fix(ai): disable bedrock mantle responses websockets

fix(ai): preserve provider-defined responses item ids (#45094 )
feat(ai): add native DeepInfra provider (#45108 )
2026-10-28 19:34:43 +08:00 · 2026-08-25 22:23:11 -05:00 · 2026-08-25 22:17:16 -05:00 · 2026-08-25 22:15:52 -05:00 · 2026-08-26 02:43:53 +00:00 · 2026-08-25 22:38:50 -04:00
235 changed files with 9573 additions and 1350 deletions
@@ -22,6 +22,36 @@ env:
  FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true

 jobs:
+  affected:
+    name: affected packages
+    runs-on: blacksmith-4vcpu-ubuntu-2404
+    outputs:
+      app: ${{ steps.packages.outputs.app }}
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
+        with:
+          token: ${{ secrets.GITHUB_TOKEN }}
+          fetch-depth: 0
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
+        with:
+          bun-version-file: package.json
+
+      - name: Find affected packages
+        id: packages
+        env:
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            echo "app=true" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          bun x turbo@2.10.2 ls --affected --filter=@opencode-ai/app --output=json > affected.json
+          bun -e 'const result = await Bun.file("affected.json").json(); console.log(`app=${result.packages.count > 0}`)' >> "$GITHUB_OUTPUT"
+
  unit:
    name: unit (${{ matrix.settings.name }})
    strategy:
@@ -41,6 +71,7 @@ jobs:
        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
        with:
          token: ${{ secrets.GITHUB_TOKEN }}
+          fetch-depth: 0

      - name: Setup Node
        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
@@ -80,9 +111,16 @@ jobs:

      - name: Run unit tests
        timeout-minutes: 20
-        run: GITHUB_ACTIONS=false bun turbo test
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            GITHUB_ACTIONS=false bun turbo test
+            exit 0
+          fi
+          GITHUB_ACTIONS=false bun turbo test --affected
        env:
          OPENCODE_EXPERIMENTAL_DISABLE_FILEWATCHER: ${{ runner.os == 'Windows' && 'true' || 'false' }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}

      - name: Verify published codemode package
        if: runner.os == 'Linux'
@@ -92,8 +130,15 @@ jobs:
      - name: Verify packed workerd SDK
        if: runner.os == 'Linux'
        timeout-minutes: 15
-        working-directory: packages/sdk
-        run: bun run verify:package
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            bun turbo verify:package --filter=@opencode-ai/sdk
+            exit 0
+          fi
+          bun turbo verify:package --affected --filter=@opencode-ai/sdk
+        env:
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}

      - name: Verify compiled service lifecycle
        if: always()
@@ -133,7 +178,7 @@ jobs:

  e2e:
    name: e2e (${{ matrix.settings.name }})
-    if: github.ref_name != 'v2' && github.head_ref != 'v2'
+    needs: affected
    strategy:
      fail-fast: false
      matrix:
@@ -144,32 +189,38 @@ jobs:
            host: blacksmith-4vcpu-windows-2025
    runs-on: ${{ matrix.settings.host }}
    env:
+      E2E_ENABLED: ${{ needs.affected.outputs.app == 'true' && github.ref_name != 'v2' && github.head_ref != 'v2' }}
      PLAYWRIGHT_BROWSERS_PATH: ${{ github.workspace }}/.playwright-browsers
    defaults:
      run:
        shell: bash
    steps:
      - name: Checkout repository
+        if: env.E2E_ENABLED == 'true'
        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
        with:
          token: ${{ secrets.GITHUB_TOKEN }}

      - name: Setup Node
+        if: env.E2E_ENABLED == 'true'
        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
        with:
          # Playwright 1.59 hangs while extracting Chromium with Node 24.16.
          node-version: "24.15"

      - name: Setup Bun
+        if: env.E2E_ENABLED == 'true'
        uses: ./.github/actions/setup-bun

      - name: Read Playwright version
+        if: env.E2E_ENABLED == 'true'
        id: playwright-version
        run: |
          version=$(node -e 'console.log(require("./package.json").workspaces.catalog["@playwright/test"])')
          echo "version=$version" >> "$GITHUB_OUTPUT"

      - name: Cache Playwright browsers
+        if: env.E2E_ENABLED == 'true'
        id: playwright-cache
        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
        with:
@@ -177,23 +228,24 @@ jobs:
          key: ${{ runner.os }}-${{ runner.arch }}-playwright-${{ steps.playwright-version.outputs.version }}-chromium

      - name: Install Playwright system dependencies
-        if: runner.os == 'Linux'
+        if: env.E2E_ENABLED == 'true' && runner.os == 'Linux'
        working-directory: packages/app
        run: bunx playwright install-deps chromium

      - name: Install Playwright browsers
-        if: steps.playwright-cache.outputs.cache-hit != 'true'
+        if: env.E2E_ENABLED == 'true' && steps.playwright-cache.outputs.cache-hit != 'true'
        working-directory: packages/app
        run: bunx playwright install chromium

      - name: Run app e2e tests
+        if: env.E2E_ENABLED == 'true'
        run: bun --cwd packages/app test:e2e:local
        env:
          CI: true
        timeout-minutes: 30

      - name: Upload Playwright artifacts
-        if: always()
+        if: always() && env.E2E_ENABLED == 'true'
        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
        with:
          name: playwright-${{ matrix.settings.name }}-${{ github.run_attempt }}
@@ -346,9 +346,7 @@
        "@ai-sdk/amazon-bedrock": "4.0.112",
        "@ai-sdk/anthropic": "3.0.82",
        "@ai-sdk/azure": "3.0.88",
-        "@ai-sdk/cerebras": "2.0.41",
        "@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",
@@ -357,7 +355,6 @@
        "@ai-sdk/perplexity": "3.0.26",
        "@ai-sdk/provider": "3.0.8",
        "@ai-sdk/provider-utils": "4.0.23",
-        "@ai-sdk/togetherai": "2.0.41",
        "@ai-sdk/vercel": "2.0.39",
        "@aws-sdk/credential-providers": "3.1057.0",
        "@ff-labs/fff-bun": "0.10.5",
@@ -667,6 +664,7 @@
      "dependencies": {
        "@opencode-ai/client": "workspace:*",
        "@opencode-ai/core": "workspace:*",
+        "@opencode-ai/plugin": "workspace:*",
        "@opencode-ai/schema": "workspace:*",
        "@opencode-ai/server": "workspace:*",
        "@opencode-ai/util": "workspace:*",
@@ -973,6 +971,7 @@
      "dependencies": {
        "@effect/opentelemetry": "catalog:",
        "@effect/platform-node": "catalog:",
+        "@effect/platform-node-shared": "catalog:",
        "@npmcli/arborist": "catalog:",
        "@npmcli/config": "10.8.1",
        "@opentelemetry/api": "1.9.0",
@@ -1178,8 +1177,6 @@

    "@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=="],
@@ -1206,8 +1203,6 @@

    "@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.23", "", { "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-z8GlDaCmRSDlqkMF2f4/RFgWxdarvIbyuk+m6WXT1LYgsnGiXRJGTD2Z1+SDl3LqtFuRtGX1aghYvQLoHL/9pg=="],

-    "@ai-sdk/togetherai": ["@ai-sdk/togetherai@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-k3p9e3k0/gpDDyTtvafsK4HYR4D/aUQW/kzCwWo1+CzdBU84i4L14gWISC/mv6tgSicMXHcEUd521fPufQwNlg=="],
-
    "@ai-sdk/vercel": ["@ai-sdk/vercel@2.0.39", "", { "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" } }, "sha512-8eu3ljJpkCTP4ppcyYB+NcBrkcBoSOFthCSgk5VnjaxnDaOJFaxnPwfddM7wx3RwMk2CiK1O61Px/LlqNc7QkQ=="],

    "@ai-sdk/xai": ["@ai-sdk/xai@3.0.123", "", { "dependencies": { "@ai-sdk/openai-compatible": "2.0.69", "@ai-sdk/provider": "3.0.15", "@ai-sdk/provider-utils": "4.0.46" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-WNASvd1C516oh2qYIj9EvAVPdU+Abads8DQWU6p9lQtvFFeGh8QW+3LDOARZd1GCINUFfw5yadEK845SMQKLsA=="],
@@ -5912,10 +5907,6 @@

    "@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=="],
@@ -5954,10 +5945,6 @@

    "@ai-sdk/perplexity/@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/togetherai/@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/togetherai/@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/vercel/@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/vercel/@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=="],
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-rIg+QSwF0kmPRnlPrpV7IlPTrHAxvoBOjdtoFOcUnKc=",
-    "aarch64-linux": "sha256-74rMnqiGy7gAUDWr9HWBc6Sn3kxh/hDfm5kp+I4fjok=",
-    "aarch64-darwin": "sha256-ImASKYxUQDHzH/UuwXUBsMNkitAH5wDYESlevEy9qeA=",
-    "x86_64-darwin": "sha256-txPhXfZOSoQdNXPfddKSdKqdjEWrb9GkBb38h+pZIuo="
+    "x86_64-linux": "sha256-oQnV96kE3lIqsQaaUrH4tiEX8/5xvBWizXMGxayFSHo=",
+    "aarch64-linux": "sha256-Hkl1xdCQ7voAllwtyrm3TjQT9cfej29fvPIP3lH7zVo=",
+    "aarch64-darwin": "sha256-s0WHRB13qcD0KlgWNXO7gLpsDOnfB8xny81GnN5YeQc=",
+    "x86_64-darwin": "sha256-fnYi1AxCrnO3byW9keDfBH2ueCfGZdaweOrV6AerrGs="
  }
 }
@@ -157,9 +157,9 @@ const PROVIDERS: ReadonlyArray<Provider> = [
    id: "togetherai",
    label: "TogetherAI",
    tier: "compatible",
-    note: "Existing OpenAI-compatible text/tool recorded tests",
-    vars: [{ name: "TOGETHER_AI_API_KEY" }],
-    validate: (env) => validateBearer("https://api.together.xyz/v1/models", Redacted.make(env.TOGETHER_AI_API_KEY)),
+    note: "Native Together AI text/tool recorded tests",
+    vars: [{ name: "TOGETHER_API_KEY" }],
+    validate: (env) => validateBearer("https://api.together.xyz/v1/models", Redacted.make(env.TOGETHER_API_KEY)),
  },
  {
    id: "minimax",
@@ -200,8 +200,8 @@ const PROVIDERS: ReadonlyArray<Provider> = [
  {
    id: "cerebras",
    label: "Cerebras",
-    tier: "optional",
-    note: "OpenAI-compatible bridge",
+    tier: "compatible",
+    note: "Native Cerebras text/tool/tool-loop recorded tests",
    vars: [{ name: "CEREBRAS_API_KEY" }],
    validate: (env) => validateBearer("https://api.cerebras.ai/v1/models", Redacted.make(env.CEREBRAS_API_KEY)),
  },
@@ -36,7 +36,7 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // Protocols whose wire format ignores inline cache markers (OpenAI's implicit
 // prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
 // whole policy pass for these — emitting hints would be harmless but pointless.
-const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "bedrock-converse", "openrouter"])
+const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "google-vertex-messages", "bedrock-converse", "openrouter"])

 const makeHint = (ttlSeconds: number | undefined): CacheHint =>
  ttlSeconds !== undefined ? new CacheHint({ type: "ephemeral", ttlSeconds }) : new CacheHint({ type: "ephemeral" })
@@ -744,9 +744,12 @@ const endsInServerToolUse = (message: LLMRequest["messages"][number]) => {
  return message.role === "assistant" && last?.type === "tool-call" && last.providerExecuted === true
 }

-const canUseNativeSystemUpdate = (messages: LLMRequest["messages"], index: number) => {
-  const previous = messages[index - 1]
-  const next = messages[index + 1]
+const canUseNativeSystemUpdate = (request: LLMRequest, index: number) => {
+  const previous = request.messages[index - 1]
+  const next = request.messages[index + 1]
+  // Vertex currently rejects/404s for a system message after local tool results,
+  // so fold it into the user tool-result turn across continuations and history.
+  if (request.model.route.id === "google-vertex-messages" && previous?.role === "tool") return false
  return (
    previous !== undefined &&
    previous.role !== "system" &&
@@ -793,7 +796,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
    if (message.role === "system") {
      if (splitsLocalToolResults(request.messages, index))
        return yield* invalid("Anthropic Messages system updates cannot split a local tool call from its tool result")
-      if (supportsNativeSystemUpdates(request) && canUseNativeSystemUpdate(request.messages, index)) {
+      if (supportsNativeSystemUpdates(request) && canUseNativeSystemUpdate(request, index)) {
        messages.push(yield* lowerNativeSystemUpdate(message, breakpoints))
        continue
      }
@@ -79,10 +79,60 @@ const OpenResponsesReasoningItem = Schema.Struct({
  encrypted_content: optionalNull(Schema.String),
 })

-const OpenResponsesItemReference = Schema.Struct({
-  type: Schema.tag("item_reference"),
-  id: Schema.String,
-})
+const OpenResponsesWebSearchCall = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.tag("web_search_call"),
+    id: Schema.String,
+    status: Schema.optional(Schema.String),
+    action: optionalNull(JsonObject),
+  }),
+  [JsonObject],
+)
+
+const OpenResponsesFileSearchCall = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.tag("file_search_call"),
+    id: Schema.String,
+    status: Schema.optional(Schema.String),
+    queries: Schema.optional(Schema.Array(Schema.String)),
+    results: optionalNull(Schema.Array(JsonObject)),
+  }),
+  [JsonObject],
+)
+
+const OpenResponsesCodeInterpreterCall = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.tag("code_interpreter_call"),
+    id: Schema.String,
+    status: Schema.optional(Schema.String),
+    code: optionalNull(Schema.String),
+    container_id: optionalNull(Schema.String),
+    outputs: optionalNull(Schema.Array(JsonObject)),
+  }),
+  [JsonObject],
+)
+
+const OpenResponsesMCPCall = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.tag("mcp_call"),
+    id: Schema.String,
+    status: Schema.optional(Schema.String),
+    server_label: Schema.optional(Schema.String),
+    name: Schema.optional(Schema.String),
+    arguments: Schema.optional(Schema.String),
+    output: optionalNull(Schema.String),
+    error: Schema.optional(Schema.Unknown),
+  }),
+  [JsonObject],
+)
+
+export const HostedToolItem = Schema.Union([
+  OpenResponsesWebSearchCall,
+  OpenResponsesFileSearchCall,
+  OpenResponsesCodeInterpreterCall,
+  OpenResponsesMCPCall,
+])
+export type HostedToolItem = Schema.Schema.Type<typeof HostedToolItem>

 // `function_call_output.output` accepts either a plain string or an ordered
 // array of content items so tools can return images and files in addition to text.
@@ -111,7 +161,6 @@ export const InputItem = Schema.Union([
    phase: Schema.optionalKey(MessagePhase),
  }),
  OpenResponsesReasoningItem,
-  OpenResponsesItemReference,
  Schema.Struct({
    type: Schema.tag("function_call"),
    id: Schema.optionalKey(Schema.String),
@@ -124,10 +173,17 @@ export const InputItem = Schema.Union([
    call_id: Schema.String,
    output: OpenResponsesFunctionCallOutput,
  }),
+  HostedToolItem,
 ])
 type OpenResponsesInputItem = Schema.Schema.Type<typeof InputItem>
+export type ExtendedHostedToolItem = {
+  readonly type: string
+  readonly id: string
+  readonly [key: string]: unknown
+}
 type LoweredInputItem =
  | OpenResponsesInputItem
+  | ExtendedHostedToolItem
  | {
      readonly type: "message"
      readonly id?: string
@@ -140,7 +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
 }
@@ -288,6 +344,7 @@ export const Event = Schema.StructWithRest(
    arguments: Schema.optional(Schema.String),
    text: Schema.optional(Schema.String),
    item_id: Schema.optional(Schema.String),
+    output_index: Schema.optional(Schema.Number),
    summary_index: Schema.optional(Schema.Number),
    item: Schema.optional(StreamItem),
    response: Schema.optional(
@@ -315,9 +372,6 @@ 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" | "reference"
-
 export interface Extension {
  readonly id: string
  readonly name: string
@@ -326,10 +380,7 @@ export interface Extension {
    readonly media: ProviderShared.NormalizedMedia
    readonly request: LLMRequest
  }) => MediaInput | 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
+  readonly lowerHostedToolItem?: (item: unknown) => ExtendedHostedToolItem | undefined
 }

 const BASE: Extension = { id: ADAPTER, name: NAME }
@@ -341,10 +392,10 @@ export interface ParserState {
  readonly tools: ToolStream.State<string>
  readonly hasFunctionCall: boolean
  readonly lifecycle: Lifecycle.State
+  readonly outputItems: Readonly<Record<number, string>>
  readonly messageItems: ReadonlySet<string>
  readonly messagePhases: Readonly<Record<string, MessagePhase | null>>
  readonly reasoningItems: Readonly<Record<string, ReasoningStreamItem>>
-  readonly store: boolean | undefined
 }

 type ReasoningSummaryStatus = "active" | "can-conclude" | "concluded"
@@ -389,53 +440,37 @@ 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,
  }
@@ -526,10 +561,7 @@ const lowerToolResultOutput = Effect.fnUntraced(function* (
 })

 const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (request: LLMRequest, extension: Extension) {
-  const system: LoweredInputItem[] =
-    request.system.length === 0 ? [] : [{ role: "system", content: ProviderShared.joinText(request.system) }]
-  const input: LoweredInputItem[] = [...system]
-  const store = OpenResponsesOptions.resolve(request).store
+  const input: LoweredInputItem[] = []
  const providerMetadataKey = request.model.route.providerMetadataKey ?? "openresponses"

  for (const message of request.messages) {
@@ -552,16 +584,14 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
    if (message.role === "assistant") {
      const content: TextPart[] = []
      const reasoningItems: Record<string, OpenResponsesReasoningInput> = {}
-      const reasoningReferences = new Set<string>()
-      const hostedToolReferences = new Set<string>()
+      const hostedToolItems = new Set<string>()
      const flushText = () => {
        if (content.length === 0) return
        const groups = content.reduce<
          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)
@@ -586,51 +616,51 @@ 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
-          if (store !== false) {
-            if (!reasoningReferences.has(reasoning.id)) input.push({ type: "item_reference", id: reasoning.id })
-            reasoningReferences.add(reasoning.id)
-            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) {
          flushText()
          const id = itemID(part.providerMetadata, providerMetadataKey)
-          const reference = acceptsItemID(extension, "reference", id) ? id : undefined
-          if (store !== false && reference && !hostedToolReferences.has(reference))
-            input.push({ type: "item_reference", id: reference })
-          if (store === false) {
-            // The server is not storing this exchange, so the tool outcome has to
-            // travel in the input. Non-content results degrade to their text form.
-            const content: ReadonlyArray<Content> =
-              part.result.type === "content"
+          const hosted =
+            part.result.type !== "json"
+              ? undefined
+              : Schema.is(HostedToolItem)(part.result.value)
                ? part.result.value
-                : [{ type: "text", text: ProviderShared.toolResultText(part) }]
-            input.push({
-              role: "user",
-              content: yield* Effect.forEach(content, (item) =>
-                lowerHostedToolResultContentItem(item, request, extension),
-              ),
-            })
+                : extension.lowerHostedToolItem?.(part.result.value)
+          if (id !== undefined && hosted?.id === id) {
+            if (!hostedToolItems.has(id)) {
+              input.push(hosted)
+              hostedToolItems.add(id)
+            }
+            continue
          }
-          if (reference) hostedToolReferences.add(reference)
+          const content: ReadonlyArray<Content> =
+            part.result.type === "content"
+              ? part.result.value
+              : [{ type: "text", text: ProviderShared.toolResultText(part) }]
+          input.push({
+            role: "user",
+            content: yield* Effect.forEach(content, (item) =>
+              lowerHostedToolResultContentItem(item, request, extension),
+            ),
+          })
          continue
        }
        return yield* ProviderShared.unsupportedContent(extension.name, "assistant", [
@@ -655,22 +685,16 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
    }
  }

-  // With store:false, Responses APIs only accept previous reasoning items when the
-  // complete item has encrypted state. Summary blocks for one item may carry
-  // that state only on the last block, so filter after they have been joined.
-  return store === false
-    ? input.filter(
-        (item) => !("type" in item) || item.type !== "reasoning" || typeof item.encrypted_content === "string",
-      )
-    : input
+  return input
 })

 const lowerOptions = (request: LLMRequest) => {
  const options = OpenResponsesOptions.resolve(request)
+  const instructions = ProviderShared.joinText(request.system)
  const cacheKey = ProviderShared.promptCacheKey(request)
  const parallelToolCalls = resolveParallelToolCalls(request)
  return {
-    ...(options.instructions ? { instructions: options.instructions } : {}),
+    ...(instructions ? { instructions } : {}),
    ...(options.store !== undefined ? { store: options.store } : {}),
    ...(options.metadata ? { metadata: options.metadata } : {}),
    ...(options.safetyIdentifier ? { safety_identifier: options.safetyIdentifier } : {}),
@@ -818,6 +842,9 @@ const onOutputTextDone = (state: ParserState, event: Event, id: string): StepRes
  return [{ ...state, lifecycle: Lifecycle.textEnd(state.lifecycle, events, id) }, events]
 }

+export const outputItemID = (state: ParserState, event: Event) =>
+  event.output_index === undefined ? event.item_id : (state.outputItems[event.output_index] ?? event.item_id)
+
 export const onReasoningDelta = (state: ParserState, event: Event, itemID: string): StepResult => {
  const item = state.reasoningItems[itemID]
  if (!event.delta || !item) return [state, NO_EVENTS]
@@ -893,25 +920,23 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
      events,
    ]
  }
-  if (item?.type !== "function_call" || !item.id) return [state, NO_EVENTS]
-  const metadata = providerMetadata(state, { itemId: item.id })
+  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 events: LLMEvent[] = []
  const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
  return [
    {
      ...state,
      lifecycle,
-      tools: ToolStream.start(state.tools, item.id, {
-        id: item.call_id ?? item.id,
+      tools: ToolStream.start(state.tools, id, {
+        id: item.call_id,
        name: item.name ?? "",
        input: item.arguments ?? "",
        providerMetadata: metadata,
      }),
    },
-    [
-      ...events,
-      LLMEvent.toolInputStart({ id: item.call_id ?? item.id, name: item.name ?? "", providerMetadata: metadata }),
-    ],
+    [...events, LLMEvent.toolInputStart({ id: item.call_id, name: item.name ?? "", providerMetadata: metadata })],
  ]
 }

@@ -966,31 +991,21 @@ const onReasoningSummaryPartDone = (state: ParserState, event: Event): StepResul
  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
  const item = state.reasoningItems[event.item_id]
  if (!item) return [state, NO_EVENTS]
-  const events: LLMEvent[] = []
  return [
    {
      ...state,
-      lifecycle:
-        state.store !== false
-          ? Lifecycle.reasoningEnd(
-              state.lifecycle,
-              events,
-              `${event.item_id}:${event.summary_index}`,
-              providerMetadata(state, { itemId: event.item_id }),
-            )
-          : state.lifecycle,
      reasoningItems: {
        ...state.reasoningItems,
        [event.item_id]: {
          ...item,
          summaryParts: {
            ...item.summaryParts,
-            [event.summary_index]: state.store !== false ? "concluded" : "can-conclude",
+            [event.summary_index]: "can-conclude",
          },
        },
      },
    },
-    events,
+    NO_EVENTS,
  ]
 }

@@ -1053,18 +1068,19 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
  }

  if (item.type === "function_call") {
-    if (!item.id || !item.call_id || !item.name) return [state, NO_EVENTS] satisfies StepResult
-    const tools = state.tools[item.id]
+    if (!item.call_id || !item.name) return [state, NO_EVENTS] satisfies StepResult
+    const id = item.id ?? item.call_id
+    const tools = state.tools[id]
      ? state.tools
-      : ToolStream.start(state.tools, item.id, {
+      : ToolStream.start(state.tools, id, {
          id: item.call_id,
          name: item.name,
-          providerMetadata: providerMetadata(state, { itemId: item.id }),
+          providerMetadata: item.id ? providerMetadata(state, { itemId: item.id }) : undefined,
        })
    const result =
      item.arguments === undefined
-        ? yield* ToolStream.finish(state.id, tools, item.id)
-        : yield* ToolStream.finishWithInput(state.id, tools, item.id, item.arguments)
+        ? yield* ToolStream.finish(state.id, tools, id)
+        : yield* ToolStream.finishWithInput(state.id, tools, id, item.arguments)
    const events: LLMEvent[] = []
    const resultEvents = result.events ?? []
    const lifecycle = resultEvents.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
@@ -1118,10 +1134,11 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
          event.response?.output ?? [],
          () => [state, NO_EVENTS] satisfies StepResult,
          ([current, events], item) => {
+            const id = item.id ?? (item.type === "function_call" ? item.call_id : undefined)
            if (
-              !item.id ||
-              ((item.type !== "function_call" || !current.tools[item.id]) &&
-                (item.type !== "reasoning" || !current.reasoningItems[item.id]))
+              !id ||
+              ((item.type !== "function_call" || !current.tools[id]) &&
+                (item.type !== "reasoning" || !current.reasoningItems[id]))
            )
              return Effect.succeed([current, events] satisfies StepResult)
            return onOutputItemDone(current, { type: "response.output_item.done", item }).pipe(
@@ -1201,7 +1218,11 @@ 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, event: Event) => {
+export const step = (state: ParserState, input: Event) => {
+  const event =
+    input.item_id && outputItemID(state, input) !== input.item_id
+      ? { ...input, item_id: outputItemID(state, input) }
+      : input
  if (event.type === "response.output_text.delta" || event.type === "response.output_text.done") {
    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(
@@ -1243,7 +1264,15 @@ export const step = (state: ParserState, event: Event) => {
  if (event.type === "response.output_item.added") {
    if (event.item?.type === "message" && !event.item.id)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
-    return Effect.succeed(onOutputItemAdded(state, event))
+    const id = event.item?.id ?? (event.item?.type === "function_call" ? event.item.call_id : undefined)
+    return Effect.succeed(
+      onOutputItemAdded(
+        event.output_index !== undefined && id
+          ? { ...state, outputItems: { ...state.outputItems, [event.output_index]: id } }
+          : state,
+        event,
+      ),
+    )
  }
  if (event.type === "response.function_call_arguments.delta" || event.type === "response.function_call_arguments.done")
    return event.item_id
@@ -1278,10 +1307,10 @@ export const initial = (request: LLMRequest, extension: Extension = BASE): Parse
  hasFunctionCall: false,
  tools: ToolStream.empty<string>(),
  lifecycle: Lifecycle.initial(),
+  outputItems: {},
  messageItems: new Set<string>(),
  messagePhases: {},
  reasoningItems: {},
-  store: OpenResponsesOptions.resolve(request).store,
 })

 export const protocol = Protocol.make({
@@ -278,6 +278,7 @@ interface LoweringOptions {
  readonly cacheControl?: (
    cache: CacheHint | undefined,
  ) => Schema.Schema.Type<typeof OpenAIChatCacheControl> | undefined
+  readonly toolCallID?: (id: string) => string
 }

 const lowerTool = (
@@ -304,8 +305,8 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
    tool: (name) => ({ type: "function" as const, function: { name } }),
  })

-const lowerToolCall = (part: ToolCallPart): OpenAIChatAssistantToolCall => ({
-  id: part.id,
+const lowerToolCall = (part: ToolCallPart, options: LoweringOptions): OpenAIChatAssistantToolCall => ({
+  id: options.toolCallID?.(part.id) ?? part.id,
  type: "function",
  function: {
    name: part.name,
@@ -363,8 +364,9 @@ const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (

 const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(function* (
  message: OpenAIChatRequestMessage,
-  configuredField?: string,
-  options: LoweringOptions = {},
+  configuredField: string | undefined,
+  requireReasoning: boolean,
+  options: LoweringOptions,
 ) {
  const content: TextPart[] = []
  const reasoning: ReasoningPart[] = []
@@ -381,7 +383,7 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
      continue
    }
    if (part.type === "tool-call") {
-      toolCalls.push(lowerToolCall(part))
+      toolCalls.push(lowerToolCall(part, options))
      continue
    }
  }
@@ -391,15 +393,17 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
  const nativeReasoning = openAICompatibleReasoningContent(message.native?.openaiCompatible)
  const fullyStructured = reasoning.every((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))
  const field = (() => {
-    if (configuredField !== undefined) return configuredField
-    if (reasoning.length === 0) return undefined
+    if (configuredField !== undefined && (requireReasoning || reasoning.length > 0 || nativeReasoning !== undefined))
+      return configuredField
+    if (reasoning.length === 0) return requireReasoning ? "reasoning_content" : undefined
    if (observedField !== undefined) return observedField
    if (nativeReasoning !== undefined) return "reasoning_content"
-    if (!fullyStructured) return "reasoning_content"
+    if (!fullyStructured || requireReasoning) return "reasoning_content"
  })()
  const reasoningText = (() => {
-    if (configuredField !== undefined) return reasoning.length === 0 ? (nativeReasoning ?? "") : text
-    if (reasoning.length === 0) return nativeReasoning
+    if (configuredField !== undefined)
+      return reasoning.length === 0 ? (nativeReasoning ?? (requireReasoning ? "" : undefined)) : text
+    if (reasoning.length === 0) return nativeReasoning ?? (requireReasoning ? "" : undefined)
    return text
  })()
  const cached = message.content.findLast((part) => "cache" in part && part.cache !== undefined)
@@ -427,7 +431,7 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
    if (part.result.type !== "content") {
      messages.push({
        role: "tool",
-        tool_call_id: part.id,
+        tool_call_id: options.toolCallID?.(part.id) ?? part.id,
        content: ProviderShared.toolResultText(part),
        cache_control: options.cacheControl?.(part.cache),
      })
@@ -437,7 +441,7 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
    const text = content.filter((item) => item.type === "text").map((item) => item.text)
    messages.push({
      role: "tool",
-      tool_call_id: part.id,
+      tool_call_id: options.toolCallID?.(part.id) ?? part.id,
      content: text.join("\n"),
      cache_control: options.cacheControl?.(part.cache),
    })
@@ -453,11 +457,13 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (

 const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
  message: OpenAIChatRequestMessage,
-  reasoningField?: string,
-  options: LoweringOptions = {},
+  reasoningField: string | undefined,
+  requireReasoning: boolean,
+  options: LoweringOptions,
 ) {
  if (message.role === "user") return [yield* lowerUserMessage(message, options)]
-  if (message.role === "assistant") return [yield* lowerAssistantMessage(message, reasoningField, options)]
+  if (message.role === "assistant")
+    return [yield* lowerAssistantMessage(message, reasoningField, requireReasoning, options)]
  return (yield* lowerToolMessages(message, options)).messages
 })

@@ -478,12 +484,37 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
          ]
        : [{ role: "system", content: ProviderShared.joinText(request.system) }]
  const messages = [...system]
+  const modelID = request.model.id.toLowerCase()
+  const requireReasoning =
+    request.model.compatibility?.requireReasoning ??
+    (request.model.compatibility?.reasoningField !== undefined ||
+      request.model.provider === "deepseek" ||
+      request.model.route.endpoint.baseURL?.toLowerCase().includes("deepseek.com") ||
+      modelID.includes("deepseek"))
+  const reasoningField = request.model.compatibility?.reasoningField
+  const mistral = ["mistral", "devstral", "codestral", "pixtral", "mixtral"].some((family) => modelID.includes(family))
+  const lowering = {
+    ...options,
+    toolCallID: (id: string) => {
+      if (mistral) return id.replace(/[^a-zA-Z0-9]/g, "").slice(0, 9).padEnd(9, "0")
+      if (modelID.includes("claude")) return id.replace(/[^a-zA-Z0-9_-]/g, "_")
+      if (request.model.provider === "openai" || request.model.provider === "azure" || modelID.startsWith("openai/"))
+        return id.slice(0, 40)
+      return id
+    },
+  }
+  const requireAssistantAfterTool = request.model.compatibility?.requireAssistantAfterTool ?? mistral
+  const bridgeTools = () => {
+    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool") messages.push({ role: "assistant", content: "Done." })
+  }
  const pendingImages: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
  const flushImages = () => {
    if (pendingImages.length === 0) return
+    bridgeTools()
    messages.push({ role: "user", content: pendingImages.splice(0) })
  }
  for (const message of request.messages) {
+    if (message.role === "user") bridgeTools()
    if (message.role === "system") {
      const part = yield* ProviderShared.wrappedSystemUpdate("OpenAI Chat", message)
      if (pendingImages.length > 0) {
@@ -526,14 +557,16 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
        )
      continue
    }
+    if (message.role === "assistant" && message.content.every((part) => part.type === "text" && part.text.trim() === ""))
+      continue
    if (message.role === "tool") {
-      const lowered = yield* lowerToolMessages(message, options)
+      const lowered = yield* lowerToolMessages(message, lowering)
      messages.push(...lowered.messages)
      pendingImages.push(...lowered.images)
      continue
    }
    flushImages()
-    messages.push(...(yield* lowerMessage(message, request.model.compatibility?.reasoningField, options)))
+    messages.push(...(yield* lowerMessage(message, reasoningField, requireReasoning, lowering)))
  }
  flushImages()
  return messages
@@ -7,7 +7,7 @@ import { Protocol } from "../route/protocol.js"
 import { HttpTransport } from "../route/transport/index.js"
 import { LLMRequest, type JsonSchema, type ToolDefinition } from "../schema/index.js"
 import { OpenResponses } from "./open-responses.js"
-import { optionalArray, ProviderShared } from "./shared.js"
+import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
 import { OpenAIImage } from "./utils/openai-image.js"
 import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
 import { ToolSchemaProjection } from "./utils/tool-schema.js"
@@ -32,6 +32,40 @@ const OpenAIResponsesImageGenerationTool = Schema.Struct({
  size: Schema.optional(OpenAIImage.Size),
 })

+const OpenAIResponsesHostedToolItem = Schema.Union([
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("computer_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      call_id: Schema.optional(Schema.String),
+      action: optionalNull(JsonObject),
+      pending_safety_checks: Schema.optional(Schema.Array(JsonObject)),
+    }),
+    [JsonObject],
+  ),
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("web_search_preview_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      action: optionalNull(JsonObject),
+    }),
+    [JsonObject],
+  ),
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("image_generation_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      result: optionalNull(Schema.String),
+      output_format: Schema.optional(Schema.Literals(["png", "jpeg", "webp"])),
+      revised_prompt: optionalNull(Schema.String),
+    }),
+    [JsonObject],
+  ),
+])
+
 const OpenAIResponsesTools = Schema.Union([OpenResponses.Tool, OpenAIResponsesImageGenerationTool])

 const OpenAIResponsesToolChoice = Schema.Union([
@@ -41,6 +75,7 @@ const OpenAIResponsesToolChoice = Schema.Union([

 const OpenAIResponsesCoreFields = {
  ...OpenResponses.coreFields,
+  input: Schema.Array(Schema.Union([OpenResponses.InputItem, OpenAIResponsesHostedToolItem])),
  tools: optionalArray(OpenAIResponsesTools),
  tool_choice: Schema.optional(OpenAIResponsesToolChoice),
 }
@@ -51,28 +86,10 @@ const OpenAIResponsesBody = Schema.Struct({
 })
 export type OpenAIResponsesBody = Schema.Schema.Type<typeof OpenAIResponsesBody>

-// Replayed items are paired with stored server state by id, so a foreign or
-// synthetic token can fail request validation even when `call_id` pairing is
-// intact. Only resend ids in each item kind's own grammar; hosted tool
-// references keep generic validation because every hosted tool mints its own
-// prefix. The same allowlist approach codex uses before resending history
-// (codex-rs core/src/client.rs, `prepare_response_items_for_request`).
-const ITEM_ID_PREFIXES: Record<OpenResponses.ItemKind, ReadonlyArray<string>> = {
-  message: ["msg_"],
-  reasoning: ["rs_"],
-  "function-call": ["fc_"],
-  // Every hosted tool mints its own id prefix, so references keep generic
-  // validation only.
-  reference: [],
-}
-
 const extension = {
  id: ADAPTER,
  name: NAME,
-  acceptsItemID: (kind: OpenResponses.ItemKind, id: string) => {
-    const prefixes = ITEM_ID_PREFIXES[kind]
-    return prefixes.length === 0 || prefixes.some((prefix) => id.startsWith(prefix))
-  },
+  lowerHostedToolItem: (item: unknown) => (Schema.is(OpenAIResponsesHostedToolItem)(item) ? item : undefined),
 } satisfies OpenResponses.Extension

 const nativeImageToolInput = (tool: ToolDefinition) => {
@@ -105,6 +122,8 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>, tool
        : { type: "function" as const, name },
  })

+const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenAIResponsesBody))
+
 const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request: LLMRequest) {
  const body = yield* OpenResponses.fromRequestWithExtension(
    LLMRequest.update(request, { tools: [], toolChoice: undefined }),
@@ -112,7 +131,7 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
  )
  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  const parallelToolCalls = OpenResponses.resolveParallelToolCalls(request)
-  return {
+  return yield* decodeBody({
    ...body,
    ...(parallelToolCalls === undefined ? {} : { parallel_tool_calls: parallelToolCalls }),
    tools:
@@ -123,7 +142,7 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
          ),
    tool_choice:
      body.tool_choice ?? (request.toolChoice ? yield* lowerToolChoice(request.toolChoice, request.tools) : undefined),
-  } satisfies OpenAIResponsesBody
+  })
 })

 const hostedToolResult = Effect.fn("OpenAIResponses.hostedToolResult")(function* (item: ResponsesHostedTools.Item) {
@@ -166,7 +185,9 @@ const HOSTED_TOOLS = {
 const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
  if (event.type === "response.reasoning_text.delta")
    return event.item_id
-      ? Effect.succeed(OpenResponses.onReasoningDelta(state, event, event.item_id))
+      ? Effect.succeed(
+          OpenResponses.onReasoningDelta(state, event, OpenResponses.outputItemID(state, event) ?? event.item_id),
+        )
      : ProviderShared.eventError(ADAPTER, `${event.type} is missing item_id`)
  if (event.type === "response.output_item.done" && event.item && ResponsesHostedTools.isItem(event.item, HOSTED_TOOLS))
    return ResponsesHostedTools.onDone(state, event.item, HOSTED_TOOLS)
@@ -210,10 +210,9 @@ export const errorText = (error: unknown) => {
 * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
 * decoder, optionally filters named events, and drops empty / `[DONE]`
 * keep-alive events so the protocol event schema sees one JSON string per
- * element. The SSE channel emits a
- * `Retry` control event on its error channel; we drop it here (we don't
- * implement client-driven retries). Decoder failures become provider output
- * errors so the public error channel stays `AIError`.
+ * element. Retry control events are ignored without interrupting the stream.
+ * Decoder failures become provider output errors so the public error channel
+ * stays `AIError`.
 */
 export const sseFraming = (
  bytes: Stream.Stream<Uint8Array, AIError>,
@@ -221,9 +220,23 @@ export const sseFraming = (
 ): Stream.Stream<string, AIError> =>
  bytes.pipe(
    Stream.decodeText(),
-    Stream.pipeThroughChannel(Sse.decode()),
-    Stream.catchTag("Retry", () => Stream.empty),
-    Stream.catchTag("SseError", (error) => Stream.fail(eventError("sse", error.message))),
+    Stream.mapAccumEffect(
+      () => {
+        const output: Sse.Event[] = []
+        return {
+          output,
+          parser: Sse.makeParser((event) => {
+            if (event._tag === "Event") output.push(event)
+          }),
+        }
+      },
+      (state, chunk) =>
+        Effect.gen(function* () {
+          const error = state.parser.feed(chunk)
+          if (error) return yield* eventError("sse", error.message)
+          return [state, state.output.splice(0)] as const
+        }),
+    ),
    Stream.filter(
      (event) =>
        (events === undefined || events.has(event.event)) &&
@@ -56,7 +56,6 @@ export const StreamOptions = Schema.Struct({
 })

 export const Options = Schema.Struct({
-  instructions: Schema.optional(Schema.String),
  store: Schema.optional(Schema.Boolean),
  metadata: Schema.optional(Schema.Record(Schema.String, Schema.String)),
  safetyIdentifier: Schema.optional(Schema.String),
@@ -1,15 +1,52 @@
+import { Effect, Schema } from "effect"
 import { Protocol } from "../route/protocol.js"
+import type { LLMRequest } from "../schema/index.js"
 import { OpenResponses } from "./open-responses.js"
+import { JsonObject, optionalNull, ProviderShared } from "./shared.js"
 import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"

 const ADAPTER = "xai-responses"
 const NAME = "xAI Responses"

+const XAIResponsesHostedToolItem = Schema.Union([
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("x_search_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      action: optionalNull(JsonObject),
+    }),
+    [JsonObject],
+  ),
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.tag("image_generation_call"),
+      id: Schema.String,
+      status: Schema.optional(Schema.String),
+      result: Schema.optional(Schema.Unknown),
+      error: Schema.optional(Schema.Unknown),
+    }),
+    [JsonObject],
+  ),
+])
+
+const XAIResponsesBody = Schema.Struct({
+  ...OpenResponses.coreFields,
+  input: Schema.Array(Schema.Union([OpenResponses.InputItem, XAIResponsesHostedToolItem])),
+  stream: Schema.Literal(true),
+})
+
 const extension = {
  id: ADAPTER,
  name: NAME,
+  lowerHostedToolItem: (item: unknown) => (Schema.is(XAIResponsesHostedToolItem)(item) ? item : undefined),
 } satisfies OpenResponses.Extension

+const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(XAIResponsesBody))
+const fromRequest = Effect.fn("XAIResponses.fromRequest")(function* (request: LLMRequest) {
+  return yield* decodeBody(yield* OpenResponses.fromRequestWithExtension(request, extension))
+})
+
 const HOSTED_TOOLS = {
  web_search_call: { name: "web_search", input: (item) => item.action ?? {} },
  x_search_call: { name: "x_search", input: (item) => item.action ?? {} },
@@ -35,7 +72,10 @@ const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {

 export const protocol = Protocol.make({
  id: ADAPTER,
-  body: OpenResponses.protocol.body,
+  body: {
+    schema: XAIResponsesBody,
+    from: fromRequest,
+  },
  stream: {
    event: OpenResponses.protocol.stream.event,
    initial: (request) => OpenResponses.initial(request, extension),
@@ -29,6 +29,11 @@ export interface Settings extends ProviderPackage.Settings {
 const responsesRoute = OpenAIResponses.route.with({
  id: "bedrock-mantle-responses",
  provider: id,
+  transport: OpenAIResponses.channelTransport({
+    id: "bedrock-mantle-responses",
+    name: "Bedrock Mantle Responses",
+    enabled: () => false,
+  }),
 })

 const chatRoute = OpenAIChat.route.with({
@@ -0,0 +1,58 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("cerebras")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "cerebras-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.cerebras.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.cerebras.baseURL },
+    auth: AuthOptions.bearer(input, "CEREBRAS_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning", supportsStore: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
@@ -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)
@@ -3,8 +3,10 @@ export * as AnthropicCompatible from "./anthropic-compatible.js"
 export * as AmazonBedrock from "./amazon-bedrock.js"
 export * as AmazonBedrockMantle from "./amazon-bedrock-mantle.js"
 export * as Azure from "./azure.js"
+export * as Cerebras from "./cerebras.js"
 export * as Cloudflare from "./cloudflare.js"
 export { CloudflareAIGateway, CloudflareWorkersAI } from "./cloudflare.js"
+export * as DeepInfra from "./deepinfra.js"
 export * as Google from "./google.js"
 export * as GoogleVertex from "./google-vertex.js"
 export * as GoogleVertexChat from "./google-vertex-chat.js"
@@ -14,5 +16,6 @@ export * as OpenAI from "./openai.js"
 export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
 export * as OpenRouter from "./openrouter.js"
+export * as TogetherAI from "./togetherai.js"
 export * as XAI from "./xai.js"
 export * as ZAI from "./zai.js"
@@ -0,0 +1,58 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("togetherai")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "togetherai-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.togetherai.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.togetherai.baseURL },
+    auth: AuthOptions.bearer(input, ["TOGETHER_API_KEY", "TOGETHER_AI_API_KEY"]),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", supportsStore: false, supportsStrictMode: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
@@ -153,8 +153,11 @@ export class LanguageModelCompatibility extends Schema.Class<LanguageModelCompat
 )({
  toolSchema: Schema.optional(LanguageModelToolSchemaCompatibility),
  reasoningField: Schema.optional(Schema.String),
+  /** Require every assistant message to include its reasoning field, even when empty. */
+  requireReasoning: Schema.optional(Schema.Boolean),
  maxTokensField: Schema.optional(LanguageModelMaxTokensFieldCompatibility),
  requireFinishReason: Schema.optional(Schema.Boolean),
+  requireAssistantAfterTool: Schema.optional(Schema.Boolean),
  supportsStore: Schema.optional(Schema.Boolean),
  supportsUsageInStreaming: Schema.optional(Schema.Boolean),
  supportsStrictMode: Schema.optional(Schema.Boolean),
@@ -3,7 +3,7 @@ import { Effect } from "effect"
 import { CacheHint, LLM, Message } from "../src/index.js"
 import { Auth } from "../src/route.js"
 import { compileRequest } from "../src/route/client.js"
-import { AmazonBedrock } from "../src/providers.js"
+import { AmazonBedrock, GoogleVertexMessages } from "../src/providers.js"
 import * as AnthropicMessages from "../src/protocols/anthropic-messages.js"
 import * as Gemini from "../src/protocols/gemini.js"
 import * as OpenAIChat from "../src/protocols/openai-chat.js"
@@ -86,6 +86,27 @@ describe("applyCachePolicy", () => {
    }),
  )

+  it.effect("'auto' emits Anthropic cache markers on Vertex", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: GoogleVertexMessages.configure({ accessToken: "test", location: "global", project: "test" }).model(
+            "claude-opus-4-8",
+          ),
+          system: "You are concise.",
+          tools: [{ name: "lookup", description: "Look up a value", inputSchema: { type: "object", properties: {} } }],
+          prompt: "hi",
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        tools: [{ name: "lookup", cache_control: { type: "ephemeral" } }],
+        system: [{ type: "text", text: "You are concise.", cache_control: { type: "ephemeral" } }],
+        messages: [{ role: "user", content: [{ type: "text", text: "hi", cache_control: { type: "ephemeral" } }] }],
+      })
+    }),
+  )
+
  it.effect("'auto' is a no-op on OpenAI (implicit caching protocol)", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "cerebras",
+    "route": "cerebras-chat",
+    "transport": "http",
+    "model": "gpt-oss-120b",
+    "tags": ["prefix:cerebras-chat", "provider:cerebras", "text", "golden"],
+    "name": "cerebras-chat/cerebras-gpt-oss-120b-text",
+    "recordedAt": "2026-08-25T23:55:27.619Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.cerebras.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-oss-120b\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":256}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"role\":\"assistant\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\"The\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" user says: \\\"Reply exactly with: Hello!\\\" So we must output exactly\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" \\\"Hello!\\\" with no extra characters, no formatting\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\". Ensure no extra spaces or new\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\"lines? Probably just \\\"Hello!\\\".\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" Usually we output exactly that.\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"content\":\"Hello!\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\",\"usage\":{\"total_tokens\":142,\"completion_tokens\":58,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0,\"reasoning_tokens\":46},\"prompt_tokens\":84,\"prompt_tokens_details\":{\"cached_tokens\":0}},\"time_info\":{\"created\":1787702127.645281,\"queue_time\":0.003817115,\"prompt_time\":0.001587193,\"completion_time\":0.029805929,\"total_time\":0.036823272705078125}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "cerebras",
+    "route": "cerebras-chat",
+    "transport": "http",
+    "model": "gpt-oss-120b",
+    "tags": ["prefix:cerebras-chat", "provider:cerebras", "tool", "tool-call", "golden"],
+    "name": "cerebras-chat/cerebras-gpt-oss-120b-tool-call",
+    "recordedAt": "2026-08-25T23:55:28.454Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.cerebras.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-oss-120b\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":512}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"role\":\"assistant\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\"We\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\" need to call the function get_weather with city \\\"\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\"Paris\\\".\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"name\":\"get_weather\",\"arguments\":\"\"},\"type\":\"function\",\"id\":\"3d860cefe\",\"index\":0}]},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},\"type\":\"function\",\"index\":0}]},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\",\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\",\"usage\":{\"total_tokens\":174,\"completion_tokens\":37,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0,\"reasoning_tokens\":13},\"prompt_tokens\":137,\"prompt_tokens_details\":{\"cached_tokens\":0}},\"time_info\":{\"created\":1787702127.8019717,\"queue_time\":0.31196235,\"prompt_time\":0.005234764,\"completion_time\":0.020198402,\"total_time\":0.702225923538208}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -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"
+      }
+    }
+  ]
+}
@@ -32,7 +32,7 @@
        {
          "direction": "client",
          "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Call get_weather once, then reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Call get_weather once, then reply exactly: Paris is sunny.\"}"
        },
        {
          "direction": "server",
@@ -62,7 +62,7 @@
        {
          "direction": "client",
          "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"type\":\"function_call_output\",\"call_id\":\"call_ws_weather\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"previous_response_id\":\"resp_ws_tool_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"type\":\"function_call_output\",\"call_id\":\"call_ws_weather\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"previous_response_id\":\"resp_ws_tool_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Call get_weather once, then reply exactly: Paris is sunny.\"}"
        },
        {
          "direction": "server",
@@ -32,7 +32,7 @@
        {
          "direction": "client",
          "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
        },
        {
          "direction": "server",
@@ -81,7 +81,7 @@
        {
          "direction": "client",
          "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Beta.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Beta.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
        },
        {
          "direction": "server",
@@ -32,7 +32,7 @@
        {
          "direction": "client",
          "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
        },
        {
          "direction": "server",
@@ -81,7 +81,7 @@
        {
          "direction": "client",
          "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"previous_response_id\":\"resp_ws_rejection_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"previous_response_id\":\"resp_ws_rejection_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
        },
        {
          "direction": "server",
@@ -91,7 +91,7 @@
        {
          "direction": "client",
          "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
        },
        {
          "direction": "server",
@@ -26,7 +26,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true,\"instructions\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"}"
      },
      "response": {
        "status": 200,
@@ -18,7 +18,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true,\"instructions\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"}"
      },
      "response": {
        "status": 200,
@@ -25,7 +25,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true,\"instructions\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"}"
      },
      "response": {
        "status": 200,
@@ -43,7 +43,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]},{\"type\":\"reasoning\",\"id\":\"rs_0ad67c31d9ddad95016a869efbd02487d1a51eed850e6f87f5\",\"summary\":[],\"encrypted_content\":\"gAAAAABqhp79U9UPKmTdmo9tmdil0C2KXpkFqUc4MNkYHT53Lzos9omncFPg76QzUmmSOdBcajisWBEo-xiTCvhp135uACUq8TJcdw4DluieYq6dWszijy28PFFfeO-6MmHwi7zeln1Z202zErJUEyuf1bML68VAeam5PqlMLG-a4-pmnWiH2ExWKibTUX37QoMQoArrkccJOCmxwDflV_kWDPMFxQVDfeMg9fd1gVv2u-x1Mjk0b9mJDOq0Fe5Gh-IkpWzfXgZTdptFmCM75cksvs61Rqsx6P33czal-LSixEF0WMizCvbMQmqKGs7MKGMeoa6j6vWOnB3ICIbv6FShnSaTpZWJFwejvOurkfuxa-2q6xVDZsBoQCgMWPHsqLxwAo1JKdfBk0pMvSuvpw2BRxykUZ1ULCYJ-BypST65292-EuSZFIuXPMPir-_raSCTsgsZNMscDG6ll3qksDTDS6_o5NutD7Ra-WZzaUe_HQlSLKLACTc4qv2EK1QoC4aYv4goxkTSx17WhS2D86lILgkUd-TIHjJ6iR3uxSNx7YeBNxiJgddIAEjAaSrdF-WDouSNT9k3efd5HhT3zahIOMKgb3XIQzFOYWfWgea5-SbaIdKwne9hU0QyhcBQs6yoifSg-fJZtahbPb-GCDYnOLlH-bV94vldoccb-2P1JdB3jaLj5tJUecfr2H4qiu8MgkPj0TkwYNbJynYmJo9H5Lm-XJ9gfzIXzJh0arKwsS4gwDLf4J3LOEF3WEW3mknOjjb9PrLmHRYXQQh9tTiX9ILPZpbufkyCurTUMQgWiSCitXBC6FoLXRHilSmb-6_avBnlUMziMfey-FkKvRfiPox6BaJrnOq6SGlOv11y7EKvzrn29la7HKPygYenDAkyq2mq0Zk2nLWNmJcv9sQTBrkBdFMmJYPi2J2im8XD5MmAjEL8R4FCBHoPIIZ6pENQykvH8PhpWKuzF5gJlY3Vwz4iJ0Qb9TrNI0hzBoI1U0LeB5FJ2HgjZQwCFF5x3ubh72xrUsFpuyYyYPa8GDT0Bo-LW_IlJ_mN4EwI5Nk9n-8Bt015yxsfpa5YaDeCeQFcdj8SD0UAd7QWtGACpzKcIj1-vJJU7OiwscV_v1dLvoiEe1ehI9jcvPn28TgHlo_dippe0iMN4FAm1Bf8vtWVMFDvfV1rPv1pAFFnSa9XqFszD5Exo_xzcQEoKXvQv3OnUtoiM4Db4uadClazLjoep2TQgHcJBVbTbLySTVPmok4ROFQZsU_mq4vu2M__d8HOjADfIIYz5VLVQKNpo0Hv_QkT2bn56Q==\"},{\"type\":\"function_call\",\"id\":\"fc_0ad67c31d9ddad95016a869efd126887d1b7e2f17f155cb5dc\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]},{\"type\":\"reasoning\",\"id\":\"rs_0ad67c31d9ddad95016a869efbd02487d1a51eed850e6f87f5\",\"summary\":[],\"encrypted_content\":\"gAAAAABqhp79U9UPKmTdmo9tmdil0C2KXpkFqUc4MNkYHT53Lzos9omncFPg76QzUmmSOdBcajisWBEo-xiTCvhp135uACUq8TJcdw4DluieYq6dWszijy28PFFfeO-6MmHwi7zeln1Z202zErJUEyuf1bML68VAeam5PqlMLG-a4-pmnWiH2ExWKibTUX37QoMQoArrkccJOCmxwDflV_kWDPMFxQVDfeMg9fd1gVv2u-x1Mjk0b9mJDOq0Fe5Gh-IkpWzfXgZTdptFmCM75cksvs61Rqsx6P33czal-LSixEF0WMizCvbMQmqKGs7MKGMeoa6j6vWOnB3ICIbv6FShnSaTpZWJFwejvOurkfuxa-2q6xVDZsBoQCgMWPHsqLxwAo1JKdfBk0pMvSuvpw2BRxykUZ1ULCYJ-BypST65292-EuSZFIuXPMPir-_raSCTsgsZNMscDG6ll3qksDTDS6_o5NutD7Ra-WZzaUe_HQlSLKLACTc4qv2EK1QoC4aYv4goxkTSx17WhS2D86lILgkUd-TIHjJ6iR3uxSNx7YeBNxiJgddIAEjAaSrdF-WDouSNT9k3efd5HhT3zahIOMKgb3XIQzFOYWfWgea5-SbaIdKwne9hU0QyhcBQs6yoifSg-fJZtahbPb-GCDYnOLlH-bV94vldoccb-2P1JdB3jaLj5tJUecfr2H4qiu8MgkPj0TkwYNbJynYmJo9H5Lm-XJ9gfzIXzJh0arKwsS4gwDLf4J3LOEF3WEW3mknOjjb9PrLmHRYXQQh9tTiX9ILPZpbufkyCurTUMQgWiSCitXBC6FoLXRHilSmb-6_avBnlUMziMfey-FkKvRfiPox6BaJrnOq6SGlOv11y7EKvzrn29la7HKPygYenDAkyq2mq0Zk2nLWNmJcv9sQTBrkBdFMmJYPi2J2im8XD5MmAjEL8R4FCBHoPIIZ6pENQykvH8PhpWKuzF5gJlY3Vwz4iJ0Qb9TrNI0hzBoI1U0LeB5FJ2HgjZQwCFF5x3ubh72xrUsFpuyYyYPa8GDT0Bo-LW_IlJ_mN4EwI5Nk9n-8Bt015yxsfpa5YaDeCeQFcdj8SD0UAd7QWtGACpzKcIj1-vJJU7OiwscV_v1dLvoiEe1ehI9jcvPn28TgHlo_dippe0iMN4FAm1Bf8vtWVMFDvfV1rPv1pAFFnSa9XqFszD5Exo_xzcQEoKXvQv3OnUtoiM4Db4uadClazLjoep2TQgHcJBVbTbLySTVPmok4ROFQZsU_mq4vu2M__d8HOjADfIIYz5VLVQKNpo0Hv_QkT2bn56Q==\"},{\"type\":\"function_call\",\"id\":\"fc_0ad67c31d9ddad95016a869efd126887d1b7e2f17f155cb5dc\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true,\"instructions\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"}"
      },
      "response": {
        "status": 200,
@@ -21,7 +21,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"gpt-4o-mini\",\"input\":[{\"role\":\"system\",\"content\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,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\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true}"
+        "body": "{\"model\":\"gpt-4o-mini\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,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\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true,\"instructions\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"}"
      },
      "response": {
        "status": 200,
@@ -21,7 +21,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"grok-4.5\",\"input\":[{\"role\":\"system\",\"content\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,JVBERi0xLjQKMSAwIG9iago8PCAvVHlwZSAvQ2F0YWxvZyAvUGFnZXMgMiAwIFIgPj4KZW5kb2JqCjIgMCBvYmoKPDwgL1R5cGUgL1BhZ2VzIC9LaWRzIFszIDAgUl0gL0NvdW50IDEgPj4KZW5kb2JqCjMgMCBvYmoKPDwgL1R5cGUgL1BhZ2UgL1BhcmVudCAyIDAgUiAvTWVkaWFCb3ggWzAgMCA2MTIgNzkyXSAvUmVzb3VyY2VzIDw8IC9Gb250IDw8IC9GMSA1IDAgUiA+PiA+PiAvQ29udGVudHMgNCAwIFIgPj4KZW5kb2JqCjQgMCBvYmoKPDwgL0xlbmd0aCA3NSA+PgpzdHJlYW0KQlQKL0YxIDE4IFRmCjcyIDcyMCBUZAooUERGIGNhc3NldHRlIHZlcmlmaWNhdGlvbiBjb2RlOiBPUkNISUQtNzM5MSkgVGoKRVQKZW5kc3RyZWFtCmVuZG9iago1IDAgb2JqCjw8IC9UeXBlIC9Gb250IC9TdWJ0eXBlIC9UeXBlMSAvQmFzZUZvbnQgL0hlbHZldGljYSA+PgplbmRvYmoKeHJlZgowIDYKMDAwMDAwMDAwMCA2NTUzNSBmIAowMDAwMDAwMDA5IDAwMDAwIG4gCjAwMDAwMDAwNTggMDAwMDAgbiAKMDAwMDAwMDExNSAwMDAwMCBuIAowMDAwMDAwMjQxIDAwMDAwIG4gCjAwMDAwMDAzNjUgMDAwMDAgbiAKdHJhaWxlcgo8PCAvU2l6ZSA2IC9Sb290IDEgMCBSID4+CnN0YXJ0eHJlZgo0MzUKJSVFT0YK\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true}"
+        "body": "{\"model\":\"grok-4.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,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\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true,\"instructions\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"}"
      },
      "response": {
        "status": 200,
@@ -26,6 +26,9 @@ describe("provider package entrypoints", () => {
      import("@opencode-ai/ai/providers/amazon-bedrock/mantle"),
      import("@opencode-ai/ai/providers/amazon-bedrock/mantle/chat"),
      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"),
    ])

    for (const module of modules) expect(module.model).toBeFunction()
@@ -35,6 +38,24 @@ 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")
@@ -5,6 +5,7 @@ import { CacheHint, LLM, AIError, LLMRequest, Message, ToolCallPart, ToolDefinit
 import { Auth, LLMClient } 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"
 import { continuationRequest, nativeAnthropicMessagesContinuation } from "../continuation-scenarios.js"
 import { it } from "../lib/effect.js"
 import { dynamicResponse, fixedResponse } from "../lib/http.js"
@@ -27,6 +28,12 @@ const compileUnsignedReasoning = (model: LLMRequest["model"]) =>
    }),
  )

+const vertexOpus48 = GoogleVertexMessages.configure({
+  accessToken: "test",
+  location: "global",
+  project: "test",
+}).model("claude-opus-4-8")
+
 const request = LLM.request({
  id: "req_1",
  model,
@@ -286,6 +293,149 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("keeps a terminal Vertex system update in the tool-result turn", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+            {
+              type: "text",
+              text: "<system-update>\nOperator update.\n</system-update>",
+              cache_control: undefined,
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("preserves folded tool-result system updates across multi-turn Vertex history", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+            Message.assistant("Acknowledged."),
+            Message.user("Next step."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+            {
+              type: "text",
+              text: "<system-update>\nOperator update.\n</system-update>",
+              cache_control: undefined,
+            },
+          ],
+        },
+        { role: "assistant", content: [{ type: "text", text: "Acknowledged." }] },
+        { role: "user", content: [{ type: "text", text: "Next step." }] },
+      ])
+    }),
+  )
+
+  it.effect("keeps a terminal direct Anthropic system update native", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: opus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+          ],
+        },
+        {
+          role: "system",
+          content: [{ type: "text", text: "Operator update.", cache_control: undefined }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps an ordinary terminal Vertex system update native", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [Message.user("Before."), Message.system("Operator update.")],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: [{ type: "text", text: "Before." }] },
+        {
+          role: "system",
+          content: [{ type: "text", text: "Operator update.", cache_control: undefined }],
+        },
+      ])
+    }),
+  )
+
  it.effect("rejects a system update between a local tool call and its result", () =>
    Effect.gen(function* () {
      const error = yield* compileRequest(
@@ -1,11 +1,11 @@
 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 { 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"

@@ -59,6 +59,27 @@ describe("Amazon Bedrock Mantle provider", () => {
    }),
  )

+  it.effect("keeps Responses on HTTP when a WebSocket executor is available", () =>
+    Effect.gen(function* () {
+      const model = AmazonBedrockMantle.configure({ apiKey: "test-key" }).responses("openai.gpt-oss-120b")
+
+      yield* LLMClient.generate(LLM.request({ model, prompt: "Hi" }), {
+        webSocket: { execute: () => Effect.die("unexpected Mantle WebSocket request") },
+      }).pipe(
+        Effect.provide(
+          dynamicResponse((input) => {
+            expect(input.request.url).toBe("https://bedrock-mantle.us-east-1.api.aws/v1/responses")
+            return Effect.succeed(
+              input.respond(sseEvents({ type: "response.completed", response: {} }), {
+                headers: { "content-type": "text/event-stream" },
+              }),
+            )
+          }),
+        ),
+      )
+    }),
+  )
+
  it.effect("supports bearer authentication and custom base URLs", () =>
    Effect.gen(function* () {
      const seen: Array<{ readonly url: string; readonly authorization: string | undefined }> = []
@@ -83,6 +104,39 @@ 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(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({
@@ -1,5 +1,6 @@
 import * as Anthropic from "../../src/providers/anthropic.js"
 import * as AnthropicCompatible from "../../src/providers/anthropic-compatible.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"
@@ -47,14 +48,16 @@ const cloudflareWorkersAITools = cloudflareWorkers.model("@cf/openai/gpt-oss-20b
 const deepseek = OpenAICompatible.deepseek
  .configure({ apiKey: process.env.DEEPSEEK_API_KEY ?? "fixture" })
  .model("deepseek-chat")
-const together = OpenAICompatible.togetherai
-  .configure({
-    apiKey: process.env.TOGETHER_AI_API_KEY ?? "fixture",
-  })
-  .model("meta-llama/Llama-3.3-70B-Instruct-Turbo")
+const together = TogetherAI.configure({
+  apiKey: process.env.TOGETHER_API_KEY ?? process.env.TOGETHER_AI_API_KEY ?? "fixture",
+}).model("meta-llama/Llama-3.3-70B-Instruct-Turbo")
+const cerebras = Cerebras.configure({ apiKey: process.env.CEREBRAS_API_KEY ?? "fixture" }).model("gpt-oss-120b")
 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")
@@ -193,8 +196,27 @@ describeRecordedGoldenScenarios([
    name: "TogetherAI Llama 3.3 70B",
    prefix: "openai-compatible-chat",
    model: together,
-    requires: ["TOGETHER_AI_API_KEY"],
-    scenarios: ["text", "tool-call"],
+    requires: ["TOGETHER_API_KEY"],
+    scenarios: [
+      {
+        id: "text",
+        cassette: "openai-compatible-chat/togetherai-streams-text",
+        prompt: "Reply with exactly: Hello!",
+        maxTokens: 20,
+      },
+      { id: "tool-call", cassette: "openai-compatible-chat/togetherai-streams-tool-call" },
+    ],
+  },
+  {
+    name: "Cerebras GPT OSS 120B",
+    prefix: "cerebras-chat",
+    model: cerebras,
+    requires: ["CEREBRAS_API_KEY"],
+    scenarios: [
+      { id: "text", maxTokens: 256, temperature: false },
+      { id: "tool-call", maxTokens: 512, temperature: false },
+      { id: "tool-loop", maxTokens: 512, temperature: false, timeout: 30_000 },
+    ],
  },
  {
    name: "Groq Llama 3.3 70B",
@@ -203,6 +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",
@@ -0,0 +1,184 @@
+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, DeepInfra, TogetherAI } 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("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")
+      const cerebras = Cerebras.configure({ apiKey: "fixture" }).model("qwen-3-235b-a22b")
+
+      expect(together).toMatchObject({
+        provider: "togetherai",
+        compatibility: { maxTokensField: "max_tokens", supportsStore: false, supportsStrictMode: false },
+        route: { id: "togetherai-chat", protocol: "openai-chat" },
+      })
+      expect(together.route.endpoint.baseURL).toBe("https://api.together.xyz/v1")
+      expect(cerebras).toMatchObject({
+        provider: "cerebras",
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning", supportsStore: false },
+        route: { id: "cerebras-chat", protocol: "openai-chat" },
+      })
+      expect(cerebras.route.endpoint.baseURL).toBe("https://api.cerebras.ai/v1")
+    }),
+  )
+
+  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(
+        LLM.request({
+          model: TogetherAI.configure({ apiKey: "fixture", baseURL: "https://gateway.example/v1" }).model("llama"),
+          prompt: "Use a tool.",
+          generation: { maxTokens: 32 },
+          tools: [
+            ToolDefinition.make({ name: "lookup", description: "Look up data", inputSchema: { type: "object" } }),
+          ],
+          providerOptions: { store: true },
+        }),
+      )
+
+      expect(together.body).toMatchObject({
+        max_tokens: 32,
+        stream_options: { include_usage: true },
+        tools: [{ function: { name: "lookup" } }],
+      })
+      expect(together.body).not.toHaveProperty("max_completion_tokens")
+      expect(together.body).not.toHaveProperty("store")
+      expect(together.body.tools?.[0]?.function).not.toHaveProperty("strict")
+
+      const cerebras = yield* compileRequest(
+        LLM.request({
+          model: Cerebras.configure({ apiKey: "fixture", baseURL: "https://gateway.example/v1" }).model("qwen"),
+          generation: { maxTokens: 48 },
+          messages: [
+            Message.user("Think first."),
+            Message.assistant([
+              { type: "reasoning", text: "A deliberate thought." },
+              { type: "text", text: "An answer." },
+            ]),
+            Message.user("Continue."),
+          ],
+          providerOptions: { store: true },
+        }),
+      )
+
+      expect(cerebras.body).toMatchObject({
+        max_tokens: 48,
+        messages: [
+          { role: "user", content: "Think first." },
+          { role: "assistant", content: "An answer.", reasoning: "A deliberate thought." },
+          { role: "user", content: "Continue." },
+        ],
+      })
+      expect(cerebras.body).not.toHaveProperty("max_completion_tokens")
+      expect(cerebras.body).not.toHaveProperty("store")
+      expect(cerebras.body.messages[1]).not.toHaveProperty("reasoning_content")
+    }),
+  )
+
+  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]) {
+        const selected = native.model("provider-model", {
+          apiKey: "fixture",
+          baseURL: "https://gateway.example/v1",
+          headers: { "x-application": "opencode" },
+          body: { service_tier: "priority" },
+          providerOptions: { reasoningEffort: "high" },
+        })
+
+        expect(selected.route.endpoint.baseURL).toBe("https://gateway.example/v1")
+        expect(selected.route.defaults.headers).toEqual({ "x-application": "opencode" })
+        expect(selected.route.defaults.http?.body).toEqual({ service_tier: "priority" })
+        expect(selected.route.defaults.providerOptions).toEqual({ reasoningEffort: "high" })
+      }
+    }),
+  )
+
+  it.effect("resolves provider environment credentials and preserves deprecated Together credentials", () =>
+    Effect.gen(function* () {
+      const scenarios = [
+        {
+          model: TogetherAI.configure().model("llama"),
+          env: { TOGETHER_API_KEY: "together-primary", TOGETHER_AI_API_KEY: "together-legacy" },
+          token: "together-primary",
+          url: "https://api.together.xyz/v1/chat/completions",
+        },
+        {
+          model: TogetherAI.configure().model("llama"),
+          env: { TOGETHER_AI_API_KEY: "together-legacy" },
+          token: "together-legacy",
+          url: "https://api.together.xyz/v1/chat/completions",
+        },
+        {
+          model: Cerebras.configure().model("qwen"),
+          env: { CEREBRAS_API_KEY: "cerebras-secret" },
+          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",
+        },
+      ]
+
+      yield* Effect.forEach(scenarios, (scenario) =>
+        LLM.generate(LLM.request({ model: scenario.model, prompt: "Say hello." })).pipe(
+          Effect.provide(
+            dynamicResponse((input) =>
+              Effect.gen(function* () {
+                const request = yield* HttpClientRequest.toWeb(input.request).pipe(Effect.orDie)
+                expect(request.url).toBe(scenario.url)
+                expect(request.headers.get("authorization")).toBe(`Bearer ${scenario.token}`)
+                return input.respond(
+                  sseEvents(
+                    { id: "chatcmpl_fixture", choices: [{ delta: { content: "Hello" }, finish_reason: null }] },
+                    { id: "chatcmpl_fixture", choices: [{ delta: {}, finish_reason: "stop" }] },
+                  ),
+                  { headers: { "content-type": "text/event-stream" } },
+                )
+              }),
+            ),
+          ),
+          Effect.provide(ConfigProvider.layer(ConfigProvider.fromEnv({ env: scenario.env }))),
+          Effect.tap((response) => Effect.sync(() => expect(response.text).toBe("Hello"))),
+        ),
+      )
+    }),
+  )
+})
@@ -85,6 +85,28 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("omits empty and whitespace-only assistant messages", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("Before."),
+            Message.assistant([]),
+            Message.assistant(""),
+            Message.assistant(" \n\t "),
+            Message.assistant("After."),
+          ],
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: "Before." },
+        { role: "assistant", content: "After." },
+      ])
+    }),
+  )
+
  it.effect("replays canonical reasoning as OpenAI-compatible reasoning_content", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -147,6 +169,56 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("preserves observed reasoning fields when reasoning is required", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: LanguageModel.update(model, { compatibility: { requireReasoning: true } }),
+          messages: [
+            Message.assistant([
+              {
+                type: "reasoning",
+                text: "thinking",
+                providerMetadata: { openai: { reasoningField: "reasoning_text" } },
+              },
+              { type: "text", text: "Hello" },
+            ]),
+            Message.assistant("Done"),
+          ],
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: "Hello", reasoning_text: "thinking" },
+        { role: "assistant", content: "Done", reasoning_content: "" },
+      ])
+    }),
+  )
+
+  it.effect("omits empty configured reasoning fields when reasoning is explicitly optional", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: LanguageModel.update(model, {
+            compatibility: { reasoningField: "reasoning_text", requireReasoning: false },
+          }),
+          messages: [
+            Message.assistant([
+              { type: "reasoning", text: "thinking" },
+              { type: "text", text: "Hello" },
+            ]),
+            Message.assistant("Done"),
+          ],
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: "Hello", reasoning_text: "thinking" },
+        { role: "assistant", content: "Done" },
+      ])
+    }),
+  )
+
  it.effect("rejects reasoning fields that conflict with assistant message fields", () =>
    Effect.gen(function* () {
      const error = yield* compileRequest(
@@ -366,6 +438,35 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("limits OpenAI and Azure Chat tool call IDs to 40 characters", () =>
+    Effect.gen(function* () {
+      const id = `call_${"a".repeat(48)}`
+      const models = [
+        model,
+        Azure.configure({ baseURL: "https://opencode-test.openai.azure.com/openai/", apiKey: "test" }).chat("gpt-4o"),
+      ]
+
+      yield* Effect.forEach(models, (selected) =>
+        Effect.gen(function* () {
+          const prepared = yield* compileRequest(
+            LLM.request({
+              model: selected,
+              messages: [
+                Message.assistant([ToolCallPart.make({ id, name: "lookup", input: {} })]),
+                Message.tool({ id, name: "lookup", result: "Sunny" }),
+              ],
+            }),
+          )
+
+          expect(prepared.body.messages).toMatchObject([
+            { role: "assistant", tool_calls: [{ id: id.slice(0, 40) }] },
+            { role: "tool", tool_call_id: id.slice(0, 40) },
+          ])
+        }),
+      )
+    }),
+  )
+
  it.effect("preserves structured tool errors for the model", () =>
    Effect.gen(function* () {
      const error = { error: { type: "unknown", message: "Tool execution interrupted" } }
@@ -431,6 +532,30 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("bridges image tool results before their synthetic user message when required", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: LanguageModel.update(model, { compatibility: { requireAssistantAfterTool: true } }),
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_image", name: "read", input: {} })]),
+            Message.tool({
+              id: "call_image",
+              name: "read",
+              result: {
+                type: "content",
+                value: [{ type: "file", uri: "data:image/png;base64,AAECAw==", mime: "image/png", name: "pixel.png" }],
+              },
+            }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.messages.map((message) => message.role)).toEqual(["assistant", "tool", "assistant", "user"])
+      expect(prepared.body.messages[2]).toEqual({ role: "assistant", content: "Done." })
+    }),
+  )
+
  it.effect("orders parallel tool responses before one aggregated vision message", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -238,6 +238,135 @@ describe("OpenAI-compatible Chat route", () => {
    }),
  )

+  it.effect("normalizes tool call IDs for the selected model family", () =>
+    Effect.gen(function* () {
+      const longID = `call_${"a".repeat(48)}`
+      const cases = [
+        { provider: "custom", model: "mistral-small", id: "toolu_01CBhTTz95qkd9LJMdC9sf8t", expected: "toolu01CB" },
+        { provider: "custom", model: "devstral-small", id: "abc", expected: "abc000000" },
+        { provider: "custom", model: "codestral-latest", id: "toolu_01CBhTTz95", expected: "toolu01CB" },
+        { provider: "custom", model: "pixtral-large", id: "toolu_01CBhTTz95", expected: "toolu01CB" },
+        { provider: "custom", model: "open-mixtral-8x22b", id: "toolu_01CBhTTz95", expected: "toolu01CB" },
+        { provider: "gateway", model: "anthropic/claude-sonnet-4", id: "call|item/+", expected: "call_item__" },
+        { provider: "gateway", model: "openai/gpt-4o", id: longID, expected: longID.slice(0, 40) },
+        { provider: "custom", model: "ordinary-model", id: "call|item/+", expected: "call|item/+" },
+        { provider: "mistral", model: "zai-glm-5-2", id: "call_long_identifier", expected: "call_long_identifier" },
+      ]
+
+      yield* Effect.forEach(cases, (item) =>
+        Effect.gen(function* () {
+          const prepared = yield* compileRequest(
+            LLM.request({
+              model: OpenAICompatibleChat.route
+                .with({ provider: item.provider, endpoint: { baseURL: "https://api.custom.test/v1" } })
+                .model({ id: item.model }),
+              messages: [
+                Message.assistant([ToolCallPart.make({ id: item.id, name: "lookup", input: {} })]),
+                Message.tool({ id: item.id, name: "lookup", result: { type: "content", value: [] } }),
+              ],
+            }),
+          )
+
+          expect(prepared.body.messages).toMatchObject([
+            { role: "assistant", tool_calls: [{ id: item.expected }] },
+            { role: "tool", tool_call_id: item.expected },
+          ])
+        }),
+      )
+    }),
+  )
+
+  it.effect("bridges tool results for Mistral-family models and honors compatibility overrides", () =>
+    Effect.gen(function* () {
+      const cases = [
+        { id: "mistral-small", bridge: true },
+        { id: "devstral-small", bridge: true },
+        { id: "codestral-latest", bridge: true },
+        { id: "pixtral-large", bridge: true },
+        { id: "open-mixtral-8x22b", bridge: true },
+        { id: "ordinary-model", bridge: false },
+        { id: "ordinary-model", override: true, bridge: true },
+        { id: "mistral-small", override: false, bridge: false },
+      ] as const
+
+      yield* Effect.forEach(cases, (item) =>
+        Effect.gen(function* () {
+          const selected = OpenAICompatibleChat.route
+            .with({ provider: "custom", endpoint: { baseURL: "https://api.custom.test/v1" } })
+            .model({
+              id: item.id,
+              compatibility: "override" in item ? { requireAssistantAfterTool: item.override } : undefined,
+            })
+          const prepared = yield* compileRequest(
+            LLM.request({
+              model: selected,
+              messages: [
+                Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+                Message.tool({ id: "call_1", name: "lookup", result: "Sunny" }),
+                Message.user("What next?"),
+              ],
+            }),
+          )
+
+          expect(prepared.body.messages.map((message) => message.role)).toEqual(
+            item.bridge ? ["assistant", "tool", "assistant", "user"] : ["assistant", "tool", "user"],
+          )
+          if (item.bridge) expect(prepared.body.messages[2]).toEqual({ role: "assistant", content: "Done." })
+        }),
+      )
+    }),
+  )
+
+  it.effect("requires reasoning for DeepSeek models, providers, and endpoints unless explicitly overridden", () =>
+    Effect.gen(function* () {
+      const cases = [
+        { id: "DeepSeek-V3", provider: "custom", baseURL: "https://api.custom.test/v1", required: true },
+        { id: "custom-model", provider: "deepseek", baseURL: "https://api.custom.test/v1", required: true },
+        { id: "custom-model", provider: "custom", baseURL: "https://API.DeepSeek.COM/v1", required: true },
+        { id: "ordinary-model", provider: "custom", baseURL: "https://api.custom.test/v1", required: false },
+        {
+          id: "ordinary-model",
+          provider: "custom",
+          baseURL: "https://api.custom.test/v1",
+          compatibility: { requireReasoning: true, reasoningField: "reasoning" },
+          required: true,
+          field: "reasoning",
+        },
+        {
+          id: "deepseek-chat",
+          provider: "deepseek",
+          baseURL: "https://api.deepseek.com/v1",
+          compatibility: { requireReasoning: false },
+          required: false,
+        },
+      ] as const
+
+      yield* Effect.forEach(cases, (item) =>
+        Effect.gen(function* () {
+          const selected = OpenAICompatibleChat.route
+            .with({ provider: item.provider, endpoint: { baseURL: item.baseURL } })
+            .model({ id: item.id, compatibility: "compatibility" in item ? item.compatibility : undefined })
+          const prepared = yield* compileRequest(
+            LLM.request({
+              model: selected,
+              messages: [
+                Message.assistant("Hello"),
+                Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+                Message.tool({ id: "call_1", name: "lookup", result: "Sunny" }),
+              ],
+            }),
+          )
+          const field = "field" in item ? item.field : "reasoning_content"
+
+          for (const message of prepared.body.messages.filter((message) => message.role === "assistant")) {
+            if (item.required) expect(message).toHaveProperty(field, "")
+            else expect(message).not.toHaveProperty(field)
+          }
+        }),
+      )
+    }),
+  )
+
  it.effect("posts to the configured compatible endpoint and parses text usage", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
@@ -47,10 +47,8 @@ describe("Open Responses-compatible route", () => {
      })
      expect(prepared.body).toEqual({
        model: "example-model",
-        input: [
-          { role: "system", content: "You are concise." },
-          { role: "user", content: [{ type: "input_text", text: "Say hello." }] },
-        ],
+        input: [{ role: "user", content: [{ type: "input_text", text: "Say hello." }] }],
+        instructions: "You are concise.",
        stream: true,
        store: false,
        include: ["reasoning.encrypted_content"],
@@ -84,10 +82,12 @@ describe("Open Responses-compatible route", () => {
      const prepared = yield* compileRequest(
        LLM.request({
          model,
+          system: "Initial instructions.",
          messages: [Message.user("Before."), Message.system("Operator update."), Message.assistant("After.")],
        }),
      )

+      expect(prepared.body.instructions).toBe("Initial instructions.")
      expect(prepared.body.input).toEqual([
        { role: "user", content: [{ type: "input_text", text: "Before." }] },
        { role: "developer", content: "Operator update." },
@@ -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: "Dropped.",
-                providerMetadata: { openresponses: { itemId: `m${"a".repeat(64)}` } },
+                text: "Long.",
+                providerMetadata: { openresponses: { itemId: `history_${"a".repeat(64)}` } },
              },
+              {
+                type: "text",
+                text: "Opaque.",
+                providerMetadata: { openresponses: { itemId: "provider_value/with+symbols" } },
+              },
+              { type: "text", text: "No suffix.", providerMetadata: { openresponses: { itemId: "msg_" } } },
+              { type: "text", text: "No prefix.", providerMetadata: { openresponses: { itemId: "_item" } } },
            ]),
          ],
        }),
@@ -218,13 +222,128 @@ 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." },
+          ],
        },
      ])
    }),
  )

+  it.effect("replays only shared hosted tool items", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+        provider: "example",
+      }).model("example-model")
+      const items = [
+        { type: "web_search_call", id: "ws_1", status: "completed" },
+        { type: "x_search_call", id: "x_search_1", status: "completed" },
+        { type: "future_call", id: "future_1", status: "completed" },
+        { type: "file_search_call", id: "fs_1", queries: "not-an-array" },
+      ]
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: items.map((item) =>
+            Message.assistant({
+              type: "tool-result",
+              id: item.id,
+              name: item.type,
+              result: { type: "json", value: item },
+              providerExecuted: true,
+              providerMetadata: { openresponses: { itemId: item.id } },
+            }),
+          ),
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        items[0],
+        { role: "user", content: [{ type: "input_text", text: JSON.stringify(items[1]) }] },
+        { role: "user", content: [{ type: "input_text", text: JSON.stringify(items[2]) }] },
+        { role: "user", content: [{ type: "input_text", text: JSON.stringify(items[3]) }] },
+      ])
+    }),
+  )
+
+  it.effect("routes response deltas by output index", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+      }).model("example-model")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Say hello." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 2, item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", output_index: 2, item_id: "wrong_message", delta: "Indexed" },
+              { type: "response.output_item.done", output_index: 2, item: { type: "message", id: "msg_1" } },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        { type: "text", text: "Indexed", providerMetadata: { openresponses: { itemId: "msg_1" } } },
+      ])
+    }),
+  )
+
+  it.effect("streams function calls without optional item ids through the shared baseline", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+        provider: "example",
+      }).model("example-model")
+      const item = { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" }
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Look it up." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 1, item },
+              {
+                type: "response.function_call_arguments.delta",
+                output_index: 1,
+                item_id: "opaque_item",
+                delta: '{"query":"shared"}',
+              },
+              {
+                type: "response.output_item.done",
+                output_index: 1,
+                item: { ...item, arguments: '{"query":"complete"}' },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "complete" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+    }),
+  )
+
  it.effect("finalizes pending function calls from completed response output", () =>
    Effect.gen(function* () {
      const model = configure({
@@ -136,6 +136,7 @@ describe("OpenAI Responses WebSocket recorded", () => {
      expect(channel.opens()).toBe(1)
      expect(channel.sent).toHaveLength(2)
      expect(channel.sent[1]).toMatchObject({
+        instructions: "Call get_weather once, then reply exactly: Paris is sunny.",
        previous_response_id: expect.any(String),
        input: [{ type: "function_call_output", call_id: call.id, output: expect.any(String) }],
      })
@@ -167,8 +168,8 @@ describe("OpenAI Responses WebSocket recorded", () => {
      expect(channel.opens()).toBe(2)
      expect(channel.sent[1]).not.toHaveProperty("previous_response_id")
      expect(channel.sent[1]).toMatchObject({
+        instructions: "Follow the user's exact reply instruction.",
        input: [
-          { role: "system", content: "Follow the user's exact reply instruction." },
          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Alpha." }] },
          { role: "assistant", content: [{ type: "output_text", text: "Alpha." }] },
          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Beta." }] },
@@ -204,8 +205,8 @@ describe("OpenAI Responses WebSocket recorded", () => {
      expect(channel.sent[1]).toHaveProperty("previous_response_id", expect.any(String))
      expect(channel.sent[2]).not.toHaveProperty("previous_response_id")
      expect(channel.sent[2]).toMatchObject({
+        instructions: "Follow the user's exact reply instruction.",
        input: [
-          { role: "system", content: "Follow the user's exact reply instruction." },
          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Ready." }] },
          { role: "assistant", content: [{ type: "output_text", text: "Ready." }] },
          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Recovered." }] },
@@ -112,10 +112,8 @@ describe("OpenAI Responses route", () => {

      expect(prepared.body).toEqual({
        model: "gpt-4.1-mini",
-        input: [
-          { role: "system", content: "You are concise." },
-          { role: "user", content: [{ type: "input_text", text: "Say hello." }] },
-        ],
+        input: [{ role: "user", content: [{ type: "input_text", text: "Say hello." }] }],
+        instructions: "You are concise.",
        store: false,
        include: ["reasoning.encrypted_content"],
        stream: true,
@@ -469,7 +467,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("continues a tool call with only the new tool output", () =>
+  it.effect("continues an item-id-less tool call with only the new tool output", () =>
    Effect.gen(function* () {
      const firstRequest = {
        type: "response.create",
@@ -485,7 +483,6 @@ describe("OpenAI Responses route", () => {
          type: "response.output_item.done",
          item: {
            type: "function_call",
-            id: "fc_1",
            status: "completed",
            call_id: "call_1",
            name: "weather",
@@ -830,7 +827,9 @@ describe("OpenAI Responses route", () => {
          webSocket: {
            execute: (exchange) =>
              Effect.gen(function* () {
-                yield* exchange.driver.create(undefined).pipe(Effect.flatMap((create) => Ref.set(message, create.message)))
+                yield* exchange.driver
+                  .create(undefined)
+                  .pipe(Effect.flatMap((create) => Ref.set(message, create.message)))
                return { frames: exchange.fallback(), complete: Effect.void }
              }),
          },
@@ -1595,8 +1594,8 @@ describe("OpenAI Responses route", () => {
      )

      expect(prepared.body).toMatchObject({
+        instructions: "You are concise. Continue from the provided history.",
        input: [
-          { role: "system", content: "You are concise. Continue from the provided history." },
          {
            role: "user",
            content: [
@@ -2051,6 +2050,204 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("routes assistant text by output index when its item id disagrees", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 2, item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", output_index: 2, item_id: "wrong_message", delta: "Indexed" },
+              { type: "response.output_item.done", output_index: 2, item: { type: "message", id: "msg_1" } },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("Indexed")
+      expect(response.message.content).toEqual([
+        { type: "text", text: "Indexed", providerMetadata: { openai: { itemId: "msg_1" } } },
+      ])
+    }),
+  )
+
+  it.effect("routes interleaved function calls by output index", () =>
+    Effect.gen(function* () {
+      const first = { type: "function_call", id: "fc_1", call_id: "call_1", name: "first", arguments: "" }
+      const second = { type: "function_call", id: "fc_2", call_id: "call_2", name: "second", arguments: "" }
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 1, item: first },
+              { type: "response.output_item.added", output_index: 3, item: second },
+              { type: "response.function_call_arguments.delta", output_index: 1, item_id: "fc_2", delta: '{"a":' },
+              { type: "response.function_call_arguments.delta", output_index: 3, item_id: "fc_1", delta: '{"b":' },
+              {
+                type: "response.function_call_arguments.done",
+                output_index: 3,
+                item_id: "fc_1",
+                arguments: '{"b":2}',
+              },
+              {
+                type: "response.function_call_arguments.done",
+                output_index: 1,
+                item_id: "fc_2",
+                arguments: '{"a":1}',
+              },
+              { type: "response.output_item.done", output_index: 1, item: { ...first, arguments: '{"a":1}' } },
+              { type: "response.output_item.done", output_index: 3, item: { ...second, arguments: '{"b":2}' } },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter((event) => event.type === "tool-input-delta")).toMatchObject([
+        { id: "call_1", text: '{"a":' },
+        { id: "call_2", text: '{"b":' },
+        { id: "call_2", text: "2}" },
+        { id: "call_1", text: "1}" },
+      ])
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "first", input: { a: 1 } }),
+        expect.objectContaining({ id: "call_2", name: "second", input: { b: 2 } }),
+      ])
+    }),
+  )
+
+  it.effect("routes item-id-less function arguments by output index and prefers item completion", () =>
+    Effect.gen(function* () {
+      const item = { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" }
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 2, item },
+              {
+                type: "response.function_call_arguments.delta",
+                output_index: 2,
+                item_id: "opaque_delta",
+                delta: '{"query":"streamed"}',
+              },
+              {
+                type: "response.function_call_arguments.done",
+                output_index: 2,
+                item_id: "opaque_done",
+                arguments: '{"query":"arguments-done"}',
+              },
+              {
+                type: "response.output_item.done",
+                output_index: 2,
+                item: { ...item, arguments: '{"query":"output-item-done"}' },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter((event) => event.type === "tool-input-delta")).toMatchObject([
+        { id: "call_1", text: '{"query":"streamed"}' },
+      ])
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "output-item-done" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+    }),
+  )
+
+  it.effect("routes reasoning summary events by output index", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                output_index: 4,
+                item: { type: "reasoning", id: "rs_1" },
+              },
+              {
+                type: "response.reasoning_summary_part.added",
+                output_index: 4,
+                item_id: "wrong_reasoning",
+                summary_index: 0,
+              },
+              {
+                type: "response.reasoning_summary_text.delta",
+                output_index: 4,
+                item_id: "wrong_reasoning",
+                summary_index: 0,
+                delta: "Thinking",
+              },
+              {
+                type: "response.reasoning_summary_part.done",
+                output_index: 4,
+                item_id: "wrong_reasoning",
+                summary_index: 0,
+              },
+              {
+                type: "response.output_item.done",
+                output_index: 4,
+                item: { type: "reasoning", id: "rs_1", encrypted_content: "state" },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Thinking")
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Thinking",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "state" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("routes native reasoning text deltas by output index", () =>
+    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: "rs_1" } },
+              { type: "response.reasoning_text.delta", output_index: 1, item_id: "wrong_reasoning", delta: "Raw" },
+              { type: "response.output_item.done", output_index: 1, item: { type: "reasoning", id: "rs_1" } },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Raw")
+    }),
+  )
+
+  it.effect("falls back to item ids when an output index was not registered", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", output_index: 9, item_id: "msg_1", delta: "Fallback" },
+              { type: "response.output_item.done", item: { type: "message", id: "msg_1" } },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("Fallback")
+    }),
+  )
+
  it.effect("rejects output text events without the spec-required item id", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
@@ -2070,6 +2267,25 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("requires item ids even when their output index is known", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 0, item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", output_index: 0, delta: "Missing item ID" },
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      expect(error.message).toContain("response.output_text.delta is missing item_id")
+    }),
+  )
+
  it.effect("ignores deltas without a matching output item", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
@@ -2120,13 +2336,25 @@ describe("OpenAI Responses route", () => {
  it.effect("rejects function argument events without the spec-required item id", () =>
    Effect.gen(function* () {
      const events = [
-        { type: "response.function_call_arguments.delta", delta: "{}" },
-        { type: "response.function_call_arguments.done", arguments: "{}" },
+        { type: "response.function_call_arguments.delta", output_index: 0, delta: "{}" },
+        { type: "response.function_call_arguments.done", output_index: 0, arguments: "{}" },
      ]

      for (const event of events) {
        const error = yield* LLMClient.generate(request).pipe(
-          Effect.provide(fixedResponse(sseEvents(event, { type: "response.completed", response: { id: "resp_1" } }))),
+          Effect.provide(
+            fixedResponse(
+              sseEvents(
+                {
+                  type: "response.output_item.added",
+                  output_index: 0,
+                  item: { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" },
+                },
+                event,
+                { type: "response.completed", response: { id: "resp_1" } },
+              ),
+            ),
+          ),
          Effect.flip,
        )

@@ -2580,7 +2808,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("closes reasoning summary parts when storage is not disabled", () =>
+  it.effect("preserves final reasoning metadata when storage is enabled", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(LLMRequest.update(request, { providerOptions: { store: true } })).pipe(
        Effect.provide(
@@ -2598,7 +2826,7 @@ describe("OpenAI Responses route", () => {
              { type: "response.reasoning_summary_part.done", item_id: "rs_1", summary_index: 1 },
              {
                type: "response.output_item.done",
-                item: { type: "reasoning", id: "rs_1", encrypted_content: null },
+                item: { type: "reasoning", id: "rs_1", encrypted_content: "encrypted-state" },
              },
              { type: "response.completed", response: { id: "resp_1" } },
            ),
@@ -2608,7 +2836,11 @@ describe("OpenAI Responses route", () => {

      expect(response.events.filter((event) => event.type === "reasoning-end")).toEqual([
        { type: "reasoning-end", id: "rs_1:0", providerMetadata: { openai: { itemId: "rs_1" } } },
-        { type: "reasoning-end", id: "rs_1:1", providerMetadata: { openai: { itemId: "rs_1" } } },
+        {
+          type: "reasoning-end",
+          id: "rs_1:1",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        },
      ])
    }),
  )
@@ -2713,7 +2945,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("references stored reasoning items by id", () =>
+  it.effect("replays complete reasoning items when storage is enabled", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
        LLM.request({
@@ -2723,7 +2955,7 @@ describe("OpenAI Responses route", () => {
              {
                type: "reasoning",
                text: "Checked the previous diff.",
-                providerMetadata: { openai: { itemId: "rs_1" } },
+                providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
              },
            ]),
          ],
@@ -2731,12 +2963,20 @@ describe("OpenAI Responses route", () => {
        }),
      )

-      expect(prepared.body.input).toEqual([{ type: "item_reference", id: "rs_1" }])
+      expect(prepared.body.input).toEqual([
+        {
+          type: "reasoning",
+          id: "rs_1",
+          summary: [{ type: "summary_text", text: "Checked the previous diff." }],
+          encrypted_content: "encrypted-state",
+        },
+      ])
    }),
  )

-  it.effect("references stored provider-executed hosted tool results by id", () =>
+  it.effect("replays complete hosted tool items when storage is enabled", () =>
    Effect.gen(function* () {
+      const item = { type: "web_search_call", id: "ws_1", status: "completed" }
      const prepared = yield* compileRequest(
        LLM.request({
          model,
@@ -2753,7 +2993,7 @@ describe("OpenAI Responses route", () => {
                type: "tool-result",
                id: "ws_1",
                name: "web_search",
-                result: { type: "json", value: { type: "web_search_call", id: "ws_1", status: "completed" } },
+                result: { type: "json", value: item },
                providerExecuted: true,
                providerMetadata: { openai: { itemId: "ws_1" } },
              },
@@ -2765,14 +3005,15 @@ describe("OpenAI Responses route", () => {
      )

      expect(prepared.body.input).toEqual([
-        { type: "item_reference", id: "ws_1" },
+        item,
        { role: "user", content: [{ type: "input_text", text: "Continue." }] },
      ])
    }),
  )

-  it.effect("continues stateless hosted tool results with their text form", () =>
+  it.effect("replays stateless hosted tool results as native provider items", () =>
    Effect.gen(function* () {
+      const item = { type: "web_search_call", id: "ws_1", status: "completed" }
      const prepared = yield* compileRequest(
        LLM.request({
          model,
@@ -2790,7 +3031,7 @@ describe("OpenAI Responses route", () => {
                type: "tool-result",
                id: "ws_1",
                name: "web_search",
-                result: { type: "json", value: { type: "web_search_call", id: "ws_1", status: "completed" } },
+                result: { type: "json", value: item },
                providerExecuted: true,
                providerMetadata: { openai: { itemId: "ws_1" } },
              },
@@ -2803,6 +3044,74 @@ describe("OpenAI Responses route", () => {

      expect(prepared.body.input).toEqual([
        { role: "user", content: [{ type: "input_text", text: "Search." }] },
+        { type: "web_search_call", id: "ws_1", status: "completed" },
+        { role: "user", content: [{ type: "input_text", text: "Continue." }] },
+      ])
+    }),
+  )
+
+  it.effect("replays OpenAI hosted tool extensions but rejects foreign and unknown items", () =>
+    Effect.gen(function* () {
+      const items = [
+        { type: "computer_call", id: "computer_1", status: "completed", action: { type: "click", x: 1, y: 2 } },
+        { type: "x_search_call", id: "x_search_1", status: "completed" },
+        { type: "future_call", id: "future_1", status: "completed" },
+      ]
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: items.map((item) =>
+            Message.assistant({
+              type: "tool-result",
+              id: item.id,
+              name: item.type,
+              result: { type: "json", value: item },
+              providerExecuted: true,
+              providerMetadata: { openai: { itemId: item.id } },
+            }),
+          ),
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        items[0],
+        { role: "user", content: [{ type: "input_text", text: JSON.stringify(items[1]) }] },
+        { role: "user", content: [{ type: "input_text", text: JSON.stringify(items[2]) }] },
+      ])
+    }),
+  )
+
+  it.effect("preserves foreign hosted tool results as portable message content when storage is enabled", () =>
+    Effect.gen(function* () {
+      const item = { type: "web_search_call", id: "ws_1", status: "completed" }
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: xaiModel,
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({
+                id: "ws_1",
+                name: "web_search",
+                input: { query: "effect 4" },
+                providerExecuted: true,
+                providerMetadata: { openai: { itemId: "ws_1" } },
+              }),
+              {
+                type: "tool-result",
+                id: "ws_1",
+                name: "web_search",
+                result: { type: "json", value: item },
+                providerExecuted: true,
+                providerMetadata: { openai: { itemId: "ws_1" } },
+              },
+            ]),
+            Message.user("Continue."),
+          ],
+          providerOptions: { store: true },
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
        {
          role: "user",
          content: [{ type: "input_text", text: '{"type":"web_search_call","id":"ws_1","status":"completed"}' }],
@@ -2812,38 +3121,79 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("drops replayed item ids outside the server's grammar", () =>
+  it.effect("does not replay hosted tool items whose result id differs from provider metadata", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              {
+                type: "tool-result",
+                id: "ws_1",
+                name: "web_search",
+                result: { type: "json", value: { type: "web_search_call", id: "ws_other", status: "completed" } },
+                providerExecuted: true,
+                providerMetadata: { openai: { itemId: "ws_1" } },
+              },
+            ]),
+          ],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        {
+          role: "user",
+          content: [{ type: "input_text", text: '{"type":"web_search_call","id":"ws_other","status":"completed"}' }],
+        },
+      ])
+    }),
+  )
+
+  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_" } },
+              }),
            ]),
          ],
        }),
@@ -2852,41 +3202,70 @@ 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"}',
+        },
      ])
    }),
  )

-  it.effect("keeps well-formed hosted references and drops malformed ones under storage", () =>
+  it.effect("falls back to portable hosted results when stored item metadata is malformed", () =>
    Effect.gen(function* () {
-      const hostedResult = (itemId: string) => [
-        ToolCallPart.make({
-          id: itemId,
-          name: "web_search",
-          input: { query: "effect 4" },
-          providerExecuted: true,
-          providerMetadata: { openai: { itemId } },
-        }),
-        {
-          type: "tool-result" as const,
-          id: itemId,
-          name: "web_search",
-          result: { type: "json" as const, value: { status: "completed" } },
-          providerExecuted: true as const,
-          providerMetadata: { openai: { itemId } },
-        },
-      ]
+      const hostedResult = (itemId: string) => {
+        const item = { type: "web_search_call", id: itemId, status: "completed" }
+        return [
+          ToolCallPart.make({
+            id: itemId,
+            name: "web_search",
+            input: { query: "effect 4" },
+            providerExecuted: true,
+            providerMetadata: { openai: { itemId } },
+          }),
+          {
+            type: "tool-result" as const,
+            id: itemId,
+            name: "web_search",
+            result: { type: "json" as const, value: item },
+            providerExecuted: true as const,
+            providerMetadata: { openai: { itemId } },
+          },
+        ]
+      }
      const prepared = yield* compileRequest(
        LLM.request({
          model,
@@ -2895,7 +3274,13 @@ describe("OpenAI Responses route", () => {
        }),
      )

-      expect(prepared.body.input).toEqual([{ type: "item_reference", id: "ws_1" }])
+      expect(prepared.body.input).toEqual([
+        { type: "web_search_call", id: "ws_1", status: "completed" },
+        {
+          role: "user",
+          content: [{ type: "input_text", text: '{"type":"web_search_call","id":"bad ref","status":"completed"}' }],
+        },
+      ])
    }),
  )

@@ -2941,6 +3326,43 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("preserves foreign hosted images as portable image content when storage is enabled", () =>
+    Effect.gen(function* () {
+      const item = { type: "image_generation_call", id: "ig_1", status: "completed", result: "AQID" }
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: xaiModel,
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({
+                id: "ig_1",
+                name: "image_generation",
+                input: {},
+                providerExecuted: true,
+                providerMetadata: { openai: { itemId: "ig_1" } },
+              }),
+              ToolResultPart.make({
+                id: "ig_1",
+                name: "image_generation",
+                result: {
+                  type: "content",
+                  value: [{ type: "file", uri: "data:image/png;base64,AQID", mime: "image/png" }],
+                },
+                providerExecuted: true,
+                providerMetadata: { openai: { itemId: "ig_1" } },
+              }),
+            ]),
+          ],
+          providerOptions: { store: true },
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        { role: "user", content: [{ type: "input_image", image_url: "data:image/png;base64,AQID" }] },
+      ])
+    }),
+  )
+
  it.effect("joins streamed summary blocks into one continuation reasoning item", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -2979,7 +3401,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("skips non-persisted reasoning ids without encrypted state", () =>
+  it.effect("replays stateless reasoning without encrypted state", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
        LLM.request({
@@ -3009,6 +3431,12 @@ describe("OpenAI Responses route", () => {
      expect(prepared.body).toMatchObject({
        input: [
          { role: "user", content: [{ type: "input_text", text: "What changed?" }] },
+          {
+            type: "reasoning",
+            id: "rs_1",
+            summary: [{ type: "summary_text", text: "Checked the previous diff." }],
+            encrypted_content: null,
+          },
          { role: "assistant", content: [{ type: "output_text", text: "The parser changed." }] },
          { role: "user", content: [{ type: "input_text", text: "Summarize it." }] },
        ],
@@ -3117,6 +3545,43 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("finalizes and replays a completed function call without an optional item id", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.done",
+                item: { type: "function_call", call_id: "call_1", name: "lookup", arguments: '{"query":"weather"}' },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "weather" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            response.message,
+            Message.tool({ id: "call_1", name: "lookup", resultType: "json", result: { forecast: "sunny" } }),
+          ],
+        }),
+      )
+      expect(prepared.body.input).toEqual([
+        { type: "function_call", call_id: "call_1", name: "lookup", arguments: '{"query":"weather"}' },
+        { type: "function_call_output", call_id: "call_1", output: '{"forecast":"sunny"}' },
+      ])
+    }),
+  )
+
  it.effect("emits only missing function arguments from the arguments done event", () =>
    Effect.gen(function* () {
      const body = sseEvents(
@@ -3341,6 +3806,37 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("reconciles an item-id-less pending function call from completed response output", () =>
+    Effect.gen(function* () {
+      const item = { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" }
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 0, item },
+              {
+                type: "response.function_call_arguments.delta",
+                output_index: 0,
+                item_id: "opaque_delta",
+                delta: '{"query":"partial',
+              },
+              {
+                type: "response.completed",
+                response: { id: "resp_1", output: [{ ...item, arguments: '{"query":"complete"}' }] },
+              },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "complete" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+      expect(response.events.filter(LLMEvent.is.toolInputEnd)).toHaveLength(1)
+    }),
+  )
+
  it.effect("lets completed response output override arguments done", () =>
    Effect.gen(function* () {
      const body = sseEvents(
@@ -3644,13 +4140,15 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("decodes image generation output as image content", () =>
+  it.effect("replays hosted image results as portable content regardless of storage", () =>
    Effect.gen(function* () {
      const item = {
        type: "image_generation_call",
        id: "ig_1",
        status: "completed",
        result: "AQID",
+        action: "generate",
+        output_format: "png",
      }
      const response = yield* LLMClient.generate(request).pipe(
        Effect.provide(
@@ -3663,6 +4161,9 @@ describe("OpenAI Responses route", () => {
        ),
      )

+      expect(response.events.find(LLMEvent.is.toolCall)).toMatchObject({
+        providerMetadata: { openai: { itemId: "ig_1" } },
+      })
      expect(response.events.find(LLMEvent.is.toolResult)).toMatchObject({
        id: "ig_1",
        name: "image_generation",
@@ -3671,7 +4172,52 @@ describe("OpenAI Responses route", () => {
          type: "content",
          value: [{ type: "file", uri: "data:image/png;base64,AQID", mime: "image/png" }],
        },
+        providerMetadata: { openai: { itemId: "ig_1" } },
      })
+
+      const prepared = yield* Effect.forEach([false, true], (store) =>
+        compileRequest(LLM.request({ model, messages: [response.message], providerOptions: { store } })),
+      )
+      expect(prepared.map((request) => request.body.input)).toEqual([
+        [{ role: "user", content: [{ type: "input_image", image_url: "data:image/png;base64,AQID" }] }],
+        [{ role: "user", content: [{ type: "input_image", image_url: "data:image/png;base64,AQID" }] }],
+      ])
+    }),
+  )
+
+  it.effect("preserves failed hosted tool results as portable error content", () =>
+    Effect.gen(function* () {
+      const item = {
+        type: "web_search_call",
+        id: "ws_failed",
+        status: "failed",
+        error: { code: "search_failed", message: "Search unavailable" },
+      }
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.done", item },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.find(LLMEvent.is.toolResult)).toMatchObject({
+        result: { type: "error", value: item.error },
+        providerMetadata: { openai: { itemId: "ws_failed" } },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model, messages: [response.message], providerOptions: { store: true } }),
+      )
+      expect(prepared.body.input).toEqual([
+        {
+          role: "user",
+          content: [{ type: "input_text", text: '{"code":"search_failed","message":"Search unavailable"}' }],
+        },
+      ])
    }),
  )

@@ -1,6 +1,6 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
-import { LLM, LLMEvent } from "../../src/index.js"
+import { LLM, LLMEvent, Message } from "../../src/index.js"
 import { XAI } from "../../src/providers.js"
 import { OpenResponses } from "../../src/protocols/open-responses.js"
 import { OpenAIResponses } from "../../src/protocols/openai-responses.js"
@@ -14,9 +14,9 @@ import { sseEvents } from "../lib/sse.js"
 const model = XAI.configure({ apiKey: "test", baseURL: "https://api.x.ai/v1" }).responses("grok-4.6")

 describe("xAI Responses route", () => {
-  it.effect("extends the Open Responses baseline directly", () =>
+  it.effect("composes the Open Responses baseline with xAI extensions", () =>
    Effect.gen(function* () {
-      expect(XAIResponses.protocol.body).toBe(OpenResponses.protocol.body)
+      expect(XAIResponses.protocol.body).not.toBe(OpenResponses.protocol.body)
      expect(XAIResponses.protocol.body).not.toBe(OpenAIResponses.protocol.body)

      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Hello" }))
@@ -70,16 +70,106 @@ describe("xAI Responses route", () => {
    }),
  )

-  it.effect("parses xAI hosted tool items", () =>
+  it.effect("routes xAI reasoning summaries by output index", () =>
    Effect.gen(function* () {
-      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Search X" })).pipe(
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think" })).pipe(
        Effect.provide(
          fixedResponse(
            sseEvents(
              {
-                type: "response.output_item.done",
-                item: { type: "x_search_call", id: "x_search_1", status: "completed", action: { query: "news" } },
+                type: "response.output_item.added",
+                output_index: 3,
+                item: { type: "reasoning", id: "reasoning_1" },
              },
+              {
+                type: "response.reasoning_summary_text.delta",
+                output_index: 3,
+                item_id: "wrong_reasoning",
+                summary_index: 0,
+                delta: "Considering.",
+              },
+              {
+                type: "response.output_item.done",
+                output_index: 3,
+                item: { type: "reasoning", id: "reasoning_1", encrypted_content: "opaque" },
+              },
+              { type: "response.completed", response: { id: "response_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Considering.")
+      expect(response.message.content.find((part) => part.type === "reasoning")).toMatchObject({
+        providerMetadata: { xai: { itemId: "reasoning_1", reasoningEncryptedContent: "opaque" } },
+      })
+    }),
+  )
+
+  it.effect("replays xAI hosted tool items when continuing with the same provider", () =>
+    Effect.gen(function* () {
+      const item = { type: "x_search_call", id: "x_search_1", status: "completed", action: { query: "news" } }
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              {
+                type: "tool-result",
+                id: "x_search_1",
+                name: "x_search",
+                result: { type: "json", value: item },
+                providerExecuted: true,
+                providerMetadata: { xai: { itemId: "x_search_1" } },
+              },
+            ]),
+          ],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([item])
+    }),
+  )
+
+  it.effect("replays shared and xAI hosted tool items but rejects OpenAI extensions", () =>
+    Effect.gen(function* () {
+      const items = [
+        { type: "web_search_call", id: "ws_1", status: "completed" },
+        { type: "image_generation_call", id: "ig_1", status: "completed", result: "AQID" },
+        { type: "computer_call", id: "computer_1", status: "completed" },
+      ]
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: items.map((item) =>
+            Message.assistant({
+              type: "tool-result",
+              id: item.id,
+              name: item.type,
+              result: { type: "json", value: item },
+              providerExecuted: true,
+              providerMetadata: { xai: { itemId: item.id } },
+            }),
+          ),
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        items[0],
+        items[1],
+        { role: "user", content: [{ type: "input_text", text: JSON.stringify(items[2]) }] },
+      ])
+    }),
+  )
+
+  it.effect("parses xAI hosted tool items", () =>
+    Effect.gen(function* () {
+      const item = { type: "x_search_call", id: "x_search_1", status: "completed", action: { query: "news" } }
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Search X" })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.done", item },
              { type: "response.completed", response: { id: "response_1" } },
            ),
          ),
@@ -91,6 +181,11 @@ describe("xAI Responses route", () => {
        name: "x_search",
        input: { query: "news" },
        providerExecuted: true,
+        providerMetadata: { xai: { itemId: "x_search_1" } },
+      })
+      expect(response.events.find(LLMEvent.is.toolResult)).toMatchObject({
+        result: { type: "json", value: item },
+        providerMetadata: { xai: { itemId: "x_search_1" } },
      })
    }),
  )
@@ -20,6 +20,7 @@ type ScenarioInput =
      readonly name?: string
      readonly cassette?: string
      readonly tags?: ReadonlyArray<string>
+      readonly prompt?: string
      readonly maxTokens?: number
      readonly temperature?: number | false
      readonly timeout?: number
@@ -87,6 +88,7 @@ const runTarget = (target: TargetInput) => {
            yield* runGoldenScenario(input.id, {
              id: `recorded_${kebab(target.name).replaceAll("-", "_")}_${input.id.replaceAll("-", "_")}`,
              model: target.model,
+              prompt: input.prompt,
              maxTokens: input.maxTokens,
              temperature: input.temperature,
            })
@@ -164,6 +164,7 @@ export const expectGoldenWeatherToolLoop = (events: ReadonlyArray<LLMEvent>) =>
 export interface GoldenScenarioContext {
  readonly id: string
  readonly model: LanguageModel
+  readonly prompt?: string
  readonly maxTokens?: number
  readonly temperature?: number | false
 }
@@ -298,7 +299,7 @@ const runGeneratedConversation = (context: GoldenScenarioContext, steps: Readonl

 const runTextScenario = (context: GoldenScenarioContext) =>
  runGeneratedConversation(context, [
-    user("Reply exactly with: Hello!"),
+    user(context.prompt ?? "Reply exactly with: Hello!"),
    assistant.expectText(/^Hello!?$/, {
      system: "You are concise.",
      maxTokens: context.maxTokens ?? 40,
@@ -102,6 +102,38 @@ describe("AI.Usage", () => {
    expect(error.reason._tag).toBe("InvalidProviderOutput")
  })

+  test("sseFraming ignores retry directives without ending the stream", async () => {
+    const encoder = new TextEncoder()
+    const frames = await Effect.runPromise(
+      ProviderShared.sseFraming(
+        Stream.make(
+          encoder.encode("retry: 1000\n\n"),
+          encoder.encode('data: {"first":true}\n\n'),
+          encoder.encode("retry: 2000\n\n"),
+          encoder.encode('data: {"second":true}\n\n'),
+        ).pipe(Stream.rechunk(1)),
+      ).pipe(Stream.runCollect),
+    )
+
+    expect(Array.from(frames)).toEqual(['{"first":true}', '{"second":true}'])
+  })
+
+  test("sseFraming preserves event data around retry directives", async () => {
+    const encoder = new TextEncoder()
+    const frames = await Effect.runPromise(
+      ProviderShared.sseFraming(
+        Stream.make(
+          encoder.encode("event: update\ndata: first\n"),
+          encoder.encode("retry: 1000\n"),
+          encoder.encode("data: second\n\n"),
+        ).pipe(Stream.rechunk(1)),
+        new Set(["update"]),
+      ).pipe(Stream.runCollect),
+    )
+
+    expect(Array.from(frames)).toEqual(["first\nsecond"])
+  })
+
  test("visibleOutputTokens clamps reasoning > output to zero", () => {
    expect(new Usage({ outputTokens: 10, reasoningTokens: 4 }).visibleOutputTokens).toBe(6)
    expect(new Usage({ outputTokens: 10 }).visibleOutputTokens).toBe(10)
@@ -500,7 +500,7 @@ export const dict = {
  "context.stats.lastActivity": "Last Activity",

  "context.usage.tokens": "Tokens",
-  "context.usage.usage": "Usage",
+  "context.usage.usage": "Context Usage",
  "context.usage.cost": "Cost",
  "context.usage.clickToView": "Click to view context",
  "context.usage.view": "View context usage",
@@ -92,6 +92,15 @@ export function createTimelineVirtualizer(input: Input) {
    initialOffset: () => (input.pinned() ? Number.MAX_SAFE_INTEGER : 0),
    initialMeasurementsCache: initialMeasurements,
    estimateSize: () => fallbackItemSize,
+    // Do not replace this with TanStack's default measurer: without a ResizeObserver entry,
+    // it returns the cached height instead of reading the element (TanStack/virtual#1183).
+    // Restored sessions, deferred tools, and rewrapped content can then keep stale heights;
+    // our fixed-height, overflow-clipped rows will hide their content. Keep observer entries
+    // on the cheap precomputed path, but make explicit measurements read the real height.
+    measureElement: (element, entry) => {
+      const box = entry?.borderBoxSize[0]
+      return box ? Math.round(box.blockSize) : element.offsetHeight
+    },
    scrollToFn: (offset, options, instance) => {
      if (virtualContent) virtualContent.style.height = `${instance.getTotalSize()}px`
      elementScroll(offset, options, instance)
@@ -11,7 +11,8 @@ import { useLanguage } from "@/runtime/i18n/language"
 import { getProjectAvatarVariant, type LocalProject } from "@/shell/state/layout"
 import { ServerConnection } from "@/runtime/server/registry"
 import { LocationProvider } from "@/workspaces/location"
-import { displayName, getProjectAvatarSource } from "@/shell/layout/helpers"
+import { displayName } from "@/shell/layout/helpers"
+import { ProjectIcon } from "@/shell/layout/project-icon"
 import { createEditProjectModel } from "./project-model"
 import { ProjectSettingsExtensions } from "./project-extensions"
 import { SettingsServerDataScope } from "@/settings/server-scope"
@@ -57,11 +58,7 @@ function ProjectSettingsDialog(props: { project: LocalProject; server: ServerCon
        <Tabs.List>
          <div class="project-settings-nav">
            <Tabs.Trigger value="general">
-              <ProjectAvatar
-                fallback={projectName()}
-                variant={getProjectAvatarVariant(props.project.icon?.color)}
-                class="!size-4 shrink-0"
-              />
+              <ProjectIcon project={props.project} class="!size-4 shrink-0" />
              <span class="truncate">{projectName()}</span>
            </Tabs.Trigger>
            <Tabs.Trigger value="scripts">
@@ -114,14 +111,14 @@ function ProjectSettingsDialog(props: { project: LocalProject; server: ServerCon
                    onDragLeave={model.dragLeave}
                    onClick={model.iconClick}
                  >
-                    <ProjectAvatar
+                    <ProjectIcon
+                      project={props.project}
                      fallback={model.store.name || model.defaultName()}
-                      src={getProjectAvatarSource(props.project.id, {
+                      icon={{
                        color: model.store.color,
                        url: props.project.icon?.url,
                        override: model.store.iconOverride,
-                      })}
-                      variant={getProjectAvatarVariant(model.store.color)}
+                      }}
                      class="!size-16 [&_[data-slot=project-avatar-surface]]:!rounded-[6px] [&_[data-slot=project-avatar-surface]]:!text-[32px]"
                    />
                    <span
@@ -1,13 +1,12 @@
 import { Component, For, Show, createMemo, createSignal } from "solid-js"
 import { IconButton } from "@opencode-ai/ui/icon-button"
 import { Icon } from "@opencode-ai/ui/icon"
-import { ProjectAvatar } from "@opencode-ai/ui/project-avatar"
 import { useDialog } from "@opencode-ai/ui/context/dialog"
 import { useLanguage } from "@/runtime/i18n/language"
 import { useGlobal } from "@/runtime/server/runtime"
-import { getProjectAvatarVariant } from "@/shell/state/layout"
 import { ServerConnection, serverName } from "@/runtime/server/registry"
 import { displayName } from "@/shell/layout/helpers"
+import { ProjectIcon } from "@/shell/layout/project-icon"
 import { InlineServerSelect } from "@/settings/server-select"
 import { DialogEditProject } from "./project-dialog"
 import "@/settings/settings.css"
@@ -40,14 +39,13 @@ export const SettingsProjects: Component = () => {

  const ProjectRow: Component<{ project: ProjectItem; server: ServerConnection.Any }> = (props) => {
    const name = () => displayName(props.project)
-    const color = () => getProjectAvatarVariant(props.project.icon?.color)
    return (
      <div
        class="group flex items-center justify-between gap-5 px-4 py-2.5 rounded-lg bg-v2-background-bg-base shadow-[var(--v2-elevation-raised)] cursor-pointer transition-all hover:bg-v2-background-bg-layer-01"
        onClick={() => openProjectSettings(props.project, props.server)}
      >
        <div class="flex items-center gap-2.5 min-w-0 flex-1">
-          <ProjectAvatar fallback={name()} variant={color()} class="shrink-0" />
+          <ProjectIcon project={props.project} class="shrink-0" />
          <span class="text-13-medium text-v2-text-text-base truncate">{name()}</span>
        </div>
        <div class="flex items-center gap-2 shrink-0 opacity-0 group-hover:opacity-100 transition-opacity">
@@ -0,0 +1,24 @@
+import { ProjectAvatar, type ProjectAvatarProps } from "@opencode-ai/ui/project-avatar"
+import { splitProps } from "solid-js"
+import { displayName, getProjectAvatarSource } from "@/shell/layout/helpers"
+import { getProjectAvatarVariant, type LocalProject } from "@/shell/state/layout"
+
+type ProjectIconProps = Omit<ProjectAvatarProps, "fallback" | "src" | "variant"> & {
+  project: LocalProject
+  fallback?: string
+  icon?: LocalProject["icon"]
+}
+
+export function ProjectIcon(props: ProjectIconProps) {
+  const [local, rest] = splitProps(props, ["project", "fallback", "icon"])
+  const icon = () => local.icon ?? local.project.icon
+
+  return (
+    <ProjectAvatar
+      {...rest}
+      fallback={local.fallback ?? displayName(local.project)}
+      src={getProjectAvatarSource(local.project.id, icon())}
+      variant={getProjectAvatarVariant(icon()?.color)}
+    />
+  )
+}
@@ -6,6 +6,7 @@ import { LocalProvider } from "@/providers/models/selection"
 import type { ServerConnection } from "@/runtime/server/registry"
 import { sessionHref } from "@/shell/routes/session"
 import { useData } from "@/runtime/server/current"
+import { useServerSDK } from "@/runtime/server/client"
 import { useTabs } from "@/shell/tabs/tabs"

 export function SessionUIProvider(
@@ -17,6 +18,7 @@ export function SessionUIProvider(
  const navigate = useNavigate()
  const params = useParams()
  const data = useData()
+  const serverSDK = useServerSDK()
  const tabs = useTabs()
  const directory = () => props.directory
  const href = (sessionID: string) => sessionHref(props.server, sessionID)
@@ -53,6 +55,7 @@ export function SessionUIProvider(
      data={sessionUIData()}
      directory={directory()}
      sessionID={params.id}
+      shellOutput={(input) => serverSDK.api.shell.output(input)}
      onNavigateToSession={navigateToSession}
      onSessionHref={href}
    >
@@ -113,6 +113,32 @@ test("reactive count updates preserve measured row sizes", () => {
  })
 })

+test("explicit measurement refreshes a cached row size with a custom measurer", () => {
+  const root = document.createElement("div")
+  const element = document.createElement("div")
+  element.dataset.index = "0"
+  Object.defineProperty(element, "offsetHeight", { value: 120 })
+
+  const virtualizer = new Virtualizer<HTMLDivElement, HTMLDivElement>({
+    count: 1,
+    estimateSize: () => 60,
+    initialRect: { width: 400, height: 200 },
+    getScrollElement: () => root,
+    scrollToFn: () => {},
+    observeElementRect: () => {},
+    observeElementOffset: () => {},
+    measureElement: (node) => node.offsetHeight,
+  })
+
+  virtualizer.getTotalSize()
+  virtualizer.resizeItem(0, 60)
+  virtualizer._willUpdate()
+  virtualizer.measureElement(element)
+
+  expect(virtualizer.itemSizeCache.get(0)).toBe(120)
+  expect(virtualizer.getTotalSize()).toBe(120)
+})
+
 test("initial rect projects rows before a scroll element connects", () => {
  createRoot((dispose) => {
    const virtualizer = createVirtualizer<HTMLDivElement, HTMLDivElement>({
@@ -9,6 +9,7 @@ import type { BunPlugin } from "bun"
 import pkg from "../package.json"
 import { buildAppArchive } from "./app-assets"
 import { verifyArtifact, verifySimulationGraph } from "./verify-artifact"
+import { resolveOpencodePty } from "./opencode-pty"

 const dir = path.resolve(import.meta.dirname, "..")
 const binary = "opencode2"
@@ -78,6 +79,23 @@ const appAssetsPlugin: BunPlugin = {
 }

 for (const item of targets) {
+  const opencodePty = await resolveOpencodePty({
+    platform: item.os,
+    arch: item.arch,
+    ...(item.os === "linux" ? { libc: item.abi ?? "glibc" } : {}),
+  })
+  const opencodePtyPlugin: BunPlugin = {
+    name: "opencode-pty-binary",
+    setup(build) {
+      build.onLoad({ filter: /persistent-pty[/\\]pty-binding\.ts$/ }, () => ({
+        loader: "js",
+        contents: opencodePty
+          ? `import file from ${JSON.stringify(opencodePty.source)} with { type: "file" }
+export default { path: file, version: ${JSON.stringify(opencodePty.version)}, sha256: ${JSON.stringify(opencodePty.sha256)} }`
+          : "export default undefined",
+      }))
+    },
+  }
  const simulationInputs = new Set<string>()
  const simulationGraphPlugin: BunPlugin = {
    name: "opencode-simulation-graph",
@@ -105,7 +123,7 @@ for (const item of targets) {
  const result = await Bun.build({
    entrypoints: ["./src/index.ts"],
    tsconfig: "./tsconfig.json",
-    plugins: [appAssetsPlugin, solidPlugin, parcelWatcherPlugin, simulationGraphPlugin],
+    plugins: [appAssetsPlugin, solidPlugin, parcelWatcherPlugin, opencodePtyPlugin, simulationGraphPlugin],
    external: ["node-gyp"],
    format: "esm",
    minify: true,
@@ -5,6 +5,7 @@ import { fileURLToPath } from "node:url"
 import { getNodeAssets } from "@opentui/core/node-assets"
 import { attentionSoundAssets, type NodeTarget, photonWasmAsset, shellParserWasmAssets } from "../src/node/target"
 import { collectFiles } from "./files"
+import { resolveOpencodePty } from "./opencode-pty"

 const dir = path.resolve(import.meta.dirname, "..")

@@ -18,6 +19,11 @@ export type NodeAsset = {
 }

 export async function collectNodeAssets(target: NodeTarget) {
+  const opencodePty = await resolveOpencodePty({
+    platform: target.platform,
+    arch: target.arch,
+    ...(target.platform === "linux" ? { libc: "glibc" as const } : {}),
+  })
  const ptyEntry = fileURLToPath(import.meta.resolve(target.nodePtyPackage))
  const ptyRoot = path.resolve(path.dirname(ptyEntry), "..")
  const assets: NodeAsset[] = [
@@ -41,6 +47,7 @@ export async function collectNodeAssets(target: NodeTarget) {
      key,
      source: path.resolve(dir, "../ui/src/assets/audio", path.basename(key)),
    })),
+    ...(opencodePty && target.opencodePtyAsset ? [{ key: target.opencodePtyAsset, source: opencodePty.source }] : []),
    ...(await collectFiles(ptyRoot))
      .filter((relative) => !relative.endsWith(".map") && !relative.endsWith(".pdb"))
      .map((relative) => ({
@@ -0,0 +1,102 @@
+import { spawnSync } from "node:child_process"
+import { createHash } from "node:crypto"
+import { mkdir, mkdtemp, readFile, rename, rm, writeFile } from "node:fs/promises"
+import os from "node:os"
+import path from "node:path"
+
+const VERSION = "0.1.5"
+const RELEASE = `https://github.com/anomalyco/opencode-pty/releases/download/v${VERSION}`
+const SHA256 = {
+  "aarch64-apple-darwin": "d5156e44a6783381aadbd968dbd27c1d83e7e0f1b6042c7c934e6d33541d334f",
+  "aarch64-unknown-linux-gnu": "075d99ffb269cbd0846d3d404fdee93965a53cd6eaf046dbd1064785a7ce9351",
+  "aarch64-unknown-linux-musl": "22fb55c944ff05fbe03e84de67333e9fd037ad4e04ffc93d8a3f0b2193c29421",
+  "x86_64-apple-darwin": "773e363b5385c1bd56021e69ada95132efd615ed5b9c3734f878ad644ae22b01",
+  "x86_64-unknown-linux-gnu": "d9cac2a7c09d013188f696c45ded5eb5764d308e52dd31cb2de68bf4fc675624",
+  "x86_64-unknown-linux-musl": "2a176302de3d24f8ae3fbacf0b4afce7b4af3e00abd619906187a487b5e50bd6",
+} as const
+
+export type OpencodePtyAsset = {
+  readonly source: string
+  readonly version: string
+  readonly sha256: string
+}
+
+type Target = {
+  readonly platform: string
+  readonly arch: string
+  readonly libc?: "glibc" | "musl"
+}
+
+const pending = new Map<string, Promise<OpencodePtyAsset | undefined>>()
+
+export function resolveOpencodePty(target: Target) {
+  const rustTarget = targetName(target)
+  if (!rustTarget) return Promise.resolve(undefined)
+  const existing = pending.get(rustTarget)
+  if (existing) return existing
+  const result = acquire(rustTarget).catch((error) => {
+    pending.delete(rustTarget)
+    throw error
+  })
+  pending.set(rustTarget, result)
+  return result
+}
+
+async function acquire(target: keyof typeof SHA256): Promise<OpencodePtyAsset> {
+  const root = path.resolve(import.meta.dirname, "../.cache/opencode-pty", VERSION, target)
+  const executable = path.join(root, "opencode-pty")
+  const cached = await readFile(executable).catch(() => undefined)
+  if (cached)
+    return {
+      source: executable,
+      version: VERSION,
+      sha256: createHash("sha256").update(cached).digest("hex"),
+    }
+
+  await mkdir(root, { recursive: true })
+  const archiveName = `opencode-pty-${VERSION}-${target}.tar.gz`
+  const response = await fetch(`${RELEASE}/${archiveName}`)
+  if (!response.ok) throw new Error(`Failed to download ${archiveName}: ${response.status}`)
+  const archive = new Uint8Array(await response.arrayBuffer())
+  const actual = createHash("sha256").update(archive).digest("hex")
+  if (actual !== SHA256[target]) throw new Error(`Checksum mismatch for ${archiveName}`)
+
+  const temporary = await mkdtemp(path.join(os.tmpdir(), "opencode-pty-build-"))
+  try {
+    const archivePath = path.join(temporary, archiveName)
+    await writeFile(archivePath, archive)
+    run("tar", ["-xzf", archivePath, "-C", temporary])
+    const source = path.join(temporary, `opencode-pty-${VERSION}-${target}`, "opencode-pty")
+    const bytes = await readFile(source)
+    const staged = path.join(root, `opencode-pty.${process.pid}.${crypto.randomUUID()}.tmp`)
+    await writeFile(staged, bytes, { flag: "wx", mode: 0o755 })
+    await rename(staged, executable).catch(async (error) => {
+      await rm(staged, { force: true })
+      if (!(await readFile(executable).catch(() => undefined))) throw error
+    })
+    const installed = await readFile(executable)
+    return {
+      source: executable,
+      version: VERSION,
+      sha256: createHash("sha256").update(installed).digest("hex"),
+    }
+  } finally {
+    await rm(temporary, { recursive: true, force: true })
+  }
+}
+
+function targetName(target: Target): keyof typeof SHA256 | undefined {
+  const arch = target.arch === "arm64" ? "aarch64" : target.arch === "x64" ? "x86_64" : undefined
+  if (!arch) return undefined
+  if (target.platform === "darwin") return arch === "aarch64" ? "aarch64-apple-darwin" : "x86_64-apple-darwin"
+  if (target.platform === "linux" && target.libc === "musl")
+    return arch === "aarch64" ? "aarch64-unknown-linux-musl" : "x86_64-unknown-linux-musl"
+  if (target.platform === "linux") return arch === "aarch64" ? "aarch64-unknown-linux-gnu" : "x86_64-unknown-linux-gnu"
+  return undefined
+}
+
+function run(command: string, args: readonly string[]) {
+  const result = spawnSync(command, args, { stdio: "inherit" })
+  if (result.error) throw result.error
+  if (result.status !== 0) throw new Error(`${command} exited with status ${result.status ?? "unknown"}`)
+}
@@ -0,0 +1,27 @@
+import { Schema } from "effect"
+import { format } from "prettier"
+import { Info, SchemaURL } from "../src/config/schema"
+
+const target = process.argv[2]
+if (!target) throw new Error("A schema output path is required")
+
+const document = Schema.toJsonSchemaDocument(Info)
+const content = await format(
+  JSON.stringify({
+    $schema: "https://json-schema.org/draft/2020-12/schema",
+    $id: SchemaURL,
+    ...document.schema,
+    ...(Object.keys(document.definitions).length ? { $defs: document.definitions } : {}),
+  }),
+  { parser: "json", printWidth: 120 },
+)
+
+if (process.argv.includes("--check")) {
+  if ((await Bun.file(target).text()) !== content) {
+    console.error("Generated CLI config schema is stale. Run `bun run generate` from packages/www.")
+    process.exit(1)
+  }
+  process.exit(0)
+}
+
+await Bun.write(target, content)
@@ -79,6 +79,7 @@ const Root = Spec.make(typeof OPENCODE_CLI_NAME === "string" ? OPENCODE_CLI_NAME
      commands: [
        Spec.make("agents", { description: "List all agents" }),
        Spec.make("config", { description: "List configuration sources" }),
+        Spec.make("paths", { description: "Show global paths (data, config, cache, state)" }),
      ],
    }),
    Spec.make("console", {
@@ -180,7 +181,7 @@ const Root = Spec.make(typeof OPENCODE_CLI_NAME === "string" ? OPENCODE_CLI_NAME
        Spec.make("add", {
          description: "Install a plugin and add it to the global configuration",
          params: {
-            package: Argument.string("package").pipe(Argument.withDescription("npm registry package specifier")),
+            package: Argument.string("package").pipe(Argument.withDescription("npm registry or Git package specifier")),
          },
        }),
        Spec.make("remove", {
@@ -0,0 +1,17 @@
+import { EOL } from "os"
+import { Effect } from "effect"
+import { Global } from "@opencode-ai/util/global"
+import { Commands } from "../../commands"
+import { Runtime } from "../../../framework/runtime"
+
+export default Runtime.handler(
+  Commands.commands.debug.commands.paths,
+  Effect.fn("cli.debug.paths")(function* () {
+    const global = yield* Global.Service
+    process.stdout.write(
+      Object.entries(global)
+        .map(([key, value]) => `${key.padEnd(10)} ${value}${EOL}`)
+        .join(""),
+    )
+  }),
+)
@@ -13,10 +13,8 @@ import { Config } from "../../../config"
 export default Runtime.handler(
  Commands.commands.plugin.commands.add,
  Effect.fn("cli.plugin.add")(function* (input) {
-    if (!(yield* Effect.promise(() => Npm.isRegistryPackage(input.package))))
-      return yield* Effect.fail(
-        new Error("Plugin target must be an npm registry package name, version, tag, or semver range"),
-      )
+    if (!(yield* Effect.promise(() => Npm.isInstallablePackage(input.package))))
+      return yield* Effect.fail(new Error("Plugin target must be an npm registry package or Git package specifier"))
    const npm = yield* Npm.Service
    const installed = yield* npm.add(input.package, { subpaths: ["server", ""] })
    const tui = yield* npm.resolve(input.package, { subpaths: ["tui"] })
@@ -7,7 +7,7 @@ import { produce, type Draft } from "immer"
 import { applyEdits, modify, parse, type ParseError } from "jsonc-parser"
 import path from "path"
 import { ConfigMigration } from "./migrate"
-import { Info } from "./schema"
+import { Info, SchemaURL } from "./schema"

 export * from "./schema"

@@ -87,7 +87,9 @@ export const layer = Layer.effect(
          const next = produce(current, update)
          const edits = changes(current, next)
          if (!edits.length) return current
-          const text = yield* fs.readFileString(file).pipe(Effect.orElseSucceed(() => "{}"))
+          const text = yield* fs
+            .readFileString(file)
+            .pipe(Effect.orElseSucceed(() => JSON.stringify({ $schema: SchemaURL }, null, 2)))
          const updated = edits.reduce(
            (text, edit) =>
              applyEdits(
@@ -7,7 +7,7 @@ import { Effect, FileSystem, Option, Schema } from "effect"
 import { randomUUID } from "crypto"
 import { applyEdits, createScanner, modify, parse, parseTree, type Node, type ParseError } from "jsonc-parser"
 import path from "path"
-import { Info } from "./schema"
+import { Info, SchemaURL } from "./schema"

 const decodeV1 = Schema.decodeUnknownOption(TuiConfigV1.Info)
 const decodeInfo = Schema.decodeUnknownOption(Info)
@@ -100,8 +100,9 @@ export const run = Effect.fn("cli.config.migrate")(function* (input: {
  const legacyValue = yield* readJson(path.join(input.config, "tui.json"))
  const legacy = Option.getOrUndefined(decodeV1(legacyValue))
  const kv = yield* readJson(path.join(input.state, "kv.json"))
-  const migrated = migrateV1(legacy, kv ?? {})
-  if (!Object.keys(migrated).length) return
+  const values = migrateV1(legacy, kv ?? {})
+  if (!Object.keys(values).length) return
+  const migrated = { $schema: SchemaURL, ...values }

  const result = yield* persist(JSON.stringify(migrated, null, 2) + "\n", migrated)
  if (result.cause === undefined)
@@ -1,5 +1,10 @@
 import { Config } from "@opencode-ai/tui/config"
 import { Schema } from "effect"

-export const Info = Schema.Struct({ ...Config.Info.fields })
+export const SchemaURL = "https://opencode.ai/v2/cli.json"
+
+export const Info = Schema.Struct({
+  $schema: Schema.optional(Schema.String).annotate({ description: "JSON Schema for CLI configuration" }),
+  ...Config.Info.fields,
+})
 export type Info = Schema.Schema.Type<typeof Info>
@@ -27,6 +27,7 @@ const Handlers = Runtime.handlers(Commands, {
  debug: {
    agents: () => import("./commands/handlers/debug/agents"),
    config: () => import("./commands/handlers/debug/config"),
+    paths: () => import("./commands/handlers/debug/paths"),
  },
  console: {
    login: () => import("./commands/handlers/console/login"),
@@ -13,6 +13,7 @@ export function nodeTarget(platform: string, arch: string) {
  const parcelWatcherPackage = `@parcel/watcher-${targetPlatform}-${targetArch}${targetPlatform === "linux" ? "-glibc" : ""}`
  const fffPackage = `@ff-labs/fff-bin-${targetPlatform}-${targetArch}${targetPlatform === "linux" ? "-gnu" : ""}`
  const fffFfiPackage = `@yuuang/ffi-rs-${targetPlatform}-${targetArch}${targetPlatform === "linux" ? "-gnu" : targetPlatform === "win32" ? "-msvc" : ""}`
+  const opencodePtyAsset = targetPlatform === "win32" ? undefined : "opencode-pty/opencode-pty"

  return {
    platform: targetPlatform,
@@ -25,6 +26,7 @@ export function nodeTarget(platform: string, arch: string) {
    fffAsset: `${fffPackage}/${targetPlatform === "darwin" ? "libfff_c.dylib" : targetPlatform === "win32" ? "fff_c.dll" : "libfff_c.so"}`,
    fffFfiPackage,
    fffFfiAsset: `${fffFfiPackage}/ffi-rs.${targetPlatform}-${targetArch}${targetPlatform === "linux" ? "-gnu" : targetPlatform === "win32" ? "-msvc" : ""}.node`,
+    opencodePtyAsset,
  }
 }

@@ -1,7 +1,7 @@
 import { NodeFileSystem } from "@effect/platform-node"
 import { Flock } from "@opencode-ai/util/flock"
 import { Global } from "@opencode-ai/util/global"
-import { Effect, FileSystem, Option } from "effect"
+import { Effect, FileSystem, Option, Schema } from "effect"
 import { expect, test } from "bun:test"
 import { parse } from "jsonc-parser"
 import path from "path"
@@ -17,6 +17,67 @@ function run<A, E>(directory: string, effect: Effect.Effect<A, E, Config.Service
  )
 }

+test("generates reusable keybind schemas and preserves descriptions and numeric constraints", () => {
+  const document = Schema.toJsonSchemaDocument(Config.Info)
+  expect(document).toHaveProperty(["definitions", "TuiKeybind.BindingValue"])
+  expect(document).toHaveProperty(["schema", "properties", "keybinds", "anyOf", 0, "properties", "app.exit"], {
+    anyOf: [{ $ref: "#/$defs/TuiKeybind.BindingValue" }, { type: "null" }],
+    description: "Exit the application",
+  })
+  expect(document).toHaveProperty(["schema", "properties", "scroll", "anyOf", 0, "properties", "speed", "anyOf", 0], {
+    type: "number",
+    minimum: 0.001,
+  })
+  expect(document).toHaveProperty(
+    ["schema", "properties", "attention", "anyOf", 0, "properties", "volume", "anyOf", 0],
+    { type: "number", minimum: 0, maximum: 1 },
+  )
+})
+
+test("includes the published schema when creating cli.json", async () => {
+  const directory = await Bun.$`mktemp -d`.text().then((value) => value.trim())
+
+  try {
+    const config = await run(
+      directory,
+      Effect.gen(function* () {
+        const service = yield* Config.Service
+        return yield* service.update((draft) => {
+          draft.animations = false
+        })
+      }),
+    )
+
+    expect(config).toEqual({ $schema: "https://opencode.ai/v2/cli.json", animations: false })
+    expect(await Bun.file(path.join(directory, "cli.json")).json()).toEqual(config)
+  } finally {
+    await Bun.$`rm -rf ${directory}`
+  }
+})
+
+test("preserves the schema in an existing cli.json", async () => {
+  const directory = await Bun.$`mktemp -d`.text().then((value) => value.trim())
+  const file = path.join(directory, "cli.json")
+  await Bun.write(file, JSON.stringify({ $schema: "https://opencode.ai/v2/cli.json", animations: true }))
+
+  try {
+    const config = await run(
+      directory,
+      Effect.gen(function* () {
+        const service = yield* Config.Service
+        return yield* service.update((draft) => {
+          draft.animations = false
+        })
+      }),
+    )
+
+    expect(config).toEqual({ $schema: "https://opencode.ai/v2/cli.json", animations: false })
+    expect(await Bun.file(file).json()).toEqual(config)
+  } finally {
+    await Bun.$`rm -rf ${directory}`
+  }
+})
+
 test("migrates tui and kv config into cli.json", async () => {
  const directory = await Bun.$`mktemp -d`.text().then((value) => value.trim())
  await Bun.write(
@@ -73,6 +134,7 @@ test("migrates tui and kv config into cli.json", async () => {
    )

    expect(config).toMatchObject({
+      $schema: "https://opencode.ai/v2/cli.json",
      theme: { name: "legacy", mode: "light" },
      keybinds: {
        leader: "ctrl+o",
@@ -128,8 +190,14 @@ test("migrates before the first update and does not remigrate afterward", async
      }),
    )

-    expect(config).toEqual({ theme: { name: "legacy" }, animations: false, mouse: false })
+    expect(config).toEqual({
+      $schema: "https://opencode.ai/v2/cli.json",
+      theme: { name: "legacy" },
+      animations: false,
+      mouse: false,
+    })
    expect(await Bun.file(path.join(directory, "cli.json")).json()).toEqual({
+      $schema: "https://opencode.ai/v2/cli.json",
      theme: { name: "legacy" },
      animations: false,
      mouse: false,
@@ -0,0 +1,66 @@
+import { describe, expect, test } from "bun:test"
+import fs from "node:fs/promises"
+import os from "node:os"
+import path from "node:path"
+
+describe("debug paths command", () => {
+  test("is included in troubleshooting help", async () => {
+    const [debug, paths] = await Promise.all([cli(["debug", "--help"]), cli(["debug", "paths", "--help"])])
+
+    expect(debug.exitCode).toBe(0)
+    expect(debug.stdout).toContain("paths")
+    expect(debug.stdout).toContain("Show global paths (data, config, cache, state)")
+    expect(paths.exitCode).toBe(0)
+    expect(paths.stdout).toContain("opencode debug paths [flags]")
+  })
+
+  test("prints resolved global paths without starting a server", async () => {
+    const root = await fs.mkdtemp(path.join(os.tmpdir(), "opencode-debug-paths-"))
+
+    try {
+      const result = await cli(["debug", "paths"], {
+        XDG_DATA_HOME: path.join(root, "data"),
+        XDG_CONFIG_HOME: path.join(root, "config"),
+        XDG_CACHE_HOME: path.join(root, "cache"),
+        XDG_STATE_HOME: path.join(root, "state"),
+      })
+      const paths = Object.fromEntries(
+        result.stdout
+          .trim()
+          .split("\n")
+          .map((line) => line.trim().split(/\s+/, 2)),
+      )
+
+      expect({ exitCode: result.exitCode, stderr: result.stderr }).toEqual({ exitCode: 0, stderr: "" })
+      expect(paths).toMatchObject({
+        home: os.homedir(),
+        data: path.join(root, "data", "opencode"),
+        config: path.join(root, "config", "opencode"),
+        cache: path.join(root, "cache", "opencode"),
+        state: path.join(root, "state", "opencode"),
+        bin: path.join(root, "cache", "opencode", "bin"),
+        log: path.join(root, "data", "opencode", "log"),
+        repos: path.join(root, "data", "opencode", "repos"),
+      })
+      expect(paths.tmp).toBeTruthy()
+      expect(await Bun.file(path.join(root, "state", "opencode", "service-local.json")).exists()).toBe(false)
+    } finally {
+      await fs.rm(root, { recursive: true, force: true })
+    }
+  })
+})
+
+async function cli(args: string[], env?: Record<string, string>) {
+  const child = Bun.spawn([process.execPath, "run", path.join(import.meta.dir, "../src/index.ts"), ...args], {
+    cwd: path.join(import.meta.dir, ".."),
+    env: { ...process.env, ...env },
+    stdout: "pipe",
+    stderr: "pipe",
+  })
+  const [stdout, stderr, exitCode] = await Promise.all([
+    new Response(child.stdout).text(),
+    new Response(child.stderr).text(),
+    child.exited,
+  ])
+  return { stdout, stderr, exitCode }
+}
@@ -8,6 +8,7 @@ test("collects each SEA asset key once", async () => {
  const keys = assets.map((asset) => asset.key)

  expect(new Set(keys).size).toBe(keys.length)
+  if (process.platform !== "win32") expect(keys.filter((key) => key === "opencode-pty/opencode-pty")).toHaveLength(1)
  expect(assets.filter((asset) => asset.key === shellParserWasmAssets.runtime)).toEqual([
    {
      key: shellParserWasmAssets.runtime,
@@ -120,6 +120,7 @@ function nodePrelude(input: NodeBuildInput) {
    input.target.platform === "darwin"
      ? `${input.target.nodePtyPackage}/prebuilds/darwin-${input.target.arch}/spawn-helper`
      : undefined
+  const opencodePtyAsset = input.target.opencodePtyAsset
  const promiseModule = `const sdk = globalThis[Symbol.for("opencode.plugin.v2.promise")]
 if (!sdk) throw new Error("OpenCode Promise plugin SDK is unavailable")
 export const Agent = sdk.Agent
@@ -200,13 +201,17 @@ if (__ocIsSea()) {
 const __ocAssetRoot = __ocIsSea()
  ? __ocPath.join(__ocCacheRoot, ${JSON.stringify(`${input.assetHash}-${input.target.platform}-${input.target.arch}`)})
  : __ocFileURLToPath(new URL("./assets/", import.meta.url))
+const __ocPersistentPty = ${JSON.stringify(opencodePtyAsset)}
 if (__ocIsSea()) {
+  const __ocPtySpawnHelper = ${JSON.stringify(nodePtySpawnHelper)}
  for (const __ocKey of __ocAssetKeys()) {
    const __ocTarget = __ocPath.join(__ocAssetRoot, __ocKey)
    if (__ocExists(__ocTarget)) continue
    __ocMkdir(__ocPath.dirname(__ocTarget), { recursive: true })
    const __ocTemporary = \`${"${__ocTarget}"}.${"${process.pid}"}.${"${crypto.randomUUID()}"}.tmp\`
    __ocWrite(__ocTemporary, new Uint8Array(__ocRawAsset(__ocKey)))
+    if ((__ocKey === __ocPtySpawnHelper || __ocKey === __ocPersistentPty) && process.platform !== "win32")
+      __ocChmod(__ocTemporary, 0o755)
    try {
      __ocRename(__ocTemporary, __ocTarget)
    } catch (__ocError) {
@@ -214,8 +219,6 @@ if (__ocIsSea()) {
      if (!__ocExists(__ocTarget)) throw __ocError
    }
  }
-  const __ocPtySpawnHelper = ${JSON.stringify(nodePtySpawnHelper)}
-  if (__ocPtySpawnHelper) __ocChmod(__ocPath.join(__ocAssetRoot, __ocPtySpawnHelper), 0o755)
 }
 process.env.OPENCODE_NODE_ASSETS_DIR = __ocAssetRoot
 process.env.OTUI_ASSET_ROOT = __ocAssetRoot
@@ -227,6 +230,7 @@ process.env.OPENCODE_TREE_SITTER_BASH_WASM_PATH = __ocPath.join(__ocAssetRoot, $
 process.env.OPENCODE_TREE_SITTER_POWERSHELL_WASM_PATH = __ocPath.join(__ocAssetRoot, ${JSON.stringify(shellParserWasmAssets.powershell)})
 process.env.FFF_BINARY_PATH = __ocPath.join(__ocAssetRoot, ${JSON.stringify(input.target.fffAsset)})
 process.env.OPENCODE_FFF_FFI_PATH = __ocPath.join(__ocAssetRoot, ${JSON.stringify(input.target.fffFfiAsset)})
+if (__ocPersistentPty && !process.env.OPENCODE_PTY_BIN) process.env.OPENCODE_PTY_BIN = __ocPath.join(__ocAssetRoot, __ocPersistentPty)
 try {
  globalThis.__OPENCODE_FFF_FFI = require(process.env.OPENCODE_FFF_FFI_PATH)
 } catch {}
@@ -1333,6 +1333,15 @@ export type CredentialUpdateOperation<E = never> = (
  input: CredentialUpdateInput,
 ) => Effect.Effect<CredentialUpdateOutput, E>

+export type CredentialActivateInput = {
+  readonly credentialID: Credential.ID
+  readonly location?: { readonly directory?: string | undefined; readonly workspace?: string | undefined } | undefined
+}
+export type CredentialActivateOutput = void
+export type CredentialActivateOperation<E = never> = (
+  input: CredentialActivateInput,
+) => Effect.Effect<CredentialActivateOutput, E>
+
 export type CredentialRemoveInput = {
  readonly credentialID: Credential.ID
  readonly location?: { readonly directory?: string | undefined; readonly workspace?: string | undefined } | undefined
@@ -1344,6 +1353,7 @@ export type CredentialRemoveOperation<E = never> = (

 export interface CredentialApi<E = never> {
  readonly update: CredentialUpdateOperation<E>
+  readonly activate: CredentialActivateOperation<E>
  readonly remove: CredentialRemoveOperation<E>
 }

@@ -138,6 +138,8 @@ import type {
  McpResourceCatalogOutput,
  CredentialUpdateInput,
  CredentialUpdateOutput,
+  CredentialActivateInput,
+  CredentialActivateOutput,
  CredentialRemoveInput,
  CredentialRemoveOutput,
  ProjectListOutput,
@@ -935,6 +937,14 @@ const EndpointCredentialUpdate = (raw: RawClient["server.credential"]) => (input
    }).pipe(Effect.mapError(mapClientError)),
  )

+const EndpointCredentialActivate = (raw: RawClient["server.credential"]) => (input: CredentialActivateInput) =>
+  preserveEffect<CredentialActivateOutput>()(
+    raw["credential.activate"]({
+      params: { credentialID: input["credentialID"] },
+      query: { location: input["location"] },
+    }).pipe(Effect.mapError(mapClientError)),
+  )
+
 const EndpointCredentialRemove = (raw: RawClient["server.credential"]) => (input: CredentialRemoveInput) =>
  preserveEffect<CredentialRemoveOutput>()(
    raw["credential.remove"]({
@@ -945,6 +955,7 @@ const EndpointCredentialRemove = (raw: RawClient["server.credential"]) => (input

 const adaptGroupCredential = (raw: RawClient["server.credential"]) => ({
  update: EndpointCredentialUpdate(raw),
+  activate: EndpointCredentialActivate(raw),
  remove: EndpointCredentialRemove(raw),
 })

@@ -4,10 +4,12 @@ export type AgentApi = Client["agent"]
 export type CommandApi = Client["command"]
 export type ConfigApi = Client["config"]
 export type EventApi = Client["event"]
+export type GenerateApi = Client["generate"]
 export type IntegrationApi = Client["integration"]
 export type McpApi = Client["mcp"]
 export type ModelApi = Client["model"]
 export type PluginApi = Client["plugin"]
+export type PermissionApi = Client["permission"]
 export type ProviderApi = Client["provider"]
 export type ReferenceApi = Client["reference"]
 export type WebSearchApi = Client["websearch"]
@@ -132,6 +132,8 @@ import type {
  McpResourceCatalogOutput,
  CredentialUpdateInput,
  CredentialUpdateOutput,
+  CredentialActivateInput,
+  CredentialActivateOutput,
  CredentialRemoveInput,
  CredentialRemoveOutput,
  ProjectListOutput,
@@ -1270,6 +1272,18 @@ export function make(options: ClientOptions) {
          },
          requestOptions,
        ),
+      activate: (input: CredentialActivateInput, requestOptions?: RequestOptions) =>
+        request<CredentialActivateOutput>(
+          {
+            method: "POST",
+            path: `/api/credential/${encodeURIComponent(input.credentialID)}/activate`,
+            query: { location: input["location"] },
+            successStatus: 204,
+            declaredStatuses: [401, 400],
+            empty: true,
+          },
+          requestOptions,
+        ),
      remove: (input: CredentialRemoveInput, requestOptions?: RequestOptions) =>
        request<CredentialRemoveOutput>(
          {
@@ -476,7 +476,7 @@ export type PromptFileAttachment = {

 export type PromptAgentAttachment = { name: string; mention?: PromptMention }

-export type PromptSkillAttachment = { id: string; name: string; mention?: PromptMention }
+export type PromptSkillAttachment = { id: string; name: string; text?: string; mention?: PromptMention }

 export type SessionMessageAssistantText = { type: "text"; text: string; state?: SessionMessageProviderState }

@@ -823,6 +823,24 @@ export type ModelsDevRefreshed = {
  data: {}
 }

+export type CredentialUpdated = {
+  id: string
+  created: number
+  metadata?: { [x: string]: any }
+  type: "credential.updated"
+  location?: LocationRef
+  data: {}
+}
+
+export type CredentialSwitched = {
+  id: string
+  created: number
+  metadata?: { [x: string]: any }
+  type: "credential.switched"
+  location?: LocationRef
+  data: { integrationID: string; credentialID: string | null }
+}
+
 export type IntegrationUpdated = {
  id: string
  created: number
@@ -832,15 +850,6 @@ export type IntegrationUpdated = {
  data: {}
 }

-export type IntegrationConnectionUpdated = {
-  id: string
-  created: number
-  metadata?: { [x: string]: any }
-  type: "integration.connection.updated"
-  location?: LocationRef
-  data: { integrationID: string }
-}
-
 export type CatalogUpdated = {
  id: string
  created: number
@@ -1310,8 +1319,10 @@ export type ToolContent1 = ToolTextContent | ToolFileContent1

 export type ModelCompatibility = {
  reasoningField?: ModelReasoningField
+  requireReasoning?: boolean
  maxTokensField?: ModelMaxTokensField
  requireFinishReason?: boolean
+  requireAssistantAfterTool?: boolean
 }

 export type ModelCost = {
@@ -1407,6 +1418,24 @@ export type Project = {
  sandboxes: Array<string>
 }

+export type ProjectUpdated = {
+  id: string
+  created: number
+  metadata?: { [x: string]: any }
+  type: "project.updated"
+  location?: LocationRef
+  data: {
+    id: string
+    canonical: string
+    vcs?: ProjectVcs
+    name?: string
+    icon?: ProjectIcon
+    commands?: ProjectCommands
+    time: ProjectTime
+    sandboxes: Array<string>
+  }
+}
+
 export type FormAnswer = { [x: string]: FormValue }

 export type PermissionRequest = {
@@ -1417,6 +1446,7 @@ export type PermissionRequest = {
  save?: Array<string>
  metadata?: { [x: string]: JsonValue }
  source?: PermissionSource
+  message?: string
 }

 export type PermissionAsked = {
@@ -1433,6 +1463,7 @@ export type PermissionAsked = {
    save?: Array<string>
    metadata?: { [x: string]: any }
    source?: PermissionSource
+    message?: string
  }
 }

@@ -2102,14 +2133,16 @@ export type SessionMessagesResponse = {
 export type IntegrationInfo = {
  id: string
  name: string
+  metadata?: { [x: string]: any }
  methods: Array<IntegrationMethod>
  connections: Array<ConnectionInfo>
 }

 export type V2Event =
  | ModelsDevRefreshed
+  | CredentialUpdated
+  | CredentialSwitched
  | IntegrationUpdated
-  | IntegrationConnectionUpdated
  | CatalogUpdated
  | AgentUpdated
  | SessionCreated
@@ -2164,6 +2197,7 @@ export type V2Event =
  | PermissionReplied
  | PluginAdded
  | PluginUpdated
+  | ProjectUpdated
  | WorktreeUpdated
  | WorktreeResolved
  | CommandUpdated
@@ -2706,6 +2740,7 @@ export type SessionImportInput = {
          readonly skills?: ReadonlyArray<{
            readonly id: string
            readonly name: string
+            readonly text?: string
            readonly mention?: { readonly start: number; readonly end: number; readonly text: string }
          }>
          readonly type: "user"
@@ -2981,6 +3016,7 @@ export type SessionImportInput = {
          readonly skills?: ReadonlyArray<{
            readonly id: string
            readonly name: string
+            readonly text?: string
            readonly mention?: { readonly start: number; readonly end: number; readonly text: string }
          }>
          readonly type: "user"
@@ -3256,6 +3292,7 @@ export type SessionImportInput = {
          readonly skills?: ReadonlyArray<{
            readonly id: string
            readonly name: string
+            readonly text?: string
            readonly mention?: { readonly start: number; readonly end: number; readonly text: string }
          }>
          readonly type: "user"
@@ -4349,6 +4386,15 @@ export type CredentialUpdateInput = {

 export type CredentialUpdateOutput = void

+export type CredentialActivateInput = {
+  readonly credentialID: { readonly credentialID: string }["credentialID"]
+  readonly location?: {
+    readonly location?: { readonly directory?: string | undefined; readonly workspace?: string | undefined } | undefined
+  }["location"]
+}
+
+export type CredentialActivateOutput = void
+
 export type CredentialRemoveInput = {
  readonly credentialID: { readonly credentialID: string }["credentialID"]
  readonly location?: {
@@ -521,6 +521,9 @@ export function createData(config: CreateDataInput) {
        void result.location.vcs.sync().catch((error) => console.error("Failed to preload VCS info", error))
        void result.project.sync().catch((error) => console.error("Failed to preload projects", error))
        return
+      case "project.updated":
+        setStore("project", "info", event.data.id, reconcile(event.data))
+        return
      case "session.created":
        sessionOutbox.delete(event.data.sessionID)
        result.session.invalidate(event.data.sessionID)
@@ -1045,6 +1048,36 @@ export function createData(config: CreateDataInput) {
        return
    }

+    if (event.type === "credential.updated" || event.type === "credential.switched") {
+      Object.keys(store.location).forEach((key) => {
+        const ref = JSON.parse(key) as [string, string | null]
+        const location = { directory: ref[0], workspaceID: ref[1] ?? undefined }
+        if (event.type === "credential.updated") {
+          result.location.integration.invalidate(location)
+          void result.location.integration.sync(location)
+          return
+        }
+        setStore("location", key, (data) => ({
+          ...data,
+          integration: data?.integration?.map((integration) => {
+            if (integration.id !== event.data.integrationID) return integration
+            const active = integration.connections.find(
+              (connection) => connection.type === "credential" && connection.id === event.data.credentialID,
+            )
+            if (!active) return integration
+            return {
+              ...integration,
+              connections: [active, ...integration.connections.filter((connection) => connection !== active)],
+            }
+          }),
+        }))
+        result.location.model.invalidate(location)
+        result.location.provider.invalidate(location)
+        void Promise.all([result.location.model.sync(location), result.location.provider.sync(location)])
+      })
+      return
+    }
+
    if (!event.location) return
    const location = event.location
    switch (event.type) {
@@ -26,6 +26,7 @@ test("exposes every standard HTTP API group", () => {
    "skill",
    "event",
    "pty",
+    "experimental",
    "shell",
    "reference",
    "worktree",
@@ -187,6 +188,24 @@ test("experimental wellknown integration add uses the public HTTP contract", asy
  expect(await request?.json()).toEqual({ url: "https://example.com" })
 })

+test("credential.activate uses the public HTTP contract", async () => {
+  let request: Request | undefined
+  const client = OpenCode.make({
+    baseUrl: "http://localhost:3000",
+    fetch: async (input, init) => {
+      request = input instanceof Request ? input : new Request(input, init)
+      return new Response(null, { status: 204 })
+    },
+  })
+
+  await client.credential.activate({ credentialID: "cred_work", location: { directory: "/tmp/project" } })
+
+  expect(request?.method).toBe("POST")
+  expect(request?.url).toBe(
+    "http://localhost:3000/api/credential/cred_work/activate?location%5Bdirectory%5D=%2Ftmp%2Fproject",
+  )
+})
+
 test("integration connections optionally submit a form answer", async () => {
  const requests: Request[] = []
  const client = OpenCode.make({
@@ -1,7 +1,7 @@
 import { expect, test } from "bun:test"
 import { createRoot } from "solid-js"
 import { createData, type CreateDataInput } from "../src/solid"
-import { OpenCode, type OpenCodeEvent, type SessionInfo } from "../src/promise"
+import { OpenCode, type OpenCodeEvent, type Project, type SessionInfo } from "../src/promise"

 const session = (viewed: number): SessionInfo => ({
  id: "ses_refresh",
@@ -72,6 +72,185 @@ test("revalidates after an event overtakes an active session read", async () =>
  }
 })

+test("updates authoritative cached project metadata from live events", async () => {
+  const listeners = new Set<Parameters<CreateDataInput["event"]["listen"]>[0]>()
+  const original: Project = {
+    id: "project_renamed",
+    canonical: "/projects/original",
+    name: "Original custom name",
+    time: { created: 1, updated: 1 },
+    sandboxes: [],
+  }
+  const unrelated: Project = {
+    id: "project_unrelated",
+    canonical: "/projects/unrelated",
+    name: "Unrelated project",
+    time: { created: 1, updated: 1 },
+    sandboxes: [],
+  }
+  let requests = 0
+  const api = OpenCode.make({
+    baseUrl: "http://opencode.local",
+    fetch: async (input, init) => {
+      const request = input instanceof Request ? input : new Request(input, init)
+      if (!request.url.endsWith("/api/project")) throw new Error(`Unexpected request: ${request.url}`)
+      requests++
+      return Response.json([original, unrelated])
+    },
+  })
+  const event: CreateDataInput["event"] = {
+    on: () => () => {},
+    listen(handler) {
+      listeners.add(handler)
+      return () => listeners.delete(handler)
+    },
+  }
+  const setup = createRoot((dispose) => ({
+    data: createData({ api: () => api, directory: "/projects/original", event }),
+    dispose,
+  }))
+
+  try {
+    await setup.data.project.sync()
+    expect(setup.data.project.get(original.id)).toEqual(original)
+
+    const updated: OpenCodeEvent = {
+      id: "evt_project_renamed",
+      created: 2,
+      type: "project.updated",
+      data: {
+        ...original,
+        canonical: "/projects/renamed",
+        name: "Updated custom name",
+        time: { ...original.time, updated: 2 },
+      },
+    }
+    listeners.forEach((listener) => listener({ name: updated.type, details: updated }))
+
+    expect(setup.data.project.get(original.id)?.canonical).toBe("/projects/renamed")
+    expect(setup.data.project.get(original.id)?.name).toBe("Updated custom name")
+    expect(setup.data.project.get(unrelated.id)).toEqual(unrelated)
+    expect(requests).toBe(1)
+
+    const reset: OpenCodeEvent = {
+      id: "evt_project_name_reset",
+      created: 3,
+      type: "project.updated",
+      data: {
+        id: original.id,
+        canonical: "/projects/renamed-again",
+        time: { ...original.time, updated: 3 },
+        sandboxes: [],
+      },
+    }
+    listeners.forEach((listener) => listener({ name: reset.type, details: reset }))
+
+    expect(setup.data.project.get(original.id)?.canonical).toBe("/projects/renamed-again")
+    expect(setup.data.project.get(original.id)?.name).toBeUndefined()
+    expect(setup.data.project.get(unrelated.id)).toEqual(unrelated)
+    expect(requests).toBe(1)
+  } finally {
+    setup.dispose()
+  }
+})
+
+test("refreshes global credential events across every loaded location and workspace", async () => {
+  const listeners = new Set<Parameters<CreateDataInput["event"]["listen"]>[0]>()
+  const requests: URL[] = []
+  const api = OpenCode.make({
+    baseUrl: "http://opencode.local",
+    fetch: async (input, init) => {
+      const request = input instanceof Request ? input : new Request(input, init)
+      const url = new URL(request.url)
+      requests.push(url)
+      const directory = url.searchParams.get("location[directory]") ?? "/project"
+      return Response.json({
+        location: {
+          directory,
+          workspaceID: url.searchParams.get("location[workspace]") ?? undefined,
+          project: { id: "project", directory, canonical: directory },
+        },
+        data: [],
+      })
+    },
+  })
+  const setup = createRoot((dispose) => ({
+    data: createData({
+      api: () => api,
+      directory: "/project",
+      event: {
+        on: () => () => {},
+        listen(handler) {
+          listeners.add(handler)
+          return () => listeners.delete(handler)
+        },
+      },
+      connection: { status: () => "connected" },
+    }),
+    dispose,
+  }))
+  const locations = [{ directory: "/project" }, { directory: "/other", workspaceID: "workspace-other" }]
+
+  try {
+    await Promise.all(
+      locations.flatMap((location) => [
+        setup.data.location.integration.sync(location),
+        setup.data.location.model.sync(location),
+        setup.data.location.provider.sync(location),
+      ]),
+    )
+    requests.length = 0
+
+    const updated: OpenCodeEvent = {
+      id: "evt_credential.updated",
+      created: 1,
+      type: "credential.updated",
+      data: {},
+    }
+    listeners.forEach((listener) => listener({ name: updated.type, details: updated }))
+    await wait(() => requests.length === 2)
+    expect(
+      requests.map((url) => [
+        url.pathname,
+        url.searchParams.get("location[directory]"),
+        url.searchParams.get("location[workspace]"),
+      ]),
+    ).toEqual([
+      ["/api/integration", "/project", null],
+      ["/api/integration", "/other", "workspace-other"],
+    ])
+    requests.length = 0
+
+    for (const credentialID of ["credential", null]) {
+      const switched: OpenCodeEvent = {
+        id: `evt_credential.switched.${credentialID}`,
+        created: 2,
+        type: "credential.switched",
+        data: { credentialID, integrationID: "integration" },
+      }
+      listeners.forEach((listener) => listener({ name: switched.type, details: switched }))
+      await wait(() => requests.length === 4)
+      expect(
+        requests.map((url) => [
+          url.pathname,
+          url.searchParams.get("location[directory]"),
+          url.searchParams.get("location[workspace]"),
+        ]),
+      ).toEqual(
+        expect.arrayContaining([
+          ["/api/model", "/project", null],
+          ["/api/provider", "/project", null],
+          ["/api/model", "/other", "workspace-other"],
+          ["/api/provider", "/other", "workspace-other"],
+        ]),
+      )
+      requests.length = 0
+    }
+  } finally {
+    setup.dispose()
+  }
+})
+
 test("reports optimistic sessions as creating until the request settles", async () => {
  const release = Promise.withResolvers<void>()
  const api = OpenCode.make({
@@ -22,7 +22,7 @@
    "benchmark:location": "bun run script/benchmark-location.ts",
    "build": "bun run script/build.ts",
    "update-models-snapshot": "bun run script/update-models-snapshot.ts",
-    "test": "bun test --only-failures",
+    "test": "bun run script/test.ts",
    "typecheck": "tsgo -b tsconfig.json tsconfig.tests.json"
  },
  "exports": {
@@ -41,6 +41,12 @@
      "node": "./src/pty/pty.node.ts",
      "default": "./src/pty/pty.bun.ts"
    },
+    "#persistent-pty-binary": {
+      "workerd": "./src/persistent-pty/binary.workerd.ts",
+      "bun": "./src/persistent-pty/binary.bun.ts",
+      "node": "./src/persistent-pty/binary.node.ts",
+      "default": "./src/persistent-pty/binary.bun.ts"
+    },
    "#fff": {
      "workerd": "./src/filesystem/fff.workerd.ts",
      "bun": "./src/filesystem/fff.bun.ts",
@@ -96,9 +102,7 @@
    "@ai-sdk/amazon-bedrock": "4.0.112",
    "@ai-sdk/anthropic": "3.0.82",
    "@ai-sdk/azure": "3.0.88",
-    "@ai-sdk/cerebras": "2.0.41",
    "@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",
@@ -107,7 +111,6 @@
    "@ai-sdk/perplexity": "3.0.26",
    "@ai-sdk/provider": "3.0.8",
    "@ai-sdk/provider-utils": "4.0.23",
-    "@ai-sdk/togetherai": "2.0.41",
    "@ai-sdk/vercel": "2.0.39",
    "@aws-sdk/credential-providers": "3.1057.0",
    "@lydell/node-pty": "catalog:",
@@ -0,0 +1,52 @@
+import fs from "fs/promises"
+import path from "path"
+import { tmpdir } from "../test/fixture/tmpdir"
+
+await using directory = await tmpdir("oc-")
+const home = directory.path
+const temporary = path.join(home, "tmp")
+await fs.mkdir(temporary)
+
+const environment = {
+  ...Object.fromEntries(
+    Object.entries(process.env).filter(([name]) => {
+      if (process.env.RECORD === "true" && name === "OPENAI_API_KEY") return true
+      if (
+        /^(?:AWS|AZURE|GOOGLE|GCP|GCLOUD|VERTEX|OPENAI|ANTHROPIC|GEMINI|XAI|CLOUDFLARE|CF_AIG|SNOWFLAKE|AICORE|GITLAB|NPM_CONFIG)_/i.test(
+          name,
+        )
+      ) {
+        return false
+      }
+      return !/(?:^|_)(?:API_KEY|AUTHORIZATION|TOKEN|SECRET|PASSWORD|CREDENTIALS?)$/i.test(name)
+    }),
+  ),
+  HOME: home,
+  OPENCODE_TEST_HOME: home,
+  XDG_CONFIG_HOME: path.join(home, ".config"),
+  XDG_DATA_HOME: path.join(home, ".local", "share"),
+  XDG_CACHE_HOME: path.join(home, ".cache"),
+  XDG_STATE_HOME: path.join(home, ".local", "state"),
+  OPENCODE_CONFIG_DIR: path.join(home, ".config", "opencode"),
+  OPENCODE_CONFIG: undefined,
+  OPENCODE_CONFIG_CONTENT: undefined,
+  TMPDIR: temporary,
+  ...(process.platform === "win32" ? { USERPROFILE: home, TMP: temporary, TEMP: temporary } : {}),
+}
+
+const child = Bun.spawn({
+  cmd: [process.execPath, "test", "--only-failures", ...process.argv.slice(2)],
+  env: environment,
+  stdin: "inherit",
+  stdout: "inherit",
+  stderr: "inherit",
+})
+const interrupt = () => child.kill("SIGINT")
+const terminate = () => child.kill("SIGTERM")
+process.once("SIGINT", interrupt)
+process.once("SIGTERM", terminate)
+const result = await child.exited
+process.off("SIGINT", interrupt)
+process.off("SIGTERM", terminate)
+
+process.exitCode = result
@@ -53,6 +53,18 @@ export function map(input: MapInput): Mapping | undefined {
          ...mapOpenAIOptions(input.settings),
        },
      }
+    case "@ai-sdk/cerebras":
+    case "@ai-sdk/deepinfra":
+    case "@ai-sdk/togetherai":
+      return {
+        package: `@opencode-ai/ai/providers/${input.packageName.slice("@ai-sdk/".length)}`,
+        settings: {
+          ...baseSettings,
+          ...mapAPIKey(input.settings),
+          ...mapProviderOptions(input.settings, ["apiKey", "baseURL", "fetch", "headers", "name"]),
+        },
+        ...(isStringRecord(input.settings.headers) ? { headers: input.settings.headers } : {}),
+      }
    case "@ai-sdk/google":
      return {
        package: "@opencode-ai/ai/providers/google",
@@ -34,6 +34,7 @@ import {
 import { Auth, Endpoint, RequestExecutor, type AnyRoute } from "@opencode-ai/ai/route"
 import { ProviderShared } from "@opencode-ai/ai/protocols/shared"
 import { Cause, Context, Effect, Layer, Option, Schema, Scope, Stream } from "effect"
+import { makeParser } from "effect/unstable/encoding/Sse"
 import type { ID, Info } from "./model.js"
 import { Provider } from "./provider.js"
 import { State } from "./state.js"
@@ -65,15 +66,23 @@ function wrapSSE(res: Response, ms: number, ctl: AbortController) {
  if (!res.headers.get("content-type")?.includes("text/event-stream")) return res

  const reader = res.body.getReader()
+  const decoder = new TextDecoder()
+  let deadline: number | undefined
+  const parser = makeParser((event) => {
+    if (event._tag === "Event") deadline = Date.now() + ms
+  })
  const body = new ReadableStream<Uint8Array>({
    async pull(ctrl) {
+      const expires = deadline ?? Date.now() + ms
+      deadline = expires
      const part = await new Promise<Awaited<ReturnType<typeof reader.read>>>((resolve, reject) => {
+        const remaining = Math.max(0, expires - Date.now())
        const id = setTimeout(() => {
          const err = new Error("SSE read timed out")
          ctl.abort(err)
          void reader.cancel(err)
          reject(err)
-        }, ms)
+        }, remaining)

        reader.read().then(
          (part) => {
@@ -92,6 +101,7 @@ function wrapSSE(res: Response, ms: number, ctl: AbortController) {
        return
      }

+      parser.feed(decoder.decode(part.value, { stream: true }))
      ctrl.enqueue(part.value)
    },
    async cancel(reason) {
@@ -89,6 +89,7 @@ export interface PublishOptions {
  readonly id?: Event.ID
  readonly metadata?: Record<string, unknown>
  readonly location?: Location.Ref
+  readonly global?: boolean
  /** Local operational projection committed atomically with a new durable event. Not replayed or serialized. */
  readonly commit?: (seq: number) => Effect.Effect<void>
 }
@@ -446,11 +447,12 @@ export function configured(options?: Options) {
        function publish<D extends Event.Definition>(definition: D, data: Event.Data<D>, options?: PublishOptions) {
          return Effect.gen(function* () {
            const serviceLocation = Option.getOrUndefined(yield* Effect.serviceOption(Location.Service))
-            const location =
-              options?.location ??
-              (serviceLocation
-                ? { directory: serviceLocation.directory, workspaceID: serviceLocation.workspaceID }
-                : undefined)
+            const location = options?.global
+              ? undefined
+              : (options?.location ??
+                (serviceLocation
+                  ? { directory: serviceLocation.directory, workspaceID: serviceLocation.workspaceID }
+                  : undefined))
            return yield* publishEvent(
              definition,
              {
@@ -480,11 +482,12 @@ export function configured(options?: Options) {
                    }),
                  )
                }
-                const location =
-                  options?.location ??
-                  (serviceLocation
-                    ? { directory: serviceLocation.directory, workspaceID: serviceLocation.workspaceID }
-                    : undefined)
+                const location = options?.global
+                  ? undefined
+                  : (options?.location ??
+                    (serviceLocation
+                      ? { directory: serviceLocation.directory, workspaceID: serviceLocation.workspaceID }
+                      : undefined))
                return {
                  definition,
                  aggregateID,
@@ -16,7 +16,6 @@ import {
  type Entry,
  Event,
 } from "@opencode-ai/schema/config"
-import { Integration } from "@opencode-ai/schema/integration"
 import { isRecord } from "@opencode-ai/ai/utils/record"
 import { Credential } from "./credential.js"
 import { Bus } from "./bus.js"
@@ -159,9 +158,7 @@ export const layer = (options?: Options) =>
      const loadWellknownEntry = Effect.fnUntraced(function* (entry: WellKnown.Entry) {
        const auth = entry.manifest.auth
        if (!auth) return []
-        const credential = (yield* credentials.list(entry.integrationID)).findLast(
-          (credential) => credential.value.type === "key",
-        )
+        const credential = (yield* credentials.list(entry.integrationID)).at(-1)
        if (!credential || credential.value.type !== "key") return []
        const variables = { [auth.env]: credential.value.key }
        const configs = yield* wellknown
@@ -343,7 +340,7 @@ export const layer = (options?: Options) =>
        ),
        Effect.forkScoped({ startImmediately: true }),
      )
-      yield* bus.subscribe(Integration.Event.ConnectionUpdated).pipe(
+      yield* bus.subscribe(Credential.Event.Switched).pipe(
        Stream.filterEffect((event) =>
          wellknown.entries().pipe(
            Effect.map((entries) => entries.some((entry) => entry.integrationID === event.data.integrationID)),
@@ -1,10 +1,11 @@
 export * as Credential from "./credential.js"

-import { asc, eq } from "drizzle-orm"
+import { asc, desc, eq } from "drizzle-orm"
 import { Context, Effect, Layer, Schema } from "effect"
 import { Credential } from "@opencode-ai/schema/credential"
 import { Integration } from "@opencode-ai/schema/integration"
 import { Database } from "./database/database.js"
+import { Bus } from "./bus.js"
 import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
 import { CredentialTable } from "./credential/sql.js"

@@ -20,6 +21,8 @@ export type Key = Credential.Key
 export const Value = Credential.Value
 export type Value = Credential.Value

+export const Event = Credential.Event
+
 export class Info extends Schema.Class<Info>("Credential.Info")({
  id: ID,
  integrationID: Integration.ID,
@@ -34,12 +37,14 @@ export interface Interface {
  readonly list: (integrationID: Integration.ID) => Effect.Effect<Info[]>
  /** Returns one stored credential by ID. */
  readonly get: (id: ID) => Effect.Effect<Info | undefined>
-  /** Replaces any credential for an integration and returns the new record. */
+  /** Creates a credential for an integration and returns the new record. */
  readonly create: (input: {
    readonly integrationID: Integration.ID
    readonly value: Value
    readonly label?: string
  }) => Effect.Effect<Info>
+  /** Selects a stored credential for its integration. */
+  readonly activate: (id: ID) => Effect.Effect<void>
  /** Updates the label or secret value of a stored credential. */
  readonly update: (id: ID, updates: Partial<Pick<Info, "label" | "value">>) => Effect.Effect<void>
  /** Removes a stored credential. */
@@ -52,6 +57,7 @@ const layer = Layer.effect(
  Service,
  Effect.gen(function* () {
    const db = (yield* Database.Service).db
+    const bus = yield* Bus.Service
    const decode = Schema.decodeUnknownSync(Value)
    const stored = (row: typeof CredentialTable.$inferSelect) => {
      if (!row.integration_id) return
@@ -73,7 +79,7 @@ const layer = Layer.effect(
        db
          .select()
          .from(CredentialTable)
-          .orderBy(asc(CredentialTable.time_created))
+          .orderBy(asc(CredentialTable.active), asc(CredentialTable.time_created), asc(CredentialTable.id))
          .all()
          .pipe(Effect.orDie, Effect.map(storedRows)),
      ),
@@ -82,7 +88,7 @@ const layer = Layer.effect(
          .select()
          .from(CredentialTable)
          .where(eq(CredentialTable.integration_id, integrationID))
-          .orderBy(asc(CredentialTable.time_created))
+          .orderBy(asc(CredentialTable.active), asc(CredentialTable.time_created), asc(CredentialTable.id))
          .all()
          .pipe(Effect.orDie, Effect.map(storedRows)),
      ),
@@ -101,7 +107,8 @@ const layer = Layer.effect(
          .transaction((tx) =>
            Effect.gen(function* () {
              yield* tx
-                .delete(CredentialTable)
+                .update(CredentialTable)
+                .set({ active: false })
                .where(eq(CredentialTable.integration_id, credential.integrationID))
                .run()
              yield* tx
@@ -111,27 +118,117 @@ const layer = Layer.effect(
                  integration_id: credential.integrationID,
                  label: credential.label,
                  value: credential.value,
+                  active: true,
                })
                .run()
            }),
          )
          .pipe(Effect.orDie)
+        yield* bus.publish(Event.Updated, {}, { global: true })
+        yield* bus.publish(
+          Event.Switched,
+          { integrationID: credential.integrationID, credentialID: credential.id },
+          { global: true },
+        )
        return credential
      }),
+      activate: Effect.fn("Credential.activate")(function* (id) {
+        const integrationID = yield* db
+          .transaction((tx) =>
+            Effect.gen(function* () {
+              const credential = yield* tx.select().from(CredentialTable).where(eq(CredentialTable.id, id)).get()
+              if (!credential?.integration_id) return
+              const active = yield* tx
+                .select({ id: CredentialTable.id })
+                .from(CredentialTable)
+                .where(eq(CredentialTable.integration_id, credential.integration_id))
+                .orderBy(desc(CredentialTable.active), desc(CredentialTable.time_created), desc(CredentialTable.id))
+                .get()
+              if (active?.id === id) return
+              yield* tx
+                .update(CredentialTable)
+                .set({ active: false })
+                .where(eq(CredentialTable.integration_id, credential.integration_id))
+                .run()
+              yield* tx.update(CredentialTable).set({ active: true }).where(eq(CredentialTable.id, id)).run()
+              return credential.integration_id
+            }),
+          )
+          .pipe(Effect.orDie)
+        if (integrationID) yield* bus.publish(Event.Switched, { integrationID, credentialID: id }, { global: true })
+      }),
      update: Effect.fn("Credential.update")(function* (id, updates) {
-        if (!updates.label && !updates.value) return
+        if (updates.label === undefined && updates.value === undefined) return
+        const credential = yield* db
+          .select({ integrationID: CredentialTable.integration_id, label: CredentialTable.label })
+          .from(CredentialTable)
+          .where(eq(CredentialTable.id, id))
+          .get()
+          .pipe(Effect.orDie)
+        if (!credential?.integrationID) return
+        if (updates.label === credential.label && updates.value === undefined) return
        yield* db
          .update(CredentialTable)
          .set({ label: updates.label, value: updates.value })
          .where(eq(CredentialTable.id, id))
          .run()
          .pipe(Effect.orDie)
+        if (updates.label !== undefined && updates.label !== credential.label)
+          yield* bus.publish(Event.Updated, {}, { global: true })
      }),
      remove: Effect.fn("Credential.remove")(function* (id) {
-        yield* db.delete(CredentialTable).where(eq(CredentialTable.id, id)).run().pipe(Effect.orDie)
+        const removed = yield* db
+          .transaction((tx) =>
+            Effect.gen(function* () {
+              const credential = yield* tx.select().from(CredentialTable).where(eq(CredentialTable.id, id)).get()
+              if (!credential) return
+              const active = credential.integration_id
+                ? yield* tx
+                    .select({ id: CredentialTable.id })
+                    .from(CredentialTable)
+                    .where(eq(CredentialTable.integration_id, credential.integration_id))
+                    .orderBy(desc(CredentialTable.active), desc(CredentialTable.time_created), desc(CredentialTable.id))
+                    .get()
+                : undefined
+              yield* tx.delete(CredentialTable).where(eq(CredentialTable.id, id)).run()
+              if (!credential.integration_id || active?.id !== id) return { switched: false as const }
+              const replacement = yield* tx
+                .select({ id: CredentialTable.id })
+                .from(CredentialTable)
+                .where(eq(CredentialTable.integration_id, credential.integration_id))
+                .orderBy(desc(CredentialTable.time_created), desc(CredentialTable.id))
+                .get()
+              if (replacement) {
+                yield* tx
+                  .update(CredentialTable)
+                  .set({ active: false })
+                  .where(eq(CredentialTable.integration_id, credential.integration_id))
+                  .run()
+                yield* tx
+                  .update(CredentialTable)
+                  .set({ active: true })
+                  .where(eq(CredentialTable.id, replacement.id))
+                  .run()
+              }
+              return {
+                switched: true as const,
+                integrationID: credential.integration_id,
+                credentialID: replacement?.id ?? null,
+              }
+            }),
+          )
+          .pipe(Effect.orDie)
+        if (!removed) return
+        yield* bus.publish(Event.Updated, {}, { global: true })
+        if (removed.switched)
+          yield* bus.publish(
+            Event.Switched,
+            { integrationID: removed.integrationID, credentialID: removed.credentialID },
+            { global: true },
+          )
      }),
    })
  }),
 )

-export const node = makeGlobalNode({ service: Service, layer, deps: [Database.node] })
+export const node = makeGlobalNode({ service: Service, layer, deps: [Database.node, Bus.node] })
@@ -1,14 +0,0 @@
-/* oxlint-disable */
-import type { MigrationConfig } from "drizzle-orm/migrator"
-import { readMigrationFiles } from "drizzle-orm/migrator"
-import type { AnyRelations } from "drizzle-orm/relations"
-import { migrate as coreMigrate } from "../sqlite-core/effect/session.js"
-import type { EffectSQLiteDatabase } from "./driver.js"
-
-export function migrate<TRelations extends AnyRelations>(
-  db: EffectSQLiteDatabase<TRelations>,
-  config: MigrationConfig,
-) {
-  const migrations = readMigrationFiles(config)
-  return coreMigrate(migrations, db.session, config)
-}
@@ -1,6 +1,5 @@
 export { EffectLogger } from "drizzle-orm/effect-core"
 export * from "./effect-sqlite/driver.js"
 export * from "./effect-sqlite/session.js"
-export { migrate } from "./effect-sqlite/migrator.js"

 export * as EffectDrizzleSqlite from "./index.js"
@@ -5,13 +5,10 @@ import type { SqlError } from "effect/unstable/sql/SqlError"
 import type { EffectCacheShape } from "drizzle-orm/cache/core/cache-effect"
 import { NoopCache, strategyFor } from "drizzle-orm/cache/core/cache"
 import type { WithCacheConfig } from "drizzle-orm/cache/core/types"
-import { MigratorInitError } from "drizzle-orm/effect-core/errors"
 import { EffectDrizzleQueryError, EffectTransactionRollbackError } from "drizzle-orm/effect-core/errors"
 import type { EffectLoggerShape } from "drizzle-orm/effect-core/logger"
 import type { QueryEffectHKTBase, QueryEffectKind } from "drizzle-orm/effect-core/query-effect"
 import { entityKind, is } from "drizzle-orm/entity"
-import type { MigrationConfig, MigrationMeta } from "drizzle-orm/migrator"
-import { getMigrationsToRun } from "drizzle-orm/migrator.utils"
 import type {
  AnyRelations,
  EmptyRelations,
@@ -20,17 +17,14 @@ import type {
 } from "drizzle-orm/relations"
 import { makeJitRqbMapper } from "drizzle-orm/relations"
 import type { PreparedQuery } from "drizzle-orm/session"
-import { fillPlaceholders, type Query, type SQL, sql } from "drizzle-orm/sql/sql"
+import { fillPlaceholders, type Query, type SQL } from "drizzle-orm/sql/sql"
 import type { SQLiteDialect } from "drizzle-orm/sqlite-core/dialect"
 import type { SelectedFieldsOrdered } from "drizzle-orm/sqlite-core/query-builders/select.types"
 import type { PreparedQueryConfig, SQLiteExecuteMethod, SQLiteTransactionConfig } from "drizzle-orm/sqlite-core/session"
-import { upgradeIfNeeded } from "../../up-migrations/effect-sqlite.js"
 import { assertUnreachable, makeJitQueryMapper, type RowsMapper } from "drizzle-orm/utils"
 import { mapResultRow, resolveNullableObjectPaths } from "../../internal/drizzle-utils.js"
 import { SQLiteEffectDatabase } from "./db.js"

-type MigrationConfigWithInit = MigrationConfig & { init?: boolean }
-
 type SQLiteEffectExecuteMethod = SQLiteExecuteMethod | "values"

 export class SQLiteEffectPreparedQuery<
@@ -429,69 +423,3 @@ export abstract class SQLiteEffectTransaction<
    return new EffectTransactionRollbackError()
  }
 }
-
-export const migrate = Effect.fn("migrate")(function* <TEffectHKT extends QueryEffectHKTBase>(
-  migrations: MigrationMeta[],
-  session: SQLiteEffectSession<TEffectHKT>,
-  config: string | MigrationConfigWithInit,
-) {
-  const migrationsTable =
-    typeof config === "string" ? "__drizzle_migrations" : (config.migrationsTable ?? "__drizzle_migrations")
-
-  const { newDb } = yield* upgradeIfNeeded(migrationsTable, session, migrations)
-
-  if (newDb) {
-    yield* session.run(sql`
-		CREATE TABLE IF NOT EXISTS ${sql.identifier(migrationsTable)} (
-			id INTEGER PRIMARY KEY,
-			hash text NOT NULL,
-			created_at numeric,
-			name text,
-			applied_at TEXT
-		)
-	`)
-  }
-
-  const dbMigrations = yield* session.all<{ id: number; hash: string; created_at: string; name: string | null }>(
-    sql`SELECT id, hash, created_at, name FROM ${sql.identifier(migrationsTable)}`,
-  )
-
-  if (typeof config === "object" && config.init) {
-    if (dbMigrations.length) {
-      return yield* new MigratorInitError({ exitCode: "databaseMigrations" })
-    }
-
-    if (migrations.length > 1) {
-      return yield* new MigratorInitError({ exitCode: "localMigrations" })
-    }
-
-    const [migration] = migrations
-    if (!migration) return
-
-    yield* session.run(
-      sql`insert into ${sql.identifier(
-        migrationsTable,
-      )} ("hash", "created_at", "name", "applied_at") values(${migration.hash}, ${migration.folderMillis}, ${migration.name}, ${new Date().toISOString()})`,
-    )
-
-    return
-  }
-
-  const migrationsToRun = getMigrationsToRun({ localMigrations: migrations, dbMigrations })
-  if (migrationsToRun.length === 0) return
-
-  yield* session.transaction((tx) =>
-    Effect.gen(function* () {
-      for (const migration of migrationsToRun) {
-        for (const stmt of migration.sql) {
-          yield* tx.run(sql.raw(stmt))
-        }
-        yield* tx.run(
-          sql`insert into ${sql.identifier(
-            migrationsTable,
-          )} ("hash", "created_at", "name", "applied_at") values(${migration.hash}, ${migration.folderMillis}, ${migration.name}, ${new Date().toISOString()})`,
-        )
-      }
-    }),
-  )
-})
@@ -1,102 +0,0 @@
-/* oxlint-disable */
-import * as Effect from "effect/Effect"
-import type { SqlError } from "effect/unstable/sql/SqlError"
-import { EffectDrizzleError } from "drizzle-orm/effect-core/errors"
-import type { QueryEffectHKTBase } from "drizzle-orm/effect-core/query-effect"
-import type { MigrationMeta } from "drizzle-orm/migrator"
-import { sql } from "drizzle-orm/sql/sql"
-import type { SQLiteEffectSession } from "../sqlite-core/effect/session.js"
-import {
-  buildSQLiteMigrationBackfillStatements,
-  prepareSQLiteMigrationBackfill,
-  type SQLiteMigrationTableRow,
-} from "./sqlite.js"
-import { GET_VERSION_FOR, MIGRATIONS_TABLE_VERSIONS, type UpgradeResult } from "./utils.js"
-
-const migrationUpgradeError = (cause: unknown) =>
-  new EffectDrizzleError({
-    message:
-      typeof cause === "object" && cause !== null && "message" in cause && typeof cause.message === "string"
-        ? cause.message
-        : String(cause),
-    cause,
-  })
-
-export const upgradeIfNeeded: <TEffectHKT extends QueryEffectHKTBase>(
-  migrationsTable: string,
-  session: SQLiteEffectSession<TEffectHKT>,
-  localMigrations: MigrationMeta[],
-) => Effect.Effect<UpgradeResult, EffectDrizzleError | TEffectHKT["error"] | SqlError, TEffectHKT["context"]> =
-  Effect.fn("upgradeIfNeeded")(function* <TEffectHKT extends QueryEffectHKTBase>(
-    migrationsTable: string,
-    session: SQLiteEffectSession<TEffectHKT>,
-    localMigrations: MigrationMeta[],
-  ) {
-    const tableExists = yield* session.all(
-      sql`SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = ${migrationsTable}`,
-    )
-
-    if (tableExists.length === 0) {
-      return { newDb: true }
-    }
-
-    const rows = yield* session.all<{ column_name: string }>(
-      sql`SELECT name as column_name FROM pragma_table_info(${migrationsTable})`,
-    )
-
-    const version = GET_VERSION_FOR.sqlite(rows.map((r) => r.column_name))
-
-    for (let v = version; v < MIGRATIONS_TABLE_VERSIONS.sqlite; v++) {
-      const upgradeFn = upgradeFunctions[v]
-      if (!upgradeFn) {
-        return yield* new EffectDrizzleError({
-          message: `No upgrade path from migration table version ${v} to ${v + 1}`,
-          cause: { version: v },
-        })
-      }
-      yield* upgradeFn(migrationsTable, session, localMigrations)
-    }
-
-    return { newDb: false }
-  })
-
-const upgradeFunctions: Record<
-  number,
-  <TEffectHKT extends QueryEffectHKTBase>(
-    migrationsTable: string,
-    session: SQLiteEffectSession<TEffectHKT>,
-    localMigrations: MigrationMeta[],
-  ) => Effect.Effect<void, EffectDrizzleError | TEffectHKT["error"] | SqlError, TEffectHKT["context"]>
-> = {
-  0: upgradeFromV0,
-}
-
-function upgradeFromV0<TEffectHKT extends QueryEffectHKTBase>(
-  migrationsTable: string,
-  session: SQLiteEffectSession<TEffectHKT>,
-  localMigrations: MigrationMeta[],
-): Effect.Effect<void, EffectDrizzleError | TEffectHKT["error"] | SqlError, TEffectHKT["context"]> {
-  return Effect.gen(function* () {
-    const table = sql`${sql.identifier(migrationsTable)}`
-
-    const dbRows = yield* session.all<SQLiteMigrationTableRow>(
-      sql`SELECT id, hash, created_at FROM ${table} ORDER BY id ASC`,
-    )
-    const statements = yield* Effect.try({
-      try: () =>
-        buildSQLiteMigrationBackfillStatements(
-          migrationsTable,
-          prepareSQLiteMigrationBackfill(dbRows, localMigrations),
-        ),
-      catch: migrationUpgradeError,
-    })
-
-    yield* session.transaction((tx) =>
-      Effect.gen(function* () {
-        for (const statement of statements) {
-          yield* tx.run(statement)
-        }
-      }),
-    )
-  })
-}
@@ -1,105 +0,0 @@
-/* oxlint-disable */
-import type { MigrationMeta } from "drizzle-orm/migrator"
-import { type SQL, sql } from "drizzle-orm/sql/sql"
-
-/** @internal */
-export type SQLiteMigrationTableRow = { id: number | null; hash: string; created_at: number }
-
-type SQLiteMigrationBackfillEntry = {
-  name: string
-  selector:
-    | { column: "id"; value: number }
-    | { column: "created_at"; value: number }
-    | { column: "hash"; value: string }
-}
-
-function unmatchedMigrationError(unmatched: SQLiteMigrationTableRow[]) {
-  return new Error(
-    `While upgrading your database migrations table we found ${unmatched.length} (${unmatched
-      .map((it) => `[id: ${it.id}, created_at: ${it.created_at}]`)
-      .join(
-        ", ",
-      )}) migrations in the database that do not match any local migration. This means that some migrations were applied to the database but are missing from the local environment`,
-  )
-}
-
-/** @internal */
-export function prepareSQLiteMigrationBackfill(
-  dbRows: SQLiteMigrationTableRow[],
-  localMigrations: MigrationMeta[],
-): SQLiteMigrationBackfillEntry[] {
-  const sortedLocalMigrations = [...localMigrations].sort((a, b) =>
-    a.folderMillis !== b.folderMillis ? a.folderMillis - b.folderMillis : (a.name ?? "").localeCompare(b.name ?? ""),
-  )
-  const byMillis = new Map<number, MigrationMeta[]>()
-  const byHash = new Map<string, MigrationMeta>()
-  for (const migration of sortedLocalMigrations) {
-    if (!byMillis.has(migration.folderMillis)) {
-      byMillis.set(migration.folderMillis, [])
-    }
-    byMillis.get(migration.folderMillis)!.push(migration)
-    byHash.set(migration.hash, migration)
-  }
-
-  const toApply: SQLiteMigrationBackfillEntry[] = []
-  const unmatched: SQLiteMigrationTableRow[] = []
-
-  for (const dbRow of dbRows) {
-    const stringified = String(dbRow.created_at)
-    const millis = Number(stringified.substring(0, stringified.length - 3) + "000")
-    const candidates = byMillis.get(millis)
-
-    const matchedByMillis = candidates?.length === 1 ? candidates[0] : undefined
-    const matchedByCandidateHash =
-      candidates && candidates.length > 1
-        ? candidates.find((candidate) => candidate.hash && dbRow.hash && candidate.hash === dbRow.hash)
-        : undefined
-    const matchedByHash = matchedByMillis || matchedByCandidateHash ? undefined : byHash.get(dbRow.hash)
-    const matched = matchedByMillis ?? matchedByCandidateHash ?? matchedByHash
-
-    if (matched) {
-      toApply.push({
-        name: matched.name,
-        selector:
-          dbRow.id !== null
-            ? { column: "id", value: dbRow.id }
-            : matchedByMillis
-              ? { column: "created_at", value: dbRow.created_at }
-              : { column: "hash", value: dbRow.hash },
-      })
-      continue
-    }
-
-    unmatched.push(dbRow)
-  }
-
-  if (unmatched.length > 0) {
-    throw unmatchedMigrationError(unmatched)
-  }
-
-  return toApply
-}
-
-/** @internal */
-export function buildSQLiteMigrationBackfillStatements(
-  migrationsTable: string,
-  backfillEntries: SQLiteMigrationBackfillEntry[],
-) {
-  const table = sql`${sql.identifier(migrationsTable)}`
-  const statements: SQL[] = [
-    sql`ALTER TABLE ${table} ADD COLUMN ${sql.identifier("name")} text`,
-    sql`ALTER TABLE ${table} ADD COLUMN ${sql.identifier("applied_at")} TEXT`,
-  ]
-
-  for (const backfillEntry of backfillEntries) {
-    const updateQuery = sql`UPDATE ${table} SET ${sql.identifier("name")} = ${backfillEntry.name}, ${sql.identifier(
-      "applied_at",
-    )} = NULL WHERE`
-
-    updateQuery.append(sql` ${sql.identifier(backfillEntry.selector.column)} = ${backfillEntry.selector.value}`)
-
-    statements.push(updateQuery)
-  }
-
-  return statements
-}
@@ -1,45 +0,0 @@
-/* oxlint-disable */
-export interface UpgradeResult {
-  newDb: boolean
-}
-
-export const MIGRATIONS_TABLE_VERSIONS = {
-  sqlite: 1,
-  pg: 1,
-  effect: 1,
-  mysql: 1,
-  mssql: 1,
-  cockroach: 1,
-  singlestore: 1,
-} as const
-
-export const GET_VERSION_FOR = {
-  mysql: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  pg: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  effect: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  mssql: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  cockroach: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  singlestore: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  sqlite: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-} as const
@@ -66,12 +66,39 @@ export function applyOnly(db: Database, input: Migration[]) {
      if (
        yield* db.get(sql`SELECT name FROM sqlite_master WHERE type = 'table' AND name = ${"__drizzle_migrations"}`)
      ) {
-        yield* db.run(sql`
-          INSERT OR IGNORE INTO ${sql.identifier("migration")} (id, time_completed)
-          SELECT name, ${Date.now()}
-          FROM ${sql.identifier("__drizzle_migrations")}
-          WHERE name IS NOT NULL
-        `)
+        const named = (yield* db.all<{ name: string }>(
+          sql`SELECT name FROM pragma_table_info('__drizzle_migrations')`,
+        )).some((column) => column.name === "name")
+
+        if (named) {
+          yield* db.run(sql`
+            INSERT OR IGNORE INTO ${sql.identifier("migration")} (id, time_completed)
+            SELECT name, ${Date.now()}
+            FROM ${sql.identifier("__drizzle_migrations")}
+            WHERE name IS NOT NULL
+          `)
+        }
+
+        if (!named) {
+          const entries = yield* db.all<{ created_at: number; prefix: string | null }>(sql`
+            SELECT created_at, strftime('%Y%m%d%H%M%S', created_at / 1000, 'unixepoch') AS prefix
+            FROM ${sql.identifier("__drizzle_migrations")}
+            WHERE created_at IS NOT NULL
+          `)
+
+          for (const entry of entries) {
+            const migration = input.find((item) => item.id.startsWith(`${entry.prefix}_`))
+            if (!migration) {
+              return yield* Effect.die(
+                new Error(`Legacy migration timestamp ${entry.created_at} does not match any known migration`),
+              )
+            }
+            yield* db.run(sql`
+              INSERT OR IGNORE INTO ${sql.identifier("migration")} (id, time_completed)
+              VALUES (${migration.id}, ${Date.now()})
+            `)
+          }
+        }
        completed = new Set(
          (yield* db.all<{ id: string }>(sql`SELECT id FROM ${sql.identifier("migration")}`)).map((row) => row.id),
        )
@@ -5,6 +5,9 @@ const migration: DatabaseMigration.Migration = {
  id: "20260410174513_workspace-name",
  up(tx) {
    return Effect.gen(function* () {
+      const columns = yield* tx.all<{ name: string }>(`PRAGMA table_info(\`workspace\`)`)
+      const name = columns.some((column) => column.name === "name") ? "`name`" : "''"
+
      yield* tx.run(`PRAGMA foreign_keys=OFF;`)
      yield* tx.run(`
        CREATE TABLE \`__new_workspace\` (
@@ -19,7 +22,7 @@ const migration: DatabaseMigration.Migration = {
        );
      `)
      yield* tx.run(
-        `INSERT INTO \`__new_workspace\`(\`id\`, \`type\`, \`branch\`, \`name\`, \`directory\`, \`extra\`, \`project_id\`) SELECT \`id\`, \`type\`, \`branch\`, \`name\`, \`directory\`, \`extra\`, \`project_id\` FROM \`workspace\`;`,
+        `INSERT INTO \`__new_workspace\`(\`id\`, \`type\`, \`branch\`, \`name\`, \`directory\`, \`extra\`, \`project_id\`) SELECT \`id\`, \`type\`, \`branch\`, ${name}, \`directory\`, \`extra\`, \`project_id\` FROM \`workspace\`;`,
      )
      yield* tx.run(`DROP TABLE \`workspace\`;`)
      yield* tx.run(`ALTER TABLE \`__new_workspace\` RENAME TO \`workspace\`;`)
@@ -166,6 +166,8 @@ export interface Interface extends State.Transformable<Draft> {
      /** User-facing label for the stored credential. */
      readonly label?: string
    }) => Effect.Effect<void, AuthorizationError>
+    /** Selects a stored credential as the active integration connection. */
+    readonly activate: (credentialID: Credential.ID) => Effect.Effect<void>
    /** Updates a stored credential exposed as a connection. */
    readonly update: (
      credentialID: Credential.ID,
@@ -329,6 +331,17 @@ const layer = Layer.effect(
      finalize: () => bus.publish(Integration.Event.Updated, {}).pipe(Effect.asVoid),
    })

+    const createCredential = Effect.fnUntraced(function* (input: Parameters<Credential.Interface["create"]>[0]) {
+      if (input.label !== undefined) return yield* credentials.create(input)
+      const name = state.get().integrations.get(input.integrationID)?.ref.name ?? input.integrationID
+      const labels = new Set((yield* credentials.list(input.integrationID)).map((credential) => credential.label))
+      const label =
+        Array.from({ length: labels.size + 1 }, (_, index) => (index === 0 ? name : `${name} ${index + 1}`)).find(
+          (candidate) => !labels.has(candidate),
+        ) ?? name
+      return yield* credentials.create({ ...input, label })
+    })
+
    const resolveConnections = (entry: Entry | undefined, saved: readonly Credential.Info[]) => {
      const credentials = saved
        .map((credential) => ({
@@ -348,6 +361,7 @@ const layer = Layer.effect(
      Info.make({
        id: entry.ref.id,
        name: entry.ref.name,
+        ...(entry.ref.metadata === undefined ? {} : { metadata: entry.ref.metadata }),
        methods: entry.methods,
        connections,
      })
@@ -394,7 +408,7 @@ const layer = Layer.effect(
              ?.implementations.get(attempt.methodID)
            const persistence = yield* Effect.sync(() => attempt.label ?? implementation?.label?.(exit.value)).pipe(
              Effect.flatMap((label) =>
-                credentials.create({
+                createCredential({
                  integrationID: attempt.integrationID,
                  label,
                  value: exit.value,
@@ -421,8 +435,6 @@ const layer = Layer.effect(
            // Persisting attempts cannot be cancelled, expired, or claimed again.
            yield* SynchronizedRef.update(attempts, (current) => new Map(current).set(attemptID, terminal))
            if (Exit.isFailure(persistence)) yield* Effect.failCause(persistence.cause)
-            yield* bus.publish(Integration.Event.ConnectionUpdated, { integrationID: attempt.integrationID })
-            yield* bus.publish(Integration.Event.Updated, {})
          }).pipe(Effect.ensuring(close(attempt.scope)))
        }),
      )
@@ -452,13 +464,11 @@ const layer = Layer.effect(
            return
          }

-          const persistence = yield* credentials
-            .create({
-              integrationID: attempt.integrationID,
-              label: attempt.label,
-              value: Credential.Key.make({ type: "key", key: exit.value }),
-            })
-            .pipe(Effect.asVoid, Effect.exit)
+          const persistence = yield* createCredential({
+            integrationID: attempt.integrationID,
+            label: attempt.label,
+            value: Credential.Key.make({ type: "key", key: exit.value }),
+          }).pipe(Effect.asVoid, Effect.exit)
          const settledAt = yield* Clock.currentTimeMillis
          const terminal: TerminalCommandAttempt = Exit.isSuccess(persistence)
            ? {
@@ -476,9 +486,6 @@ const layer = Layer.effect(
              }
          yield* SynchronizedRef.update(commandAttempts, (current) => new Map(current).set(attemptID, terminal))
          yield* close(attempt.scope)
-          if (Exit.isFailure(persistence)) return
-          yield* bus.publish(Integration.Event.ConnectionUpdated, { integrationID: attempt.integrationID })
-          yield* bus.publish(Integration.Event.Updated, {})
        }),
      )
    })
@@ -702,7 +709,7 @@ const layer = Layer.effect(
          if (!method.form && Object.keys(answer).length > 0) {
            return yield* new AuthorizationError({ cause: new Error("Key method does not accept a form answer") })
          }
-          yield* credentials.create({
+          yield* createCredential({
            integrationID: input.integrationID,
            label: input.label,
            value: Credential.Key.make({
@@ -711,25 +718,21 @@ const layer = Layer.effect(
              ...(Object.keys(answer).length > 0 ? { configuration: answer } : {}),
            }),
          })
-          yield* bus.publish(Integration.Event.ConnectionUpdated, { integrationID: input.integrationID })
-          yield* bus.publish(Integration.Event.Updated, {})
        }),
-        update: Effect.fn("Integration.connection.update")(function* (credentialID, updates) {
+        activate: Effect.fn("Integration.connection.activate")(function* (credentialID) {
          const credential = yield* credentials.get(credentialID)
-          yield* credentials.update(credentialID, updates)
-          if (credential) {
-            yield* bus.publish(Integration.Event.ConnectionUpdated, { integrationID: credential.integrationID })
-          }
-          yield* bus.publish(Integration.Event.Updated, {})
-        }),
-        remove: Effect.fn("Integration.connection.remove")(function* (credentialID) {
-          const credential = yield* credentials.get(credentialID)
-          yield* credentials.remove(credentialID)
-          if (credential) {
-            yield* bus.publish(Integration.Event.ConnectionUpdated, { integrationID: credential.integrationID })
-          }
-          yield* bus.publish(Integration.Event.Updated, {})
+          if (!credential) return
+          const active = resolveConnections(
+            state.get().integrations.get(credential.integrationID),
+            yield* credentials.list(credential.integrationID),
+          )[0]
+          if (active?.type === "credential" && active.id === credentialID) return
+          yield* credentials.activate(credentialID)
        }),
+        update: Effect.fn("Integration.connection.update")((credentialID, updates) =>
+          credentials.update(credentialID, updates),
+        ),
+        remove: Effect.fn("Integration.connection.remove")((credentialID) => credentials.remove(credentialID)),
      },
      oauth: {
        connect: connectOAuth,
@@ -140,8 +140,7 @@ export type Draft = {
  remove: (server: ServerName | string) => void
 }

-const cloneConfig = (config: Mcp.ServerConfig) =>
-  structuredClone(config) as Types.DeepMutable<Mcp.ServerConfig>
+const cloneConfig = (config: Mcp.ServerConfig) => structuredClone(config) as Types.DeepMutable<Mcp.ServerConfig>

 export interface Interface extends State.Transformable<Draft> {
  readonly servers: () => Effect.Effect<ServerInfo[]>
@@ -229,6 +228,7 @@ export const layer = (options?: Options) =>
          .transform((draft) => {
            draft.update(integrationID, (ref) => {
              ref.name = name
+              ref.metadata = { source: "mcp" }
            })
            draft.method.update({
              integrationID,
@@ -264,8 +264,7 @@ export const layer = (options?: Options) =>
          // No browser during connect: an auth-gated server surfaces needs_auth instead of opening a browser.
          onRedirect: () => {},
        }
-        const stored = yield* credentials.list(entry.integrationID)
-        const found = stored.find((credential) => credential.value.type === "oauth")
+        const found = (yield* credentials.list(entry.integrationID)).at(-1)
        if (!found || found.value.type !== "oauth")
          // No stored credential yet: an empty in-memory store still lets the SDK run the auth handshake, which
          // ends in UnauthorizedError -> needs_auth. Returning no provider instead would let the transport throw
@@ -288,8 +287,8 @@ export const layer = (options?: Options) =>
          // Drop a credential the SDK rejected so the next connect cleanly reports needs_auth — but only if it is
          // still the stored one. Rotating servers hand out a fresh refresh token per use, so a concurrent
          // connection may have already replaced ours; deleting then would discard the newer valid credential and
-          // strand every connection in needs_auth until a manual re-auth. Uses the raw credential service (no
-          // integration event) to avoid re-triggering the reconnect subscriber mid-connect.
+          // strand every connection in needs_auth until a manual re-auth. Credential deletion notifies all locations;
+          // reconnects remain serialized by the server lock.
          invalidate: async (scope) => {
            if (scope === "verifier" || scope === "discovery") return
            const oauth = await readOAuthCredential()
@@ -673,7 +672,7 @@ export const layer = (options?: Options) =>
          }).pipe(locks.withLock(name))
        })
      fork(
-        bus.subscribe(Integration.Event.ConnectionUpdated).pipe(
+        bus.subscribe(Credential.Event.Switched).pipe(
          Stream.filter((event) => owned.has(event.data.integrationID)),
          Stream.runForEach((event) => Effect.sync(() => fork(reconnect(event.data.integrationID)))),
          Effect.ignore,
@@ -687,9 +686,7 @@ export const layer = (options?: Options) =>
              config === false ? [] : [[name, cloneConfig(config)] as const],
            ),
          ),
-          removed: new Set(
-            Array.from(overrides).flatMap(([name, config]) => (config === false ? [name] : [])),
-          ),
+          removed: new Set(Array.from(overrides).flatMap(([name, config]) => (config === false ? [name] : []))),
        }),
        draft: (draft) => ({
          list: () => Array.from(draft.servers),
@@ -333,8 +333,11 @@ function usesAPIKeyAuth(packageName: string | undefined) {
  return (
    name === "@ai-sdk/openai" ||
    name === "@ai-sdk/anthropic" ||
+    name === "@ai-sdk/cerebras" ||
+    name === "@ai-sdk/deepinfra" ||
    name === "@ai-sdk/openai-compatible" ||
    name === "@ai-sdk/google" ||
+    name === "@ai-sdk/togetherai" ||
    name === "@ai-sdk/xai" ||
    name === "@openrouter/ai-sdk-provider" ||
    name === "@ai-sdk/azure" ||
@@ -342,8 +345,11 @@ function usesAPIKeyAuth(packageName: string | undefined) {
    name?.startsWith("@opencode-ai/ai/providers/openai/") === true ||
    name === "@opencode-ai/ai/providers/anthropic" ||
    name === "@opencode-ai/ai/providers/anthropic-compatible" ||
+    name === "@opencode-ai/ai/providers/cerebras" ||
+    name === "@opencode-ai/ai/providers/deepinfra" ||
    name === "@opencode-ai/ai/providers/openai-compatible" ||
    name === "@opencode-ai/ai/providers/google" ||
+    name === "@opencode-ai/ai/providers/togetherai" ||
    name === "@opencode-ai/ai/providers/xai" ||
    name === "@opencode-ai/ai/providers/openrouter" ||
    name === "@opencode-ai/ai/providers/azure" ||
@@ -11,6 +11,7 @@ import { SessionSchema } from "./session/schema.js"
 import { SessionStore } from "./session/store.js"
 import { Wildcard } from "./util/wildcard.js"
 import { PermissionSaved } from "./permission/saved.js"
+import { PluginHooks } from "./plugin/hooks.js"

 const PermissionEffect = Permission.Effect
 export { PermissionEffect as Effect }
@@ -70,9 +71,10 @@ export class BlockedError extends Schema.TaggedError<BlockedError>()("Permission
  rules: Permission.Ruleset,
  permission: Schema.String,
  resources: Schema.Array(Schema.String),
+  reason: Schema.String.pipe(Schema.optional),
 }) {
  override get message() {
-    return `Permission denied: ${this.permission}`
+    return this.reason ?? `Permission denied: ${this.permission}`
  }
 }

@@ -99,11 +101,6 @@ export function merge(...rulesets: Permission.Ruleset[]): Permission.Ruleset {
 }

 export interface Interface {
-  readonly allowsAll: (input: {
-    readonly sessionID: SessionSchema.ID
-    readonly action: string
-    readonly agent?: Agent.ID
-  }) => Effect.Effect<boolean, SessionErrors.NotFoundError>
  readonly ask: (input: AssertInput) => Effect.Effect<AskResult, SessionErrors.NotFoundError>
  readonly assert: (input: AssertInput) => Effect.Effect<void, Error | SessionErrors.NotFoundError>
  readonly reply: (input: ReplyInput) => Effect.Effect<void, NotFoundError>
@@ -128,6 +125,7 @@ const layer = Layer.effect(
    const agents = yield* Agent.Service
    const sessions = yield* SessionStore.Service
    const saved = yield* PermissionSaved.Service
+    const hooks = yield* PluginHooks.Service
    const pending = new Map<ID, Pending>()

    yield* Effect.addFinalizer(() =>
@@ -159,24 +157,6 @@ const layer = Layer.effect(
      return agent?.permissions ?? missingAgentPermissions
    })

-    const allowsAll = Effect.fnUntraced(function* (input: {
-      readonly sessionID: SessionSchema.ID
-      readonly action: string
-      readonly agent?: Agent.ID
-    }) {
-      const rules = yield* configured(input.sessionID, input.agent)
-      const relevant = rules.filter((rule) => Wildcard.match(input.action, rule.action))
-      for (let index = relevant.length - 1; index >= 0; index--) {
-        const rule = relevant[index]
-        if (rule.resource !== "*") {
-          if (rule.effect !== "allow") return false
-          continue
-        }
-        return rule.effect === "allow"
-      }
-      return false
-    })
-
    function denied(input: Pick<Request, "action" | "resources">, rules: Permission.Ruleset) {
      return input.resources.some((resource) => evaluate(input.action, resource, rules).effect === "deny")
    }
@@ -191,10 +171,19 @@ const layer = Layer.effect(
      const all = [...rules, ...(yield* savedRules())]
      const effects = input.resources.map((resource) => evaluate(input.action, resource, all).effect)
      const effect: Permission.Effect = effects.includes("deny") ? "deny" : effects.includes("ask") ? "ask" : "allow"
-      return { effect, rules: all }
+      const event = yield* hooks.trigger("permission", "evaluate", {
+        sessionID: input.sessionID,
+        agent: input.agent,
+        action: input.action,
+        resources: input.resources,
+        metadata: input.metadata,
+        source: input.source,
+        effect,
+      })
+      return { effect: event.effect, message: event.message, rules: all }
    })

-    function request(input: AssertInput): Request {
+    function request(input: AssertInput, message?: string): Request {
      return {
        id: input.id ?? ID.create(),
        sessionID: input.sessionID,
@@ -203,6 +192,7 @@ const layer = Layer.effect(
        save: input.save,
        metadata: input.metadata,
        source: input.source,
+        message,
      }
    }

@@ -223,39 +213,42 @@ const layer = Layer.effect(

    const ask = Effect.fn("Permission.ask")(function* (input: AssertInput) {
      const result = yield* evaluateInput(input)
-      const value = request(input)
+      const value = request(input, result.message)
      if (result.effect === "ask") yield* create(value, input.agent)
      return { id: value.id, effect: result.effect }
    })

    const assert = Effect.fn("Permission.assert")((input: AssertInput) =>
-      Effect.uninterruptibleMask((restore) =>
-        Effect.gen(function* () {
-          const result = yield* evaluateInput(input)
-          if (result.effect === "deny") {
-            return yield* new BlockedError({
-              rules: relevant(input, result.rules),
-              permission: input.action,
-              resources: input.resources,
-            })
-          }
-          if (result.effect === "allow") return
-          const item = yield* create(request(input), input.agent)
-          return yield* restore(Deferred.await(item.deferred)).pipe(
-            // Deliberate defect tunnel: leaves wrap execution in blanket `mapError`, which
-            // must not convert a user's decline into model-facing tool output. The decline
-            // resurfaces as a typed failure at SessionModelRequest.executeTool. A decline
-            // WITH feedback (CorrectedError) intentionally stays typed so the leaf can turn
-            // it into ToolFailure and the model continues.
-            Effect.catchTag("Permission.DeclinedError", (error) => Effect.die(error)),
-            Effect.ensuring(
-              Effect.sync(() => {
-                pending.delete(item.request.id)
-              }),
-            ),
-          )
-        }),
-      ),
+      Effect.gen(function* () {
+        const result = yield* evaluateInput(input)
+        return yield* Effect.uninterruptibleMask((restore) =>
+          Effect.gen(function* () {
+            if (result.effect === "deny") {
+              return yield* new BlockedError({
+                rules: relevant(input, result.rules),
+                permission: input.action,
+                resources: input.resources,
+                reason: result.message,
+              })
+            }
+            if (result.effect === "allow") return
+            const item = yield* create(request(input, result.message), input.agent)
+            return yield* restore(Deferred.await(item.deferred)).pipe(
+              // Deliberate defect tunnel: leaves wrap execution in blanket `mapError`, which
+              // must not convert a user's decline into model-facing tool output. The decline
+              // resurfaces as a typed failure at SessionModelRequest.executeTool. A decline
+              // WITH feedback (CorrectedError) intentionally stays typed so the leaf can turn
+              // it into ToolFailure and the model continues.
+              Effect.catchTag("Permission.DeclinedError", (error) => Effect.die(error)),
+              Effect.ensuring(
+                Effect.sync(() => {
+                  pending.delete(item.request.id)
+                }),
+              ),
+            )
+          }),
+        )
+      }),
    )

    const reply = Effect.fn("Permission.reply")((input: ReplyInput) =>
@@ -337,12 +330,12 @@ const layer = Layer.effect(
      return Array.from(pending.values(), (item) => item.request).filter((request) => request.sessionID === sessionID)
    })

-    return Service.of({ allowsAll, ask, assert, reply, get, forSession, list })
+    return Service.of({ ask, assert, reply, get, forSession, list })
  }),
 )

 export const node = makeLocationNode({
  service: Service,
  layer,
-  deps: [Bus.node, Location.node, Agent.node, SessionStore.node, PermissionSaved.node],
+  deps: [Bus.node, Location.node, Agent.node, SessionStore.node, PermissionSaved.node, PluginHooks.node],
 })
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Aiden Cline	481defdf4e	fix(ai): disable bedrock mantle responses websockets	2026-08-25 22:23:11 -05:00
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]	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]	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
Kit Langton	4b71ae6a0d	feat(core): support git plugin packages (#45110 )	2026-08-25 22:02:49 -04:00
opencode-agent[bot]	e211b6f30e	test: run only affected unit suites (#45034 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 09:51:44 +08:00
Kit Langton	ab4621e437	fix(sdk): keep packaged Effect runtime coherent (#45122 )	2026-08-26 01:51:36 +00:00
Kit Langton	97ba700fac	fix(tui): preserve prompt metadata visibility across sessions (#45116 )	2026-08-25 21:37:58 -04:00
Luke Parker	03fb5c6c67	fix(app): prevent clipped virtual timeline rows (#45115 )	2026-08-26 01:33:32 +00:00
Luke Parker	5a54eb4afc	fix(app): stream running shell tool output (#45106 )	2026-08-26 00:58:15 +00:00
Kit Langton	c4dcf72e13	fix(tui): detect clipped transcript bottom (#45100 )	2026-08-25 20:46:52 -04:00
opencode-agent[bot]	27e0de6b23	chore: update nix node_modules hashes	2026-08-26 00:41:39 +00:00
Kit Langton	73d7b1d4c1	fix(tui): preserve interrupted Mermaid diagrams (#45102 )	2026-08-26 00:25:00 +00:00
Kit Langton	690ad8e8bd	test(core): isolate host configuration and credentials (#44845 )	2026-08-25 20:22:21 -04:00
Aiden Cline	6c97be6974	feat(ai): add native Cerebras and Together AI providers (#45098 )	2026-08-25 19:20:47 -05:00
Dax Raad	6cd1ffac50	chore: synchronize bun lockfile	2026-08-25 19:18:17 -04:00
Aiden Cline	f08c234890	fix(ai): ignore SSE retry directives without ending streams (#45093 )	2026-08-25 18:17:45 -05:00
Kit Langton	7f2b052db6	refactor(core): remove unused Drizzle migration framework	2026-08-25 19:16:26 -04:00
Dax Raad	297a3328c6	feat(tui): group MCP integrations in connection dialog	2026-08-25 19:14:12 -04:00
Kit Langton	7f9e5e91ab	feat(tui): add experimental session preview tabs (#45021 )	2026-08-25 18:51:36 -04:00
Aiden Cline	3726e3254d	fix(ai): enable Vertex Anthropic prompt caching (#45088 )	2026-08-25 17:44:32 -05:00
Aiden Cline	24605d048f	fix(ai): send responses instructions at top level (#45085 )	2026-08-25 17:34:40 -05:00
Aiden Cline	0a84625618	fix(ai): accept responses calls without item ids (#45081 )	2026-08-25 17:29:57 -05:00
Major Hayden	2e7f06a155	fix(ai): preserve Vertex Anthropic tool continuations (#43498 ) Signed-off-by: Major Hayden <major@mhtx.net>	2026-08-25 17:22:51 -05:00
Aiden Cline	c2a3b813a0	fix(ai): require reasoning fields for deepseek assistants (#45075 )	2026-08-25 17:13:17 -05:00
opencode-agent[bot]	b79cad5ec8	fix(core): respect automatic compaction opt-out on overflow (#45036 ) Co-authored-by: rekram1-node <rekram1-node@users.noreply.github.com>	2026-08-25 16:58:39 -05:00
opencode-agent[bot]	61dd296161	fix(www): generate CLI schema without tracking it (#45080 ) Co-authored-by: neriousy <34747899+neriousy@users.noreply.github.com>	2026-08-25 23:54:07 +02:00
Filip	66a790c624	chore(www): regenerate documentation artifacts (#45077 )	2026-08-25 23:33:01 +02:00
Aiden Cline	0ae3aac317	fix(ai): replay responses history independently of storage (#45050 )	2026-08-25 16:28:22 -05:00
Dax	eb1ac54d73	feat(tui): support multiple integration accounts (#45072 )	2026-08-25 17:03:16 -04:00
opencode-agent[bot]	d2100a51f1	feat(cli): restore debug paths command (#45063 ) Co-authored-by: rekram1-node <rekram1-node@users.noreply.github.com>	2026-08-25 23:01:43 +02:00
opencode-agent[bot]	d4803ffe38	feat(cli): publish generated CLI configuration schema (#45070 ) Co-authored-by: neriousy <34747899+neriousy@users.noreply.github.com>	2026-08-25 22:49:45 +02:00
Kit Langton	db7837814c	fix(project): refresh renamed projects across live clients (#45043 )	2026-08-25 20:13:10 +00:00
opencode-agent[bot]	543a4f4912	fix(ci): preserve required e2e matrix checks (#45065 ) Co-authored-by: neriousy <34747899+neriousy@users.noreply.github.com>	2026-08-25 22:11:43 +02:00
opencode-agent[bot]	d562f6df1e	fix(core): recover legacy database migration history (#45062 ) Co-authored-by: neriousy <34747899+neriousy@users.noreply.github.com>	2026-08-25 22:02:26 +02:00
Aiden Cline	263a442a6c	fix(ai): normalize chat tool call ids (#45056 )	2026-08-25 14:39:00 -05:00
Filip	82947af8e3	fix(core): ignore SSE comment heartbeats (#43626 )	2026-08-25 21:34:08 +02:00
Filip	e1afdaac52	fix(core): fall back on oversized websocket requests v2 (#43100 )	2026-08-25 21:27:08 +02:00
Aiden Cline	a5829431b0	fix(ai): bridge tool results for mistral family models (#45051 )	2026-08-25 13:40:35 -05:00
Aiden Cline	1ca82d154c	fix(ai): omit empty chat assistant messages (#45046 )	2026-08-25 12:56:24 -05:00
James Long	bc1f67e518	feat(cli): embed persistent PTY service binaries (#44970 )	2026-08-25 13:19:48 -04:00
Kit Langton	1f45962c84	fix(tui): support inline session rename and title regeneration (#45023 )	2026-08-25 13:10:22 -04:00
Kit Langton	88242e21a8	fix(tui): stop completed exploration groups from spinning	2026-08-25 13:09:03 -04:00
opencode-agent[bot]	0fd719067d	feat(plugin): add permission review hooks (#45003 ) Co-authored-by: R44VC0RP <R44VC0RP@users.noreply.github.com> Co-authored-by: nexxeln <nexxeln@users.noreply.github.com> Co-authored-by: thdxr <thdxr@users.noreply.github.com>	2026-08-25 16:57:17 +00:00
Aiden Cline	c94a4913c0	fix(ai): preserve unencrypted reasoning history (#45032 )	2026-08-25 11:44:44 -05:00
opencode-agent[bot]	6c32ba81e2	test: skip unaffected app e2e tests (#45030 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 00:33:14 +08:00
Kit Langton	938a82226a	test(tui): synchronize tab close frame before pointer moves (#45017 )	2026-08-25 12:30:20 -04:00
Aiden Cline	c46b76b58e	fix(ai): route response events by output index (#45013 )	2026-08-25 11:27:40 -05:00
opencode-agent[bot]	fbb3730fdd	chore: update nix node_modules hashes	2026-08-25 16:26:05 +00:00
opencode-agent[bot]	96cff7bb7a	fix(app): clarify context usage label (#45012 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-26 00:20:00 +08:00
Kit Langton	578f8d637a	fix(server): pin shared Effect platform runtime (#45016 )	2026-08-25 12:19:24 -04:00