock(seed: string[]) {
    prompts,
    changes,
    log,
+    messages,
+    emit,
    events: () => events.splice(0),
    onPrompt: (input: { sessionID: string; body: Record<string, unknown> }) => {
      prompts.push(input.body)
@@ -126,10 +131,11 @@ async function openSession(page: Page, mock: ReturnType<typeof createQueueMock>,
        directory,
        title: "Session queue regression",
        version: "dev",
+        model: { id: "queue-model", providerID: "opencode" },
        time: { created: 1700000000000, updated: 1700000000000 },
      },
    ],
-    pageMessages: () => ({ items: [] }),
+    pageMessages: () => ({ items: mock.messages }),
    sessionStatus: () => ({ [sessionID]: { type: "running" } }),
    inbox: () => mock.rows.map((row) => ({ ...row, payload: { ...row.payload } })),
    onPrompt: mock.onPrompt,
@@ -201,12 +207,15 @@ test("editing restores the existing draft and replaces only the original queue p
  await view.input.fill("my in-progress draft")
  await original.click()
  await expect(view.input).toHaveText("tighten the error copy")
+  await expect(view.input).toBeFocused()
  await view.input.press("Escape")
  await expect(view.input).toHaveText("my in-progress draft")

  await original.click()
  await expect(view.input).toHaveText("tighten the error copy")
+  await expect(view.input).toBeFocused()
  await view.input.fill("tighten the error copy and add a retry hint")
+  await expect(view.input).toHaveText("tighten the error copy and add a retry hint")
  await view.input.press("Enter")

  await expect(view.rows.locator('[data-action="session-queue-edit"]')).toHaveText([
@@ -224,3 +233,130 @@ test("editing restores the existing draft and replaces only the original queue p
  expect(mock.changes.map((change) => change.action)).toEqual(["cancel", "cancel", "cancel"])
  expect(mock.log[0]).toBe("prompt:queue")
 })
+
+for (const delivery of ["steer", "queue"] as const) {
+  test(`keeps finished tools above a pending ${delivery === "queue" ? "queue-to-steer" : "steer"} follow-up`, async ({
+    page,
+  }, testInfo) => {
+    const model = { id: "queue-model", providerID: "opencode" }
+    const userID = "msg_queue_initial_user"
+    const assistantID = "msg_queue_continued_assistant"
+    const followUp = "U2: Also check the retry path."
+    const mock = createQueueMock(
+      [],
+      [
+        { id: userID, type: "user", text: "U1: Inspect the queue ordering.", time: { created: 1700000000000 } },
+        {
+          id: "msg_queue_initial_assistant",
+          type: "assistant",
+          agent: "build",
+          model,
+          content: [{ type: "text", text: "A1: I will inspect the current implementation." }],
+          finish: "tool-calls",
+          time: { created: 1700000000001, completed: 1700000000002 },
+        },
+      ],
+    )
+    const view = await openSession(page, mock, delivery)
+    const transcript = page.locator("[data-timeline-virtual-content]")
+    const thinking = transcript.locator('[data-timeline-row="Thinking"]')
+    await expect(transcript.getByText("A1: I will inspect the current implementation.", { exact: true })).toBeVisible()
+    await expect(thinking).toBeVisible()
+    await expect(view.input).toBeEditable()
+    await view.input.fill(followUp)
+    await view.input.press("Enter")
+    await expect.poll(() => mock.rows.map((row) => row.delivery)).toEqual([delivery])
+    await expect(view.input).toHaveText("")
+
+    const inboxID = mock.rows[0].id
+    const pending = transcript.locator(`[data-timeline-row="UserMessage"][data-message-id="${inboxID}"]`)
+    if (delivery === "queue") {
+      const queued = view.rows.filter({ hasText: followUp })
+      await expect(queued).toBeVisible()
+      await expect(pending).toHaveCount(0)
+      await queued.hover()
+      await queued.getByRole("button", { name: "Steer", exact: true }).click()
+      await expect.poll(() => mock.changes).toEqual([{ inboxID, action: "steer" }])
+    }
+    await expect(view.rows).toHaveCount(0)
+    await expect(pending).toContainText(followUp)
+
+    // The next assistant step still belongs to U1: U2 has been admitted, not delivered.
+    mock.emit("session.step.started", { sessionID, assistantMessageID: assistantID, agent: "build", model })
+    for (const tool of [
+      { id: "tool_queue_read", name: "read", input: { path: "src/queue.ts" } },
+      { id: "tool_queue_grep", name: "grep", input: { pattern: "retry", path: "src" } },
+    ]) {
+      const ref = { sessionID, assistantMessageID: assistantID, id: tool.id }
+      mock.emit("session.tool.input.started", { ...ref, name: tool.name })
+      mock.emit("session.tool.input.ended", { ...ref, text: JSON.stringify(tool.input) })
+      mock.emit("session.tool.called", { ...ref, input: tool.input, executed: true })
+      mock.emit("session.tool.success", {
+        ...ref,
+        content: [{ type: "text", text: "Inspection complete." }],
+        executed: true,
+      })
+    }
+    mock.emit("session.step.ended", {
+      sessionID,
+      assistantMessageID: assistantID,
+      finish: "tool-calls",
+      cost: 0,
+      tokens: { input: 100, output: 20, reasoning: 0, cache: { read: 0, write: 0 } },
+    })
+    const tools = page.locator('[data-timeline-part-ids="tool_queue_read,tool_queue_grep"]')
+    await expect(tools).toBeVisible()
+    await expect(tools).toContainText(/Used\s*Read, Grep/)
+    await expect(tools.locator('[data-component="tag"]')).toHaveText("2")
+    await expect(thinking).toBeVisible()
+    await expect(pending).toBeVisible()
+    expect(mock.rows.map((row) => ({ id: row.id, delivery: row.delivery }))).toEqual([
+      { id: inboxID, delivery: "steer" },
+    ])
+    await transcript.screenshot({ path: testInfo.outputPath("pending-steer.png") })
+
+    // Soft assertions let delivery run too, even when the pending ordering regresses.
+    await expect
+      .soft(tools.or(thinking).or(pending))
+      .toHaveText([/Used\s*Read, Grep/, /Thinking/, /U2: Also check the retry path\./])
+    await expect
+      .soft(transcript.locator('[data-timeline-row="AssistantPart"]').filter({ has: tools }))
+      .toHaveAttribute("data-message-id", userID)
+    await expect
+      .configure({ soft: true })
+      .poll(async () => {
+        const boxes = await Promise.all([tools.boundingBox(), thinking.boundingBox(), pending.boundingBox()])
+        return (
+          boxes.every((box) => box !== null) &&
+          boxes[0]!.y + boxes[0]!.height <= boxes[1]!.y &&
+          boxes[1]!.y + boxes[1]!.height <= boxes[2]!.y
+        )
+      })
+      .toBe(true)
+
+    mock.rows.splice(0, 1)
+    mock.emit("session.inbox.delivered", { sessionID, inboxID })
+    await expect(thinking).toHaveAttribute("data-message-id", inboxID)
+    await expect(pending).toHaveCount(1)
+    await expect(transcript.locator('[data-timeline-row="UserMessage"]')).toHaveCount(2)
+    await expect(transcript.locator('[data-timeline-row="AssistantPart"]').filter({ has: tools })).toHaveAttribute(
+      "data-message-id",
+      userID,
+    )
+
+    const later = { sessionID, assistantMessageID: "msg_queue_follow_up_assistant" }
+    mock.emit("session.step.started", { ...later, agent: "build", model })
+    mock.emit("session.text.started", { ...later, ordinal: 0 })
+    mock.emit("session.text.ended", { ...later, ordinal: 0, text: "A3: Now checking the retry path for U2." })
+    const response = transcript
+      .locator('[data-timeline-row="AssistantPart"]')
+      .filter({ hasText: "A3: Now checking the retry path for U2." })
+    await expect(response).toHaveAttribute("data-message-id", inboxID)
+    await expect(tools.or(pending).or(response).or(thinking)).toHaveText([
+      /Used\s*Read, Grep/,
+      /U2: Also check the retry path\./,
+      /A3: Now checking the retry path for U2\./,
+      /Thinking/,
+    ])
+  })
+}
@@ -1,16 +1,43 @@
 import { expect, test } from "@playwright/test"
-import { assistantMessage, setupTimeline, shell, userMessage } from "../performance/timeline-stability/fixture"
+import {
+  assistantMessage,
+  setupTimeline,
+  shell,
+  textPart,
+  userMessage,
+} from "../performance/timeline-stability/fixture"

 test("space activates a focused timeline button instead of scrolling", async ({ page }) => {
  const shellID = "prt_space_button_shell"
  await setupTimeline(page, {
-    messages: [userMessage(), assistantMessage([shell(shellID, "completed", lines(5))])],
+    messages: [
+      userMessage(),
+      assistantMessage([
+        shell(shellID, "completed", lines(5)),
+        textPart(
+          "prt_space_following",
+          "Following content leaves room to focus the command away from the bottom. ".repeat(40),
+        ),
+      ]),
+    ],
    settings: { shellToolPartsExpanded: false },
    reducedMotion: true,
+    seedHistory: true,
  })
  const scroller = page.locator(".scroll-view__viewport", { has: page.locator("[data-timeline-row]") })
-  const trigger = page.locator(`[data-timeline-part-id="${shellID}"] [data-slot="collapsible-trigger"]`)
+  const trigger = page.getByRole("button", { name: "Used Shell" })
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight))
+    .toBeGreaterThan(300)
+  await trigger.scrollIntoViewIfNeeded()
+  await scroller.hover()
+  await page.mouse.wheel(0, -100)
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeGreaterThan(50)
+  await expect(trigger).toBeInViewport()
  await trigger.focus()
+  await expect(trigger).toBeFocused()
  const before = await scroller.evaluate((element) => element.scrollTop)
  await trigger.press("Space")
  await expect(trigger).toHaveAttribute("aria-expanded", "true")
@@ -40,7 +40,7 @@ test.describe("regression: session timeline context group resize", () => {
    expect(samples.at(-1)?.expanded).toBe("true")
  })

-  test("paints a stable exploring to explored transition", async ({ page }) => {
+  test("keeps a grouped tool summary stable as its calls complete", async ({ page }) => {
    const events: OpenCodeEvent[] = []
    await page.setViewportSize({ width: 1400, height: 900 })
    await mockServer(page, events, [
@@ -55,13 +55,12 @@ test.describe("regression: session timeline context group resize", () => {
    await devtools.send("Emulation.setCPUThrottlingRate", { rate: 4 })
    const context = page.locator(`[data-timeline-part-ids="${contextIDs.join(",")}"]`).first()
    await expectAppVisible(context)
-    await expect(context.locator('[data-component="tool-status-title"]')).toHaveAttribute("aria-label", "Exploring")
+    await expect(context.getByRole("button")).toHaveAccessibleName("Used Read, Glob, Grep, List")

    const contextSelector = `[data-timeline-part-ids="${contextIDs.join(",")}"]`
    const regions = defineVisualRegions({
      status: {
-        selector: `${contextSelector} [data-component="tool-status-title"]`,
-        opacitySelectors: ['[data-slot="tool-status-active"]', '[data-slot="tool-status-done"]'],
+        selector: `${contextSelector} [data-component="context-tool-group-trigger"]`,
      },
      context: { selector: contextSelector, closest: '[data-timeline-row="AssistantPart"]' },
      following: {
@@ -89,7 +88,7 @@ test.describe("regression: session timeline context group resize", () => {
      await page.waitForTimeout(delay)
    }

-    await expect(context.locator('[data-component="tool-status-title"]')).toHaveAttribute("aria-label", "Explored")
+    await expect(context.getByRole("button")).toHaveAccessibleName("Used Read, Glob, Grep, List")
    await page.waitForTimeout(700)
    const trace = await stopVisualProbe<keyof typeof regions>(page)
    const labels = trace.samples
@@ -108,7 +107,7 @@ test.describe("regression: session timeline context group resize", () => {
      ]),
    )

-    expect(labels).toEqual(["Exploring", "Explored"])
+    expect(labels).toEqual(["Used Read, Glob, Grep, List"])
    expect(issues, JSON.stringify(trace.samples, null, 2)).toEqual([])
  })
 })
@@ -209,13 +208,7 @@ function turn(index: number, target: boolean, status: "running" | "completed" =
  const content: SessionMessageAssistant["content"] = target
    ? [
        toolContent(
-          contextTool(
-            contextIDs[0]!,
-            assistantID,
-            "read",
-            { path: "src/recent-a.ts", offset: 0, limit: 120 },
-            status,
-          ),
+          contextTool(contextIDs[0]!, assistantID, "read", { path: "src/recent-a.ts", offset: 0, limit: 120 }, status),
        ),
        toolContent(contextTool(contextIDs[1]!, assistantID, "glob", { path: directory, pattern: "**/*.ts" }, status)),
        toolContent(
@@ -8,21 +8,6 @@ import {
  userText,
 } from "../performance/timeline-stability/fixture"

-test("renders completed write content", async ({ page }) => {
-  const id = "prt_file_projection_write"
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(id, "write", "completed", { path: "src/write.ts", content: "export const written = true\n" }),
-      ]),
-    ],
-    settings: { editToolPartsExpanded: true },
-  })
-
-  await expect(page.locator(`[data-timeline-part-id="${id}"] [data-component="write-content"]`)).toBeVisible()
-})
-
 test("renders a completed single-file patch", async ({ page }) => {
  const id = "prt_file_projection_single_patch"
  await setupTimeline(page, {
@@ -83,6 +68,7 @@ test("keeps an expanded file diff header at the same viewport position", async (
  const before = Array.from({ length: 80 }, (_, index) => `export const value${index} = ${index}\n`).join("")
  const after = before.replaceAll(" = ", " = compute(").replaceAll("\n", ")\n")
  await setupTimeline(page, {
+    settings: { editToolPartsExpanded: true },
    messages: [
      userMessage([userText("Preceding context ".repeat(120))]),
      assistantMessage([
@@ -1,53 +0,0 @@
-import { expect, test } from "@playwright/test"
-import { assistantMessage, setupTimeline, toolPart, userMessage } from "../performance/timeline-stability/fixture"
-import { createTwoFilesPatch } from "diff"
-
-test("keeps patch file disclosures independent", async ({ page }) => {
-  const patchID = "prt_nested_patch"
-  const files = [patchFile("src/a.ts", "modified"), patchFile("src/b.ts", "added"), patchFile("src/old.ts", "deleted")]
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(
-          patchID,
-          "patch",
-          "completed",
-          { patchText: "Update three files" },
-          { metadata: { files } },
-        ),
-      ]),
-    ],
-    settings: { editToolPartsExpanded: true },
-  })
-  const wrapper = page.locator(`[data-timeline-part-id="${patchID}"]`)
-  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 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") {
-  const before = status === "added" ? "" : source(false)
-  const after = status === "deleted" ? "" : source(true)
-  return {
-    file,
-    status,
-    patch: createTwoFilesPatch(`a/${file}`, `b/${file}`, before, after),
-    additions: status === "deleted" ? 0 : 4,
-    deletions: status === "added" ? 0 : 3,
-  }
-}
-
-function source(changed: boolean) {
-  return Array.from({ length: 12 }, (_, index) => `export const value${index} = ${changed ? index + 1 : index}\n`).join(
-    "",
-  )
-}
@@ -26,7 +26,9 @@ for (const expanded of [false, true]) {
      messages: [userMessage(), assistantMessage([shell(id, "completed", lines(3))])],
      settings: { shellToolPartsExpanded: expanded },
    })
-    const trigger = page.locator(`[data-timeline-part-id="${id}"] [data-slot="collapsible-trigger"]`)
+    const trigger = expanded
+      ? page.locator(`[data-timeline-part-id="${id}"] [data-slot="collapsible-trigger"]`)
+      : page.getByRole("button", { name: "Used Shell" })
    await expect(trigger).toHaveAttribute("aria-expanded", String(expanded))
    await trigger.click()
    await expect(trigger).toHaveAttribute("aria-expanded", String(!expanded))
@@ -120,6 +122,7 @@ test("transitions thinking and hidden reasoning through busy to idle", async ({
  await expect(page.locator('[data-timeline-row="Thinking"]')).toBeVisible()
  await expect(page.getByText("Inspecting stability", { exact: true })).toBeVisible()
  await expect(page.locator(`[data-timeline-part-id="${reasoningID}"]`)).toHaveCount(0)
+  await expect(page.locator(`[data-timeline-part-id="${renderedPartID(reasoningID)}"]`)).toHaveCount(0)
  await timeline.send(partUpdated(shell("prt_reasoning_shell", "running")), 160)
  await expect(page.locator('[data-timeline-row="Thinking"]')).toBeVisible()
  await timeline.send(partUpdated(shell("prt_reasoning_shell", "completed", "done")), 180)
@@ -127,6 +130,7 @@ test("transitions thinking and hidden reasoning through busy to idle", async ({
  await timeline.send(status("idle"), 300)
  await expect(page.locator('[data-timeline-row="Thinking"]')).toHaveCount(0)
  await expect(page.locator(`[data-timeline-part-id="${reasoningID}"]`)).toHaveCount(0)
+  await expect(page.locator(`[data-timeline-part-id="${renderedPartID(reasoningID)}"]`)).toHaveCount(0)
 })

 test("moves busy through retry and recovery to final idle content", async ({ page }) => {
@@ -1,12 +1,9 @@
 import { expect, test } from "@playwright/test"
 import { assistantMessage, setupTimeline, toolPart, userMessage } from "../performance/timeline-stability/fixture"

-for (const profile of [
-  { locale: "de", label: "Erkundung abgeschlossen" },
-  { locale: "ar", label: "تم الاستكشاف" },
-] as const) {
-  test(`projects translated context status in ${profile.locale}`, async ({ page }) => {
-    const ids = [`prt_locale_${profile.locale}_01_read`, `prt_locale_${profile.locale}_02_glob`]
+for (const locale of ["de", "ar"] as const) {
+  test(`projects localized tool names with an English fallback in ${locale}`, async ({ page }) => {
+    const ids = [`prt_locale_${locale}_01_read`, `prt_locale_${locale}_02_glob`]
    await setupTimeline(page, {
      messages: [
        userMessage(),
@@ -15,11 +12,12 @@ for (const profile of [
          toolPart(ids[1]!, "glob", "completed", { path: ".", pattern: "**/*.ts" }),
        ]),
      ],
-      locale: profile.locale,
+      locale,
    })

    const group = page.locator(`[data-timeline-part-ids="${ids.join(",")}"]`)
-    await expect(group.locator('[data-component="tool-status-title"]')).toHaveAttribute("aria-label", profile.label)
-    await expect(page.locator("html")).toHaveAttribute("lang", profile.locale)
+    await expect(group.getByRole("button")).toHaveAccessibleName(/^Used /)
+    await expect(group.locator('[data-component="tag"]')).toHaveText("2")
+    await expect(page.locator("html")).toHaveAttribute("lang", locale)
  })
 }
@@ -122,6 +122,7 @@ test("updates running compactions to failed and cancelled boundaries", async ({

  await timeline.send(compactionStarted({ sessionID, reason: "auto", recent: "" }))
  await timeline.send(compactionDelta({ sessionID, text: "Partial summary that should be discarded." }))
+  await expect(page.getByText("Partial summary that should be discarded.", { exact: true })).toBeVisible()
  await timeline.send(
    compactionFailed({
      sessionID,
@@ -140,6 +141,9 @@ test("updates running compactions to failed and cancelled boundaries", async ({
  await expect(failed).not.toContainText("Partial summary that should be discarded.")

  await timeline.send(compactionStarted({ sessionID, reason: "manual", recent: "" }))
+  await expect(compactions).toHaveCount(2)
+  await timeline.send(compactionDelta({ sessionID, text: "Summary before cancellation." }))
+  await expect(page.getByText("Summary before cancellation.", { exact: true })).toBeVisible()
  await timeline.send(
    compactionFailed({
      sessionID,
@@ -152,88 +156,7 @@ test("updates running compactions to failed and cancelled boundaries", async ({
  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()
+  await expect(cancelled).not.toContainText("Summary before cancellation.")
 })

 test("moves blocking work to the background with Ctrl+B", async ({ page }) => {
@@ -271,11 +194,6 @@ test("moves blocking work to the background with Ctrl+B", async ({ page }) => {
  await request
 })

-test("waits for completion before labeling requested background work", async ({ page }) => {
-  await setupTimeline(page, { sessionMessages: [user, assistant(false, true, undefined, true)] })
-  await expect(page.locator('[data-component="task-tool-card"]')).not.toContainText("(background)")
-})
-
 test("navigates from a running subagent card and hides background controls in the child", async ({ page }) => {
  const childID = "ses_running_child"
  await setupTimeline(page, {
@@ -7,86 +7,14 @@ import {
  toolPart,
  userMessage,
  userText,
-  type PartSeed,
 } from "../performance/timeline-stability/fixture"

 test.describe("session timeline projection", () => {
-  test("renders every admitted tool family and hides timeline-only exclusions", async ({ page }) => {
-    const parts = [
-      toolPart("prt_01_read", "read", "completed", { path: "src/a.ts" }),
-      toolPart("prt_02_glob", "glob", "completed", { path: ".", pattern: "**/*.ts" }),
-      toolPart("prt_03_grep", "grep", "completed", { path: ".", pattern: "value" }),
-      toolPart("prt_04_list", "list", "completed", { path: "src" }),
-      toolPart("prt_webfetch", "webfetch", "completed", { url: "https://example.com" }),
-      toolPart(
-        "prt_websearch",
-        "websearch",
-        "completed",
-        { query: "timeline stability" },
-        { output: "https://example.com/result" },
-      ),
-      toolPart("prt_task", "subagent", "completed", {
-        description: "Inspect timeline",
-        agent: "explore",
-        prompt: "Inspect the timeline implementation.",
-      }),
-      toolPart(
-        "prt_bash",
-        "shell",
-        "completed",
-        { command: "printf stable" },
-        { output: "stable", title: "printf stable" },
-      ),
-      editPart("prt_edit"),
-      toolPart("prt_write", "write", "completed", { path: "src/new.ts", content: "export const stable = true\n" }),
-      patchPart("prt_patch"),
-      toolPart("prt_todo", "todowrite", "completed", { todos: [{ content: "Hidden", status: "pending" }] }),
-      toolPart(
-        "prt_question",
-        "question",
-        "completed",
-        { questions: [{ question: "Keep stable?", header: "Stability", options: [] }] },
-        { metadata: { answers: [["Yes"]] } },
-      ),
-      toolPart("prt_skill", "skill", "completed", { name: "stability" }),
-      toolPart("prt_custom", "custom_mcp_tool", "completed", { target: "timeline", count: 2 }),
-    ]
-    await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
-
-    await expect(
-      page.locator('[data-timeline-part-ids="prt_01_read,prt_02_glob,prt_03_grep,prt_04_list"]'),
-    ).toBeVisible()
-    for (const id of [
-      "prt_webfetch",
-      "prt_websearch",
-      "prt_task",
-      "prt_bash",
-      "prt_edit",
-      "prt_write",
-      "prt_patch",
-      "prt_question",
-      "prt_skill",
-      "prt_custom",
-    ]) {
-      await expect(page.locator(`[data-timeline-part-id="${id}"]`).first(), id).toBeVisible()
-    }
-    const patch = page.locator('[data-timeline-part-id="prt_patch"]')
-    await expect(patch.getByText("1 file", { exact: true })).toBeVisible()
-    await expect(patch.getByRole("button", { name: "Patch 1 file", exact: true })).toHaveCount(0)
-    await expect(patch.getByRole("button")).toHaveCount(1)
-    await expect(patch.locator('[data-scope="apply-patch"] button[aria-expanded="false"]')).toHaveCount(1)
-    await expect(patch.locator('[data-slot="message-part-title-filename"]')).toHaveCount(0)
-    await expect(patch.locator('[data-slot="message-part-actions"]')).toHaveCount(0)
-    const edit = page.locator('[data-timeline-part-id="prt_edit"]')
-    await expect(edit.locator('[data-component="apply-patch-tool"]')).toBeVisible()
-    await expect(edit.locator('[data-slot="basic-tool-tool-title"]')).toContainText("Edit")
-    await expect(page.locator('[data-timeline-part-id="prt_todo"]')).toHaveCount(0)
-  })
-
  test("combines adjacent patch calls and repeated files into one group", async ({ page }) => {
    const first = "prt_patch_first"
    const second = "prt_patch_second"
    const timeline = await setupTimeline(page, {
+      settings: { editToolPartsExpanded: true },
      messages: [
        userMessage(),
        assistantMessage([
@@ -153,43 +81,6 @@ test.describe("session timeline projection", () => {
    await expect(page.locator(`[data-timeline-part-id="${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(
      [
@@ -231,25 +122,6 @@ test.describe("session timeline projection", () => {
    await expect(page.locator('[data-timeline-row="TurnGap"]')).toBeVisible()
  })

-  test("renders interruption independently when the turn is not compacted", async ({ page }) => {
-    const user = userMessage()
-    const before = assistantMessage([{ id: "prt_before", type: "text", text: "Before" }], {
-      id: "msg_1001_before",
-      error: { type: "MessageAbortedError", message: "Stopped" },
-    })
-    const after = assistantMessage([{ id: "prt_after", type: "text", text: "After" }], {
-      id: "msg_1002_after",
-      created: 1700000003000,
-    })
-    await setupTimeline(page, { messages: [user, before, after] })
-
-    await expect(page.getByText("Interrupted", { exact: true })).toBeVisible()
-    const rows = await page
-      .locator('[data-timeline-row="AssistantPart"], [data-timeline-row="TurnDivider"]')
-      .evaluateAll((elements) => elements.map((element) => element.getAttribute("data-timeline-row")))
-    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"
@@ -286,77 +158,8 @@ test.describe("session timeline projection", () => {
    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">[] = [
-      userText(text, { id: "prt_user_rich" }),
-      {
-        id: "prt_user_image",
-        type: "file",
-        mime: "image/png",
-        filename: "pixel.png",
-        url: "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg==",
-      },
-      {
-        id: "prt_user_attachment",
-        type: "file",
-        mime: "application/json",
-        filename: "tsconfig.json",
-        url: "data:application/json;base64,e30=",
-      },
-      {
-        id: "prt_user_reference",
-        type: "file",
-        mime: "text/plain",
-        filename: "a.ts",
-        url: "src/a.ts",
-        source: { type: "file", path: "src/a.ts", text: { value: "@src/a.ts", start: 18, end: 27 } },
-      },
-      {
-        id: "prt_user_agent",
-        type: "agent",
-        name: "explore",
-        source: { value: "@explore", start: 4, end: 12 },
-      },
-    ]
-    await setupTimeline(page, { messages: [userMessage(parts), assistantMessage()] })
-
-    await expect(page.getByAltText("pixel.png")).toBeVisible()
-    await expect(page.getByText("tsconfig.json")).toBeVisible()
-    await expect(page.getByText("@src/a.ts", { exact: true })).toBeVisible()
-    await expect(page.getByText("@explore", { exact: true })).toBeVisible()
-  })
 })

-function editPart(id: string) {
-  return toolPart(
-    id,
-    "edit",
-    "completed",
-    { path: "src/a.ts", oldString: "export const value = 1", newString: "export const value = 2" },
-    {
-      metadata: {
-        files: [patchFile("src/a.ts", "modified")],
-      },
-    },
-  )
-}
-
-function patchPart(id: string) {
-  return toolPart(
-    id,
-    "patch",
-    "completed",
-    { patchText: "Update the projected files" },
-    {
-      metadata: {
-        files: [patchFile("src/a.ts", "modified")],
-      },
-    },
-  )
-}
-
 function patchFile(file: string, status: "added" | "modified" | "deleted") {
  return {
    file,
@@ -7,29 +7,12 @@ import {
  renderedPartID,
  setupTimeline,
  shell,
+  toolPart,
  status,
  textPart,
-  toolPart,
  userMessage,
 } from "../performance/timeline-stability/fixture"

-test("groups singleton and separated context operations at correct boundaries", async ({ page }) => {
-  const parts = [
-    toolPart("prt_boundary_01_read", "read", "completed", { path: "src/a.ts" }),
-    textPart("prt_boundary_02_text", "Boundary text"),
-    toolPart("prt_boundary_03_glob", "glob", "completed", { path: ".", pattern: "**/*.ts" }),
-    toolPart("prt_boundary_04_grep", "grep", "completed", { path: ".", pattern: "stable" }),
-    shell("prt_boundary_05_shell", "completed", "done"),
-    toolPart("prt_boundary_06_list", "list", "completed", { path: "src" }),
-  ]
-  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
-
-  await expect(page.locator('[data-timeline-part-ids="prt_boundary_01_read"]')).toBeVisible()
-  await expect(page.locator('[data-timeline-part-ids="prt_boundary_03_glob,prt_boundary_04_grep"]')).toBeVisible()
-  await expect(page.locator('[data-timeline-part-ids="prt_boundary_06_list"]')).toBeVisible()
-  await expect(page.locator('[data-timeline-row="AssistantPart"]')).toHaveCount(5)
-})
-
 test("reducer-hardening: converges when idle arrives before final part and message completion", async ({ page }) => {
  const textID = "prt_event_order_text"
  const assistant = assistantMessage([textPart(textID, "Partial")], { completed: false })
@@ -44,3 +27,100 @@ test("reducer-hardening: converges when idle arrives before final part and messa
    "Final after early idle",
  )
 })
+
+test("expands a mixed collapsed tool stack without expanding its individual calls", async ({ page }) => {
+  const parts = [
+    shell("prt_stack_shell_1", "completed", "first"),
+    toolPart("prt_stack_explore", "subagent", "completed", {
+      agent: "explore",
+      description: "Inspect the project",
+      prompt: "Explore the project",
+    }),
+    toolPart("prt_stack_patch", "patch", "completed", { patchText: "Update src/value.ts" }),
+    shell("prt_stack_shell_2", "completed", "second"),
+  ]
+  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
+
+  const group = page.locator(
+    '[data-timeline-part-ids="prt_stack_shell_1,prt_stack_explore,prt_stack_patch,prt_stack_shell_2"]',
+  )
+  const summary = group.getByRole("button", { name: "Used Shell, Explore, Patch" })
+  await expect(summary).toHaveAttribute("aria-expanded", "false")
+  await expect(summary).toHaveCSS("height", "28px")
+  await expect(group.locator('[data-component="tag"]')).toHaveText("4")
+  await summary.click()
+  await expect(summary).toHaveAttribute("aria-expanded", "true")
+  await expect(group.locator('[data-slot="context-tool-group-item"]')).toHaveCount(4)
+  await expect(group.locator('[data-timeline-part-id="prt_stack_shell_1"]')).toBeVisible()
+  await expect(group.locator('[data-timeline-part-id="prt_stack_patch"]')).toBeVisible()
+  await expect(group.locator('[data-component="context-tool-group-list"]')).toHaveCSS("row-gap", "8px")
+  const content = group.locator(':scope > [data-component="collapsible"] > [data-slot="collapsible-content"]')
+  await expect(content).toHaveCSS("margin-left", "0px")
+  await expect(content).toHaveCSS("padding-left", "12px")
+  await expect.poll(() => content.evaluate((element) => getComputedStyle(element, "::before").content)).toBe("none")
+})
+
+test("leaves tools expanded by settings outside the collapsed stack", async ({ page }) => {
+  const parts = [
+    shell("prt_expanded_shell", "completed", "expanded"),
+    toolPart("prt_collapsed_patch", "patch", "completed", { patchText: "Update src/value.ts" }),
+    toolPart("prt_collapsed_read", "read", "completed", { path: "src/value.ts" }),
+  ]
+  await setupTimeline(page, {
+    messages: [userMessage(), assistantMessage(parts)],
+    settings: { shellToolPartsExpanded: true },
+  })
+
+  await expect(page.locator('[data-timeline-part-id="prt_expanded_shell"]')).toBeVisible()
+  const group = page.locator('[data-timeline-part-ids="prt_collapsed_patch,prt_collapsed_read"]')
+  await expect(group.getByRole("button", { name: "Used Patch, Read" })).toBeVisible()
+  await expect(group.locator('[data-component="tag"]')).toHaveText("2")
+  await expect(page.locator('[data-timeline-spacing="tool"]')).toHaveCSS("padding-top", "8px")
+})
+
+test("keeps failed search calls and their error cards inside the collapsed stack", async ({ page }) => {
+  const parts = [
+    toolPart(
+      "prt_error_glob",
+      "glob",
+      "error",
+      { path: "C:/Users", pattern: "*.ts" },
+      {
+        error: "Invalid tool input",
+      },
+    ),
+    toolPart(
+      "prt_error_grep",
+      "grep",
+      "error",
+      { path: "C:/Users", pattern: "value" },
+      {
+        error: "Search timed out after 30 seconds",
+      },
+    ),
+  ]
+  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
+
+  const group = page.locator('[data-timeline-part-ids="prt_error_glob,prt_error_grep"]')
+  const summary = group.getByRole("button", { name: "Used Glob, Grep" })
+  await expect(group.locator('[data-component="tag"]')).toHaveText("2")
+  await summary.click()
+  await expect(group.locator('[data-kind="tool-error-card"]')).toHaveCount(2)
+  const glob = group.locator('[data-timeline-part-id="prt_error_glob"]')
+  await expect(glob).toContainText("Invalid tool input")
+  await expect(glob.locator('[data-component="tool-error-card-icon"]')).toBeVisible()
+  await expect(glob.locator('[data-component="tool-error-card-icon"] use')).toHaveAttribute(
+    "href",
+    "#opencode-v2-icon-circle-exclamation",
+  )
+  await expect
+    .poll(() =>
+      glob
+        .locator('[data-kind="tool-error-card"]')
+        .evaluate((element) => getComputedStyle(element, "::before").display),
+    )
+    .toBe("none")
+  await expect(group.locator('[data-timeline-part-id="prt_error_grep"]')).toContainText(
+    "Search timed out after 30 seconds",
+  )
+})
@@ -7,32 +7,11 @@ import {
  userMessage,
 } from "../performance/timeline-stability/fixture"

-test("renders every tool error outcome without leaking hidden tools", async ({ page }) => {
-  const ordinary = ["shell", "edit", "write", "patch", "webfetch", "websearch", "subagent", "skill", "mcp_probe"]
-  const parts = ordinary.map((tool, index) =>
-    toolPart(`prt_error_${index}`, tool, "error", errorInput(tool), { error: `${tool} failed visibly` }),
-  )
-  parts.push(
-    toolPart("prt_question_dismissed", "question", "error", questionInput(), {
-      error: "The user dismissed this question",
-    }),
-    toolPart("prt_question_error", "question", "error", questionInput(), { error: "Question transport failed" }),
-    toolPart("prt_todo_error", "todowrite", "error", { todos: [] }, { error: "Hidden todo failure" }),
-  )
-  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
-
-  await expect(page.locator('[data-kind="tool-error-card"]')).toHaveCount(ordinary.length + 1)
-  await expect(page.getByText(/dismissed/i)).toBeVisible()
-  await expect(page.locator('[data-timeline-part-id="prt_todo_error"]')).toHaveCount(0)
-  for (let index = 0; index < ordinary.length; index++) {
-    await expect(page.locator(`[data-timeline-part-id="prt_error_${index}"]`)).toBeVisible()
-  }
-})
-
 test("transitions shell and question through running error outcomes", async ({ page }) => {
  const shellID = "prt_transition_error_shell"
  const questionID = "prt_transition_error_question"
  const timeline = await setupTimeline(page, {
+    settings: { shellToolPartsExpanded: true },
    messages: [
      userMessage(),
      assistantMessage(
@@ -44,7 +23,6 @@ test("transitions shell and question through running error outcomes", async ({ p
      ),
    ],
  })
-  await timeline.waitForPart(shellID)
  await expect(page.locator(`[data-timeline-part-id="${questionID}"]`)).toHaveCount(0)
  await timeline.send(partUpdated(toolPart(shellID, "shell", "running", { command: "exit 1" })), 120)
  await timeline.send(partUpdated(toolPart(questionID, "question", "running", questionInput())), 180)
@@ -68,6 +46,7 @@ test("preserves surviving grouped patch state when its first patch fails", async
  const failed = "prt_grouped_patch_failed"
  const surviving = "prt_grouped_patch_surviving"
  const timeline = await setupTimeline(page, {
+    settings: { editToolPartsExpanded: true },
    messages: [
      userMessage(),
      assistantMessage(
@@ -134,98 +113,52 @@ test("preserves surviving grouped patch state when its first patch fails", async
    .toBeGreaterThanOrEqual(-0.5)
 })

-test("labels all web search provider variants", async ({ page }) => {
+test("groups instruction files loaded by the same read", async ({ page }) => {
+  const id = "prt_read_instructions"
+  await setupTimeline(page, {
+    messages: [
+      userMessage(),
+      assistantMessage([
+        toolPart(
+          id,
+          "read",
+          "completed",
+          { path: "src/a.ts" },
+          { metadata: { loaded: ["AGENTS.md", "packages/app/AGENTS.md", "packages/ui/AGENTS.md"] } },
+        ),
+      ]),
+    ],
+  })
+
+  const tool = page.locator(`[data-timeline-part-id="${id}"]`)
+  const loaded = tool.locator('[data-component="tool-loaded-item"]')
+  await expect(loaded).toHaveCount(1)
+  await expect(loaded).toHaveAttribute("aria-label", "Loaded AGENTS.md, packages/app/AGENTS.md, packages/ui/AGENTS.md")
+  await expect(loaded.locator('[data-slot="tool-loaded-value"]')).toHaveText(
+    "AGENTS.md, packages/app/AGENTS.md, packages/ui/AGENTS.md",
+  )
+  await expect(loaded.locator('[data-slot="tool-loaded-kind"]')).toHaveCount(0)
+})
+
+test("groups only consecutive successful skill tools", async ({ page }) => {
  const parts = [
-    toolPart(
-      "prt_search_parallel",
-      "websearch",
-      "completed",
-      { query: "parallel" },
-      { metadata: { provider: "parallel" } },
-    ),
-    toolPart("prt_search_exa", "websearch", "completed", { query: "exa" }, { metadata: { provider: "exa" } }),
-    toolPart("prt_search_generic", "websearch", "completed", { query: "generic" }),
+    toolPart("prt_skill_first", "skill", "completed", { id: "ocpr" }),
+    toolPart("prt_skill_second", "skill", "completed", { id: "effect" }),
+    toolPart("prt_skill_third", "skill", "completed", { id: "ui-pr-screenshots" }),
+    toolPart("prt_skill_break", "read", "completed", { path: "src/a.ts" }),
+    toolPart("prt_skill_last", "skill", "completed", { id: "opencode" }),
  ]
  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })

-  await expect(page.getByRole("button", { name: /Parallel Web Search/ })).toBeVisible()
-  await expect(page.getByRole("button", { name: /Exa Web Search/ })).toBeVisible()
-  await expect(page.getByRole("button", { name: /^Web Search/ })).toBeVisible()
-})
+  const group = page.locator(`[data-timeline-part-ids="${parts.map((part) => part.id).join(",")}"]`)
+  await group.getByRole("button").click()

-test("labels completed searches with result counts", async ({ page }) => {
-  const glob = "prt_glob_count"
-  const grep = "prt_grep_count"
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(glob, "glob", "completed", { path: ".", pattern: "**/*.ts" }, { metadata: { count: 1 } }),
-        toolPart(grep, "grep", "completed", { path: ".", pattern: "value" }, { metadata: { matches: 12 } }),
-      ]),
-    ],
-  })
-
-  const group = page.locator(`[data-timeline-part-ids="${glob},${grep}"]`)
-  await group.locator('[data-slot="collapsible-trigger"]').click()
-  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 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" })])],
-  })
-
-  const group = page.locator(`[data-timeline-part-ids="${id}"]`)
-  await group.locator('[data-slot="collapsible-trigger"]').click()
-  await expect(
-    group
-      .locator('[data-component="context-tool-group-list"] [data-component="tool-trigger"]')
-      .filter({ hasText: "Read" }),
-  ).toContainText("a.ts")
-})
-
-test("labels skill tools from IDs and result metadata", async ({ page }) => {
-  const pending = "prt_skill_id"
-  const completed = "prt_skill_name"
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(pending, "skill", "running", { id: "frontend-design" }),
-        toolPart(completed, "skill", "completed", { id: "opencode" }, { metadata: { name: "OpenCode" } }),
-      ]),
-    ],
-  })
-
-  for (const [id, name] of [
-    [pending, "frontend-design"],
-    [completed, "OpenCode"],
-  ] as const) {
-    const skill = page.locator(`[data-timeline-part-id="${id}"]`)
-    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)
-  }
+  const loaded = group.locator('[data-component="tool-loaded-item"]')
+  await expect(loaded).toHaveCount(2)
+  await expect(loaded.nth(0)).toHaveAttribute("aria-label", "Loaded ocpr, effect, ui-pr-screenshots skills")
+  await expect(loaded.nth(1)).toHaveAttribute("aria-label", "Loaded opencode skill")
 })

 function questionInput() {
  return { questions: [{ header: "Stability", question: "Keep it stable?", options: [] }] }
 }
-
-function errorInput(tool: string) {
-  if (tool === "shell") return { command: "exit 1" }
-  if (["edit", "write"].includes(tool)) return { path: "src/error.ts", content: "" }
-  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 === "skill") return { name: "failure" }
-  return { target: "failure" }
-}
@@ -0,0 +1,121 @@
+import { expect, test } from "@playwright/test"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+const directory = "C:/Projects/settings-demo"
+const sandboxes = Array.from({ length: 12 }, (_, index) => `${directory}/workspace-${index + 1}`)
+
+test.use({ viewport: { width: 1440, height: 1000 }, colorScheme: "dark" })
+
+test.beforeEach(async ({ page }) => {
+  await mockOpenCodeServer(page, {
+    directory,
+    project: {
+      id: "proj_settings_demo",
+      canonical: directory,
+      name: "Settings demo",
+      vcs: "git",
+      time: { created: 1700000000000, updated: 1700000000000 },
+      sandboxes,
+    },
+    provider: { all: [], connected: [], default: {} },
+    sessions: sandboxes.map((directory, index) => ({
+      id: `ses_settings_${index + 1}`,
+      title: `Workspace ${index + 1} session`,
+      directory,
+      projectID: "proj_settings_demo",
+      time: { created: 1700000000000, updated: 1700000000000 },
+    })),
+    pageMessages: () => ({ items: [] }),
+  })
+  await page.goto("/")
+  await page.getByRole("button", { name: "Settings", exact: true }).click()
+  await expect(page.getByTestId("settings-screen").getByRole("tab", { name: "Preferences" })).toBeVisible()
+})
+
+test("workspaces opens without waiting for inventory or sessions", async ({ page }) => {
+  const inventory = Promise.withResolvers<void>()
+  const sessions = Promise.withResolvers<void>()
+  await page.route("**/api/worktree/*", async (route) => {
+    await inventory.promise
+    await route.fallback()
+  })
+  await page.route("**/api/session?*", async (route) => {
+    if (new URL(route.request().url()).searchParams.has("directory")) await sessions.promise
+    await route.fallback()
+  })
+  const settings = page.getByTestId("settings-screen")
+  const requested = page.waitForRequest((request) => new URL(request.url()).pathname.startsWith("/api/worktree/"))
+  await settings.getByRole("tab", { name: "Workspaces", exact: true }).click()
+  await requested
+  await expect(settings.getByRole("heading", { name: "Workspaces", exact: true })).toBeVisible()
+  await expect(settings.getByRole("button", { name: "Back to app" })).toBeVisible()
+  await expect(settings.getByText("No workspaces", { exact: true })).toHaveCount(0)
+
+  inventory.resolve()
+  await expect(settings.getByText(sandboxes[0], { exact: true })).toBeVisible()
+  await expect(settings.getByText("12 workspaces", { exact: true })).toBeVisible()
+  sessions.resolve()
+  await expect(settings.getByText("Workspace 1 session", { exact: true })).toBeVisible()
+
+  const refresh = Promise.withResolvers<void>()
+  await page.route("**/api/worktree/*", async (route) => {
+    await refresh.promise
+    await route.fallback()
+  })
+  await settings.getByRole("tab", { name: "Preferences", exact: true }).click()
+  await settings.getByRole("tab", { name: "Workspaces", exact: true }).click()
+  await expect(settings.getByText("Workspace 1 session", { exact: true })).toBeVisible()
+  refresh.resolve()
+})
+
+test("extensions opens without waiting for MCPs or plugins", async ({ page }) => {
+  const mcps = Promise.withResolvers<void>()
+  const plugins = Promise.withResolvers<void>()
+  await page.route("**/api/mcp", async (route) => {
+    await mcps.promise
+    await route.fulfill({
+      json: { location: { directory }, data: [{ name: "demo-mcp", status: { status: "connected" } }] },
+    })
+  })
+  await page.route("**/api/plugin", async (route) => {
+    await plugins.promise
+    await route.fulfill({
+      json: {
+        location: { directory },
+        data: [
+          { id: "demo-plugin", source: { type: "package", package: "demo-plugin" }, status: "active", tui: false },
+        ],
+      },
+    })
+  })
+  const settings = page.getByTestId("settings-screen")
+  const requested = page.waitForRequest((request) => new URL(request.url()).pathname === "/api/mcp")
+  await settings.getByRole("tab", { name: "Extensions", exact: true }).click()
+  await requested
+  await expect(settings.getByRole("heading", { name: "Extensions", exact: true })).toBeVisible()
+  await expect(settings.getByRole("button", { name: "Back to app" })).toBeVisible()
+  await settings.getByRole("tab", { name: "Plugins", exact: true }).click()
+  await expect(settings.getByRole("tab", { name: "Plugins", exact: true })).toHaveAttribute("aria-selected", "true")
+  plugins.resolve()
+  await expect(settings.getByText("demo-plugin", { exact: true })).toBeVisible()
+  mcps.resolve()
+  await settings.getByRole("tab", { name: "MCPs", exact: true }).click()
+  await expect(settings.getByRole("switch", { name: "demo-mcp" })).toBeChecked()
+})
+
+test("workspace inventory uses the settings panel scroll area", async ({ page }) => {
+  const settings = page.getByTestId("settings-screen")
+  await settings.getByRole("tab", { name: "Workspaces", exact: true }).click()
+  await expect(settings.getByText("Workspace 1 session", { exact: true })).toBeVisible()
+  const list = settings.locator('[data-component="settings-list"]')
+  await expect(list).toHaveCSS("max-height", "none")
+  await expect(list).toHaveCSS("overflow-y", "visible")
+  await settings.getByText("Workspace 12 session", { exact: true }).scrollIntoViewIfNeeded()
+  await expect(settings.getByText("Workspace 12 session", { exact: true })).toBeInViewport()
+  await expect(settings.getByRole("button", { name: "Back to app" })).toBeInViewport()
+  await page.setViewportSize({ width: 390, height: 844 })
+  await expect(list).toHaveCSS("max-height", "none")
+  await expect(list).toHaveCSS("overflow-y", "visible")
+  await settings.getByText("Workspace 12 session", { exact: true }).scrollIntoViewIfNeeded()
+  await expect(settings.getByText("Workspace 12 session", { exact: true })).toBeInViewport()
+})
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
LukeParkerDev	8d419ac7af	feat(browser): add experimental desktop browser	2026-08-27 12:48:12 +10:00
Luke Parker	48d4e52143	fix(app): keep pending steers after assistant work (#45435 )	2026-08-27 02:44:18 +00:00
Aiden Cline	40cbea3c19	refactor(core): use shared state for tool registry (#45414 )	2026-08-26 21:33:00 -05:00
Luke Parker	51065122d8	fix(app): keep project extensions inside settings (#45432 )	2026-08-27 02:30:36 +00:00
Luke Parker	7507f19a00	fix(app): prevent settings loading flicker (#45427 )	2026-08-27 02:01:36 +00:00
opencode-agent[bot]	71706577c4	chore: update nix node_modules hashes	2026-08-27 01:53:01 +00:00
Kit Langton	66f326ddec	fix(tui): restore default Kitty keyboard reporting (#45423 )	2026-08-26 21:41:19 -04:00
opencode-agent[bot]andBrendonovich	83b81b39f1	fix(app): remove worktree user message accent (#45250 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-27 09:36:35 +08:00
Brendan Allan	b20e4971db	test(app): run CI e2e against production builds (#45422 )	2026-08-27 09:28:23 +08:00
James Long	c3378f2d34	fix(pty): upgrade opencode-pty to 0.1.12 (#45420 )	2026-08-26 21:15:09 -04:00
Kit Langton	d68031bc0d	fix(core): preserve renamed files when undoing changes (#45383 )	2026-08-26 20:42:40 -04:00
Kit Langton	4ac19d744a	feat(tui): make session tab status explicit (#45347 )	2026-08-26 20:12:15 -04:00
Aiden Cline	b5cb5e2a55	refactor(plugin): manage MCP servers through transforms (#45408 )	2026-08-26 18:42:28 -05:00
Luke Parker	53d171781a	fix(app): reduce catalog retention and status animation work (#45406 )	2026-08-26 23:30:49 +00:00
opencode-agent[bot]andthdxr	3184427c03	feat(plugin): expose the current location in plugin context (#45403 ) Co-authored-by: thdxr <826656+thdxr@users.noreply.github.com>	2026-08-26 19:09:16 -04:00
James Long	6b0613ac6f	fix(tui): restore session focus during permission prompts (#45391 )	2026-08-26 19:02:56 -04:00
Aiden Cline	5ac4156eb2	fix(core): preserve reasoning across model switches (#45401 )	2026-08-26 17:52:41 -05:00
Luke Parker	954cdc7bc8	fix(desktop): restore compact Windows titlebar spacing (#45397 )	2026-08-26 22:18:05 +00:00
Aiden Cline	767115cd69	fix(ai): default native Mantle to Responses (#45394 )	2026-08-26 16:47:39 -05:00
James Long	5b25ee8430	style(tui): brighten embedded terminal palette (#45393 )	2026-08-26 17:30:23 -04:00
James Long	df6317d23c	refactor(tui): simplify terminal selection state (#45387 )	2026-08-26 17:18:44 -04:00
opencode-agent[bot]	524e39db02	chore: update nix node_modules hashes	2026-08-26 20:58:41 +00:00
James Long	6170221e21	fix(pty): upgrade opencode-pty to 0.1.11 (#45380 )	2026-08-26 16:39:24 -04:00
Kit Langton	74a0b2d0ad	feat(core): make the portable shell scanner authoritative	2026-08-26 16:35:16 -04:00
opencode-agent[bot]	9d46941e70	chore: update nix node_modules hashes	2026-08-26 20:31:57 +00:00
Aiden Cline	d48a5d16ff	fix(core): avoid nested MCP code mode (#45377 )	2026-08-26 15:23:14 -05:00
Kit Langton	6c392355dd	feat(tui): render LaTeX as Unicode math (#45339 )	2026-08-26 16:14:37 -04:00
Aiden Cline	dd4ee94682	refactor(ai): remove derived HTTP request ID (#45376 )	2026-08-26 15:12:53 -05:00
Aiden Cline	009eca8b5b	test(core): shrink webfetch stress fixtures (#45375 )	2026-08-26 15:03:49 -05:00
Aiden Cline	94bd9f6c8b	fix(ai): isolate native provider metadata ownership (#45155 )	2026-08-26 14:54:12 -05:00
Aiden Cline	33909f48d7	fix(core): label migrated credentials by auth type (#45369 )	2026-08-26 14:41:31 -05:00
James Long	80653a0a1a	fix(tui): preserve resolved server directory (#45354 )	2026-08-26 15:40:00 -04:00
James Long	f7913a04d2	fix(cli): stop PTY daemon on explicit service restart (#45373 )	2026-08-26 15:39:29 -04:00
Kit Langton	cf347cd5e4	refactor(core): advance sessions before running steps (#45358 ) Centralize control dispatch and first-Step preparation in advanceToStep. Keep input delivery outside logical-Step retries and preserve queue ordering, Location handoff, context refresh, and durable settlement.	2026-08-26 15:25:13 -04:00
opencode-agent[bot]	6600d59635	chore: update nix node_modules hashes	2026-08-26 19:07:20 +00:00
Aiden Cline	8b6a2450d5	fix(core): isolate invalid tool registrations (#45325 )	2026-08-26 13:49:56 -05:00
opencode-agent[bot]andBrendonovich	1aa4046f02	test(app): isolate component coverage with storybook (#45142 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-27 02:42:27 +08:00
opencode-agent[bot]	9bd69fe847	chore: update nix node_modules hashes	2026-08-26 18:32:55 +00:00
Kit Langton	ded8a492d1	fix(core): recover background jobs after restart Persist background Job ownership and terminal results across server restarts. Resume existing subagent Sessions, admit shell cancellation notices without waking idle parents, and preserve explicit cancellation.	2026-08-26 14:17:43 -04:00
James Long	018b4c40f3	fix(pty): upgrade opencode-pty to 0.1.10 (#45352 )	2026-08-26 14:15:15 -04:00
Brendan Allan	0772b67b7a	fix(app): prevent stale service worker startup failures (#45344 )	2026-08-27 02:06:21 +08:00
Aiden Cline	a841d6d046	fix(ai): accept empty Responses stream IDs and null items (#45330 )	2026-08-26 13:04:24 -05:00
opencode-agent[bot]andBrendonovich	ab6a01d135	fix(ui): remove solid menu group label line height (#45331 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-27 00:58:57 +08:00
Kit Langton	962a6ca0e7	fix(sdk): identify workspace dependencies by version specifier (#45309 )	2026-08-26 12:55:06 -04:00
Kit Langton	2602dcd0a7	refactor(core): encapsulate physical attempt execution (#45294 ) Extract physical-attempt streaming, tool execution, and durable settlement from the Session runner. Preserve logical-Step retry and recovery policy, use tagged drain outcomes, and fix multi-click selection during auto-copy.	2026-08-26 12:41:45 -04:00
Aiden Cline	cf98ca55c9	fix(core): allow non-letter tool name prefixes (#45317 )	2026-08-26 11:21:14 -05:00
Kit Langton	fedf017e25	fix(tui): clarify code mode tool call rendering	2026-08-26 11:27:17 -04:00
Aiden Cline	f4a9b93013	feat(ai): add native Groq provider with provider-specific options (#45288 )	2026-08-26 10:12:52 -05:00
Dax Raad	cbef698861	fix(core): suppress state updates during location teardown	2026-08-26 10:48:44 -04:00
James Long	ab2d251155	fix(server): unblock persistent terminal input in Bun builds (#45287 )	2026-08-26 10:47:16 -04:00
Kit Langton	3d7ba38965	feat(tui): investigate errors in a new session	2026-08-26 10:43:40 -04:00
Aiden Cline	37a6ba893e	chore(ai): format package with prettier (#45280 )	2026-08-26 09:24:13 -05:00
Brendan Allan	6c6871fd2a	fix(app): include files in mention search (#45281 )	2026-08-26 22:18:34 +08:00
opencode-agent[bot]	1e864dd8c6	chore: update nix node_modules hashes	2026-08-26 13:49:52 +00:00
James Long	21980a4448	fix(tui): restore terminal settings and move pane toggle to sessions (#45271 )	2026-08-26 09:46:33 -04:00
James Long	9cca8dd6e0	feat(tui): add persistent session terminals (#44971 )	2026-08-26 09:30:03 -04:00
Dax	91028a690b	feat(tui): show LLM token throughput	2026-08-26 09:23:53 -04:00
Shoubhit Dash	e82aa92e64	fix(session): support assistant message content updates (#45015 )	2026-08-26 18:31:59 +05:30
opencode-agent[bot]andBrendonovich	874538d702	fix(desktop): restrict macOS app entitlements (#45258 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 12:53:00 +00:00
opencode-agent[bot]	afc26c72c0	chore: update nix node_modules hashes	2026-08-26 12:11:06 +00:00
James Long	d572a5756c	feat(cli): embed persistent PTY binaries from npm packages (#45248 )	2026-08-26 07:55:30 -04:00
opencode-agent[bot]andBrendonovich	ab95695b24	fix(app): combine loaded skill and instruction file entries (#45217 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 10:08:33 +00:00
Luke Parker	5d7c2ccfc0	feat(app): add experimental vertical session tabs (#45210 )	2026-08-26 19:40:09 +10:00
Luke Parker	24ac05868c	fix(desktop): align and theme Windows caption controls (#45208 )	2026-08-26 09:02:10 +00:00
Shoubhit Dash	4eaf533cd0	feat(plugin): add declarative vcs repository markers (#45192 )	2026-08-26 13:49:36 +05:30
Shoubhit Dash	0f7a76eff0	fix(ai): keep bedrock mantle responses on http (#45197 )	2026-08-26 13:38:27 +05:30
Luke Parker	16d731bd67	feat(app): render settings as a fullscreen surface (#45190 )	2026-08-26 07:57:36 +00:00
Luke Parker	ea582fc133	feat(app): stack collapsed tool calls (#45176 )	2026-08-26 07:56:53 +00:00
opencode-agent[bot]andBrendonovich	7e27e81bc7	fix(app): preserve timeline measurements when moving sessions (#45145 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 14:39:01 +08:00
opencode-agent[bot]andBrendonovich	667722897e	fix(app): identify session worktrees from project inventory (#45153 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 14:38:08 +08:00
opencode-agent[bot]andBrendonovich	d12dbd12a9	fix(app): prevent duplicate rows when editing queued prompts (#45169 )l Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 06:21:44 +00:00
opencode-agent[bot]andBrendonovich	b0bd0bc394	fix(app): align project settings server sections (#45163 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 04:42:26 +00:00
Dax	437df1164c	feat(plugin): expose session move	2026-08-26 00:14:12 -04:00
opencode-agent[bot]andBrendonovich	9a91e21a76	fix(app): preserve workspace choices across draft tabs (#45149 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 03:51:14 +00:00
opencode-agent[bot]	d01ac2069a	chore: update nix node_modules hashes	2026-08-26 03:32:00 +00:00
Aiden Cline	a78d3c5438	fix(ai): recover Anthropic request_too_large as overflow (#45144 )	2026-08-25 22:30:14 -05:00
David Hill	474c3588c1	fix(tui): use unicode ellipses in interface text (#45143 )	2026-08-25 22:26:15 -05:00
Kit Langton	9361117504	feat(core): add directory projects (#45107 )	2026-08-25 23:22:23 -04:00
Aiden Cline	5add6a8e19	fix(ai): preserve provider-defined responses item ids (#45094 )	2026-08-25 22:17:16 -05:00
Aiden Cline	f3c390b89e	feat(ai): add native DeepInfra provider (#45108 )	2026-08-25 22:15:52 -05:00
opencode-agent[bot]andBrendonovich	185c3e5136	fix(app): show project logos throughout settings (#45134 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 02:43:53 +00:00
Kit Langton	cb863b8ea5	fix(tui): dismiss the active interaction with ctrl-c (#45111 )	2026-08-25 22:38:50 -04:00
Kit Langton	695c043e6b	feat(core): refresh unpinned plugins on startup (#45118 )	2026-08-25 22:32:05 -04:00
opencode-agent[bot]	d53456da3b	chore: update nix node_modules hashes	2026-08-26 02:16:23 +00:00
opencode-agent[bot]andBrendonovich	64e930628d	fix(ci): compare affected packages against actual PR base (#45130 ) Co-authored-by: Brendonovich <Brendonovich@users.noreply.github.com>	2026-08-26 10:13:53 +08:00
Kit Langton	3c73ce1dc7	fix(core): materialize mentioned skills on prompts (#44840 )	2026-08-25 22:04:09 -04:00