fix(ai): keep chat reasoning in one lifecycle

fix(ai): finish chat streams at done sentinel (#45850 )
fix(ai): ignore late converse tool deltas (#45847 )
2026-08-28 12:36:15 +00:00 · 2026-08-28 01:46:25 -05:00 · 2026-08-28 00:53:53 -05:00 · 2026-08-28 00:38:01 -05:00 · 2026-08-28 00:21:17 -05:00 · 2026-08-28 00:19:15 -05:00
814 changed files with 49450 additions and 20372 deletions
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Preserve prompt cache prefixes when sessions move between locations with unchanged instructions.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Fix OpenCode Console device authorization URLs when the server returns an origin-rooted verification path.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Prompt and synthetic inbox ID reuse is now idempotent: reusing an ID within the same Session succeeds and returns the first admission, ignoring the retried payload, metadata, and delivery mode. Previously reuse with a differing payload failed with a conflict. Cross-Session and cross-type reuse still fail, and control items keep their operation-specific conflict behavior.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Nested AGENTS.md instructions are re-injected after compaction. Previously the in-memory dedup claim outlived the synthetic message that compaction dropped from model-visible history, so nested instructions were silently lost for the rest of the process lifetime. The claim now only guards in-flight loads; the synthetic message metadata in durable history is the sole lasting ledger, so any history truncation (compaction, revert) self-heals on the next read in that subtree.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/plugin": patch
---
-
-Derive Promise plugin API request and response conversion from the canonical protocol schemas.
@@ -1,8 +0,0 @@
---
-"@opencode-ai/core": minor
-"@opencode-ai/schema": minor
-"@opencode-ai/protocol": minor
-"@opencode-ai/client": minor
---
-
-Remove the unused question request API and use session forms for question tool interactions.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Apply shared Session model-request preparation to transient generation.
@@ -1,9 +0,0 @@
---
-"@opencode-ai/core": minor
-"@opencode-ai/schema": minor
-"@opencode-ai/protocol": minor
-"@opencode-ai/client": minor
---
-
-Add an opt-in portable shell permission scanner. Opaque commands use normal shell authorization without inferring
-external directories, while the default tree-sitter path remains unchanged.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Simplify interrupt continuation: the steer-scoped resume decision now lives in SessionExecution as a post-cleanup inbox check, and the run coordinator drops its continuation state machine. Wakes arriving during cancellation cleanup now restart a normal full drain, and interrupting an idle session with continue now resumes pending steering input. Recovery-applied moves now end with the same full wake as inbox-admitted moves, retrying any stranded inbox work at the new location. Interrupting with continue now also resumes a next-in-line control item: between-turn manual compaction and moves run under any drain scope, while queued prompts remain parked.
@@ -1,6 +0,0 @@
---
-"@opencode-ai/client": patch
-"@opencode-ai/plugin": patch
---
-
-Add form reply and cancellation operations that reconcile terminal forms in the local TUI projection.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Title generation and compaction summaries now build their model requests through the shared session request boundary, gaining unsupported-media filtering and image bounds while explicitly opting out of session context hooks: plugins that shape the agent conversation do not observe title or compaction requests. Title requests gain the fork-aware session prompt cache key, and compaction summaries in forked sessions reuse the fork root's prompt cache key instead of the fork's own.
@@ -1,6 +0,0 @@
---
-"@opencode-ai/plugin": patch
-"@opencode-ai/core": patch
---
-
-Add transport-neutral Session model request hooks and provider-scoped hook registration so eligible OpenAI Responses requests can prefer WebSocket without bypassing HTTP-only middleware.
@@ -42,7 +42,7 @@ jobs:
      - name: Find affected packages
        id: packages
        env:
-          TURBO_SCM_BASE: ${{ github.event.pull_request.base.sha || github.event.before }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
          TURBO_SCM_HEAD: ${{ github.sha }}
        run: |
          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
@@ -119,7 +119,7 @@ jobs:
          GITHUB_ACTIONS=false bun turbo test --affected
        env:
          OPENCODE_EXPERIMENTAL_DISABLE_FILEWATCHER: ${{ runner.os == 'Windows' && 'true' || 'false' }}
-          TURBO_SCM_BASE: ${{ github.event.pull_request.base.sha || github.event.before }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
          TURBO_SCM_HEAD: ${{ github.sha }}

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

      - name: Verify compiled service lifecycle
@@ -237,13 +237,19 @@ jobs:
        working-directory: packages/app
        run: bunx playwright install chromium

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

+      - name: Verify service worker precaching and upgrades
+        if: env.E2E_ENABLED == 'true'
+        working-directory: packages/app
+        run: bunx playwright test --config e2e/service-worker/playwright.config.ts
+        timeout-minutes: 5
+
      - name: Upload Playwright artifacts
        if: always() && env.E2E_ENABLED == 'true'
        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
@@ -41,11 +41,5 @@
    // Warn when spreading non-plain objects (Headers, class instances, etc.)
    "typescript/no-misused-spread": "warn"
  },
-  "options": {
-    "typeAware": true
-  },
-  "options": {
-    "typeAware": true
-  },
  "ignorePatterns": ["**/node_modules", "**/dist", "**/.build", "**/.sst", "**/*.d.ts", "**/sdk.gen.ts"]
 }
@@ -2,7 +2,7 @@
 exact = true
 # Only install newly resolved package versions published at least 3 days ago.
 minimumReleaseAge = 259200
-minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@brendonovich/vite-plugin-opencode", "@opencode-ai/sdk", "@opentui/core", "@opentui/core-darwin-arm64", "@opentui/core-darwin-x64", "@opentui/core-linux-arm64", "@opentui/core-linux-arm64-musl", "@opentui/core-linux-x64", "@opentui/core-linux-x64-musl", "@opentui/core-win32-arm64", "@opentui/core-win32-x64", "@opentui/keymap", "@opentui/solid", "opentui-spinner", "gitlab-ai-provider", "opencode-gitlab-auth", "@ff-labs/fff-node", "@ff-labs/fff-bun", "@ff-labs/fff-bin-darwin-arm64", "@ff-labs/fff-bin-darwin-x64", "@ff-labs/fff-bin-linux-arm64-gnu", "@ff-labs/fff-bin-linux-arm64-musl", "@ff-labs/fff-bin-linux-x64-gnu", "@ff-labs/fff-bin-linux-x64-musl", "@ff-labs/fff-bin-win32-arm64", "@ff-labs/fff-bin-win32-x64", "@pierre/diffs", "@pierre/theming", "app-builder-lib", "dmg-builder", "electron-builder", "electron-publish", "blume"]
+minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@brendonovich/vite-plugin-opencode", "@opencode-ai/sdk", "@opencode-ai/pty", "@opencode-ai/pty-darwin-arm64", "@opencode-ai/pty-darwin-x64", "@opencode-ai/pty-linux-arm64-gnu", "@opencode-ai/pty-linux-arm64-musl", "@opencode-ai/pty-linux-x64-gnu", "@opencode-ai/pty-linux-x64-musl", "@opentui/core", "@opentui/core-darwin-arm64", "@opentui/core-darwin-x64", "@opentui/core-linux-arm64", "@opentui/core-linux-arm64-musl", "@opentui/core-linux-x64", "@opentui/core-linux-x64-musl", "@opentui/core-win32-arm64", "@opentui/core-win32-x64", "@opentui/keymap", "@opentui/solid", "opentui-spinner", "gitlab-ai-provider", "opencode-gitlab-auth", "@ff-labs/fff-node", "@ff-labs/fff-bun", "@ff-labs/fff-bin-darwin-arm64", "@ff-labs/fff-bin-darwin-x64", "@ff-labs/fff-bin-linux-arm64-gnu", "@ff-labs/fff-bin-linux-arm64-musl", "@ff-labs/fff-bin-linux-x64-gnu", "@ff-labs/fff-bin-linux-x64-musl", "@ff-labs/fff-bin-win32-arm64", "@ff-labs/fff-bin-win32-x64", "@pierre/diffs", "@pierre/theming", "app-builder-lib", "dmg-builder", "electron", "electron-builder", "electron-publish", "blume", "mermaid"]

 [test]
 root = "./do-not-run-tests-from-root"
@@ -63,7 +63,8 @@ new sst.cloudflare.StaticSite("WebApp", {
  domain: "app." + domain,
  path: "packages/app",
  build: {
-    command: "bun turbo build",
+    // Preserve Sentry credentials and run source-map uploads on every deployment.
+    command: "bun run build",
    output: "./dist",
  },
 })
@@ -4,7 +4,7 @@
  bun,
  nodejs,
  darwin,
-  electron_41,
+  callPackage,
  makeWrapper,
  writableTmpDirAsHomeHook,
  autoPatchelfHook,
@@ -13,7 +13,7 @@
  opencode,
 }:
 let
-  electron = electron_41;
+  electron = callPackage ./electron.nix { };
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "opencode-desktop";
@@ -48,7 +48,7 @@ stdenv.mkDerivation (finalAttrs: {
    desktopName = "OpenCode";
    exec = "opencode-desktop %U";
    icon = "ai.opencode.desktop";
-    # Electron 41 derives X11 WM_CLASS from app.name.
+    # Electron derives X11 WM_CLASS from app.name.
    startupWMClass = "OpenCode";
    categories = [ "Development" ];
  });
@@ -67,12 +67,10 @@ stdenv.mkDerivation (finalAttrs: {
    # https://github.com/electron/electron/issues/31121
    # mac builds use a .app bundle which doesnt have this issue
    + lib.optionalString stdenv.isLinux ''
-      BASE_PATH=packages/desktop
-      FILES=(src/main/windows.ts)
-      for file in "''${FILES[@]}"; do
-        substituteInPlace $BASE_PATH/$file \
-          --replace-fail "process.resourcesPath" "'$out/opt/opencode-desktop/resources'"
-      done
+      substituteInPlace \
+        packages/desktop/src/main/windows/appearance.ts \
+        packages/desktop/src/main/service/desktop-cli.ts \
+        --replace-fail "process.resourcesPath" "'$out/opt/opencode-desktop/resources'"
    '';

  preBuild = ''
@@ -0,0 +1,13 @@
+{ callPackage, path }:
+let
+  version = (builtins.fromJSON (builtins.readFile ../packages/desktop/package.json)).devDependencies.electron;
+in
+(callPackage (path + "/pkgs/development/tools/electron/binary/generic.nix") { }) version {
+  # Electron 42.10.1 SHASUMS256.txt; update with the desktop package version.
+  aarch64-linux = "20e68d6c4e47f3ebf59de7c6b1f8b8bec6a6ebda6a451132f9b465f3f13ce467";
+  x86_64-linux = "2452b27112d92387471fa2488aafac85d79ea3f2ee1216c0abd5150d6c12362b";
+  aarch64-darwin = "ac7194a3dfd81930ba35355c01620262c1254752859b42dcb8f4b9e4d174a871";
+  x86_64-darwin = "4489aba55477a0082266cb690db1c829503ba3338048599d8fd243953df37dab";
+  # fetchzip hashes the unpacked headers, not the release tarball.
+  headers = "sha256-4eUy3BZVvxTl7KUOsxio7769lL6ag/ecbeK+qLURWMI=";
+}
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-fOM/kGJJ1cipCHQIxioDZEB7NZykpSiqgwm7gIS6THI=",
-    "aarch64-linux": "sha256-XTY2C33HjsBMWO7VIiWc2MynjJrxbTLrOJ+6pM+afI0=",
-    "aarch64-darwin": "sha256-wX6+bC18djtPZ7A9ch+wryM7tDFfrAlT0xx0QTk6EJQ=",
-    "x86_64-darwin": "sha256-dcRRX4bYq5AmG4GcVmYq/M+06dlf4KJHn+clT2JY48g="
+    "x86_64-linux": "sha256-EtUp4pHl9TyPtRrLGvk/X7kd2LuIxNxCpUwF5aLtzN4=",
+    "aarch64-linux": "sha256-m0j/pMZCguclR3/T9JmzCfi11YzmIvBFyR2bVhIO37Y=",
+    "aarch64-darwin": "sha256-nqefk68ZTUfNU15q1WkXaGsFzPNwOjCtMHpp6WrpNqM=",
+    "x86_64-darwin": "sha256-syD7hX62E4yCDV/wux1QKw4q/zZr24f99Y2mmzMJo6o="
  }
 }
@@ -39,24 +39,24 @@
      "packages/stats/*"
    ],
    "catalog": {
-      "@effect/opentelemetry": "4.0.0-rc.111",
-      "@effect/platform-node": "4.0.0-rc.111",
-      "@effect/platform-node-shared": "4.0.0-rc.111",
-      "@effect/sql-sqlite-bun": "4.0.0-rc.111",
+      "@effect/opentelemetry": "4.0.0-rc.112",
+      "@effect/platform-node": "4.0.0-rc.112",
+      "@effect/platform-node-shared": "4.0.0-rc.112",
+      "@effect/sql-sqlite-bun": "4.0.0-rc.112",
      "@npmcli/arborist": "9.4.0",
      "@types/bun": "1.3.13",
      "@types/cross-spawn": "6.0.6",
      "@octokit/rest": "22.0.0",
      "@hono/standard-validator": "0.2.0",
      "@hono/zod-validator": "0.4.2",
-      "@opentui/core": "0.5.8",
-      "@opentui/keymap": "0.5.8",
-      "@opentui/solid": "0.5.8",
+      "@opentui/core": "0.5.9",
+      "@opentui/keymap": "0.5.9",
+      "@opentui/solid": "0.5.9",
      "@tanstack/solid-virtual": "3.13.37",
-      "@shikijs/stream": "4.2.0",
+      "@shikijs/stream": "4.4.3",
      "@standard-schema/spec": "1.1.0",
      "ulid": "3.0.1",
-      "@kobalte/core": "0.13.11",
+      "@kobalte/core": "0.13.13",
      "@corvu/drawer": "0.2.4",
      "@types/luxon": "3.7.1",
      "@types/node": "24.12.2",
@@ -67,13 +67,16 @@
      "@openauthjs/openauth": "0.0.0-20250322224806",
      "@pierre/diffs": "1.2.10",
      "opentui-spinner": "0.0.7",
-      "@solid-primitives/storage": "4.3.3",
+      "@solid-primitives/event-listener": "2.4.6",
+      "@solid-primitives/media": "2.3.6",
+      "@solid-primitives/resize-observer": "2.2.0",
+      "@solid-primitives/storage": "4.4.0",
      "@tailwindcss/vite": "4.1.11",
-      "diff": "8.0.2",
-      "dompurify": "3.3.1",
+      "diff": "8.0.4",
+      "dompurify": "3.4.14",
      "drizzle-kit": "1.0.0-rc.5-ab785fc",
      "drizzle-orm": "1.0.0-rc.5-169397b",
-      "effect": "4.0.0-rc.111",
+      "effect": "4.0.0-rc.112",
      "ai": "6.0.168",
      "cross-spawn": "7.0.6",
      "hono": "4.10.7",
@@ -81,9 +84,9 @@
      "fuzzysort": "3.1.0",
      "get-east-asian-width": "1.6.0",
      "luxon": "3.6.1",
-      "marked": "18.0.7",
+      "marked": "18.0.11",
      "marked-shiki": "1.2.1",
-      "remend": "1.3.0",
+      "remend": "1.3.1",
      "@playwright/test": "1.59.1",
      "semver": "7.7.4",
      "typescript": "5.8.2",
@@ -92,17 +95,17 @@
      "remeda": "2.26.0",
      "resolve.exports": "2.0.3",
      "sst": "4.13.1",
-      "shiki": "4.2.0",
+      "shiki": "4.4.3",
      "solid-list": "0.3.0",
      "string-width": "7.2.0",
      "tailwindcss": "4.1.11",
-      "vite": "7.1.4",
+      "vite": "7.3.6",
      "@solidjs/meta": "0.29.4",
-      "@solidjs/router": "0.15.4",
+      "@solidjs/router": "1.0.0",
      "@solidjs/start": "https://pkg.pr.new/@solidjs/start@dfb2020",
-      "@sentry/solid": "10.36.0",
-      "@sentry/vite-plugin": "4.6.0",
-      "solid-js": "1.9.10",
+      "@sentry/solid": "10.71.0",
+      "@sentry/vite-plugin": "5.4.0",
+      "solid-js": "1.9.15",
      "solid-sonner": "0.3.1",
      "vite-plugin-solid": "2.11.10",
      "@lydell/node-pty": "1.2.0-beta.12"
@@ -158,7 +161,8 @@
    "@effect/platform-node-shared": "catalog:",
    "@types/bun": "catalog:",
    "@types/node": "catalog:",
-    "effect": "catalog:"
+    "effect": "catalog:",
+    "solid-js": "catalog:"
  },
  "patchedDependencies": {
    "@ai-sdk/openai-compatible@2.0.41": "patches/@ai-sdk%2Fopenai-compatible@2.0.41.patch",
@@ -166,7 +170,7 @@
    "@npmcli/agent@4.0.2": "patches/@npmcli%2Fagent@4.0.2.patch",
    "@silvia-odwyer/photon-node@0.3.4": "patches/@silvia-odwyer%2Fphoton-node@0.3.4.patch",
    "@standard-community/standard-openapi@0.2.9": "patches/@standard-community%2Fstandard-openapi@0.2.9.patch",
-    "solid-js@1.9.10": "patches/solid-js@1.9.10.patch",
+    "solid-js@1.9.15": "patches/solid-js@1.9.15.patch",
    "@ai-sdk/mistral@3.0.51": "patches/@ai-sdk%2Fmistral@3.0.51.patch",
    "gcp-metadata@8.1.2": "patches/gcp-metadata@8.1.2.patch",
    "pacote@21.5.0": "patches/pacote@21.5.0.patch",
@@ -19,6 +19,14 @@ Per-type constructors live on the type, not as top-level re-exports. Use `Messag
 - Use `testEffect(...)` from `test/lib/effect.ts` for tests requiring Effect layers.
 - Keep provider tests fixture-first. Live provider calls must stay behind `RECORD=true` and required API-key checks.

+## Errors
+
+- `AIError` wraps a union of tagged reason errors. It stores only `reason`, derives `message` from the reason, and exposes the reason as its `cause`.
+- Each reason owns its readable `message`, category-specific fields, and optional `body`, `http`, and underlying exception in `cause`.
+- `reason.body` is the sole original-response or triggering-event payload field. Preserve original text before schema decoding removes fields; do not replace the complete event with only its nested error.
+- `reason.http` describes an observed HTTP response with required `url`, `status`, and response `headers`. Do not invent status codes or derive a separate request ID from headers.
+- Reclassification and transport recovery must preserve the reason's body, HTTP context, and underlying cause. Error `message` and `cause` are non-enumerable: copy them explicitly when constructing an enriched reason with its constructor or `AIErrorReason.make`.
+
 ## Architecture

 This package is an Effect Schema-first LLM core. The Schema classes in `src/schema/` are the canonical runtime data model. Convenience functions in `src/llm.ts` are thin constructors that return those same Schema class instances; they should improve callsites without creating a second model.
@@ -36,7 +36,12 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // Protocols whose wire format ignores inline cache markers (OpenAI's implicit
 // prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
 // whole policy pass for these — emitting hints would be harmless but pointless.
-const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "google-vertex-messages", "bedrock-converse", "openrouter"])
+const RESPECTS_INLINE_HINTS = new Set([
+  "anthropic-messages",
+  "google-vertex-messages",
+  "bedrock-converse",
+  "openrouter",
+])

 const makeHint = (ttlSeconds: number | undefined): CacheHint =>
  ttlSeconds !== undefined ? new CacheHint({ type: "ephemeral", ttlSeconds }) : new CacheHint({ type: "ephemeral" })
@@ -1,7 +1,7 @@
 import { Effect, Schema } from "effect"
 import {
  HttpOptions,
-  InvalidRequestReason,
+  InvalidRequestError,
  AIError,
  ModelID,
  ProviderID,
@@ -158,9 +158,10 @@ export function generate(input: ImageRequest | ImageRequestInput) {
    try: () => (input instanceof ImageRequest ? input : request(input)),
    catch: (error) =>
      new AIError({
-        module: "Image",
-        method: "generate",
-        reason: new InvalidRequestReason({ message: error instanceof Error ? error.message : String(error) }),
+        reason: new InvalidRequestError({
+          message: error instanceof Error ? error.message : String(error),
+          cause: error,
+        }),
      }),
  }).pipe(Effect.flatMap((request) => ImageClient.generate(request as unknown as ImageRequestFor<ImageOptions>)))
 }
@@ -3,7 +3,7 @@ import { LLMClient, Service } from "./route/client.js"
 import {
  GenerationOptions,
  HttpOptions,
-  InvalidProviderOutputReason,
+  InvalidProviderOutputError,
  AIError,
  LLMEvent,
  LLMRequest,
@@ -116,9 +116,7 @@ const runGenerateObject = Effect.fn("LLM.generateObject")(function* (
  )
  if (!call || !LLMEvent.is.toolCall(call))
    return yield* new AIError({
-      module: "LLM",
-      method: "generateObject",
-      reason: new InvalidProviderOutputReason({
+      reason: new InvalidProviderOutputError({
        message: `generateObject: model did not call the forced \`${GENERATE_OBJECT_TOOL_NAME}\` tool`,
      }),
    })
@@ -126,10 +124,9 @@ const runGenerateObject = Effect.fn("LLM.generateObject")(function* (
    Effect.mapError(
      (error) =>
        new AIError({
-          module: "LLM",
-          method: "generateObject",
-          reason: new InvalidProviderOutputReason({
+          reason: new InvalidProviderOutputError({
            message: `generateObject: tool input failed schema decode: ${error.message}`,
+            cause: error,
          }),
        }),
    ),
@@ -69,14 +69,22 @@ export interface OptionsInput {
  // SDK Metadata:2649 {user_id?: string | null}
  readonly metadata?: { readonly user_id?: string | null }
  // SDK MessageCreateParamsContainer:2596 ContainerParams|string
-  readonly container?: string | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
+  readonly container?:
+    | string
+    | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
  readonly inference_geo?: string | null
  readonly inferenceGeo?: string | null
  readonly cache_control?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
  readonly cacheControl?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
  // SDK OutputConfig:2684 {effort, format: JSONOutputFormat}
-  readonly output_config?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
-  readonly outputConfig?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
+  readonly output_config?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
+  readonly outputConfig?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
 }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

-  const result = serverToolResultEvent(block)
+  const result = serverToolResultEvent(block, state.providerMetadataKey)
  if (!result) return [state, NO_EVENTS]
  const events: LLMEvent[] = []
  return [{ ...state, lifecycle: Lifecycle.stepStart(state.lifecycle, events) }, [...events, result]]
@@ -1306,7 +1330,7 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
        Lifecycle.textEnd(state.lifecycle, events, `text-${event.index}`),
        events,
        `reasoning-${event.index}`,
-        signature === undefined ? undefined : anthropicMetadata({ signature }),
+        signature === undefined ? undefined : providerMetadata(state.providerMetadataKey, { signature }),
      )
  events.push(...resultEvents)
  const reasoningSignatures = { ...state.reasoningSignatures }
@@ -1318,7 +1342,7 @@ const onMessageDelta = (
  state: ParserState,
  event: AnthropicEvent & { readonly delta?: AnthropicStreamDelta },
 ): StepResult => {
-  const usage = mergeUsage(state.usage, mapUsage(event.usage))
+  const usage = mergeUsage(state.usage, mapUsage(event.usage, state.providerMetadataKey), state.providerMetadataKey)
  return [
    {
      ...state,
@@ -1331,7 +1355,7 @@ const onMessageDelta = (
        providerMetadata:
          event.delta?.stop_sequence === null || event.delta?.stop_sequence === undefined
            ? undefined
-            : anthropicMetadata({ stopSequence: event.delta.stop_sequence }),
+            : providerMetadata(state.providerMetadataKey, { stopSequence: event.delta.stop_sequence }),
      },
    },
    NO_EVENTS,
@@ -1363,14 +1387,15 @@ const providerErrorMessage = (event: AnthropicEvent): string => {
  return message || type || "Anthropic Messages stream error"
 }

-const onError = (event: AnthropicEvent) =>
-  Effect.fail(
+const onError = (event: AnthropicEvent) => {
+  const message = providerErrorMessage(event)
+  const body = ProviderShared.encodeJson(event)
+  return Effect.fail(
    new AIError({
-      module: ADAPTER,
-      method: "stream",
-      reason: classifyProviderFailure({ message: providerErrorMessage(event), code: event.error?.type }),
+      reason: classifyProviderFailure({ message, rawBody: body }),
    }),
  )
+}

 const isKnownStreamBlockType = (type: string) =>
  type === "text" ||
@@ -1418,9 +1443,7 @@ const step = (state: ParserState, event: AnthropicEvent) => {
      if (event.index === undefined)
        return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic ${block.type} missing index`))
      if (!block.id)
-        return Effect.fail(
-          ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`),
-        )
+        return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`))
    }
    return Effect.succeed(onContentBlockStart(state, { ...event, content_block: block }))
  }
@@ -1459,7 +1482,8 @@ export const protocol = Protocol.make({
  },
  stream: {
    event: Protocol.jsonEvent(AnthropicEvent),
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
      reasoningSignatures: {},
      lifecycle: Lifecycle.initial(),
@@ -1473,10 +1497,9 @@ export const route = Route.make({
  provider: "anthropic",
  providerMetadataKey: "anthropic",
  protocol,
-  endpoint: Endpoint.path(
-    (input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH),
-    { baseURL: DEFAULT_BASE_URL },
-  ),
+  endpoint: Endpoint.path((input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH), {
+    baseURL: DEFAULT_BASE_URL,
+  }),
  auth: Auth.none,
  framing,
  headers: () => ({ "anthropic-version": "2023-06-01" }),
@@ -258,19 +258,21 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
    tool: (name) => ({ tool: { name } }) as const,
  })

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

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

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

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

  for (const message of request.messages) {
    if (message.role === "system") {
@@ -365,12 +368,18 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
          continue
        }
        if (part.type === "reasoning") {
-          const signature = reasoningSignature(part)
-          const redactedData = reasoningRedactedData(part)
+          const signature = reasoningSignature(part, providerMetadataKey)
+          const redactedData = reasoningRedactedData(part, providerMetadataKey)
          if (signature === undefined && redactedData !== undefined) {
            content.push({ reasoningContent: { redactedContent: redactedData } })
            continue
          }
+          if (signature === undefined || signature.trim().length === 0) {
+            // Interrupted streams and model switches can leave unsigned reasoning.
+            // Preserve readable history as text rather than replay invalid reasoningContent.
+            if (part.text.trim().length > 0) content.push(...textWithCache(breakpoints, part.text, part.cache))
+            continue
+          }
          content.push({ reasoningContent: { reasoningText: { text: part.text, signature } } })
          continue
        }
@@ -379,7 +388,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
          continue
        }
      }
-      messages.push({ role: "assistant", content })
+      if (content.length > 0) messages.push({ role: "assistant", content })
      continue
    }

@@ -466,7 +475,7 @@ const mapFinishReason = (reason: string): FinishReason => {

 // AWS reports inputTokens separately from cache reads and writes.
 // Bedrock does not break reasoning out of outputTokens for current models.
-const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
+const mapUsage = (usage: BedrockUsageSchema | undefined, providerMetadataKey: string): Usage | undefined => {
  if (!usage) return undefined
  const inputTokens = ProviderShared.sumTokens(
    usage.inputTokens,
@@ -480,12 +489,14 @@ const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
    cacheReadInputTokens: usage.cacheReadInputTokens,
    cacheWriteInputTokens: usage.cacheWriteInputTokens,
    totalTokens: ProviderShared.totalTokens(inputTokens, usage.outputTokens, usage.totalTokens),
-    providerMetadata: { bedrock: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
  })
 }

 interface ParserState {
+  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
+  readonly finishedTools: ReadonlySet<number>
  // Bedrock splits the finish into `messageStop` (carries `stopReason`) and
  // `metadata` (carries usage). Hold the terminal event in state so `onHalt`
  // can emit exactly one finish after both chunks have had a chance to arrive.
@@ -541,20 +552,14 @@ const step = (state: ParserState, event: BedrockEvent) =>
      const reasoning = event.contentBlockDelta.delta.reasoningContent
      const events: LLMEvent[] = []
      const redactedData = reasoning.redactedContent ?? reasoning.data
-      const providerMetadata = reasoning.signature
-        ? bedrockMetadata({ signature: reasoning.signature })
+      const metadata = reasoning.signature
+        ? providerMetadata(state.providerMetadataKey, { signature: reasoning.signature })
        : redactedData !== undefined
-          ? bedrockMetadata({ redactedData })
+          ? providerMetadata(state.providerMetadataKey, { redactedData })
          : undefined
      const lifecycle =
-        reasoning.text !== undefined || providerMetadata !== undefined
-          ? Lifecycle.reasoningDelta(
-              state.lifecycle,
-              events,
-              `reasoning-${index}`,
-              reasoning.text ?? "",
-              providerMetadata,
-            )
+        reasoning.text !== undefined || metadata !== undefined
+          ? Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text ?? "", metadata)
          : state.lifecycle
      return [
        {
@@ -570,6 +575,7 @@ const step = (state: ParserState, event: BedrockEvent) =>

    if (event.contentBlockDelta?.delta?.toolUse) {
      const index = event.contentBlockDelta.contentBlockIndex
+      if (state.finishedTools.has(index)) return [state, []] as const
      const result = ToolStream.appendExisting(
        ADAPTER,
        state.tools,
@@ -596,7 +602,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
            events,
            `reasoning-${index}`,
            state.reasoningSignatures[index]
-              ? bedrockMetadata({ signature: state.reasoningSignatures[index] })
+              ? providerMetadata(state.providerMetadataKey, { signature: state.reasoningSignatures[index] })
              : undefined,
          )
      events.push(...resultEvents)
@@ -608,6 +614,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
            state.hasToolCalls,
          lifecycle,
          tools: result.tools,
+          finishedTools: resultEvents.length > 0 ? new Set([...state.finishedTools, index]) : state.finishedTools,
          reasoningSignatures: Object.fromEntries(
            Object.entries(state.reasoningSignatures).filter(([key]) => key !== String(index)),
          ),
@@ -633,7 +640,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
    }

    if (event.metadata) {
-      const usage = mapUsage(event.metadata.usage) ?? state.pendingFinish?.usage
+      const usage = mapUsage(event.metadata.usage, state.providerMetadataKey) ?? state.pendingFinish?.usage
      return [
        {
          ...state,
@@ -647,13 +654,13 @@ const step = (state: ParserState, event: BedrockEvent) =>
    }

    if (event.exception) {
+      const message =
+        event.exception.details.message ?? event.exception.details.originalMessage ?? "Bedrock Converse stream error"
+      const body = ProviderShared.encodeJson(event)
      return yield* new AIError({
-        module: ADAPTER,
-        method: "stream",
        reason: classifyProviderFailure({
-          message:
-            event.exception.details.message ?? event.exception.details.originalMessage ?? "Bedrock Converse stream error",
-          code: event.exception.type,
+          message,
+          rawBody: body,
        }),
      })
    }
@@ -696,8 +703,10 @@ export const protocol = Protocol.make({
  },
  stream: {
    event: BedrockEvent,
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
+      finishedTools: new Set<number>(),
      pendingFinish: undefined,
      hasToolCalls: false,
      lifecycle: Lifecycle.initial(),
@@ -1,6 +1,7 @@
 import { EventStreamCodec } from "@smithy/eventstream-codec"
 import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
-import { Effect, Stream } from "effect"
+import { Effect, Encoding, Stream } from "effect"
+import { AIError, AIErrorReason } from "../schema/index.js"
 import { Framing } from "../route/framing.js"
 import { ProviderShared } from "./shared.js"

@@ -49,10 +50,14 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
            `Failed to decode Bedrock Converse event-stream frame: ${
              error instanceof Error ? error.message : String(error)
            }`,
+            Encoding.encodeBase64(view.subarray(0, totalLength)),
+            error,
          ),
      })
      cursor = { buffer: cursor.buffer, offset: cursor.offset + totalLength }

+      const payload = utf8.decode(decoded.body)
+      const body = ProviderShared.encodeJson({ headers: decoded.headers, body: payload })
      const messageType = decoded.headers[":message-type"]?.value
      if (messageType === "error") {
        const code = decoded.headers[":error-code"]?.value
@@ -61,6 +66,7 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
          route,
          [code, message].filter((value): value is string => typeof value === "string").join(": ") ||
            "Bedrock Converse event-stream error",
+          body,
        )
      }
      const eventType =
@@ -70,7 +76,6 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
            ? decoded.headers[":exception-type"]?.value
            : undefined
      if (typeof eventType !== "string") continue
-      const payload = utf8.decode(decoded.body)
      if (!payload) continue
      // The AWS event stream pads short payloads with a `p` field. Drop it
      // before handing the object to the chunk schema. JSON decode goes
@@ -80,9 +85,21 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
        route,
        payload,
        "Failed to parse Bedrock Converse event-stream payload",
+      ).pipe(
+        Effect.mapError(
+          (error) =>
+            new AIError({
+              reason: AIErrorReason.make({ ...error.reason, message: error.message, cause: error.reason.cause, body }),
+            }),
+        ),
      )) as Record<string, unknown>
      delete parsed.p
-      out.push(messageType === "exception" ? { exception: { type: eventType, details: parsed } } : { [eventType]: parsed })
+      out.push({
+        ...(messageType === "exception"
+          ? { exception: { type: eventType, details: parsed } }
+          : { [eventType]: parsed }),
+        rawBody: body,
+      })
    }
    return [cursor, out] as const
  })
@@ -95,6 +112,7 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
 */
 export const framing = (route: string): Framing.Definition<object> => ({
  id: "aws-event-stream",
+  body: (frame) => ("rawBody" in frame && typeof frame.rawBody === "string" ? frame.rawBody : undefined),
  frame: (bytes) => bytes.pipe(Stream.mapAccumEffect(() => initialFrameBuffer, consumeFrames(route))),
 })

@@ -6,6 +6,7 @@ import { Endpoint } from "../route/endpoint.js"
 import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
 import {
+  AIError,
  LLMEvent,
  Usage,
  type FinishReason,
@@ -17,6 +18,7 @@ import {
  type ToolCallPart,
  type ToolDefinition,
 } from "../schema/index.js"
+import { classifyProviderFailure } from "../provider-error.js"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
 import { GeminiToolSchema } from "./utils/gemini-tool-schema.js"
 import { Lifecycle } from "./utils/lifecycle.js"
@@ -221,6 +223,7 @@ const GeminiPromptFeedback = Schema.StructWithRest(
 type GeminiPromptFeedback = Schema.Schema.Type<typeof GeminiPromptFeedback>

 const GeminiEvent = Schema.Struct({
+  error: Schema.optional(Schema.Unknown),
  candidates: optionalNull(Schema.Array(GeminiCandidate)),
  promptFeedback: optionalNull(GeminiPromptFeedback),
  usageMetadata: optionalNull(GeminiUsage),
@@ -229,6 +232,7 @@ type GeminiEvent = Schema.Schema.Type<typeof GeminiEvent>

 interface ParserState {
  readonly route: string
+  readonly providerMetadataKey: string
  readonly finishReason?: string
  readonly hasToolCalls: boolean
  readonly promptFeedback?: GeminiPromptFeedback
@@ -285,22 +289,23 @@ const lowerUserPart = Effect.fn("Gemini.lowerUserPart")(function* (part: TextPar
  return { inlineData: { mimeType: media.mime, data: media.base64 } }
 })

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

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

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

 const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMRequest) {
  const contents: GeminiContent[] = []
+  const metadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
  const omitCallIds = omitsFunctionCallIds(request.model.id)
  const legacyToolMedia = routesLegacyToolMedia(request.model.id)
  let pendingMedia: GeminiInlineDataPart[] | undefined
@@ -342,15 +347,19 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
        if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call"]))
          return yield* ProviderShared.unsupportedContent("Gemini", "assistant", ["text", "reasoning", "tool-call"])
        if (part.type === "text") {
-          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey) })
          continue
        }
        if (part.type === "reasoning") {
-          parts.push({ text: part.text, thought: true, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({
+            text: part.text,
+            thought: true,
+            thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
+          })
          continue
        }
        if (part.type === "tool-call") {
-          const lowered = lowerToolCall(part, omitCallIds)
+          const lowered = lowerToolCall(part, omitCallIds, metadataKey)
          const signature = lowered.thoughtSignature
          parts.push({
            ...lowered,
@@ -498,7 +507,7 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
 // `cachedContentTokenCount` subset. `candidatesTokenCount` is *exclusive*
 // of `thoughtsTokenCount` — visible-only, not a total — so we sum the two
 // to produce the inclusive `outputTokens` the rest of the contract expects.
-const mapUsage = (usage: GeminiUsage | undefined) => {
+const mapUsage = (usage: GeminiUsage | undefined, metadataKey: string) => {
  if (!usage) return undefined
  // Explicit provider nulls decode as `null`; normalize to `undefined` so the
  // token arithmetic below treats them like absent counts.
@@ -519,7 +528,7 @@ const mapUsage = (usage: GeminiUsage | undefined) => {
    cacheReadInputTokens: cached,
    reasoningTokens: thoughts,
    totalTokens: ProviderShared.totalTokens(promptTokens, outputTokens, usage.totalTokenCount ?? undefined),
-    providerMetadata: { google: usage },
+    providerMetadata: providerMetadata(metadataKey, usage),
  })
 }

@@ -567,10 +576,15 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
      lifecycle,
      events,
      "reasoning-0",
-      googleMetadata({ thoughtSignature: state.reasoningSignature }),
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.reasoningSignature }),
    )
  if (state.textSignature !== undefined)
-    lifecycle = Lifecycle.textEnd(lifecycle, events, "text-0", googleMetadata({ thoughtSignature: state.textSignature }))
+    lifecycle = Lifecycle.textEnd(
+      lifecycle,
+      events,
+      "text-0",
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.textSignature }),
+    )
  Lifecycle.finish(lifecycle, events, {
    reason: {
      normalized:
@@ -579,16 +593,32 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
    },
    usage: state.usage,
    providerMetadata:
-      state.promptFeedback === undefined ? undefined : googleMetadata({ promptFeedback: state.promptFeedback }),
+      state.promptFeedback === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { promptFeedback: state.promptFeedback }),
  })
  return events
 }

 const step = (state: ParserState, event: GeminiEvent) => {
+  if (ProviderShared.isRecord(event.error) && typeof event.error.message === "string") {
+    const body = ProviderShared.encodeJson(event)
+    return Effect.fail(
+      new AIError({
+        reason: classifyProviderFailure({
+          message: event.error.message,
+          status: typeof event.error.code === "number" ? event.error.code : undefined,
+          rawBody: body,
+        }),
+      }),
+    )
+  }
  const nextState = {
    ...state,
    promptFeedback: event.promptFeedback ?? state.promptFeedback,
-    usage: event.usageMetadata ? (mapUsage(event.usageMetadata) ?? state.usage) : state.usage,
+    usage: event.usageMetadata
+      ? (mapUsage(event.usageMetadata, state.providerMetadataKey) ?? state.usage)
+      : state.usage,
  }
  const candidate = event.candidates?.[0]
  if (!candidate?.content)
@@ -632,7 +662,7 @@ const step = (state: ParserState, event: GeminiEvent) => {
          events,
          "reasoning-0",
          part.text,
-          signature ? googleMetadata({ thoughtSignature: signature }) : undefined,
+          signature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: signature }) : undefined,
        )
        continue
      }
@@ -640,14 +670,16 @@ const step = (state: ParserState, event: GeminiEvent) => {
        lifecycle,
        events,
        "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
      )
      lifecycle = Lifecycle.textDelta(
        lifecycle,
        events,
        "text-0",
        part.text,
-        textSignature ? googleMetadata({ thoughtSignature: textSignature }) : undefined,
+        textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
      )
      textSignature = undefined
      continue
@@ -667,7 +699,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
        lifecycle,
        events,
        "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
      )
      lifecycle = Lifecycle.stepStart(lifecycle, events)
      events.push(
@@ -675,8 +709,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
          id,
          name: part.functionCall.name,
          input,
-          providerMetadata:
-            part.thoughtSignature ? googleMetadata({ thoughtSignature: part.thoughtSignature }) : undefined,
+          providerMetadata: part.thoughtSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: part.thoughtSignature })
+            : undefined,
        }),
      )
      hasToolCalls = true
@@ -714,6 +749,7 @@ export const protocol = Protocol.make({
    event: Protocol.jsonEvent(GeminiEvent),
    initial: (request) => ({
      route: `${request.model.provider}/${request.model.route.id}`,
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      hasToolCalls: false,
      lifecycle: Lifecycle.initial(),
    }),
@@ -9,15 +9,7 @@ import {
  type ImageRoute,
 } from "../image.js"
 import { Auth, type Definition as AuthDefinition } from "../route/auth.js"
-import {
-  InvalidProviderOutputReason,
-  AIError,
-  Usage,
-  mergeHttpOptions,
-  mergeJsonRecords,
-  type HttpOptions,
-  type ProviderMetadata,
-} from "../schema/index.js"
+import { AIError, Usage, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema/index.js"
 import { ProviderShared } from "./shared.js"
 import { ImageInputs } from "./utils/image-input.js"

@@ -124,13 +116,6 @@ const nativeOptions = (options: GoogleImageOptions | undefined) => {
  )
 }

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

@@ -1,6 +1,7 @@
 import { Effect, Schema, Stream } from "effect"
 import { Headers } from "effect/unstable/http"
 import { Framing } from "../route/framing.js"
+import type { HttpContext } from "../schema/index.js"
 import {
  HttpTransport,
  WebSocketTransport,
@@ -60,8 +61,8 @@ const driver = (options: Options, body: string): WebSocketChannelDriver => {
    observe: (_create, frame) =>
      Effect.gen(function* () {
        const event = yield* decodeEvent(frame).pipe(
-          Effect.mapError(() =>
-            ProviderShared.eventError(options.id, `Invalid ${options.name} WebSocket event`, frame),
+          Effect.mapError((cause) =>
+            ProviderShared.eventError(options.id, `Invalid ${options.name} WebSocket event`, frame, cause),
          ),
        )
        if (terminal)
@@ -73,13 +74,13 @@ const driver = (options: Options, body: string): WebSocketChannelDriver => {
        if (event.type === "error") {
          terminal = true
          yield* OpenResponses.decodeKnownErrorEvent(event).pipe(
-            Effect.mapError(() =>
-              ProviderShared.eventError(options.id, `${options.name} returned a malformed error event`, frame),
+            Effect.mapError((cause) =>
+              ProviderShared.eventError(options.id, `${options.name} returned a malformed error event`, frame, cause),
            ),
          )
          return {
            type: "provider-failure",
-            error: OpenResponses.providerFailure(options.id, event, `${options.name} stream error`),
+            error: OpenResponses.providerFailure(event, `${options.name} stream error`, frame),
          }
        }
        if (event.type === "response.failed") {
@@ -92,7 +93,7 @@ const driver = (options: Options, body: string): WebSocketChannelDriver => {
            )
          return {
            type: "provider-failure",
-            error: OpenResponses.providerFailure(options.id, event, `${options.name} response failed`),
+            error: OpenResponses.providerFailure(event, `${options.name} response failed`, frame),
          }
        }
        if (event.type === "response.created") {
@@ -175,23 +176,37 @@ export const transport = <Body>(options: Options): Transport<Body, Prepared, str
          channel,
        }
      }),
-    execute: (prepared, request, runtime, executeOptions) => {
-      if (!executeOptions?.webSocket || !prepared.channel) return http.execute(prepared.http, request, runtime)
-      const exchange: WebSocketChannelExchange = {
-        id: request.id ?? "request",
-        connect: {
-          url: prepared.channel.url,
-          headers: prepared.channel.headers,
-          rotateAfterMs: prepared.channel.rotateAfterMs,
-        },
-        fallback: () =>
-          Stream.unwrap(
-            http.execute(prepared.http, request, runtime).pipe(Effect.map((execution) => execution.frames)),
-          ),
-        driver: prepared.channel.driver,
-      }
-      return executeOptions.webSocket.execute(exchange)
-    },
+    execute: (prepared, request, runtime, executeOptions) =>
+      Effect.gen(function* () {
+        if (!executeOptions?.webSocket || !prepared.channel) return yield* http.execute(prepared.http, request, runtime)
+        let fallbackHttp: HttpContext | undefined
+        const exchange: WebSocketChannelExchange = {
+          id: request.id ?? "request",
+          connect: {
+            url: prepared.channel.url,
+            headers: prepared.channel.headers,
+            rotateAfterMs: prepared.channel.rotateAfterMs,
+          },
+          fallback: () =>
+            Stream.unwrap(
+              http.execute(prepared.http, request, runtime).pipe(
+                Effect.map((execution) => {
+                  fallbackHttp = execution.http
+                  return execution.frames
+                }),
+              ),
+            ),
+          driver: prepared.channel.driver,
+        }
+        const execution = yield* executeOptions.webSocket.execute(exchange)
+        return {
+          frames: execution.frames,
+          complete: execution.complete,
+          get http() {
+            return fallbackHttp ?? execution.http
+          },
+        }
+      }),
  }
 }

@@ -1,4 +1,4 @@
-import { AIError, TransportReason } from "../schema/index.js"
+import { AIError, TransportError } from "../schema/index.js"
 import type { ChannelCheckpoint, ChannelObservation, WebSocketChannelDriver } from "../route/transport/index.js"
 import { Effect, Option, Schema } from "effect"
 import * as ProviderShared from "./shared.js"
@@ -99,17 +99,17 @@ const incremental = (
 const code = (event: OpenResponses.Event) => event.code || event.error?.code || event.response?.error?.code || undefined

 const rejected = (
-  input: DriverInput,
  observation: Extract<ChannelObservation, { readonly type: "provider-failure" }>,
  recovery: "retry-full" | "rotate-and-retry-full",
 ): ChannelObservation => ({
  type: "rejected",
  recovery,
  error: new AIError({
-    module: input.id,
-    method: "stream",
-    reason: new TransportReason({
+    reason: new TransportError({
      message: observation.error.message,
+      body: observation.error.reason.body,
+      http: observation.error.reason.http,
+      cause: observation.error.reason.cause,
      transport: "websocket",
      operation: "read",
      phase: "receive",
@@ -137,15 +137,16 @@ export const driver = (input: DriverInput): WebSocketChannelDriver => {
    observe: (create, frame) =>
      Effect.gen(function* () {
        const event = yield* decodeEvent(frame).pipe(
-          Effect.mapError(() => ProviderShared.eventError(input.id, `Invalid ${input.name} WebSocket event`, frame)),
+          Effect.mapError((cause) =>
+            ProviderShared.eventError(input.id, `Invalid ${input.name} WebSocket event`, frame, cause),
+          ),
        )
        const observation = yield* input.base.observe(create, frame)
        if (event.type === "response.output_item.done" && event.item) output.push(event.item)
        if (observation.type === "provider-failure") {
          const rejection = code(event)
-          if (rejection === "previous_response_not_found") return rejected(input, observation, "retry-full")
-          if (rejection === "websocket_connection_limit_reached")
-            return rejected(input, observation, "rotate-and-retry-full")
+          if (rejection === "previous_response_not_found") return rejected(observation, "retry-full")
+          if (rejection === "websocket_connection_limit_reached") return rejected(observation, "rotate-and-retry-full")
        }
        if (observation.type !== "completed") return observation
        const responseID = event.response?.id
@@ -5,7 +5,7 @@ import { Protocol } from "../route/protocol.js"
 import {
  AIError,
  LLMEvent,
-  ProviderInternalReason,
+  ProviderInternalError,
  Usage,
  type FinishReason,
  type JsonSchema,
@@ -196,7 +196,7 @@ type LoweredInputItem =
 // multiple streamed summary parts into the same item before flushing.
 type OpenResponsesReasoningInput = {
  type: "reasoning"
-  id: string
+  id?: string
  summary: Array<{ type: "summary_text"; text: string }>
  encrypted_content?: string | null
 }
@@ -346,7 +346,8 @@ export const Event = Schema.StructWithRest(
    item_id: Schema.optional(Schema.String),
    output_index: Schema.optional(Schema.Number),
    summary_index: Schema.optional(Schema.Number),
-    item: Schema.optional(StreamItem),
+    // OutputItemAdded/Done permit a null item in the Open Responses OpenAPI schema.
+    item: optionalNull(StreamItem),
    response: Schema.optional(
      Schema.StructWithRest(
        Schema.Struct({
@@ -372,9 +373,6 @@ export const Event = Schema.StructWithRest(
 )
 export type Event = Schema.Schema.Type<typeof Event>

-// Which lowered input item a persisted item id is about to be attached to.
-export type ItemKind = "message" | "reasoning" | "function-call" | "hosted-tool"
-
 export interface Extension {
  readonly id: string
  readonly name: string
@@ -384,10 +382,6 @@ export interface Extension {
    readonly request: LLMRequest
  }) => MediaInput | undefined
  readonly lowerHostedToolItem?: (item: unknown) => ExtendedHostedToolItem | undefined
-  // Optional grammar check applied before a persisted item id is resent as
-  // part of replayed history. Returning false drops the id; every lowered
-  // item treats a dropped id the same as an absent one.
-  readonly acceptsItemID?: (kind: ItemKind, id: string) => boolean
 }

 const BASE: Extension = { id: ADAPTER, name: NAME }
@@ -397,17 +391,19 @@ export interface ParserState {
  readonly name: string
  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<string>
+  // Call ids stay independent of item ids, which may be omitted or reused.
+  readonly completedTools: ReadonlySet<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 message: { readonly id: string; readonly phase: MessagePhase | null | undefined } | undefined
  readonly reasoningItems: Readonly<Record<string, ReasoningStreamItem>>
 }

 type ReasoningSummaryStatus = "active" | "can-conclude" | "concluded"

 interface ReasoningStreamItem {
+  readonly open: boolean
  readonly encryptedContent: string | null | undefined
  // Keyed by the wire protocol's numeric `summary_index`. JS object keys coerce to
  // strings, but typing the map as `Record<number, ...>` documents intent
@@ -447,53 +443,37 @@ export const lowerToolChoice = (protocolName: string, toolChoice: NonNullable<LL
    tool: (toolName) => ({ type: "function" as const, name: toolName }),
  })

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

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

-const lowerReasoning = (
-  part: ReasoningPart,
-  providerMetadataKey: string,
-  extension: Extension,
-): OpenResponsesReasoningInput | undefined => {
+const lowerReasoning = (part: ReasoningPart, providerMetadataKey: string): OpenResponsesReasoningInput | undefined => {
  const metadata = part.providerMetadata?.[providerMetadataKey]
+  if (!ProviderShared.isRecord(metadata)) return undefined
  const id = itemID(part.providerMetadata, providerMetadataKey)
-  if (!ProviderShared.isRecord(metadata) || !acceptsItemID(extension, "reasoning", id)) return undefined
  const encryptedContent =
    typeof metadata.reasoningEncryptedContent === "string" || metadata.reasoningEncryptedContent === null
      ? metadata.reasoningEncryptedContent
      : undefined
  return {
    type: "reasoning",
-    id,
+    ...(id === undefined ? {} : { id }),
    summary: part.text.length > 0 ? [{ type: "summary_text", text: part.text }] : [],
    encrypted_content: encryptedContent,
  }
@@ -614,8 +594,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
          Array<{ id: string | undefined; phase: MessagePhase | null | undefined; parts: TextPart[] }>
        >((groups, part) => {
          const metadata = part.providerMetadata?.[providerMetadataKey]
-          const rawID = itemID(part.providerMetadata, providerMetadataKey)
-          const id = acceptsItemID(extension, "message", rawID) ? rawID : undefined
+          const id = itemID(part.providerMetadata, providerMetadataKey)
          const phase = ProviderShared.isRecord(metadata) ? messagePhase(metadata.phase) : undefined
          const group = groups.at(-1)
          if (group && group.id === id && group.phase === phase) group.parts.push(part)
@@ -640,23 +619,23 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        }
        if (part.type === "reasoning") {
          flushText()
-          const reasoning = lowerReasoning(part, providerMetadataKey, extension)
+          const reasoning = lowerReasoning(part, providerMetadataKey)
          if (!reasoning) continue
-          const existing = reasoningItems[reasoning.id]
+          const existing = reasoning.id === undefined ? undefined : reasoningItems[reasoning.id]
          if (existing) {
            existing.summary.push(...reasoning.summary)
            if (typeof reasoning.encrypted_content === "string")
              existing.encrypted_content = reasoning.encrypted_content
            continue
          }
-          reasoningItems[reasoning.id] = reasoning
+          if (reasoning.id !== undefined) reasoningItems[reasoning.id] = reasoning
          input.push(reasoning)
          continue
        }
        if (part.type === "tool-call") {
          flushText()
          if (part.providerExecuted === true) continue
-          input.push(lowerToolCall(part, providerMetadataKey, extension))
+          input.push(lowerToolCall(part, providerMetadataKey))
          continue
        }
        if (part.type === "tool-result" && part.providerExecuted === true) {
@@ -668,7 +647,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
              : Schema.is(HostedToolItem)(part.result.value)
                ? part.result.value
                : extension.lowerHostedToolItem?.(part.result.value)
-          if (acceptsItemID(extension, "hosted-tool", id) && hosted?.id === id) {
+          if (id !== undefined && hosted?.id === id) {
            if (!hostedToolItems.has(id)) {
              input.push(hosted)
              hostedToolItems.add(id)
@@ -836,7 +815,7 @@ export const providerMetadata = (state: ParserState, metadata: Record<string, un
 })

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

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

@@ -849,16 +828,16 @@ const TERMINAL_TYPES = new Set(["error", "response.completed", "response.incompl
 export const terminal = (event: Event) => TERMINAL_TYPES.has(event.type)

 const onOutputTextDelta = (state: ParserState, event: Event, id: string): StepResult => {
-  if (!event.delta || !state.messageItems.has(id)) return [state, NO_EVENTS]
+  if (!event.delta || state.message?.id !== id) return [state, NO_EVENTS]
  const events: LLMEvent[] = []
-  const phase = state.messagePhases[id]
+  const phase = state.message.phase
  const metadata = providerMetadata(state, { itemId: id, ...(phase === undefined ? {} : { phase }) })
  const lifecycle = Lifecycle.textStart(state.lifecycle, events, id, metadata)
  return [{ ...state, lifecycle: Lifecycle.textDelta(lifecycle, events, id, event.delta) }, events]
 }

 const onOutputTextDone = (state: ParserState, event: Event, id: string): StepResult => {
-  if (state.messageItems.has(id)) {
+  if (state.message?.id === id) {
    if (state.lifecycle.text.has(id) || event.text === undefined) return [state, NO_EVENTS]
    return onOutputTextDelta(state, { ...event, delta: event.text }, id)
  }
@@ -869,18 +848,62 @@ const onOutputTextDone = (state: ParserState, event: Event, id: string): StepRes
 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 startReasoningSummaryPart = (state: ParserState, itemID: string, index: number): StepResult => {
  const item = state.reasoningItems[itemID]
-  if (!event.delta || !item) return [state, NO_EVENTS]
-  const index = event.summary_index ?? 0
+  if (!item?.open || index === 0 || item.summaryParts[index] !== undefined) return [state, NO_EVENTS]
+
  const events: LLMEvent[] = []
+  const lifecycle = Object.entries(item.summaryParts)
+    .filter((entry) => entry[1] !== "concluded")
+    .reduce(
+      (lifecycle, entry) =>
+        Lifecycle.reasoningEnd(lifecycle, events, `${itemID}:${entry[0]}`, providerMetadata(state, { itemId: itemID })),
+      state.lifecycle,
+    )
  return [
    {
      ...state,
-      lifecycle: Lifecycle.reasoningDelta(state.lifecycle, events, `${itemID}:${index}`, event.delta),
+      lifecycle: Lifecycle.reasoningStart(
+        lifecycle,
+        events,
+        `${itemID}:${index}`,
+        providerMetadata(state, { itemId: itemID, reasoningEncryptedContent: item.encryptedContent ?? null }),
+      ),
      reasoningItems: {
        ...state.reasoningItems,
-        [itemID]: { ...item, deltaIndexes: new Set([...item.deltaIndexes, index]) },
+        [itemID]: {
+          ...item,
+          summaryParts: {
+            ...Object.fromEntries(
+              Object.entries(item.summaryParts).map((entry) =>
+                entry[1] === "concluded" ? entry : [entry[0], "concluded" as const],
+              ),
+            ),
+            [index]: "active",
+          },
+        },
+      },
+    },
+    events,
+  ]
+}
+
+export const onReasoningDelta = (state: ParserState, event: Event, itemID: string): StepResult => {
+  const item = state.reasoningItems[itemID]
+  if (!event.delta || !item?.open) return [state, NO_EVENTS]
+  const index = event.summary_index ?? 0
+  if (item.summaryParts[index] === "concluded") return [state, NO_EVENTS]
+  const [started, emitted] = startReasoningSummaryPart(state, itemID, index)
+  const current = started.reasoningItems[itemID]
+  if (!current) return [started, emitted]
+  const events: LLMEvent[] = [...emitted]
+  return [
+    {
+      ...started,
+      lifecycle: Lifecycle.reasoningDelta(started.lifecycle, events, `${itemID}:${index}`, event.delta),
+      reasoningItems: {
+        ...started.reasoningItems,
+        [itemID]: { ...current, deltaIndexes: new Set([...current.deltaIndexes, index]) },
      },
    },
    events,
@@ -892,7 +915,7 @@ export const onReasoningDelta = (state: ParserState, event: Event, itemID: strin
 // as a single delta unless that summary index already streamed one.
 export const onReasoningDone = (state: ParserState, event: Event, itemID: string): StepResult => {
  const item = state.reasoningItems[itemID]
-  if (!item || typeof event.text !== "string") return [state, NO_EVENTS]
+  if (!item?.open || typeof event.text !== "string") return [state, NO_EVENTS]
  const index = event.summary_index ?? 0
  if (item.deltaIndexes.has(index)) return [state, NO_EVENTS]
  return onReasoningDelta(state, { ...event, delta: event.text }, itemID)
@@ -901,32 +924,48 @@ export const onReasoningDone = (state: ParserState, event: Event, itemID: string
 const reasoningMetadata = (state: ParserState, item: StreamItem & { id: string }) =>
  providerMetadata(state, { itemId: item.id, reasoningEncryptedContent: item.encrypted_content ?? null })

-// Responses APIs stream reasoning items in a stable order:
+// Responses APIs normally stream reasoning items in this order:
 //   `output_item.added` (reasoning) →
 //     `reasoning_summary_part.added` (index=0) →
 //     `reasoning_summary_text.delta` →
 //     `reasoning_summary_part.done` (index=0) →
 //     (repeat for index>0) →
 //   `output_item.done` (reasoning).
-// The handlers below rely on this ordering: `onOutputItemAdded` seeds the
-// per-item entry, `onReasoningSummaryPartAdded` for `summary_index === 0`
-// short-circuits when the entry already exists, and higher-index handlers
-// fold against the same entry. Behaviour for out-of-order events is
-// best-effort, not guaranteed.
+// `onOutputItemAdded` seeds the per-item entry, while each later part start is
+// also an implicit boundary for the previous part. This keeps the common event
+// lifecycle ordered when a compatible provider omits or delays a part-done event.
 const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
  const item = event.item
-  if (item?.type === "message" && item.id) {
+  if (item?.type === "message" && item.id !== undefined) {
+    const itemID = item.id
    const phase = messagePhase(item.phase)
+    // A new message closes earlier messages, including ones that never streamed.
+    const events: LLMEvent[] = []
+    const lifecycle = [...state.lifecycle.text]
+      .filter((id) => id !== itemID)
+      .reduce((lifecycle, id) => {
+        const openPhase = state.message?.id === id ? state.message.phase : undefined
+        return Lifecycle.textEnd(
+          lifecycle,
+          events,
+          id,
+          providerMetadata(state, { itemId: id, ...(openPhase === undefined ? {} : { phase: openPhase }) }),
+        )
+      }, state.lifecycle)
    return [
      {
        ...state,
-        messageItems: new Set([...state.messageItems, item.id]),
-        messagePhases: phase === undefined ? state.messagePhases : { ...state.messagePhases, [item.id]: phase },
+        lifecycle,
+        message: {
+          id: itemID,
+          phase: phase === undefined && state.message?.id === itemID ? state.message.phase : phase,
+        },
      },
-      NO_EVENTS,
+      events,
    ]
  }
  if (item && isReasoningItem(item)) {
+    if (state.reasoningItems[item.id] !== undefined) return [state, NO_EVENTS]
    const events: LLMEvent[] = []
    return [
      {
@@ -935,6 +974,7 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
        reasoningItems: {
          ...state.reasoningItems,
          [item.id]: {
+            open: true,
            encryptedContent: item.encrypted_content,
            summaryParts: { 0: "active" },
            deltaIndexes: new Set(),
@@ -946,7 +986,9 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
  }
  if (item?.type !== "function_call" || !item.call_id) return [state, NO_EVENTS]
  const id = item.id ?? item.call_id
-  const metadata = item.id ? providerMetadata(state, { itemId: item.id }) : undefined
+  if (Object.values(state.tools).some((tool) => tool?.id === item.call_id) || state.completedTools.has(item.call_id))
+    return [state, NO_EVENTS]
+  const metadata = item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined
  const events: LLMEvent[] = []
  const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
  return [
@@ -965,56 +1007,15 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
 }

 const onReasoningSummaryPartAdded = (state: ParserState, event: Event): StepResult => {
-  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
-  const item = state.reasoningItems[event.item_id]
-  if (!item) return [state, NO_EVENTS]
-  if (event.summary_index === 0) return [state, NO_EVENTS]
-
-  const events: LLMEvent[] = []
-  const closed = Object.entries(item.summaryParts)
-    .filter((entry) => entry[1] === "can-conclude")
-    .reduce(
-      (lifecycle, entry) =>
-        Lifecycle.reasoningEnd(
-          lifecycle,
-          events,
-          `${event.item_id}:${entry[0]}`,
-          providerMetadata(state, { itemId: event.item_id }),
-        ),
-      state.lifecycle,
-    )
-  return [
-    {
-      ...state,
-      lifecycle: Lifecycle.reasoningStart(
-        closed,
-        events,
-        `${event.item_id}:${event.summary_index}`,
-        providerMetadata(state, { itemId: event.item_id, reasoningEncryptedContent: item.encryptedContent ?? null }),
-      ),
-      reasoningItems: {
-        ...state.reasoningItems,
-        [event.item_id]: {
-          ...item,
-          summaryParts: {
-            ...Object.fromEntries(
-              Object.entries(item.summaryParts).map((entry) =>
-                entry[1] === "can-conclude" ? [entry[0], "concluded" as const] : entry,
-              ),
-            ),
-            [event.summary_index]: "active",
-          },
-        },
-      },
-    },
-    events,
-  ]
+  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
+  return startReasoningSummaryPart(state, event.item_id, event.summary_index)
 }

 const onReasoningSummaryPartDone = (state: ParserState, event: Event): StepResult => {
-  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
+  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
  const item = state.reasoningItems[event.item_id]
-  if (!item) return [state, NO_EVENTS]
+  if (!item?.open) return [state, NO_EVENTS]
+  if (item.summaryParts[event.summary_index] !== "active") return [state, NO_EVENTS]
  return [
    {
      ...state,
@@ -1037,7 +1038,7 @@ const onFunctionCallArgumentsDelta = Effect.fn("OpenResponses.onFunctionCallArgu
  state: ParserState,
  event: Event,
 ) {
-  if (!event.item_id) return [state, NO_EVENTS] satisfies StepResult
+  if (event.item_id === undefined) return [state, NO_EVENTS] satisfies StepResult
  const tool = state.tools[event.item_id]
  if (!tool) return [state, NO_EVENTS] satisfies StepResult
  const final = event.type === "response.function_call_arguments.done" ? event.arguments : undefined
@@ -1068,13 +1069,10 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
  const item = event.item
  if (!item) return [state, NO_EVENTS] satisfies StepResult

-  if (item.type === "message" && item.id) {
+  if (item.type === "message" && item.id !== undefined) {
    const itemPhase = messagePhase(item.phase)
-    const phase = itemPhase === undefined ? state.messagePhases[item.id] : itemPhase
+    const phase = itemPhase === undefined && state.message?.id === item.id ? state.message.phase : itemPhase
    const events: LLMEvent[] = []
-    const messageItems = new Set(state.messageItems)
-    messageItems.delete(item.id)
-    const { [item.id]: _phase, ...messagePhases } = state.messagePhases
    return [
      {
        ...state,
@@ -1084,8 +1082,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
          item.id,
          providerMetadata(state, { itemId: item.id, ...(phase === undefined ? {} : { phase }) }),
        ),
-        messageItems,
-        messagePhases,
+        message: state.message?.id === item.id ? undefined : state.message,
      },
      events,
    ] satisfies StepResult
@@ -1093,20 +1090,36 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (

  if (item.type === "function_call") {
    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, id, {
-          id: item.call_id,
-          name: item.name,
-          providerMetadata: item.id ? providerMetadata(state, { itemId: item.id }) : undefined,
-        })
+    const callID = item.call_id
+    if (state.completedTools.has(callID)) return [state, NO_EVENTS] satisfies StepResult
+    const metadata = item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined
+    const fallback = item.id ?? callID
+    // Match the pending tool by call id so item events that disagree on
+    // whether `item.id` is present still resolve the same call.
+    const registered =
+      state.tools[fallback] !== undefined
+        ? fallback
+        : Object.keys(state.tools).find((key) => state.tools[key]?.id === callID)
+    const id = registered ?? fallback
+    const tools =
+      registered !== undefined
+        ? state.tools
+        : ToolStream.start(state.tools, id, {
+            id: callID,
+            name: item.name,
+            providerMetadata: metadata,
+          })
    const result =
      item.arguments === undefined
        ? yield* ToolStream.finish(state.id, tools, id)
        : yield* ToolStream.finishWithInput(state.id, tools, id, item.arguments)
    const events: LLMEvent[] = []
-    const resultEvents = result.events ?? []
+    const finished = result.events ?? []
+    // A done-only call never streamed a start event, so open its lifecycle here.
+    const resultEvents =
+      registered !== undefined || finished.length === 0
+        ? finished
+        : [LLMEvent.toolInputStart({ id: callID, name: item.name, providerMetadata: metadata }), ...finished]
    const lifecycle = resultEvents.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
    events.push(...resultEvents)
    return [
@@ -1117,6 +1130,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
          resultEvents.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
          state.hasFunctionCall,
        tools: result.tools,
+        completedTools: new Set([...state.completedTools, callID]),
      },
      events,
    ] satisfies StepResult
@@ -1127,20 +1141,49 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
    const metadata = reasoningMetadata(state, item)
    const reasoningItem = state.reasoningItems[item.id]
    if (reasoningItem) {
+      if (!reasoningItem.open) return [state, NO_EVENTS] satisfies StepResult
      const lifecycle = Object.entries(reasoningItem.summaryParts)
        .filter((entry) => entry[1] === "active" || entry[1] === "can-conclude")
        .reduce(
          (lifecycle, entry) => Lifecycle.reasoningEnd(lifecycle, events, `${item.id}:${entry[0]}`, metadata),
          state.lifecycle,
        )
-      const { [item.id]: _removed, ...reasoningItems } = state.reasoningItems
-      return [{ ...state, lifecycle, reasoningItems }, events] satisfies StepResult
+      return [
+        {
+          ...state,
+          lifecycle,
+          reasoningItems: {
+            ...state.reasoningItems,
+            [item.id]: {
+              ...reasoningItem,
+              open: false,
+              encryptedContent: item.encrypted_content ?? reasoningItem.encryptedContent,
+            },
+          },
+        },
+        events,
+      ] satisfies StepResult
    }
    if (!state.lifecycle.reasoning.has(item.id)) {
      const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
      events.push(LLMEvent.reasoningStart({ id: item.id, providerMetadata: metadata }))
      events.push(LLMEvent.reasoningEnd({ id: item.id, providerMetadata: metadata }))
-      return [{ ...state, lifecycle }, events] satisfies StepResult
+      return [
+        {
+          ...state,
+          lifecycle,
+          reasoningItems: {
+            ...state.reasoningItems,
+            [item.id]: {
+              open: false,
+              encryptedContent: item.encrypted_content,
+              summaryParts: { 0: "concluded" },
+              deltaIndexes: new Set(),
+            },
+          },
+        },
+        events,
+      ] satisfies StepResult
    }
    return [
      { ...state, lifecycle: Lifecycle.reasoningEnd(state.lifecycle, events, item.id, metadata) },
@@ -1160,9 +1203,9 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
          ([current, events], item) => {
            const id = item.id ?? (item.type === "function_call" ? item.call_id : undefined)
            if (
-              !id ||
+              id === undefined ||
              ((item.type !== "function_call" || !current.tools[id]) &&
-                (item.type !== "reasoning" || !current.reasoningItems[id]))
+                (item.type !== "reasoning" || !current.reasoningItems[id]?.open))
            )
              return Effect.succeed([current, events] satisfies StepResult)
            return onOutputItemDone(current, { type: "response.output_item.done", item }).pipe(
@@ -1211,11 +1254,8 @@ const providerErrorMessage = (event: Event, nested: OpenResponsesErrorPayload |
  return message || code
 }

-export const providerFailure = (id: string, event: Event, fallback: string) => {
+export const providerFailure = (event: Event, fallback: string, body = ProviderShared.encodeJson(event)) => {
  const nested = event.error ?? event.response?.error ?? undefined
-  const code = event.code || nested?.code || undefined
-  // Keep the full raw payload on the error even when the message is a summary.
-  const body = JSON.stringify(nested ?? event) ?? ""
  const summary = providerErrorMessage(event, nested)
  const message = summary ?? (body === "{}" ? fallback : body)
  const status =
@@ -1230,25 +1270,19 @@ export const providerFailure = (id: string, event: Event, fallback: string) => {
    event.response === undefined &&
    summary === undefined &&
    status === undefined
-      ? new ProviderInternalReason({ message })
-      : classifyProviderFailure({ message, code, status, rawBody: body })
-  return new AIError({
-    module: id,
-    method: "stream",
-    body,
-    reason,
-  })
+      ? new ProviderInternalError({ message, body })
+      : classifyProviderFailure({ message, status, rawBody: body })
+  return new AIError({ reason })
 }

-const providerError = (state: ParserState, event: Event, fallback: string) => providerFailure(state.id, event, fallback)
-
 export const step = (state: ParserState, input: Event) => {
+  // The OpenAPI requires string IDs but imposes no minLength; empty is not missing.
  const event =
-    input.item_id && outputItemID(state, input) !== input.item_id
+    input.item_id !== undefined && outputItemID(state, input) !== input.item_id
      ? { ...input, item_id: outputItemID(state, input) }
      : input
  if (event.type === "response.output_text.delta" || event.type === "response.output_text.done") {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(
      event.type === "response.output_text.delta"
        ? onOutputTextDelta(state, event, event.item_id)
@@ -1257,7 +1291,7 @@ export const step = (state: ParserState, input: Event) => {
  }
  if (event.type === "response.refusal.delta" || event.type === "response.refusal.done") {
    const value = event.type === "response.refusal.delta" ? event.delta : event.refusal
-    if (!event.item_id || typeof value !== "string")
+    if (event.item_id === undefined || typeof value !== "string")
      return ProviderShared.eventError(state.id, `${event.type} is malformed`)
    return Effect.succeed(
      event.type === "response.refusal.delta"
@@ -1266,7 +1300,7 @@ export const step = (state: ParserState, input: Event) => {
    )
  }
  if (event.type === "response.reasoning.delta" || event.type === "response.reasoning_summary_text.delta") {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(onReasoningDelta(state, event, event.item_id))
  }
  if (
@@ -1274,24 +1308,31 @@ export const step = (state: ParserState, input: Event) => {
    event.type === "response.reasoning_summary_text.done" ||
    event.type === "response.reasoning_text.done"
  ) {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
    return Effect.succeed(onReasoningDone(state, event, event.item_id))
  }
  if (event.type === "response.reasoning_summary_part.added")
-    return event.item_id
+    return event.item_id !== undefined
      ? Effect.succeed(onReasoningSummaryPartAdded(state, event))
      : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
  if (event.type === "response.reasoning_summary_part.done")
-    return event.item_id
+    return event.item_id !== undefined
      ? Effect.succeed(onReasoningSummaryPartDone(state, event))
      : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
  if (event.type === "response.output_item.added") {
-    if (event.item?.type === "message" && !event.item.id)
+    if (event.item?.type === "message" && event.item.id === undefined)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
+    if (
+      event.item &&
+      isReasoningItem(event.item) &&
+      state.reasoningItems[event.item.id] === undefined &&
+      state.lifecycle.reasoning.size > 0
+    )
+      return ProviderShared.eventError(state.id, `${event.type} started reasoning before the previous item ended`)
    const id = event.item?.id ?? (event.item?.type === "function_call" ? event.item.call_id : undefined)
    return Effect.succeed(
      onOutputItemAdded(
-        event.output_index !== undefined && id
+        event.output_index !== undefined && id !== undefined
          ? { ...state, outputItems: { ...state.outputItems, [event.output_index]: id } }
          : state,
        event,
@@ -1299,20 +1340,27 @@ export const step = (state: ParserState, input: Event) => {
    )
  }
  if (event.type === "response.function_call_arguments.delta" || event.type === "response.function_call_arguments.done")
-    return event.item_id
+    return event.item_id !== undefined
      ? onFunctionCallArgumentsDelta(state, event)
      : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
  if (event.type === "response.output_item.done") {
-    if (event.item?.type === "message" && !event.item.id)
+    if (event.item?.type === "message" && event.item.id === undefined)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
    return onOutputItemDone(state, event)
  }
  if (event.type === "response.completed" || event.type === "response.incomplete") return onResponseFinish(state, event)
-  if (event.type === "response.failed") return providerError(state, event, `${state.name} response failed`)
+  if (event.type === "response.failed") return providerFailure(event, `${state.name} response failed`)
  if (event.type === "error")
    return decodeKnownErrorEvent(event).pipe(
-      Effect.mapError(() => ProviderShared.eventError(state.id, `${state.name} returned a malformed error event`)),
-      Effect.flatMap(() => providerError(state, event, `${state.name} stream error`)),
+      Effect.mapError((cause) =>
+        ProviderShared.eventError(
+          state.id,
+          `${state.name} returned a malformed error event`,
+          ProviderShared.encodeJson(event),
+          cause,
+        ),
+      ),
+      Effect.flatMap(() => providerFailure(event, `${state.name} stream error`)),
    )
  return Effect.succeed<StepResult>([state, NO_EVENTS])
 }
@@ -1330,10 +1378,10 @@ export const initial = (request: LLMRequest, extension: Extension = BASE): Parse
  providerMetadataKey: request.model.route.providerMetadataKey ?? "openresponses",
  hasFunctionCall: false,
  tools: ToolStream.empty<string>(),
+  completedTools: new Set<string>(),
  lifecycle: Lifecycle.initial(),
  outputItems: {},
-  messageItems: new Set<string>(),
-  messagePhases: {},
+  message: undefined,
  reasoningItems: {},
 })

@@ -3,14 +3,16 @@ import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client.js"
 import { Auth } from "../route/auth.js"
 import { Endpoint } from "../route/endpoint.js"
+import { Framing } from "../route/framing.js"
 import { HttpTransport } from "../route/transport/index.js"
 import { Protocol } from "../route/protocol.js"
 import {
  AIError,
-  InvalidProviderOutputReason,
+  AIErrorReason,
+  InvalidProviderOutputError,
  LLMEvent,
-  ProviderInternalReason,
-  UnknownProviderReason,
+  ProviderInternalError,
+  UnknownProviderError,
  Usage,
  type FinishReason,
  type FinishReasonDetails,
@@ -244,6 +246,8 @@ export const OpenAIChatEvent = Schema.StructWithRest(
  [Schema.Record(Schema.String, Schema.Unknown)],
 )
 export type OpenAIChatEvent = Schema.Schema.Type<typeof OpenAIChatEvent>
+const DONE = "[DONE]" as const
+const OpenAIChatStreamEvent = Schema.Union([Schema.Literal(DONE), Protocol.jsonEvent(OpenAIChatEvent)])
 type OpenAIChatRequestMessage = LLMRequest["messages"][number]

 interface PendingToolDelta {
@@ -253,6 +257,7 @@ interface PendingToolDelta {
 }

 export interface ParserState {
+  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
  readonly pendingTools: Partial<Record<number, PendingToolDelta>>
  readonly toolCallEvents: ReadonlyArray<LLMEvent>
@@ -324,17 +329,18 @@ const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart
 const openAICompatibleReasoningContent = (native: unknown) =>
  isRecord(native) && typeof native.reasoning_content === "string" ? native.reasoning_content : undefined

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

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

@@ -366,7 +372,7 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
  message: OpenAIChatRequestMessage,
  configuredField: string | undefined,
  requireReasoning: boolean,
-  options: LoweringOptions,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
  const content: TextPart[] = []
  const reasoning: ReasoningPart[] = []
@@ -388,10 +394,14 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
    }
  }
  const text = reasoning.map((part) => part.text).join("")
-  const details = reasoningDetails(reasoning, message.native?.openaiCompatible)
-  const observedField = reasoning.map(reasoningField).find((value) => value !== undefined)
+  const details = reasoningDetails(reasoning, message.native?.openaiCompatible, options.providerMetadataKey)
+  const observedField = reasoning
+    .map((part) => reasoningField(part, options.providerMetadataKey))
+    .find((value) => value !== undefined)
  const nativeReasoning = openAICompatibleReasoningContent(message.native?.openaiCompatible)
-  const fullyStructured = reasoning.every((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))
+  const fullyStructured = reasoning.every((part) =>
+    Array.isArray(part.providerMetadata?.[options.providerMetadataKey]?.reasoningDetails),
+  )
  const field = (() => {
    if (configuredField !== undefined && (requireReasoning || reasoning.length > 0 || nativeReasoning !== undefined))
      return configuredField
@@ -459,7 +469,7 @@ const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
  message: OpenAIChatRequestMessage,
  reasoningField: string | undefined,
  requireReasoning: boolean,
-  options: LoweringOptions,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
  if (message.role === "user") return [yield* lowerUserMessage(message, options)]
  if (message.role === "assistant")
@@ -495,8 +505,13 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
  const mistral = ["mistral", "devstral", "codestral", "pixtral", "mixtral"].some((family) => modelID.includes(family))
  const lowering = {
    ...options,
+    providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
    toolCallID: (id: string) => {
-      if (mistral) return id.replace(/[^a-zA-Z0-9]/g, "").slice(0, 9).padEnd(9, "0")
+      if (mistral)
+        return id
+          .replace(/[^a-zA-Z0-9]/g, "")
+          .slice(0, 9)
+          .padEnd(9, "0")
      if (modelID.includes("claude")) return id.replace(/[^a-zA-Z0-9_-]/g, "_")
      if (request.model.provider === "openai" || request.model.provider === "azure" || modelID.startsWith("openai/"))
        return id.slice(0, 40)
@@ -505,7 +520,8 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
  }
  const requireAssistantAfterTool = request.model.compatibility?.requireAssistantAfterTool ?? mistral
  const bridgeTools = () => {
-    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool") messages.push({ role: "assistant", content: "Done." })
+    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool")
+      messages.push({ role: "assistant", content: "Done." })
  }
  const pendingImages: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
  const flushImages = () => {
@@ -557,7 +573,10 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
        )
      continue
    }
-    if (message.role === "assistant" && message.content.every((part) => part.type === "text" && part.text.trim() === ""))
+    if (
+      message.role === "assistant" &&
+      message.content.every((part) => part.type === "text" && part.text.trim() === "")
+    )
      continue
    if (message.role === "tool") {
      const lowered = yield* lowerToolMessages(message, lowering)
@@ -588,7 +607,10 @@ const hasToolHistory = (messages: ReadonlyArray<LLMRequest["messages"][number]>)
 // models.dev provider naming: DeepSeek, Moonshot AI, Together AI, ZAI
 // (Zhipu + Coding Plan variants), Nvidia, Cerebras, Chutes, etc. still
 // require `max_tokens`.
-const detectMaxTokensField = (provider: string, baseURL: string | undefined): "max_tokens" | "max_completion_tokens" => {
+const detectMaxTokensField = (
+  provider: string,
+  baseURL: string | undefined,
+): "max_tokens" | "max_completion_tokens" => {
  const p = provider.toLowerCase()
  const url = (baseURL ?? "").toLowerCase()
  if (
@@ -638,7 +660,8 @@ const detectSupportsStore = (provider: string, baseURL: string | undefined): boo
  const isChutes = p === "chutes" || url.includes("chutes.ai")
  const isCloudflareWorkersAI = p === "cloudflare-workers-ai" || url.includes("api.cloudflare.com")
  const isCloudflareAiGateway = p === "cloudflare-ai-gateway" || url.includes("gateway.ai.cloudflare.com")
-  const isVercelAiGateway = p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
+  const isVercelAiGateway =
+    p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
  const isAntLing = p === "ant-ling" || url.includes("api.ant-ling.com")
  const isOpencode = p === "opencode" || url.includes("opencode.ai")
  const isNonStandard =
@@ -670,11 +693,7 @@ const detectSupportsStrictMode = (provider: string, baseURL: string | undefined)
  return !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia
 }

-const detectZaiToolStream = (
-  provider: string,
-  baseURL: string | undefined,
-  modelID: string,
-): boolean => {
+const detectZaiToolStream = (provider: string, baseURL: string | undefined, modelID: string): boolean => {
  const p = provider.toLowerCase()
  const url = (baseURL ?? "").toLowerCase()
  const isZai =
@@ -724,10 +743,10 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
  const supportsStore = request.model.compatibility?.supportsStore ?? detectSupportsStore(provider, baseURL)
  const supportsUsageInStreaming =
    request.model.compatibility?.supportsUsageInStreaming ?? detectSupportsUsageInStreaming()
-  const supportsStrictMode = request.model.compatibility?.supportsStrictMode ?? detectSupportsStrictMode(provider, baseURL)
+  const supportsStrictMode =
+    request.model.compatibility?.supportsStrictMode ?? detectSupportsStrictMode(provider, baseURL)
  const zaiToolStream =
-    request.model.compatibility?.zaiToolStream ??
-    detectZaiToolStream(provider, baseURL, request.model.id)
+    request.model.compatibility?.zaiToolStream ?? detectZaiToolStream(provider, baseURL, request.model.id)
  const hasHistory = hasToolHistory(request.messages)
  const hasActiveTools = request.tools.length > 0
  return {
@@ -769,26 +788,22 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
 // Streaming parsers are small state machines: every event returns a new state
 // plus the common `LLMEvent`s produced by that event. Tool calls are accumulated
 // because OpenAI streams JSON arguments across multiple deltas.
-const finishReasonError = (event: OpenAIChatEvent, reason: AIError["reason"]) =>
-  new AIError({
-    module: ADAPTER,
-    method: "stream",
-    body: ProviderShared.encodeJson(event),
-    reason,
-  })
-
 const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event: OpenAIChatEvent, reason: string) {
  switch (reason) {
    case "error":
-      return yield* finishReasonError(
-        event,
-        new UnknownProviderReason({ message: "Provider reported an error (finish_reason: error)" }),
-      )
+      return yield* new AIError({
+        reason: new UnknownProviderError({
+          message: "Provider reported an error (finish_reason: error)",
+          body: ProviderShared.encodeJson(event),
+        }),
+      })
    case "network_error":
-      return yield* finishReasonError(
-        event,
-        new ProviderInternalReason({ message: "Provider reported a network error (finish_reason: network_error)" }),
-      )
+      return yield* new AIError({
+        reason: new ProviderInternalError({
+          message: "Provider reported a network error (finish_reason: network_error)",
+          body: ProviderShared.encodeJson(event),
+        }),
+      })
    case "stop":
    case "end":
      return "stop" as const
@@ -812,15 +827,14 @@ const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event
 // Providers differ on cache-hit location: OpenAI uses
 // `prompt_tokens_details.cached_tokens`, DeepSeek uses
 // `prompt_cache_hit_tokens`, and Zai uses top-level `cached_tokens`.
-const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
+const mapUsage = (usage: OpenAIChatEvent["usage"], providerMetadataKey: string): Usage | undefined => {
  if (!usage) return undefined
  const input = usage.prompt_tokens ?? undefined
  const output = usage.completion_tokens ?? undefined
-  const cached =
-    (usage.prompt_tokens_details?.cached_tokens ??
-      (usage as { prompt_cache_hit_tokens?: number | null }).prompt_cache_hit_tokens ??
-      (usage as { cached_tokens?: number | null }).cached_tokens ??
-      undefined) as number | undefined
+  const cached = (usage.prompt_tokens_details?.cached_tokens ??
+    (usage as { prompt_cache_hit_tokens?: number | null }).prompt_cache_hit_tokens ??
+    (usage as { cached_tokens?: number | null }).cached_tokens ??
+    undefined) as number | undefined
  const cacheWrite = usage.prompt_tokens_details?.cache_write_tokens ?? undefined
  const reasoning = usage.completion_tokens_details?.reasoning_tokens ?? undefined
  const nonCached = ProviderShared.subtractTokens(input, ProviderShared.sumTokens(cached, cacheWrite))
@@ -832,7 +846,7 @@ const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
    cacheWriteInputTokens: cacheWrite,
    reasoningTokens: reasoning,
    totalTokens: ProviderShared.totalTokens(input, output, usage.total_tokens ?? undefined),
-    providerMetadata: { openai: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
  })
 }

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

-const reasoningMetadata = (field: ParserState["reasoningField"], details?: ReadonlyArray<unknown>) => ({
-  openai: {
+const reasoningMetadata = (
+  providerMetadataKey: string,
+  field: ParserState["reasoningField"],
+  details?: ReadonlyArray<unknown>,
+) => ({
+  [providerMetadataKey]: {
    ...(field ? { reasoningField: field } : {}),
    ...(details ? { reasoningDetails: details } : {}),
  },
@@ -918,12 +936,8 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
    if (event.error) {
      const body = ProviderShared.encodeJson(event)
      return yield* new AIError({
-        module: ADAPTER,
-        method: "stream",
-        body,
        reason: classifyProviderFailure({
          message: event.error.message,
-          code: event.error.code === undefined || event.error.code === null ? undefined : String(event.error.code),
          status: typeof event.error.code === "number" ? event.error.code : undefined,
          rawBody: body,
        }),
@@ -934,15 +948,17 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
    // Moonshot (and a few other OpenAI-compatible providers) attach usage to
    // `choice.usage` instead of the top-level `usage` field.
    const choiceUsage = (choice as unknown as { usage?: OpenAIChatEvent["usage"] })?.usage
-    const usage = mapUsage(event.usage) ?? (choiceUsage ? mapUsage(choiceUsage) : undefined) ?? state.usage
+    const usage =
+      mapUsage(event.usage, state.providerMetadataKey) ??
+      (choiceUsage ? mapUsage(choiceUsage, state.providerMetadataKey) : undefined) ??
+      state.usage
    const rawFinishReason = choice?.finish_reason
-    const finishReason =
-      rawFinishReason
-        ? {
-            normalized: yield* mapFinishReason(event, rawFinishReason),
-            raw: choice?.native_finish_reason ?? rawFinishReason,
-          }
-        : state.finishReason
+    const finishReason = rawFinishReason
+      ? {
+          normalized: yield* mapFinishReason(event, rawFinishReason),
+          raw: choice?.native_finish_reason ?? rawFinishReason,
+        }
+      : state.finishReason
    const delta = choice?.delta
    const toolDeltas = delta?.tool_calls ?? []
    let tools = state.tools
@@ -973,7 +989,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
    const detailDelta = Array.isArray(delta?.reasoning_details) ? delta.reasoning_details : undefined
    if (detailDelta !== undefined) appendReasoningDetails(state.reasoningDetails, detailDelta)
    const reasoningDetailsObserved = state.reasoningDetailsObserved || detailDelta !== undefined
-    const deltaMetadata = reasoningMetadata(reasoningField)
+    const deltaMetadata = reasoningMetadata(state.providerMetadataKey, reasoningField)
    const text = detailDelta?.length ? (detailText(detailDelta) ?? reasoning?.text) : reasoning?.text
    if (text !== undefined) lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", text, deltaMetadata)
    else if (
@@ -984,25 +1000,12 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
      lifecycle = Lifecycle.reasoningStart(lifecycle, events, "reasoning-0", deltaMetadata)
    const reasoningEmitted = state.reasoningEmitted || lifecycle.reasoning.has("reasoning-0")

-    if (delta?.content) {
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
-      )
-      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)
-    }
+    // Reasoning is one response-wide channel: it stays open alongside text and
+    // refusal output so late reasoning deltas and details join the same block,
+    // and `finishEvents` closes it once with the complete metadata.
+    if (delta?.content) lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)

-    if (delta?.refusal) {
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
-      )
-      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)
-    }
+    if (delta?.refusal) lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)

    // Compatible providers may omit indexes. Prefer durable identity, then use
    // batch position for parallel deltas or the latest call for sparse chunks.
@@ -1038,7 +1041,14 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
        "OpenAI Chat tool call delta is missing id or name",
      )
      if (ToolStream.isError(result))
-        return yield* ProviderShared.eventError(ADAPTER, result.reason.message, ProviderShared.encodeJson(event))
+        return yield* new AIError({
+          reason: AIErrorReason.make({
+            ...result.reason,
+            message: result.message,
+            cause: result.reason.cause,
+            body: ProviderShared.encodeJson(event),
+          }),
+        })
      tools = result.tools
      if (result.events.length) lifecycle = Lifecycle.stepStart(lifecycle, events)
      events.push(...result.events)
@@ -1060,6 +1070,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>

    return [
      {
+        providerMetadataKey: state.providerMetadataKey,
        tools: finished?.tools ?? tools,
        pendingTools,
        toolCallEvents: finished?.events ?? state.toolCallEvents,
@@ -1081,11 +1092,9 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
 const finishEvents = Effect.fn("OpenAIChat.finishEvents")(function* (state: ParserState) {
  if (state.finishReason === undefined && state.requireFinishReason)
    return yield* new AIError({
-      module: ADAPTER,
-      method: "stream",
-      reason: new InvalidProviderOutputReason({
-        classification: "incomplete-stream",
+      reason: new InvalidProviderOutputError({
        message: "OpenAI Chat stream ended without finish_reason",
+        classification: "incomplete-stream",
        route: ADAPTER,
      }),
    })
@@ -1102,13 +1111,21 @@ const finishEvents = Effect.fn("OpenAIChat.finishEvents")(function* (state: Pars
          state.finishReason.normalized === "stop" && hasToolCalls ? "tool-calls" : state.finishReason.normalized,
      }
    : { normalized: hasToolCalls ? ("tool-calls" as const) : ("stop" as const) }
+  // Snapshot details at publish time so the emitted event never observes later
+  // mutation of the accumulated `reasoningDetails` array.
  const metadata = reasoningMetadata(
+    state.providerMetadataKey,
    state.reasoningField,
-    state.reasoningDetailsObserved ? state.reasoningDetails : undefined,
+    state.reasoningDetailsObserved ? [...state.reasoningDetails] : undefined,
  )
  const started =
    state.reasoningDetailsObserved && !state.reasoningEmitted
-      ? Lifecycle.reasoningStart(state.lifecycle, events, "reasoning-0", reasoningMetadata(state.reasoningField))
+      ? Lifecycle.reasoningStart(
+          state.lifecycle,
+          events,
+          "reasoning-0",
+          reasoningMetadata(state.providerMetadataKey, state.reasoningField),
+        )
      : state.lifecycle
  const ended = Lifecycle.reasoningEnd(started, events, "reasoning-0", metadata)
  const lifecycle = toolCallEvents.length ? Lifecycle.stepStart(ended, events) : ended
@@ -1133,8 +1150,9 @@ export const protocol = Protocol.make({
    from: fromRequest,
  },
  stream: {
-    event: Protocol.jsonEvent(OpenAIChatEvent),
+    event: OpenAIChatStreamEvent,
    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
      pendingTools: {},
      toolCallEvents: [],
@@ -1146,12 +1164,14 @@ export const protocol = Protocol.make({
      nextToolIndex: 0,
      requireFinishReason: request.model.compatibility?.requireFinishReason ?? true,
    }),
-    step,
+    step: (state: ParserState, event) => (event === DONE ? Effect.succeed([state, []] as const) : step(state, event)),
+    terminal: (event) => event === DONE,
    onHalt: finishEvents,
  },
 })

-export const httpTransport = HttpTransport.sseJson.with<OpenAIChatBody>()
+export const framing = Framing.sseWithDone
+export const httpTransport = HttpTransport.sseJson.with<OpenAIChatBody>().with({ framing })

 export const route = Route.make({
  id: ADAPTER,
@@ -1,6 +1,5 @@
 import { Route, type RouteRoutedLanguageModelInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import * as OpenAIChat from "./openai-chat.js"

 const ADAPTER = "openai-compatible-chat"
@@ -19,7 +18,7 @@ export const route = Route.make({
  providerMetadataKey: "openai",
  protocol: OpenAIChat.protocol,
  endpoint: Endpoint.path("/chat/completions"),
-  framing: Framing.sse,
+  framing: OpenAIChat.framing,
 })

 export * as OpenAICompatibleChat from "./openai-compatible-chat.js"
@@ -9,14 +9,7 @@ import {
  type ImageRoute,
 } from "../image.js"
 import { Auth, type Definition as AuthDefinition } from "../route/auth.js"
-import {
-  InvalidProviderOutputReason,
-  AIError,
-  Usage,
-  mergeHttpOptions,
-  mergeJsonRecords,
-  type HttpOptions,
-} from "../schema/index.js"
+import { Usage, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema/index.js"
 import { ProviderShared } from "./shared.js"
 import { ImageInputs } from "./utils/image-input.js"
 import { OpenAIImage } from "./utils/openai-image.js"
@@ -84,13 +77,6 @@ const nativeOptions = (options: OpenAIImageOptions | undefined) => {
  }
 }

-const invalidOutput = (message: string) =>
-  new AIError({
-    module: ADAPTER,
-    method: "generate",
-    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
-  })
-
 const applyQuery = (url: string, query: Record<string, string> | undefined) => {
  if (!query) return url
  const next = new URL(url)
@@ -104,12 +90,12 @@ export const model = (input: ModelInput) => {
    generate: Effect.fn("OpenAIImages.generate")(function* (request: ImageRequestFor<OpenAIImageOptions>, execute) {
      const mask = request.options?.mask
      if (mask !== undefined && (request.images?.length ?? 0) === 0)
-        return yield* ImageInputs.invalid(ADAPTER, "An OpenAI image mask requires at least one input image")
+        return yield* ImageInputs.invalid("An OpenAI image mask requires at least one input image")
      const http = mergeHttpOptions(request.model.http, request.http)
      const sourceImages = request.images ?? []
      const multipartImages = yield* Effect.forEach(sourceImages, (image) => {
        if (image.type === "bytes") return Effect.succeed({ data: image.data, mediaType: image.mediaType })
-        if (image.type === "url") return ImageInputs.decodeDataUrl(image.url, ADAPTER)
+        if (image.type === "url") return ImageInputs.decodeDataUrl(image.url)
        return Effect.undefined
      })
      const multipartMask =
@@ -118,7 +104,7 @@ export const model = (input: ModelInput) => {
          : mask.type === "bytes"
            ? { data: mask.data, mediaType: mask.mediaType }
            : mask.type === "url"
-              ? yield* ImageInputs.decodeDataUrl(mask.url, ADAPTER)
+              ? yield* ImageInputs.decodeDataUrl(mask.url)
              : undefined
      const useMultipart =
        sourceImages.length > 0 &&
@@ -161,7 +147,7 @@ export const model = (input: ModelInput) => {
        return undefined
      })
      if (references.some((image) => image === undefined))
-        return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts image URLs, data URLs, bytes, and file IDs")
+        return yield* ImageInputs.invalid("OpenAI Images accepts image URLs, data URLs, bytes, and file IDs")
      const maskReference =
        mask === undefined
          ? undefined
@@ -173,7 +159,7 @@ export const model = (input: ModelInput) => {
                ? { file_id: mask.id }
                : undefined
      if (mask !== undefined && maskReference === undefined)
-        return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts masks as URLs, data URLs, bytes, or file IDs")
+        return yield* ImageInputs.invalid("OpenAI Images accepts masks as URLs, data URLs, bytes, or file IDs")
      const requestBody = mergeJsonRecords(
        {
          model: request.model.id,
@@ -209,11 +195,9 @@ const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
  options: OpenAIImageOptions | undefined,
  overlay: Record<string, unknown> | undefined,
 ) {
-  const payload = yield* response.json.pipe(
-    Effect.mapError(() => invalidOutput("Failed to read the OpenAI Images response")),
-  )
-  const decoded = yield* Schema.decodeUnknownEffect(OpenAIImageResponse)(payload).pipe(
-    Effect.mapError(() => invalidOutput("OpenAI Images returned an invalid response")),
+  const output = yield* ProviderShared.imageResponse(ADAPTER, "OpenAI Images", response)
+  const decoded = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(OpenAIImageResponse))(output.body).pipe(
+    Effect.mapError((cause) => output.invalid("OpenAI Images returned an invalid response", cause)),
  )
  const requestBody = mergeJsonRecords(nativeOptions(options), overlay)
  const format =
@@ -221,7 +205,7 @@ const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
  const images = yield* Effect.forEach(decoded.data, (item, index) => {
    if (item.b64_json)
      return Effect.fromResult(Encoding.decodeBase64(item.b64_json)).pipe(
-        Effect.mapError(() => invalidOutput(`OpenAI Images result ${index} contains invalid base64 data`)),
+        Effect.mapError((cause) => output.invalid(`OpenAI Images result ${index} contains invalid base64 data`, cause)),
        Effect.map(
          (data) =>
            new GeneratedImage({
@@ -241,9 +225,9 @@ const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
            item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
        }),
      )
-    return Effect.fail(invalidOutput(`OpenAI Images result ${index} has neither image data nor a URL`))
+    return Effect.fail(output.invalid(`OpenAI Images result ${index} has neither image data nor a URL`))
  })
-  if (images.length === 0) return yield* invalidOutput("OpenAI Images returned no images")
+  if (images.length === 0) return yield* output.invalid("OpenAI Images returned no images")
  return new ImageResponse({
    images,
    usage:
@@ -86,28 +86,10 @@ const OpenAIResponsesBody = Schema.Struct({
 })
 export type OpenAIResponsesBody = Schema.Schema.Type<typeof OpenAIResponsesBody>

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

 const nativeImageToolInput = (tool: ToolDefinition) => {
@@ -167,7 +149,9 @@ const hostedToolResult = Effect.fn("OpenAIResponses.hostedToolResult")(function*
  const isError = item.error !== undefined && item.error !== null
  if (item.type === "image_generation_call" && item.result) {
    yield* Effect.fromResult(Encoding.decodeBase64(item.result)).pipe(
-      Effect.mapError(() => ProviderShared.eventError(ADAPTER, "OpenAI Responses returned invalid image base64")),
+      Effect.mapError((cause) =>
+        ProviderShared.eventError(ADAPTER, "OpenAI Responses returned invalid image base64", undefined, cause),
+      ),
    )
    const format = item.output_format ?? "png"
    return {
@@ -202,7 +186,7 @@ const HOSTED_TOOLS = {

 const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
  if (event.type === "response.reasoning_text.delta")
-    return event.item_id
+    return event.item_id !== undefined
      ? Effect.succeed(
          OpenResponses.onReasoningDelta(state, event, OpenResponses.outputItemID(state, event) ?? event.item_id),
        )
@@ -2,11 +2,12 @@ import { Buffer } from "node:buffer"
 import { Tool } from "@opencode-ai/schema/tool"
 import { Effect, Schema, Stream } from "effect"
 import * as Sse from "effect/unstable/encoding/Sse"
-import { Headers, HttpClientRequest } from "effect/unstable/http"
+import { Headers, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
 import {
-  InvalidProviderOutputReason,
-  InvalidRequestReason,
+  InvalidProviderOutputError,
+  InvalidRequestError,
  AIError,
+  HttpContext,
  type ContentPart,
  type LLMRequest,
  type MediaPart,
@@ -96,17 +97,15 @@ export const sumTokens = (...values: ReadonlyArray<number | undefined>): number
  return values.reduce((acc: number, value) => acc + (value ?? 0), 0)
 }

-export const eventError = (route: string, message: string, raw?: string) =>
+export const eventError = (route: string, message: string, body?: string, cause?: unknown) =>
  new AIError({
-    module: "ProviderShared",
-    method: "stream",
-    reason: new InvalidProviderOutputReason({ route, message, raw }),
+    reason: new InvalidProviderOutputError({ route, message, body, cause }),
  })

 export const parseJson = (route: string, input: string, message: string) =>
  Effect.try({
    try: () => decodeJson(input),
-    catch: () => eventError(route, message, input),
+    catch: (cause) => eventError(route, message, input, cause),
  })

 /**
@@ -208,15 +207,16 @@ 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. Retry control events are ignored without interrupting the stream.
+ * decoder, optionally filters named events, and drops empty events. `[DONE]`
+ * is dropped by default or retained for protocols that use it as their stream
+ * boundary. Retry control events are ignored without interrupting the stream.
 * Decoder failures become provider output errors so the public error channel
 * stays `AIError`.
 */
 export const sseFraming = (
  bytes: Stream.Stream<Uint8Array, AIError>,
  events?: ReadonlySet<string>,
+  includeDone = false,
 ): Stream.Stream<string, AIError> =>
  bytes.pipe(
    Stream.decodeText(),
@@ -233,7 +233,7 @@ export const sseFraming = (
      (state, chunk) =>
        Effect.gen(function* () {
          const error = state.parser.feed(chunk)
-          if (error) return yield* eventError("sse", error.message)
+          if (error) return yield* eventError("sse", error.message, chunk, error)
          return [state, state.output.splice(0)] as const
        }),
    ),
@@ -241,7 +241,7 @@ export const sseFraming = (
      (event) =>
        (events === undefined || events.has(event.event)) &&
        event.data.length > 0 &&
-        (event.data !== "[DONE]" || (events !== undefined && event.event !== "message")),
+        (event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
    ),
    Stream.map((event) => event.data),
  )
@@ -249,13 +249,39 @@ export const sseFraming = (
 /**
 * Canonical invalid-request constructor shared by protocol lowering.
 */
-export const invalidRequest = (message: string) =>
+export const invalidRequest = (message: string, cause?: unknown) =>
  new AIError({
-    module: "ProviderShared",
-    method: "request",
-    reason: new InvalidRequestReason({ message }),
+    reason: new InvalidRequestError({ message, cause }),
  })

+export const imageResponse = Effect.fn("ProviderShared.imageResponse")(function* (
+  route: string,
+  name: string,
+  response: HttpClientResponse.HttpClientResponse,
+) {
+  const http = new HttpContext({ url: response.request.url, status: response.status, headers: response.headers })
+  const body = yield* response.text.pipe(
+    Effect.mapError(
+      (cause) =>
+        new AIError({
+          reason: new InvalidProviderOutputError({
+            route,
+            message: `Failed to read the ${name} response`,
+            http,
+            cause,
+          }),
+        }),
+    ),
+  )
+  return {
+    body,
+    invalid: (message: string, cause?: unknown) =>
+      new AIError({
+        reason: new InvalidProviderOutputError({ route, message, body, http, cause }),
+      }),
+  }
+})
+
 export const matchToolChoice = <Auto, None, Required, Tool>(
  route: string,
  toolChoice: NonNullable<LLMRequest["toolChoice"]>,
@@ -302,7 +328,7 @@ export const unsupportedContent = (
 export const validateWith =
  <A, I, E extends { readonly message: string }>(decode: (input: I) => Effect.Effect<A, E>) =>
  (payload: I) =>
-    decode(payload).pipe(Effect.mapError((error) => invalidRequest(error.message)))
+    decode(payload).pipe(Effect.mapError((error) => invalidRequest(error.message, error)))

 /**
 * Build an HTTP POST with a JSON body. Sets `content-type: application/json`
@@ -1,12 +1,10 @@
 import { Effect, Encoding } from "effect"
 import type { ImageInput } from "../../image.js"
-import { InvalidRequestReason, AIError } from "../../schema/index.js"
+import { InvalidRequestError, AIError } from "../../schema/index.js"

-const invalid = (module: string, message: string) =>
+const invalid = (message: string, cause?: unknown) =>
  new AIError({
-    module,
-    method: "generate",
-    reason: new InvalidRequestReason({ message }),
+    reason: new InvalidRequestError({ message, cause }),
  })

 export const dataUrl = (input: Extract<ImageInput, { readonly type: "bytes" }>) =>
@@ -14,13 +12,12 @@ export const dataUrl = (input: Extract<ImageInput, { readonly type: "bytes" }>)

 export const decodeDataUrl = (
  url: string,
-  module: string,
 ): Effect.Effect<{ readonly mediaType: string; readonly data: Uint8Array } | undefined, AIError> => {
  if (!url.startsWith("data:")) return Effect.undefined
  const match = /^data:([^;,]+);base64,(.*)$/s.exec(url)
-  if (!match) return Effect.fail(invalid(module, "Image data URLs must contain a MIME type and base64 data"))
+  if (!match) return Effect.fail(invalid("Image data URLs must contain a MIME type and base64 data"))
  return Effect.fromResult(Encoding.decodeBase64(match[2])).pipe(
-    Effect.mapError(() => invalid(module, "Image data URL contains invalid base64 data")),
+    Effect.mapError((cause) => invalid("Image data URL contains invalid base64 data", cause)),
    Effect.map((data) => ({ mediaType: match[1], data })),
  )
 }
@@ -62,22 +62,31 @@ export const reasoningEnd = (
  events: LLMEvent[],
  id: string,
  providerMetadata?: ProviderMetadata,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text?: string,
 ): State => {
  if (!state.reasoning.has(id)) return state
  const stepped = stepStart(state, events)
-  events.push(LLMEvent.reasoningEnd({ id, providerMetadata }))
+  events.push(LLMEvent.reasoningEnd({ id, text, providerMetadata }))
  const reasoning = new Set(stepped.reasoning)
  reasoning.delete(id)
  return { ...stepped, reasoning }
 }

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

 const closeOpenBlocks = (state: State, events: LLMEvent[]): State => {
@@ -29,10 +29,9 @@ export type ResponseIncludable = (typeof ResponseIncludables)[number] | (string

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

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

 export * as ResponsesHostedTools from "./responses-hosted-tools.js"
@@ -2,14 +2,7 @@ import { Effect, Encoding, Schema } from "effect"
 import { Headers, HttpClientRequest } from "effect/unstable/http"
 import { GeneratedImage, ImageModel, ImageResponse, type ImageRequestFor, type ImageRoute } from "../image.js"
 import { Auth, type Definition as AuthDefinition } from "../route/auth.js"
-import {
-  InvalidProviderOutputReason,
-  AIError,
-  Usage,
-  mergeHttpOptions,
-  mergeJsonRecords,
-  type HttpOptions,
-} from "../schema/index.js"
+import { Usage, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema/index.js"
 import { ProviderShared, optionalNull } from "./shared.js"
 import { ImageInputs } from "./utils/image-input.js"

@@ -94,13 +87,6 @@ const nativeOptions = (options: XAIImageOptions | undefined) => {
  }
 }

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

@@ -63,13 +57,6 @@ const nativeOptions = (options: ZAIImageOptions | undefined) => {
  }
 }

-const invalidOutput = (message: string) =>
-  new AIError({
-    module: ADAPTER,
-    method: "generate",
-    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
-  })
-
 const applyQuery = (url: string, query: Record<string, string> | undefined) => {
  if (!query) return url
  const next = new URL(url)
@@ -82,7 +69,7 @@ export const model = (input: ModelInput) => {
    id: ADAPTER,
    generate: Effect.fn("ZAIImages.generate")(function* (request: ImageRequestFor<ZAIImageOptions>, execute) {
      if ((request.images?.length ?? 0) > 0)
-        return yield* ImageInputs.invalid(ADAPTER, "Z.ai hosted image generation does not support image inputs")
+        return yield* ImageInputs.invalid("Z.ai hosted image generation does not support image inputs")
      const http = mergeHttpOptions(request.model.http, request.http)
      const requestBody = mergeJsonRecords(
        { model: request.model.id, prompt: request.prompt },
@@ -104,13 +91,11 @@ export const model = (input: ModelInput) => {
          HttpClientRequest.bodyText(text, "application/json"),
        ),
      )
-      const payload = yield* response.json.pipe(
-        Effect.mapError(() => invalidOutput("Failed to read the Z.ai Images response")),
+      const output = yield* ProviderShared.imageResponse(ADAPTER, "Z.ai Images", response)
+      const decoded = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(ZAIImageResponse))(output.body).pipe(
+        Effect.mapError((cause) => output.invalid("Z.ai Images returned an invalid response", cause)),
      )
-      const decoded = yield* Schema.decodeUnknownEffect(ZAIImageResponse)(payload).pipe(
-        Effect.mapError(() => invalidOutput("Z.ai Images returned an invalid response")),
-      )
-      if (decoded.data.length === 0) return yield* invalidOutput("Z.ai Images returned no images")
+      if (decoded.data.length === 0) return yield* output.invalid("Z.ai Images returned no images")
      return new ImageResponse({
        images: decoded.data.map(
          (item) =>
@@ -1,17 +1,16 @@
 import { Option, Schema } from "effect"
 import {
-  AuthenticationReason,
-  ContentPolicyReason,
-  InvalidRequestReason,
+  AuthenticationError,
+  ContentPolicyError,
+  InvalidRequestError,
  AIError,
  ProviderErrorEvent,
-  ProviderInternalReason,
-  QuotaExceededReason,
-  RateLimitReason,
-  UnknownProviderReason,
+  ProviderInternalError,
+  QuotaExceededError,
+  RateLimitError,
+  UnknownProviderError,
  type HttpContext,
  type HttpRateLimitDetails,
-  type ProviderMetadata,
 } from "./schema/index.js"

 const patterns = [
@@ -38,17 +37,19 @@ const patterns = [
  /too large for model with \d+ maximum context length/i,
  /prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i,
  /model_context_window_exceeded/i,
+  /range of input length should be/i,
  /too many tokens/i,
  /token limit exceeded/i,
+  /request_too_large/i,
 ]

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

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

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

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

@@ -59,6 +60,7 @@ export const isContextOverflowFailure = (failure: unknown) =>

 const decodeJson = Schema.decodeUnknownOption(Schema.fromJsonString(Schema.Unknown))
 const QUOTA_CODES = new Set(["insufficient_quota", "usage_not_included", "billing_error"])
+const AUTH_CODES = new Set(["authentication_error", "permission_error"])
 const SERVER_CODES = new Set([
  "api_error",
  "internal_error",
@@ -74,96 +76,91 @@ const INVALID_REQUEST_CODES = new Set(["invalid_prompt", "invalid_request_error"
 const RATE_LIMIT_TEXT = /rate increased too quickly|rate[-_\s]?limit|too[_\s]?many[_\s]?requests/i
 const QUOTA_TEXT = /insufficient[-_\s]?quota|quota[-_\s]?exceeded/i
 const CONTENT_POLICY_TEXT = /content[-_\s]?policy|content_filter|safety/i
-const NETWORK_ERROR_TEXT = /network[-_\s]error/i

 export interface ProviderFailure {
  readonly message: string
  readonly status?: number | undefined
-  readonly code?: string | undefined
  // Raw wire payload, scanned for failure signals (codes, overflow phrases)
  // that the summary message does not carry. Not shown to users.
  readonly rawBody?: string | undefined
+  // Some SDKs supply parsed error data separately from the original response text.
+  readonly data?: unknown
+  readonly http?: HttpContext | undefined
+  readonly cause?: unknown
  readonly retryAfterMs?: number | undefined
  readonly rateLimit?: HttpRateLimitDetails | undefined
-  readonly http?: HttpContext | undefined
-  readonly providerMetadata?: ProviderMetadata | undefined
 }

-// Keep HTTP failures and provider-reported stream failures on one typed path so
-// session retry policy never needs provider-specific string matching.
+// Classification records affirmative evidence about a failure. Deterministic
+// failures need positive identification (a 4xx status, quota/auth/policy
+// signals); anything unrecognized stays UnknownProvider, which the session
+// retry policy treats as retry-eligible because transient failures arrive in
+// unpredictable shapes while deterministic rejections almost always carry a
+// status or known code.
 export function classifyProviderFailure(input: ProviderFailure): AIError["reason"] {
-  const body = input.http?.body ?? input.rawBody ?? ""
-  const codes = [input.code, ...providerCodes(body), ...providerCodes(input.message)]
-    .filter((code): code is string => code !== undefined)
-    .map((code) => code.toLowerCase())
+  const details = { message: input.message, body: input.rawBody, http: input.http, cause: input.cause }
+  const body = input.rawBody ?? ""
+  const codes = [...providerCodes(input.data), ...providerCodes(body), ...providerCodes(input.message)].map((code) =>
+    code.toLowerCase(),
+  )
  // Scan the raw payload too so signals missing from the summary message
  // (e.g. overflow phrases nested in a JSON error body) still classify.
  const text = [input.message, body].filter((value) => value.length > 0).join("\n")
-  const common = { message: input.message, providerMetadata: input.providerMetadata, http: input.http }
  const clientScoped = input.status === undefined || (input.status >= 400 && input.status < 500)

  if (
    clientScoped &&
    (codes.includes("context_length_exceeded") ||
      codes.includes("model_context_window_exceeded") ||
+      codes.includes("request_too_large") ||
      isContextOverflow(text))
  )
-    return new InvalidRequestReason({ ...common, classification: "context-overflow" })
+    return new InvalidRequestError({ ...details, classification: "context-overflow" })
  if (input.status === 413 || isPayloadTooLarge(text))
-    return new InvalidRequestReason({ ...common, classification: "payload-too-large" })
-  if (CONTENT_POLICY_TEXT.test(text)) return new ContentPolicyReason(common)
+    return new InvalidRequestError({ ...details, classification: "payload-too-large" })
+  if (CONTENT_POLICY_TEXT.test(text)) return new ContentPolicyError(details)
  if (codes.some((code) => QUOTA_CODES.has(code)) || (input.status === 429 && QUOTA_TEXT.test(text)))
-    return new QuotaExceededReason(common)
-  if (input.status === 401) return new AuthenticationReason({ ...common, kind: "invalid" })
-  if (input.status === 403) return new AuthenticationReason({ ...common, kind: "insufficient-permissions" })
-  if (codes.includes("authentication_error")) return new AuthenticationReason({ ...common, kind: "invalid" })
-  if (codes.includes("permission_error"))
-    return new AuthenticationReason({ ...common, kind: "insufficient-permissions" })
+    return new QuotaExceededError(details)
+  if (input.status === 401 || input.status === 403 || codes.some((code) => AUTH_CODES.has(code)))
+    return new AuthenticationError(details)
  if (
-    codes.some((code) => code.includes("rate_limit") || code === "too_many_requests" || code === "throttlingexception")
+    input.status === 429 ||
+    codes.some(
+      (code) => code.includes("rate_limit") || code === "too_many_requests" || code === "throttlingexception",
+    ) ||
+    RATE_LIMIT_TEXT.test(text)
  )
-    return new RateLimitReason({
-      ...common,
+    return new RateLimitError({
+      ...details,
      retryAfterMs: input.retryAfterMs,
      rateLimit: input.rateLimit,
    })
-  if (RATE_LIMIT_TEXT.test(text))
-    return new RateLimitReason({
-      ...common,
-      retryAfterMs: input.retryAfterMs,
-      rateLimit: input.rateLimit,
-    })
-  if (NETWORK_ERROR_TEXT.test(text)) return new ProviderInternalReason({ ...common, status: input.status })
-  if (codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable")))
-    return new ProviderInternalReason({
-      ...common,
-      status: input.status,
+  if (
+    input.status === 408 ||
+    input.status === 409 ||
+    (input.status !== undefined && input.status >= 500) ||
+    codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable"))
+  )
+    return new ProviderInternalError({
+      ...details,
      retryAfterMs: input.retryAfterMs,
    })
-  if (input.status === 429) {
-    return new RateLimitReason({
-      ...common,
-      retryAfterMs: input.retryAfterMs,
-      rateLimit: input.rateLimit,
-    })
-  }
-  if (input.status === 408 || input.status === 409 || (input.status !== undefined && input.status >= 500))
-    return new ProviderInternalReason({
-      ...common,
-      status: input.status,
-      retryAfterMs: input.retryAfterMs,
-    })
-  if (codes.some((code) => INVALID_REQUEST_CODES.has(code))) return new InvalidRequestReason(common)
-  if (input.status === 400 || input.status === 404 || input.status === 413 || input.status === 422)
-    return new InvalidRequestReason(common)
-  return new UnknownProviderReason({ ...common, status: input.status })
+  if (codes.some((code) => INVALID_REQUEST_CODES.has(code))) return new InvalidRequestError(details)
+  // Any remaining 4xx is a deterministic rejection of this request.
+  if (input.status !== undefined && input.status >= 400 && input.status < 500) return new InvalidRequestError(details)
+  return new UnknownProviderError(details)
 }

-function providerCodes(value: string) {
-  const decoded = Option.getOrUndefined(decodeJson(value))
+function providerCodes(value: unknown) {
+  const decoded = typeof value === "string" ? Option.getOrUndefined(decodeJson(value)) : value
  if (!isRecord(decoded)) return []
  const error = isRecord(decoded.error) ? decoded.error : undefined
-  return [decoded.code, error?.code, error?.type].filter((value): value is string => typeof value === "string")
+  const response = isRecord(decoded.response) ? decoded.response : undefined
+  const responseError = response && isRecord(response.error) ? response.error : undefined
+  const exception = isRecord(decoded.exception) ? decoded.exception : undefined
+  return [decoded.code, error?.code, error?.type, error?.status, responseError?.code, exception?.type].filter(
+    (value): value is string => typeof value === "string",
+  )
 }

 function isRecord(value: unknown): value is Record<string, unknown> {
@@ -1,5 +1,5 @@
 import { Auth } from "../route/auth.js"
-import type { Route as RouteDef, RouteDefaultsInput } from "../route/client.js"
+import { Route, type RouteDefaultsInput } from "../route/client.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { OpenAIChat } from "../protocols/openai-chat.js"
 import { OpenAIResponses } from "../protocols/openai-responses.js"
@@ -23,22 +23,30 @@ export interface Settings extends ProviderPackage.Settings {
  readonly baseURL?: string
  readonly credentials?: Credentials
  readonly region?: string
+  readonly topP?: number
  readonly providerOptions?: OpenAIProviderOptionsInput
 }

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

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

 export const routes = [responsesRoute, chatRoute]

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

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

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

@@ -68,7 +68,7 @@ const protocol = {
 const route = Route.make({
  id: "google-vertex-gemini",
  provider: id,
-  providerMetadataKey: "google",
+  providerMetadataKey: "vertex",
  protocol,
  endpoint: Endpoint.path(({ request }) => {
    const model = String(request.model.id)
@@ -0,0 +1,115 @@
+import { Effect, Schema } from "effect"
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAIChat } from "../protocols/openai-chat.js"
+import { ProviderShared } from "../protocols/shared.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { Route, type RouteDefaultsInput } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Protocol } from "../route/protocol.js"
+import { ProviderID, type ModelID, type LLMRequest } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("groq")
+
+export type ProviderOptions = Pick<OpenAIProviderOptionsInput, "reasoningEffort"> & {
+  /** Controls visible reasoning on GPT-OSS; other models always use parsed reasoning. */
+  readonly includeReasoning?: boolean
+  readonly parallelToolCalls?: boolean
+  readonly serviceTier?: "on_demand" | "flex" | "auto" | "performance" | (string & {})
+  readonly user?: string
+}
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: ProviderOptions
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: ProviderOptions
+}
+
+const Options = Schema.Struct({
+  includeReasoning: Schema.optional(Schema.Boolean),
+  parallelToolCalls: Schema.optional(Schema.Boolean),
+  serviceTier: Schema.optional(Schema.String),
+  user: Schema.optional(Schema.String),
+})
+
+export const protocol = Protocol.make({
+  id: "groq-chat",
+  body: {
+    schema: Schema.Struct({
+      ...OpenAIChat.bodyFields,
+      reasoning_format: Schema.optional(Schema.Literal("parsed")),
+      include_reasoning: Schema.optional(Schema.Boolean),
+      parallel_tool_calls: Schema.optional(Schema.Boolean),
+      service_tier: Schema.optional(Schema.String),
+      user: Schema.optional(Schema.String),
+    }),
+    from: Effect.fn("Groq.fromRequest")(function* (request: LLMRequest) {
+      const options = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(
+        request.providerOptions ?? {},
+      )
+      const gptOSS = request.model.id.startsWith("openai/gpt-oss-")
+      return {
+        ...(yield* OpenAIChat.fromRequest(request)),
+        reasoning_format: gptOSS ? undefined : ("parsed" as const),
+        include_reasoning: gptOSS ? options.includeReasoning : undefined,
+        parallel_tool_calls: options.parallelToolCalls,
+        service_tier: options.serviceTier,
+        user: options.user,
+      }
+    }),
+  },
+  stream: OpenAIChat.protocol.stream,
+})
+
+export const route = Route.make({
+  id: "groq-chat",
+  provider: id,
+  providerMetadataKey: "openai",
+  protocol,
+  endpoint: Endpoint.path("/chat/completions", { baseURL: profiles.groq.baseURL }),
+  framing: OpenAIChat.framing,
+})
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.groq.baseURL },
+    auth: AuthOptions.bearer(input, "GROQ_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<ProviderOptions>({
+        id: modelID,
+        compatibility: {
+          maxTokensField: "max_completion_tokens",
+          reasoningField: "reasoning",
+          requireReasoning: false,
+          supportsStore: false,
+          supportsStrictMode: false,
+        },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, ProviderOptions>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
+
+export * as Groq from "./groq.js"
@@ -6,11 +6,13 @@ export * as Azure from "./azure.js"
 export * as Cerebras from "./cerebras.js"
 export * as Cloudflare from "./cloudflare.js"
 export { CloudflareAIGateway, CloudflareWorkersAI } from "./cloudflare.js"
+export * as DeepInfra from "./deepinfra.js"
 export * as Google from "./google.js"
 export * as GoogleVertex from "./google-vertex.js"
 export * as GoogleVertexChat from "./google-vertex-chat.js"
 export * as GoogleVertexMessages from "./google-vertex-messages.js"
 export * as GoogleVertexResponses from "./google-vertex-responses.js"
+export * as Groq from "./groq.js"
 export * as OpenAI from "./openai.js"
 export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
@@ -1,7 +1,6 @@
 import { Effect, Schema } from "effect"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import { ProviderID, type CacheHint, type ModelID } from "../schema/index.js"
@@ -164,9 +163,10 @@ const bodyOptions = (input: unknown) => {
 export const route = Route.make({
  id: ADAPTER,
  provider: profile.provider,
+  providerMetadataKey: "openrouter",
  protocol,
  endpoint: Endpoint.path("/chat/completions", { baseURL: profile.baseURL }),
-  framing: Framing.sse,
+  framing: OpenAIChat.framing,
 })

 export const routes = [route]
@@ -1,6 +1,6 @@
 import { Config, Effect, Redacted } from "effect"
 import { Headers } from "effect/unstable/http"
-import { AuthenticationReason, InvalidRequestReason, AIError, type HttpOptions } from "../schema/index.js"
+import { AuthenticationError, InvalidRequestError, AIError, type HttpOptions } from "../schema/index.js"

 export class MissingCredentialError extends Error {
  readonly _tag = "MissingCredentialError"
@@ -137,12 +137,10 @@ export function bearerHeader(name: string, source?: Secret | Credential) {
 const toAIError = (error: AuthError): AIError => {
  if (error instanceof MissingCredentialError || error instanceof Config.ConfigError) {
    return new AIError({
-      module: "Auth",
-      method: "apply",
      reason:
        error instanceof MissingCredentialError
-          ? new AuthenticationReason({ message: error.message, kind: "missing" })
-          : new InvalidRequestReason({ message: `Failed to resolve auth config: ${error.message}` }),
+          ? new AuthenticationError({ message: error.message, cause: error })
+          : new InvalidRequestError({ message: `Failed to resolve auth config: ${error.message}`, cause: error }),
    })
  }
  return error
@@ -12,13 +12,14 @@ import * as ProviderShared from "../protocols/shared.js"
 import type { ProtocolID, ProviderOptions } from "../schema/index.js"
 import {
  AIError,
+  AIErrorReason,
  GenerationOptions,
  HttpOptions,
  LLMRequest,
  LLMResponse,
  LanguageModel,
  LLMEvent,
-  InvalidProviderOutputReason,
+  InvalidProviderOutputError,
  ProviderID,
  mergeGenerationOptions,
  mergeHttpOptions,
@@ -89,6 +90,7 @@ export interface RouteDefaultsInput {
 export interface RoutePatch<Body, Prepared> extends RouteDefaultsInput {
  readonly id?: string
  readonly provider?: string | ProviderID
+  readonly providerMetadataKey?: string
  readonly auth?: Auth.Definition
  readonly transport?: Transport<Body, Prepared, unknown>
  readonly endpoint?: EndpointPatch<Body>
@@ -226,16 +228,14 @@ export interface MakeTransportInput<Body, Prepared, Frame, Event, State> {
 const streamError = (route: string, message: string, cause: Cause.Cause<unknown>) => {
  const failed = cause.reasons.find(Cause.isFailReason)?.error
  if (failed instanceof AIError) return failed
-  return ProviderShared.eventError(route, message, Cause.pretty(cause))
+  return ProviderShared.eventError(route, message, undefined, cause)
 }

 const incompleteStreamError = (route: string) =>
  new AIError({
-    module: "LLMClient",
-    method: "stream",
-    reason: new InvalidProviderOutputReason({
-      classification: "incomplete-stream",
+    reason: new InvalidProviderOutputError({
      message: "The provider response ended unexpectedly.",
+      classification: "incomplete-stream",
      route,
    }),
  })
@@ -264,11 +264,12 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
  const decodeEventEffect = Schema.decodeUnknownEffect(protocol.stream.event)
  const decodeEvent = (route: string) => (frame: Frame) =>
    decodeEventEffect(frame).pipe(
-      Effect.mapError(() =>
+      Effect.mapError((cause) =>
        ProviderShared.eventError(
          input.id,
          `Invalid ${route} stream event`,
          typeof frame === "string" ? frame : ProviderShared.encodeJson(frame),
+          cause,
        ),
      ),
    )
@@ -289,11 +290,16 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
      defaults: routeInput.defaults ?? {},
      body: protocol.body,
      with: (patch: RoutePatch<Body, Prepared>) => {
-        const { id, provider, auth, transport, endpoint, ...defaults } = patch
+        const { id, provider, providerMetadataKey, auth, transport, endpoint, ...defaults } = patch
        return build({
          ...routeInput,
          id: id ?? routeInput.id,
          provider: provider ?? routeInput.provider,
+          providerMetadataKey:
+            providerMetadataKey ??
+            (provider !== undefined && String(provider) !== String(routeInput.provider)
+              ? String(provider)
+              : routeInput.providerMetadataKey),
          auth: auth ?? routeInput.auth,
          endpoint: endpoint ? Endpoint.merge(routeInput.endpoint, endpoint) : routeInput.endpoint,
          transport: (transport as Transport<Body, Prepared, Frame> | undefined) ?? routeInput.transport,
@@ -318,19 +324,48 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
        return Stream.unwrap(
          routeInput.transport.execute(prepared, request, runtime, options).pipe(
            Effect.map((execution) => {
+              const terminal = protocol.stream.terminal
+              // Preserve assembled inputs; replace only serialized event fallbacks with their original wire data.
+              const frameError =
+                (frame: Frame, event: Frame | Event = frame) =>
+                (error: AIError) =>
+                  new AIError({
+                    reason: AIErrorReason.make({
+                      ...error.reason,
+                      message: error.reason.message,
+                      cause: error.reason.cause,
+                      body:
+                        error.reason.body !== undefined && error.reason.body !== ProviderShared.encodeJson(event)
+                          ? error.reason.body
+                          : (execution.body?.(frame) ??
+                            (typeof frame === "string" ? frame : ProviderShared.encodeJson(frame))),
+                    }),
+                  })
              const events = execution.frames.pipe(
-                Stream.mapEffect(decodeEvent(route)),
-                protocol.stream.terminal ? Stream.takeUntil(protocol.stream.terminal) : (stream) => stream,
+                Stream.mapEffect((frame) =>
+                  decodeEvent(route)(frame).pipe(
+                    Effect.catchCause((cause) =>
+                      Effect.fail(streamError(route, `Failed to decode ${route} event`, cause)),
+                    ),
+                    Effect.map((event) => ({ event, frame })),
+                    Effect.mapError(frameError(frame)),
+                  ),
+                ),
+                terminal ? Stream.takeUntil(({ event }) => terminal(event)) : (stream) => stream,
              )
              const stream = Stream.suspend(() => {
                let state = protocol.stream.initial(request)
                const parsed = events.pipe(
-                  Stream.mapEffect((event) =>
+                  Stream.mapEffect(({ event, frame }) =>
                    protocol.stream.step(state, event).pipe(
+                      Effect.catchCause((cause) =>
+                        Effect.fail(streamError(route, `Failed to parse ${route} event`, cause)),
+                      ),
                      Effect.map(([next, output]) => {
                        state = next
                        return output
                      }),
+                      Effect.mapError(frameError(frame, event)),
                    ),
                  ),
                  Stream.flatMap(Stream.fromIterable),
@@ -339,15 +374,24 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
                return onHalt
                  ? parsed.pipe(
                      Stream.concat(
-                        Stream.suspend(() =>
-                          Stream.unwrap(onHalt(state).pipe(Effect.map(Stream.fromIterable))),
-                        ),
+                        Stream.suspend(() => Stream.unwrap(onHalt(state).pipe(Effect.map(Stream.fromIterable)))),
                      ),
                    )
                  : parsed
              }).pipe(
                Stream.catchCause((cause) => Stream.fail(streamError(route, `Failed to read ${route} stream`, cause))),
                requireTerminalEvent(route),
+                Stream.mapError(
+                  (error) =>
+                    new AIError({
+                      reason: AIErrorReason.make({
+                        ...error.reason,
+                        message: error.reason.message,
+                        cause: error.reason.cause,
+                        http: error.reason.http ?? execution.http,
+                      }),
+                    }),
+                ),
              )
              return execution.complete ? stream.pipe(Stream.onEnd(execution.complete)) : stream
            }),
@@ -7,14 +7,7 @@ import {
  HttpClientRequest,
  HttpClientResponse,
 } from "effect/unstable/http"
-import {
-  HttpContext,
-  HttpRateLimitDetails,
-  HttpRequestDetails,
-  HttpResponseDetails,
-  AIError,
-  TransportReason,
-} from "../schema/index.js"
+import { HttpContext, HttpRateLimitDetails, AIError, TransportError } from "../schema/index.js"
 import { classifyProviderFailure } from "../provider-error.js"

 export interface Interface {
@@ -40,17 +33,6 @@ const headerDetails = (headers: Headers.Headers) =>
 const normalizedHeaders = (headers: Headers.Headers) =>
  Object.fromEntries(Object.entries(headers).map(([key, value]) => [key.toLowerCase(), value]))

-const requestId = (headers: Record<string, string>) => {
-  return (
-    headers["x-request-id"] ??
-    headers["request-id"] ??
-    headers["x-amzn-requestid"] ??
-    headers["x-amz-request-id"] ??
-    headers["x-goog-request-id"] ??
-    headers["cf-ray"]
-  )
-}
-
 const retryAfterMs = (headers: Record<string, string>) => {
  const millis = Number(headers["retry-after-ms"])
  if (Number.isFinite(millis)) return Math.max(0, millis)
@@ -108,24 +90,13 @@ const rateLimitDetails = (headers: Record<string, string>, retryAfter: number |
  })
 }

-const requestDetails = (request: HttpClientRequest.HttpClientRequest) =>
-  new HttpRequestDetails({
-    method: request.method,
-    url: request.url,
-    headers: headerDetails(request.headers),
-  })
-
-const responseDetails = (response: HttpClientResponse.HttpClientResponse) =>
-  new HttpResponseDetails({
+export const responseHttp = (response: HttpClientResponse.HttpClientResponse) =>
+  new HttpContext({
+    url: response.request.url,
    status: response.status,
    headers: headerDetails(response.headers),
  })

-const responseBody = (body: string | void) => {
-  if (body === undefined) return {}
-  return { body }
-}
-
 const decodeProviderBody = Schema.decodeUnknownOption(
  Schema.fromJsonString(
    Schema.Struct({
@@ -143,81 +114,46 @@ const providerMessage = (status: number, body: string | void) => {
  )
 }

-const responseHttp = (input: {
-  readonly request: HttpClientRequest.HttpClientRequest
-  readonly response: HttpClientResponse.HttpClientResponse
-  readonly body: ReturnType<typeof responseBody>
-  readonly requestId?: string | undefined
-  readonly rateLimit?: HttpRateLimitDetails | undefined
-}) =>
-  new HttpContext({
-    request: requestDetails(input.request),
-    response: responseDetails(input.response),
-    ...input.body,
-    requestId: input.requestId,
-    rateLimit: input.rateLimit,
+const statusError = (response: HttpClientResponse.HttpClientResponse) =>
+  Effect.gen(function* () {
+    if (response.status < 400) return response
+    const result = yield* response.text.pipe(Effect.result)
+    return yield* httpFailure({
+      message: providerMessage(response.status, result._tag === "Success" ? result.success : undefined),
+      url: response.request.url,
+      status: response.status,
+      responseHeaders: headerDetails(response.headers),
+      responseBody: result._tag === "Success" ? result.success : undefined,
+      cause: result._tag === "Failure" ? (result.failure.cause ?? result.failure) : undefined,
+    })
  })

-const statusError =
-  (request: HttpClientRequest.HttpClientRequest) => (response: HttpClientResponse.HttpClientResponse) =>
-    Effect.gen(function* () {
-      if (response.status < 400) return response
-      const body = yield* response.text.pipe(Effect.catch(() => Effect.void))
-      const headers = normalizedHeaders(response.headers)
-      const retryAfter = retryAfterMs(headers)
-      const rateLimit = rateLimitDetails(headers, retryAfter)
-      const details = responseBody(body)
-      return yield* new AIError({
-        module: "RequestExecutor",
-        method: "execute",
-        reason: classifyProviderFailure({
-          status: response.status,
-          message: providerMessage(response.status, body),
-          retryAfterMs: retryAfter,
-          rateLimit,
-          http: responseHttp({
-            request,
-            response,
-            body: details,
-            requestId: requestId(headers),
-            rateLimit,
-          }),
-        }),
-      })
-    })
-
-// Classifies an HTTP failure captured outside the executor (for example by the
-// AI SDK's own fetch) onto the same reason types and HttpContext that
-// executor-driven requests produce. The originating request is not available on
-// that path, so the method is assumed (language model calls are always POST),
-// request headers are empty.
-export const classifyHttpFailure = (input: {
+/** Preserve HTTP diagnostics for executor and externally captured failures alike. */
+export const httpFailure = (input: {
  readonly message: string
-  readonly url: string
+  readonly url?: string | undefined
  readonly status?: number | undefined
-  readonly code?: string | undefined
+  readonly data?: unknown
  readonly responseHeaders?: Record<string, string> | undefined
  readonly responseBody?: string | undefined
+  readonly cause?: unknown
 }) => {
  const headers = normalizedHeaders(Headers.fromInput(input.responseHeaders))
  const retryAfter = retryAfterMs(headers)
  const rateLimit = rateLimitDetails(headers, retryAfter)
-  const details = responseBody(input.responseBody)
-  return classifyProviderFailure({
-    message: input.message,
-    status: input.status,
-    code: input.code,
-    retryAfterMs: retryAfter,
-    rateLimit,
-    http: new HttpContext({
-      request: new HttpRequestDetails({ method: "POST", url: input.url, headers: {} }),
-      response:
-        input.status === undefined
-          ? undefined
-          : new HttpResponseDetails({ status: input.status, headers: headerDetails(Headers.fromInput(headers)) }),
-      ...details,
-      requestId: requestId(headers),
+  return new AIError({
+    reason: classifyProviderFailure({
+      message: input.message,
+      status: input.status,
+      data: input.data,
+      rawBody: input.responseBody,
+      retryAfterMs: retryAfter,
      rateLimit,
+      cause: input.cause,
+      http:
+        input.status === undefined || input.url === undefined
+          ? undefined
+          : new HttpContext({ url: input.url, status: input.status, headers }),
    }),
  })
 }
@@ -244,25 +180,25 @@ const httpError = (input: {
  readonly error: unknown
  readonly request: HttpClientRequest.HttpClientRequest
  readonly operation: HttpOperation
+  readonly http?: HttpContext
 }) => {
  const request = HttpClientError.isHttpClientError(input.error) ? input.error.request : input.request
  const transportError = (failure: { readonly message: string; readonly code?: string | undefined }) =>
    new AIError({
-      module: "RequestExecutor",
-      method: input.operation,
-      reason: new TransportReason({
+      reason: new TransportError({
        message: failure.message,
+        cause: source,
+        http: input.http,
        transport: "http",
        operation: input.operation,
        code: failure.code,
        url: request.url,
-        http: new HttpContext({ request: requestDetails(request) }),
      }),
    })

  const source =
    HttpClientError.isHttpClientError(input.error) && "cause" in input.error.reason
-      ? input.error.reason.cause
+      ? (input.error.reason.cause ?? input.error)
      : input.error
  const native = nativeTransportFailure(source)
  const code = native?.code
@@ -286,6 +222,13 @@ const httpError = (input: {
  })
 }

+export const responseStream = (response: HttpClientResponse.HttpClientResponse): Stream.Stream<Uint8Array, AIError> =>
+  response.stream.pipe(
+    Stream.mapError((error) =>
+      httpError({ error, request: response.request, operation: "read", http: responseHttp(response) }),
+    ),
+  )
+
 export const stream = (
  executor: Interface,
  request: HttpClientRequest.HttpClientRequest,
@@ -294,9 +237,7 @@ export const stream = (
  Stream.unwrap(
    Effect.gen(function* () {
      const response = yield* executor.execute(request, middleware)
-      return response.stream.pipe(
-        Stream.mapError((error) => httpError({ error, request: response.request, operation: "read" })),
-      )
+      return responseStream(response)
    }),
  )

@@ -309,7 +250,7 @@ export const layer: Layer.Layer<Service, never, HttpClient.HttpClient> = Layer.e
        if (!middleware)
          return yield* http.execute(request).pipe(
            Effect.mapError((error) => httpError({ error, request, operation: "request" })),
-            Effect.flatMap(statusError(request)),
+            Effect.flatMap(statusError),
          )

        const response = yield* middleware(request, (input) =>
@@ -317,7 +258,7 @@ export const layer: Layer.Layer<Service, never, HttpClient.HttpClient> = Layer.e
            .execute(input)
            .pipe(Effect.mapError((cause) => (cause instanceof Error ? cause : new Error(String(cause))))),
        ).pipe(Effect.mapError((error) => httpError({ error, request, operation: "request" })))
-        return yield* statusError(response.request)(response)
+        return yield* statusError(response)
      })
    return Service.of({
      execute: executeOnce,
@@ -8,8 +8,8 @@ import type { AIError } from "../schema/index.js"
 * `Framing` is the byte-stream-shaped seam between transport and protocol:
 *
 * - SSE (`Framing.sse`) — UTF-8 decode the body, run the SSE channel decoder,
- *   drop empty / `[DONE]` keep-alives. Each emitted frame is the JSON `data:`
- *   payload of one event.
+ *   and emit the `data:` payload of each non-empty event. The default drops
+ *   `[DONE]`; protocols that use it as a terminal select `sseWithDone`.
 * - AWS event stream — length-prefixed binary frames with CRC checksums.
 *   Each emitted frame is one parsed binary event record.
 *
@@ -19,11 +19,19 @@ import type { AIError } from "../schema/index.js"
 export interface Definition<Frame> {
  readonly id: string
  readonly frame: (bytes: Stream.Stream<Uint8Array, AIError>) => Stream.Stream<Frame, AIError>
+  /** Original wire representation when framing transforms the provider payload. */
+  readonly body?: (frame: Frame) => string | undefined
 }

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

+/** Server-Sent Events framing that retains the conventional `[DONE]` sentinel. */
+export const sseWithDone: Definition<string> = {
+  id: "sse",
+  frame: (bytes) => ProviderShared.sseFraming(bytes, undefined, true),
+}
+
 /** SSE framing restricted to protocol-recognized event names. */
 export const sseEvents = (events: ReadonlySet<string>): Definition<string> => ({
  id: "sse",
@@ -88,8 +88,13 @@ export const httpJson = <Body, Frame>(input: HttpJsonInput<Body, Frame>): HttpJs
      }
    }),
  execute: (prepared, _request, runtime) =>
-    Effect.succeed({
-      frames: prepared.framing.frame(RequestExecutor.stream(runtime.http, prepared.request, prepared.middleware)),
+    Effect.gen(function* () {
+      const response = yield* runtime.http.execute(prepared.request, prepared.middleware)
+      return {
+        frames: prepared.framing.frame(RequestExecutor.responseStream(response)),
+        http: RequestExecutor.responseHttp(response),
+        body: prepared.framing.body,
+      }
    }),
 })

@@ -3,7 +3,7 @@ import { Endpoint } from "../endpoint.js"
 import { Auth } from "../auth.js"
 import type { HttpMiddleware, Interface as RequestExecutorInterface } from "../executor.js"
 import type { WebSocketChannelExecutor } from "./websocket-channel.js"
-import type { AIError, LLMRequest } from "../../schema/index.js"
+import type { AIError, HttpContext, LLMRequest } from "../../schema/index.js"

 export interface TransportRuntime {
  readonly http: RequestExecutorInterface
@@ -11,6 +11,8 @@ export interface TransportRuntime {

 export interface TransportExecution<Frame> {
  readonly frames: Stream.Stream<Frame, AIError>
+  readonly http?: HttpContext
+  body?(frame: Frame): string | undefined
  /** Optional successful-consumption acknowledgement. HTTP leaves this absent. */
  readonly complete?: Effect.Effect<void>
 }
@@ -1,6 +1,6 @@
 import type { Effect, Scope, Stream } from "effect"
 import type { Headers } from "effect/unstable/http"
-import type { AIError } from "../../schema/index.js"
+import type { AIError, HttpContext } from "../../schema/index.js"

 export interface WebSocketChannelExecutor {
  readonly execute: (
@@ -10,6 +10,7 @@ export interface WebSocketChannelExecutor {

 export interface WebSocketChannelExecution {
  readonly frames: Stream.Stream<string, AIError>
+  readonly http?: HttpContext
  /** Commits staged state after the decoded Route stream ends successfully. */
  readonly complete: Effect.Effect<void>
 }
@@ -1,7 +1,13 @@
 import { Cause, Effect, Queue, Stream } from "effect"
 import { Headers } from "effect/unstable/http"
 import { Socket } from "effect/unstable/socket"
-import { AIError, TransportReason, type TransportOperation } from "../../schema/index.js"
+import {
+  AIError,
+  AIErrorReason,
+  TransportError,
+  type HttpContext,
+  type TransportOperation,
+} from "../../schema/index.js"
 import * as HttpTransport from "./http.js"
 import type { Transport } from "./index.js"
 import type {
@@ -17,6 +23,7 @@ export interface WebSocketRequest {
 }

 export interface WebSocketConnection {
+  readonly http?: HttpContext
  readonly sendText: (message: string) => Effect.Effect<void, AIError>
  readonly messages: Stream.Stream<string | Uint8Array, AIError>
  readonly close: Effect.Effect<void, never>
@@ -33,21 +40,22 @@ type WebSocketConstructorWithHeaders = (

 const MAX_FRAME_BYTES = 16 * 1024 * 1024
 const transportError = (
-  method: string,
  message: string,
  input: {
    readonly operation: TransportOperation
    readonly url?: string
    readonly code?: string
-    readonly phase?: TransportReason["phase"]
-    readonly delivery?: TransportReason["delivery"]
+    readonly phase?: TransportError["phase"]
+    readonly delivery?: TransportError["delivery"]
+    readonly body?: string
+    readonly cause?: unknown
  },
 ) =>
  new AIError({
-    module: "WebSocketConnector",
-    method,
-    reason: new TransportReason({
+    reason: new TransportError({
      message,
+      body: input.body,
+      cause: input.cause,
      transport: "websocket",
      operation: input.operation,
      url: input.url,
@@ -59,22 +67,16 @@ const transportError = (

 const annotateTransportError = (
  error: AIError,
-  input: { readonly phase: TransportReason["phase"]; readonly delivery: TransportReason["delivery"] },
+  input: { readonly phase: TransportError["phase"]; readonly delivery: TransportError["delivery"] },
 ) =>
  error.reason._tag === "Transport"
    ? new AIError({
-        module: error.module,
-        method: error.method,
-        reason: new TransportReason({
+        reason: new TransportError({
+          ...error.reason,
          message: error.reason.message,
-          transport: error.reason.transport,
-          operation: error.reason.operation,
-          code: error.reason.code,
-          url: error.reason.url,
-          http: error.reason.http,
+          cause: error.reason.cause,
          phase: input.phase,
          delivery: input.delivery,
-          recovery: error.reason.recovery,
        }),
      })
    : error
@@ -95,7 +97,7 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
  if (ws.readyState === globalThis.WebSocket.OPEN) return Effect.void
  if (ws.readyState === globalThis.WebSocket.CLOSING || ws.readyState === globalThis.WebSocket.CLOSED) {
    return Effect.fail(
-      transportError("open", `WebSocket closed before opening (state ${ws.readyState})`, {
+      transportError(`WebSocket closed before opening (state ${ws.readyState})`, {
        url: input.url,
        operation: "request",
        code: "closed",
@@ -124,7 +126,8 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
      cleanup()
      resume(
        Effect.fail(
-          transportError("open", `Failed to open WebSocket: ${eventMessage(event)}`, {
+          transportError(`Failed to open WebSocket: ${eventMessage(event)}`, {
+            cause: "error" in event ? (event.error ?? event) : event,
            url: input.url,
            operation: "request",
            phase: "connect",
@@ -137,7 +140,9 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
      cleanup()
      resume(
        Effect.fail(
-          transportError("open", `WebSocket closed before opening with code ${event.code}`, {
+          transportError(`WebSocket closed before opening with code ${event.code}`, {
+            body: event.reason,
+            cause: event,
            url: input.url,
            operation: "request",
            code: String(event.code),
@@ -169,7 +174,8 @@ export const toWebSocketUrl = (value: string) =>
      throw new Error(`Unsupported WebSocket URL protocol ${url.protocol}`)
    },
    catch: (error) =>
-      transportError("prepare", error instanceof Error ? error.message : "Invalid WebSocket URL", {
+      transportError(error instanceof Error ? error.message : "Invalid WebSocket URL", {
+        cause: error,
        url: value,
        operation: "request",
        code: "invalid-url",
@@ -189,7 +195,8 @@ export const open = (input: WebSocketRequest) =>
          headers: input.headers,
        }),
      catch: (error) =>
-        transportError("open", error instanceof Error ? error.message : "Failed to construct WebSocket", {
+        transportError(error instanceof Error ? error.message : "Failed to construct WebSocket", {
+          cause: error,
          url: input.url,
          operation: "request",
          phase: "connect",
@@ -214,7 +221,8 @@ export const fromWebSocket = (
      Queue.failCauseUnsafe(
        messages,
        Cause.fail(
-          transportError("message", "WebSocket message exceeds the 16 MiB limit", {
+          transportError("WebSocket message exceeds the 16 MiB limit", {
+            body: typeof message === "string" ? message : new TextDecoder().decode(message),
            url: input.url,
            operation: "read",
            code: "message-too-large",
@@ -231,7 +239,8 @@ export const fromWebSocket = (
      Queue.failCauseUnsafe(
        messages,
        Cause.fail(
-          transportError("message", "WebSocket inbound queue overflow", {
+          transportError("WebSocket inbound queue overflow", {
+            body: typeof message === "string" ? message : new TextDecoder().decode(message),
            url: input.url,
            operation: "read",
            code: "queue-overflow",
@@ -248,7 +257,8 @@ export const fromWebSocket = (
      Queue.failCauseUnsafe(
        messages,
        Cause.fail(
-          transportError("message", "Unsupported WebSocket message payload", {
+          transportError("Unsupported WebSocket message payload", {
+            cause: event,
            url: input.url,
            operation: "read",
            code: "message",
@@ -261,7 +271,8 @@ export const fromWebSocket = (
      Queue.failCauseUnsafe(
        messages,
        Cause.fail(
-          transportError("message", `WebSocket error: ${eventMessage(event)}`, {
+          transportError(`WebSocket error: ${eventMessage(event)}`, {
+            cause: "error" in event ? (event.error ?? event) : event,
            url: input.url,
            operation: "read",
            code: "message",
@@ -274,7 +285,9 @@ export const fromWebSocket = (
      Queue.failCauseUnsafe(
        messages,
        Cause.fail(
-          transportError("message", `WebSocket closed with code ${event.code}`, {
+          transportError(`WebSocket closed with code ${event.code}`, {
+            body: event.reason,
+            cause: event,
            url: input.url,
            operation: "read",
            code: String(event.code),
@@ -298,7 +311,7 @@ export const fromWebSocket = (
        Effect.suspend(() => {
          if (ws.readyState !== globalThis.WebSocket.OPEN)
            return Effect.fail(
-              transportError("sendText", `WebSocket is not open (state ${ws.readyState})`, {
+              transportError(`WebSocket is not open (state ${ws.readyState})`, {
                url: input.url,
                operation: "write",
                phase: "send",
@@ -308,7 +321,8 @@ export const fromWebSocket = (
          return Effect.try({
            try: () => ws.send(message),
            catch: (error) =>
-              transportError("sendText", error instanceof Error ? error.message : "Failed to send WebSocket message", {
+              transportError(error instanceof Error ? error.message : "Failed to send WebSocket message", {
+                cause: error,
                url: input.url,
                operation: "write",
                phase: "send",
@@ -349,10 +363,23 @@ export const makeDirect = (connector: WebSocketConnector): WebSocketChannelExecu
        (connection) => connection.close,
      )
      const create = yield* exchange.driver.create(undefined)
-      yield* connection.sendText(create.message)
+      yield* connection.sendText(create.message).pipe(
+        Effect.mapError(
+          (error) =>
+            new AIError({
+              reason: AIErrorReason.make({
+                ...error.reason,
+                message: error.reason.message,
+                cause: error.reason.cause,
+                http: error.reason.http ?? connection.http,
+              }),
+            }),
+        ),
+      )
      const decoder = new TextDecoder()
      let observed = false
      return {
+        http: connection.http,
        frames: connection.messages.pipe(
          Stream.map((message) => {
            observed = true
@@ -364,9 +391,49 @@ export const makeDirect = (connector: WebSocketConnector): WebSocketChannelExecu
              delivery: observed ? "accepted" : "ambiguous",
            }),
          ),
-          Stream.mapEffect((frame) => exchange.driver.observe(create, frame)),
+          Stream.mapEffect((frame) =>
+            exchange.driver.observe(create, frame).pipe(
+              Effect.mapError(
+                (error) =>
+                  new AIError({
+                    reason: AIErrorReason.make({
+                      ...error.reason,
+                      message: error.reason.message,
+                      cause: error.reason.cause,
+                      body: frame,
+                    }),
+                  }),
+              ),
+              Effect.map((observation) =>
+                "error" in observation
+                  ? {
+                      ...observation,
+                      error: new AIError({
+                        reason: AIErrorReason.make({
+                          ...observation.error.reason,
+                          message: observation.error.reason.message,
+                          cause: observation.error.reason.cause,
+                          body: frame,
+                        }),
+                      }),
+                    }
+                  : observation,
+              ),
+            ),
+          ),
          Stream.takeUntil(observationTerminal),
          Stream.mapEffect(observationFrame),
+          Stream.mapError(
+            (error) =>
+              new AIError({
+                reason: AIErrorReason.make({
+                  ...error.reason,
+                  message: error.reason.message,
+                  cause: error.reason.cause,
+                  http: error.reason.http ?? connection.http,
+                }),
+              }),
+          ),
        ),
        complete: Effect.void,
      }
@@ -417,7 +484,7 @@ export const json = <Body, Message>(input: JsonInput<Body, Message>): JsonTransp
    const webSocket = options?.webSocket
    if (!webSocket) {
      return Effect.fail(
-        transportError("json", "WebSocket JSON transport requires StreamOptions.webSocket", {
+        transportError("WebSocket JSON transport requires StreamOptions.webSocket", {
          url: prepared.url,
          operation: "request",
          code: "unavailable",
@@ -435,7 +502,7 @@ export const json = <Body, Message>(input: JsonInput<Body, Message>): JsonTransp
      connect: { url: prepared.url, headers: prepared.headers },
      fallback: () =>
        Stream.fail(
-          transportError("fallback", "WebSocket JSON transport does not provide HTTP fallback", {
+          transportError("WebSocket JSON transport does not provide HTTP fallback", {
            url: prepared.url,
            operation: "request",
            code: "websocket",
@@ -1,19 +1,13 @@
 import { Schema } from "effect"
 import { Tool } from "@opencode-ai/schema/tool"
 import { ModelID, ProviderID, RouteID } from "./ids.js"
-import { ProviderMetadata } from "./messages.js"

 export const ProviderFailureClassification = Schema.Literals(["context-overflow", "payload-too-large"])
 export type ProviderFailureClassification = typeof ProviderFailureClassification.Type

-export class HttpRequestDetails extends Schema.Class<HttpRequestDetails>("AI.HttpRequestDetails")({
-  method: Schema.String,
+export class HttpContext extends Schema.Class<HttpContext>("AI.HttpContext")({
  url: Schema.String,
-  headers: Schema.Record(Schema.String, Schema.String),
-}) {}
-
-export class HttpResponseDetails extends Schema.Class<HttpResponseDetails>("AI.HttpResponseDetails")({
-  status: Schema.Number,
+  status: Schema.Int.check(Schema.isBetween({ minimum: 100, maximum: 599 })),
  headers: Schema.Record(Schema.String, Schema.String),
 }) {}

@@ -24,74 +18,58 @@ export class HttpRateLimitDetails extends Schema.Class<HttpRateLimitDetails>("AI
  reset: Schema.optional(Schema.Record(Schema.String, Schema.String)),
 }) {}

-export class HttpContext extends Schema.Class<HttpContext>("AI.HttpContext")({
-  request: HttpRequestDetails,
-  response: Schema.optional(HttpResponseDetails),
-  body: Schema.optional(Schema.String),
-  bodyTruncated: Schema.optional(Schema.Boolean),
-  requestId: Schema.optional(Schema.String),
-  rateLimit: Schema.optional(HttpRateLimitDetails),
-}) {}
-
-export class InvalidRequestReason extends Schema.Class<InvalidRequestReason>("AI.Error.InvalidRequest")({
-  _tag: Schema.tag("InvalidRequest"),
+const ReasonFields = {
  message: Schema.String,
-  parameter: Schema.optional(Schema.String),
-  classification: Schema.optional(ProviderFailureClassification),
-  providerMetadata: Schema.optional(ProviderMetadata),
+  // Preserve the complete original response or triggering event before decoding narrows it.
+  body: Schema.optional(Schema.String),
  http: Schema.optional(HttpContext),
-}) {}
+  cause: Schema.optional(Schema.Defect({ includeStack: true })),
+}

-export class NoRouteReason extends Schema.Class<NoRouteReason>("AI.Error.NoRoute")({
-  _tag: Schema.tag("NoRoute"),
+export class InvalidRequestError extends Schema.TaggedError<InvalidRequestError>("AI.Error.InvalidRequest")(
+  "InvalidRequest",
+  {
+    ...ReasonFields,
+    parameter: Schema.optional(Schema.String),
+    classification: Schema.optional(ProviderFailureClassification),
+  },
+) {}
+
+export class NoRouteError extends Schema.TaggedError<NoRouteError>("AI.Error.NoRoute")("NoRoute", {
+  ...ReasonFields,
  route: RouteID,
  provider: ProviderID,
  model: ModelID,
-}) {
-  get message() {
-    return `No AI route for ${this.provider}/${this.model} using ${this.route}`
-  }
-}
-
-export class AuthenticationReason extends Schema.Class<AuthenticationReason>("AI.Error.Authentication")({
-  _tag: Schema.tag("Authentication"),
-  message: Schema.String,
-  kind: Schema.Literals(["missing", "invalid", "expired", "insufficient-permissions", "unknown"]),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
 }) {}

-export class RateLimitReason extends Schema.Class<RateLimitReason>("AI.Error.RateLimit")({
-  _tag: Schema.tag("RateLimit"),
-  message: Schema.String,
+export class AuthenticationError extends Schema.TaggedError<AuthenticationError>("AI.Error.Authentication")(
+  "Authentication",
+  ReasonFields,
+) {}
+
+export class RateLimitError extends Schema.TaggedError<RateLimitError>("AI.Error.RateLimit")("RateLimit", {
+  ...ReasonFields,
  retryAfterMs: Schema.optional(Schema.Number),
  rateLimit: Schema.optional(HttpRateLimitDetails),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
 }) {}

-export class QuotaExceededReason extends Schema.Class<QuotaExceededReason>("AI.Error.QuotaExceeded")({
-  _tag: Schema.tag("QuotaExceeded"),
-  message: Schema.String,
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class QuotaExceededError extends Schema.TaggedError<QuotaExceededError>("AI.Error.QuotaExceeded")(
+  "QuotaExceeded",
+  ReasonFields,
+) {}

-export class ContentPolicyReason extends Schema.Class<ContentPolicyReason>("AI.Error.ContentPolicy")({
-  _tag: Schema.tag("ContentPolicy"),
-  message: Schema.String,
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class ContentPolicyError extends Schema.TaggedError<ContentPolicyError>("AI.Error.ContentPolicy")(
+  "ContentPolicy",
+  ReasonFields,
+) {}

-export class ProviderInternalReason extends Schema.Class<ProviderInternalReason>("AI.Error.ProviderInternal")({
-  _tag: Schema.tag("ProviderInternal"),
-  message: Schema.String,
-  status: Schema.optional(Schema.Number),
-  retryAfterMs: Schema.optional(Schema.Number),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class ProviderInternalError extends Schema.TaggedError<ProviderInternalError>("AI.Error.ProviderInternal")(
+  "ProviderInternal",
+  {
+    ...ReasonFields,
+    retryAfterMs: Schema.optional(Schema.Number),
+  },
+) {}

 export const TransportType = Schema.Literals(["http", "websocket"])
 export type TransportType = typeof TransportType.Type
@@ -99,14 +77,12 @@ export type TransportType = typeof TransportType.Type
 export const TransportOperation = Schema.Literals(["request", "read", "write"])
 export type TransportOperation = typeof TransportOperation.Type

-export class TransportReason extends Schema.Class<TransportReason>("AI.Error.Transport")({
-  _tag: Schema.tag("Transport"),
-  message: Schema.String,
+export class TransportError extends Schema.TaggedError<TransportError>("AI.Error.Transport")("Transport", {
+  ...ReasonFields,
  transport: TransportType,
  operation: TransportOperation,
  code: Schema.optional(Schema.String),
  url: Schema.optional(Schema.String),
-  http: Schema.optional(HttpContext),
  phase: Schema.optional(
    Schema.Literals(["prepare", "queue", "connect", "send", "receive", "decode", "complete", "fallback", "close"]),
  ),
@@ -116,51 +92,40 @@ export class TransportReason extends Schema.Class<TransportReason>("AI.Error.Tra
  ),
 }) {}

-export class InvalidProviderOutputReason extends Schema.Class<InvalidProviderOutputReason>(
+export class InvalidProviderOutputError extends Schema.TaggedError<InvalidProviderOutputError>(
  "AI.Error.InvalidProviderOutput",
-)({
-  _tag: Schema.tag("InvalidProviderOutput"),
-  message: Schema.String,
+)("InvalidProviderOutput", {
+  ...ReasonFields,
  classification: Schema.optional(Schema.Literals(["incomplete-stream"])),
  route: Schema.optional(Schema.String),
-  raw: Schema.optional(Schema.String),
-  providerMetadata: Schema.optional(ProviderMetadata),
 }) {}

-export class UnknownProviderReason extends Schema.Class<UnknownProviderReason>("AI.Error.UnknownProvider")({
-  _tag: Schema.tag("UnknownProvider"),
-  message: Schema.String,
-  status: Schema.optional(Schema.Number),
-  providerMetadata: Schema.optional(ProviderMetadata),
-  http: Schema.optional(HttpContext),
-}) {}
+export class UnknownProviderError extends Schema.TaggedError<UnknownProviderError>("AI.Error.UnknownProvider")(
+  "UnknownProvider",
+  ReasonFields,
+) {}

 export const AIErrorReason = Schema.Union([
-  InvalidRequestReason,
-  NoRouteReason,
-  AuthenticationReason,
-  RateLimitReason,
-  QuotaExceededReason,
-  ContentPolicyReason,
-  ProviderInternalReason,
-  TransportReason,
-  InvalidProviderOutputReason,
-  UnknownProviderReason,
+  InvalidRequestError,
+  NoRouteError,
+  AuthenticationError,
+  RateLimitError,
+  QuotaExceededError,
+  ContentPolicyError,
+  ProviderInternalError,
+  TransportError,
+  InvalidProviderOutputError,
+  UnknownProviderError,
 ]).pipe(Schema.toTaggedUnion("_tag"))
 export type AIErrorReason = Schema.Schema.Type<typeof AIErrorReason>

 export class AIError extends Schema.TaggedError<AIError>()("AI.Error", {
-  module: Schema.String,
-  method: Schema.String,
  reason: AIErrorReason,
-  // Raw provider payload as a string, so classified failures never lose the
-  // original error detail even when the pretty message is a summary.
-  body: Schema.optional(Schema.String),
 }) {
  override readonly cause = this.reason

-  override get message() {
-    return `${this.module}.${this.method}: ${this.reason.message}`
+  override get message(): string {
+    return this.reason.message
  }
 }

@@ -112,6 +112,8 @@ export type TextDelta = Schema.Schema.Type<typeof TextDelta>
 export const TextEnd = Schema.Struct({
  type: Schema.tag("text-end"),
  id: ContentBlockID,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text: Schema.optional(Schema.String),
  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.TextEnd" })
 export type TextEnd = Schema.Schema.Type<typeof TextEnd>
@@ -134,6 +136,8 @@ export type ReasoningDelta = Schema.Schema.Type<typeof ReasoningDelta>
 export const ReasoningEnd = Schema.Struct({
  type: Schema.tag("reasoning-end"),
  id: ContentBlockID,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text: Schema.optional(Schema.String),
  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.ReasoningEnd" })
 export type ReasoningEnd = Schema.Schema.Type<typeof ReasoningEnd>
@@ -328,17 +332,32 @@ export const LLMEvent = Object.assign(llmEventTagged, {
 })
 export type LLMEvent = Schema.Schema.Type<typeof llmEventTagged>

+/** Joins deltas per fragment, letting an authoritative end value replace that fragment's accumulated deltas. */
+const joinFragments = <Delta extends { id: string; text: string }, End extends { id: string; text?: string }>(
+  events: ReadonlyArray<LLMEvent>,
+  isDelta: (event: LLMEvent) => event is Extract<LLMEvent, Delta>,
+  isEnd: (event: LLMEvent) => event is Extract<LLMEvent, End>,
+) => {
+  const order: string[] = []
+  const parts = new Map<string, string>()
+  for (const event of events) {
+    if (isDelta(event)) {
+      if (!parts.has(event.id)) order.push(event.id)
+      parts.set(event.id, (parts.get(event.id) ?? "") + event.text)
+    }
+    if (isEnd(event) && event.text !== undefined) {
+      if (!parts.has(event.id)) order.push(event.id)
+      parts.set(event.id, event.text)
+    }
+  }
+  return order.map((id) => parts.get(id)).join("")
+}
+
 const responseText = (events: ReadonlyArray<LLMEvent>) =>
-  events
-    .filter(LLMEvent.is.textDelta)
-    .map((event) => event.text)
-    .join("")
+  joinFragments(events, LLMEvent.is.textDelta, LLMEvent.is.textEnd)

 const responseReasoning = (events: ReadonlyArray<LLMEvent>) =>
-  events
-    .filter(LLMEvent.is.reasoningDelta)
-    .map((event) => event.text)
-    .join("")
+  joinFragments(events, LLMEvent.is.reasoningDelta, LLMEvent.is.reasoningEnd)

 const responseUsage = (events: ReadonlyArray<LLMEvent>) =>
  events.reduce<Usage | undefined>(
@@ -445,10 +464,11 @@ const reduceTextDelta = (state: ResponseState, event: TextDelta): ResponseState
 const reduceTextEnd = (state: ResponseState, event: TextEnd): ResponseState => {
  const current = state.textParts[event.id]
  if (!current) return state
+  const text = event.text ?? current.text
  const providerMetadata = event.providerMetadata ?? current.providerMetadata
  return {
-    ...replaceContent(state, current.contentIndex, textContent(current.text, providerMetadata)),
-    textParts: { ...state.textParts, [event.id]: { ...current, providerMetadata } },
+    ...replaceContent(state, current.contentIndex, textContent(text, providerMetadata)),
+    textParts: { ...state.textParts, [event.id]: { ...current, text, providerMetadata } },
  }
 }

@@ -478,10 +498,11 @@ const reduceReasoningDelta = (state: ResponseState, event: ReasoningDelta): Resp
 const reduceReasoningEnd = (state: ResponseState, event: ReasoningEnd): ResponseState => {
  const current = state.reasoningParts[event.id]
  if (!current) return state
+  const text = event.text ?? current.text
  const providerMetadata = event.providerMetadata ?? current.providerMetadata
  return {
-    ...replaceContent(state, current.contentIndex, reasoningContent(current.text, providerMetadata)),
-    reasoningParts: { ...state.reasoningParts, [event.id]: { ...current, providerMetadata } },
+    ...replaceContent(state, current.contentIndex, reasoningContent(text, providerMetadata)),
+    reasoningParts: { ...state.reasoningParts, [event.id]: { ...current, text, providerMetadata } },
  }
 }

@@ -579,12 +600,12 @@ export class LLMResponse extends Schema.Class<LLMResponse>("LLM.Response")({
  usage: Schema.optional(Usage),
  finishReason: FinishReasonDetails,
 }) {
-  /** Concatenated assistant text assembled from streamed `text-delta` events. */
+  /** Concatenated assistant text; each fragment's `text-end` value replaces its accumulated deltas when present. */
  get text() {
    return responseText(this.events)
  }

-  /** Concatenated reasoning text assembled from streamed `reasoning-delta` events. */
+  /** Concatenated reasoning text; each fragment's `reasoning-end` value replaces its accumulated deltas when present. */
  get reasoning() {
    return responseReasoning(this.events)
  }
@@ -1,4 +1,4 @@
-import { describe, expect } from "bun:test"
+import { describe, expect, test } from "bun:test"
 import { Effect, Schema, Stream } from "effect"
 import { LLM, LLMRequest, LLMResponse } from "../src/index.js"
 import { Route, Endpoint, LLMClient, Protocol, type FramingDef } from "../src/route.js"
@@ -148,15 +148,13 @@ describe("llm route", () => {
    }),
  )

-  it.effect("builds models from configured routes", () =>
-    Effect.gen(function* () {
-      const configured = fake.with({ provider: "fake-provider", endpoint: { baseURL: "https://fake.local" } })
+  test("builds models from configured routes", () => {
+    const configured = fake.with({ provider: "fake-provider", endpoint: { baseURL: "https://fake.local" } })

-      expect(configured.model({ id: "fake-model" })).toMatchObject({
-        provider: "fake-provider",
-      })
-    }),
-  )
+    expect(configured.model({ id: "fake-model" })).toMatchObject({
+      provider: "fake-provider",
+    })
+  })

  it.effect("does not register duplicate route ids globally", () =>
    Effect.gen(function* () {
@@ -1,10 +1,10 @@
 import { describe, expect } from "bun:test"
-import { Deferred, Effect, Fiber, Layer, Ref, Stream } from "effect"
-import { Headers, HttpClient, HttpClientError, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
-import { LLM, AIError } from "../src/index.js"
+import { Deferred, Effect, Fiber, Ref, Stream } from "effect"
+import { Headers, HttpClientError, HttpClientRequest } from "effect/unstable/http"
+import { LLM, AIError, HttpContext, InvalidProviderOutputError, TransportError } from "../src/index.js"
 import { LLMClient, RequestExecutor, WebSocketTransport, type WebSocketChannelExecutor } from "../src/route.js"
-import * as OpenAIChat from "../src/protocols/openai-chat.js"
-import * as OpenAI from "../src/providers/openai.js"
+import { route } from "../src/protocols/openai-chat.js"
+import { configure } from "../src/providers/openai.js"
 import { dynamicResponse, fixedResponse, systemError } from "./lib/http.js"
 import { deltaChunk } from "./lib/openai-chunks.js"
 import { sseEvents, sseRaw } from "./lib/sse.js"
@@ -18,66 +18,86 @@ const secretRequest = HttpClientRequest.post("https://provider.test/v1/chat?api_
  HttpClientRequest.setHeaders(Headers.fromInput({ authorization: "Bearer header-secret-456" })),
 )

-const responsesLayer = (responses: ReadonlyArray<Response>) =>
-  RequestExecutor.layer.pipe(
-    Layer.provide(
-      Layer.unwrap(
-        Effect.gen(function* () {
-          const cursor = yield* Ref.make(0)
-          return Layer.succeed(
-            HttpClient.HttpClient,
-            HttpClient.make((request) =>
-              Effect.gen(function* () {
-                const index = yield* Ref.getAndUpdate(cursor, (value) => value + 1)
-                return HttpClientResponse.fromWeb(request, responses[index] ?? responses[responses.length - 1])
-              }),
-            ),
-          )
-        }),
-      ),
-    ),
-  )
-
-const countedResponsesLayer = (attempts: Ref.Ref<number>, responses: ReadonlyArray<Response>) =>
-  RequestExecutor.layer.pipe(
-    Layer.provide(
-      Layer.unwrap(
-        Effect.gen(function* () {
-          const cursor = yield* Ref.make(0)
-          return Layer.succeed(
-            HttpClient.HttpClient,
-            HttpClient.make((request) =>
-              Effect.gen(function* () {
-                yield* Ref.update(attempts, (value) => value + 1)
-                const index = yield* Ref.getAndUpdate(cursor, (value) => value + 1)
-                return HttpClientResponse.fromWeb(request, responses[index] ?? responses[responses.length - 1])
-              }),
-            ),
-          )
-        }),
-      ),
-    ),
-  )
-
 const expectAIError = (error: unknown) => {
  expect(error).toBeInstanceOf(AIError)
  if (!(error instanceof AIError)) throw new Error("expected AIError")
+  expect(error.reason).toBeInstanceOf(Error)
+  expect(error.cause).toBe(error.reason)
  return error
 }

-const errorHttp = (error: AIError) => ("http" in error.reason ? error.reason.http : undefined)
 const largeProviderMessage = `Upstream request failed: ${"validation failed; ".repeat(1_000)}`

 describe("RequestExecutor", () => {
+  it.effect("preserves externally captured HTTP errors without inventing response context", () =>
+    Effect.sync(() => {
+      const cause = new Error("upstream request failed")
+      const body = '{"error":{"message":"Rate limited","trace":"original"}}'
+      const error = RequestExecutor.httpFailure({
+        message: "Rate limited",
+        url: request.url,
+        status: 429,
+        responseHeaders: { "Retry-After": "2", "X-Request-ID": "req_external" },
+        responseBody: body,
+        cause,
+      })
+
+      expect(error.message).toBe("Rate limited")
+      expect(error.reason).toMatchObject({ _tag: "RateLimit", retryAfterMs: 2000 })
+      expect(error.reason.body).toBe(body)
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason.http).toEqual(
+        new HttpContext({
+          url: request.url,
+          status: 429,
+          headers: { "retry-after": "2", "x-request-id": "req_external" },
+        }),
+      )
+      expect(RequestExecutor.httpFailure({ message: "No response", url: request.url }).reason.http).toBeUndefined()
+      expect(RequestExecutor.httpFailure({ message: "No URL", status: 500 }).reason.http).toBeUndefined()
+    }),
+  )
+
+  it.effect("retains the original body-read failure on an HTTP status error", () =>
+    Effect.gen(function* () {
+      const cause = new Error("response body disconnected")
+      const error = yield* Effect.gen(function* () {
+        const executor = yield* RequestExecutor.Service
+        return yield* executor.execute(request).pipe(Effect.flip)
+      }).pipe(
+        Effect.provide(
+          fixedResponse(
+            new ReadableStream({
+              start(controller) {
+                controller.error(cause)
+              },
+            }),
+            {
+              status: 503,
+              headers: { "x-request-id": "req_failed_body" },
+            },
+          ),
+        ),
+      )
+
+      expect(error.reason._tag).toBe("ProviderInternal")
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason.body).toBeUndefined()
+      expect(error.reason.http).toMatchObject({ status: 503, headers: { "x-request-id": "req_failed_body" } })
+    }),
+  )
+
  it.effect("parses response body failures at the executor seam", () =>
    Effect.gen(function* () {
      const executor = yield* RequestExecutor.Service
      const error = yield* RequestExecutor.stream(executor, secretRequest).pipe(Stream.runDrain, Effect.flip)

      expectAIError(error)
+      expect(error.message).toBe("ECONNRESET: disconnected query-secret-123 header-secret-456")
+      expect(error.reason.http).toMatchObject({ status: 200, url: secretRequest.url })
+      expect(error.reason.cause).toMatchObject({ code: "ECONNRESET" })
      expect(error.reason).toMatchObject({
        _tag: "Transport",
-        message: "ECONNRESET: disconnected query-secret-123 header-secret-456",
        transport: "http",
        operation: "read",
        code: "ECONNRESET",
@@ -85,15 +105,14 @@ describe("RequestExecutor", () => {
      })
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response(
-            new ReadableStream({
-              start(controller) {
-                controller.error(systemError("ECONNRESET", "disconnected query-secret-123 header-secret-456"))
-              },
-            }),
-          ),
-        ]),
+        fixedResponse(
+          new ReadableStream({
+            start(controller) {
+              controller.error(systemError("ECONNRESET", "disconnected query-secret-123 header-secret-456"))
+            },
+          }),
+          {},
+        ),
      ),
    ),
  )
@@ -104,23 +123,23 @@ describe("RequestExecutor", () => {
      const error = yield* RequestExecutor.stream(executor, secretRequest).pipe(Stream.runDrain, Effect.flip)

      expectAIError(error)
+      expect(error.message).toBe("ECONNRESET: socket closed")
+      expect(error.reason.cause).toBeInstanceOf(TypeError)
      expect(error.reason).toMatchObject({
        _tag: "Transport",
-        message: "ECONNRESET: socket closed",
        operation: "read",
        code: "ECONNRESET",
      })
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response(
-            new ReadableStream({
-              pull(controller) {
-                controller.error(new TypeError("fetch failed", { cause: systemError("ECONNRESET", "socket closed") }))
-              },
-            }),
-          ),
-        ]),
+        fixedResponse(
+          new ReadableStream({
+            pull(controller) {
+              controller.error(new TypeError("fetch failed", { cause: systemError("ECONNRESET", "socket closed") }))
+            },
+          }),
+          {},
+        ),
      ),
    ),
  )
@@ -133,8 +152,10 @@ describe("RequestExecutor", () => {
        .pipe(Effect.flip)

      expectAIError(error)
-      expect(error.reason.message).toBe("plugin rejected request")
-    }).pipe(Effect.provide(responsesLayer([]))),
+      expect(error.message).toBe("plugin rejected request")
+      expect(error.reason.cause).toBeInstanceOf(Error)
+      expect(error.reason.http).toBeUndefined()
+    }).pipe(Effect.provide(dynamicResponse(() => Effect.die(new Error("unexpected HTTP request"))))),
  )

  it.effect("reports the request sent by middleware", () =>
@@ -152,16 +173,11 @@ describe("RequestExecutor", () => {
        .pipe(Effect.flip)

      expectAIError(error)
+      expect(error.message).toBe("ECONNRESET: proxy disconnected proxy-secret")
+      expect(error.reason.http).toBeUndefined()
      expect(error.reason).toMatchObject({
        _tag: "Transport",
-        message: "ECONNRESET: proxy disconnected proxy-secret",
        url: "https://proxy.test/v1/chat?api_key=proxy-secret",
-        http: {
-          request: {
-            url: "https://proxy.test/v1/chat?api_key=proxy-secret",
-            headers: { authorization: "Bearer proxy-secret" },
-          },
-        },
      })
    }).pipe(
      Effect.provide(
@@ -188,11 +204,9 @@ describe("RequestExecutor", () => {
      expect(error.reason).toMatchObject({ _tag: "InvalidRequest", classification: "context-overflow" })
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response('{"error":{"code":"context_length_exceeded","message":"prompt too long"}}', {
-            status: 400,
-          }),
-        ]),
+        fixedResponse('{"error":{"code":"context_length_exceeded","message":"prompt too long"}}', {
+          status: 400,
+        }),
      ),
    ),
  )
@@ -206,9 +220,29 @@ describe("RequestExecutor", () => {
      expect(error.reason).toMatchObject({
        _tag: "InvalidRequest",
        classification: "payload-too-large",
-        http: { response: { status: 413 } },
      })
-    }).pipe(Effect.provide(responsesLayer([new Response("request too large", { status: 413 })]))),
+      expect(error.reason.http?.status).toBe(413)
+    }).pipe(Effect.provide(fixedResponse("request too large", { status: 413 }))),
+  )
+
+  it.effect("classifies Anthropic request_too_large as context overflow", () =>
+    Effect.gen(function* () {
+      const executor = yield* RequestExecutor.Service
+      const error = yield* executor.execute(request).pipe(Effect.flip)
+
+      expectAIError(error)
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidRequest",
+        classification: "context-overflow",
+      })
+      expect(error.reason.http?.status).toBe(413)
+    }).pipe(
+      Effect.provide(
+        fixedResponse('{"error":{"type":"request_too_large","message":"Request exceeds the maximum size"}}', {
+          status: 413,
+        }),
+      ),
+    ),
  )

  it.effect("does not classify ordinary invalid requests as context overflow", () =>
@@ -219,8 +253,8 @@ describe("RequestExecutor", () => {
      expectAIError(error)
      expect(error.reason).toMatchObject({ _tag: "InvalidRequest" })
      expect("classification" in error.reason ? error.reason.classification : undefined).toBeUndefined()
-      expect(error.reason.message).toBe("Provider request failed with HTTP 400")
-    }).pipe(Effect.provide(responsesLayer([new Response("invalid parameter", { status: 400 })]))),
+      expect(error.message).toBe("Provider request failed with HTTP 400")
+    }).pipe(Effect.provide(fixedResponse("invalid parameter", { status: 400 }))),
  )

  it.effect("preserves structured provider messages from large error bodies", () =>
@@ -229,20 +263,18 @@ describe("RequestExecutor", () => {
      const error = yield* executor.execute(request).pipe(Effect.flip)

      expectAIError(error)
-      expect(error.reason).toMatchObject({ _tag: "InvalidRequest", message: largeProviderMessage })
-      expect(errorHttp(error)?.body).toContain(largeProviderMessage)
-      expect(errorHttp(error)?.bodyTruncated).toBeUndefined()
+      expect(error.reason).toMatchObject({ _tag: "InvalidRequest" })
+      expect(error.message).toBe(largeProviderMessage)
+      expect(error.reason.body).toContain(largeProviderMessage)
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response(
-            JSON.stringify({
-              model: "gpt-5.6-sol",
-              error: { type: "invalid_request", message: largeProviderMessage },
-            }),
-            { status: 400 },
-          ),
-        ]),
+        fixedResponse(
+          JSON.stringify({
+            model: "test-model",
+            error: { type: "invalid_request", message: largeProviderMessage },
+          }),
+          { status: 400 },
+        ),
      ),
    ),
  )
@@ -255,9 +287,9 @@ describe("RequestExecutor", () => {
      expectAIError(error)
      expect(error.reason).toMatchObject({
        _tag: "InvalidRequest",
-        message: "Provider request failed with HTTP 400",
      })
-    }).pipe(Effect.provide(responsesLayer([new Response('{"error":{"message":"  "}}', { status: 400 })]))),
+      expect(error.message).toBe("Provider request failed with HTTP 400")
+    }).pipe(Effect.provide(fixedResponse('{"error":{"message":"  "}}', { status: 400 }))),
  )

  it.effect("classifies provider rate limits hidden behind HTTP 400", () =>
@@ -269,7 +301,7 @@ describe("RequestExecutor", () => {

          expectAIError(error)
          expect(error.reason).toMatchObject({ _tag: "RateLimit" })
-        }).pipe(Effect.provide(responsesLayer([new Response(body, { status: 400 })])))
+        }).pipe(Effect.provide(fixedResponse(body, { status: 400 })))

      yield* classify("Request rate increased too quickly")
      yield* classify('{"type":"error","error":{"type":"too_many_requests"}}')
@@ -286,7 +318,7 @@ describe("RequestExecutor", () => {

          expectAIError(error)
          expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
-        }).pipe(Effect.provide(responsesLayer([new Response(body, { status: 400 })])))
+        }).pipe(Effect.provide(fixedResponse(body, { status: 400 })))

      yield* classify('{"code":"resource_exhausted"}')
      yield* classify('{"code":"service_unavailable"}')
@@ -299,38 +331,27 @@ describe("RequestExecutor", () => {
      const error = yield* executor.execute(request).pipe(Effect.flip)

      expectAIError(error)
-      expect(error).toMatchObject({
-        reason: {
-          _tag: "RateLimit",
-          retryAfterMs: 0,
-          rateLimit: { retryAfterMs: 0 },
-          http: {
-            requestId: "req_123",
-            request: {
-              method: "POST",
-              url: "https://provider.test/v1/chat?api_key=secret&key=secret&debug=1",
-              headers: { authorization: "Bearer secret", "x-safe": "visible" },
-            },
-            response: {
-              status: 429,
-              headers: {
-                "retry-after-ms": "0",
-                "x-request-id": "req_123",
-                "x-api-key": "secret",
-              },
-            },
+      expect(error.reason).toMatchObject({
+        _tag: "RateLimit",
+        retryAfterMs: 0,
+        rateLimit: { retryAfterMs: 0 },
+        http: {
+          url: "https://provider.test/v1/chat?api_key=secret&key=secret&debug=1",
+          status: 429,
+          headers: {
+            "retry-after-ms": "0",
+            "x-request-id": "req_123",
+            "x-api-key": "secret",
          },
        },
      })
-      expect(errorHttp(error)?.body).toBe("rate limited")
+      expect(error.reason.body).toBe("rate limited")
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response("rate limited", {
-            status: 429,
-            headers: { "retry-after-ms": "0", "x-request-id": "req_123", "x-api-key": "secret" },
-          }),
-        ]),
+        fixedResponse("rate limited", {
+          status: 429,
+          headers: { "retry-after-ms": "0", "x-request-id": "req_123", "x-api-key": "secret" },
+        }),
      ),
    ),
  )
@@ -341,10 +362,9 @@ describe("RequestExecutor", () => {
      const error = yield* executor.execute(request).pipe(Effect.flip)

      expectAIError(error)
-      expect(errorHttp(error)?.request.headers["x-safe"]).toBe("visible")
-      expect(errorHttp(error)?.response?.headers["x-safe"]).toBe("response-secret")
+      expect(error.reason.http?.headers["x-safe"]).toBe("response-secret")
    }).pipe(
-      Effect.provide(responsesLayer([new Response("bad", { status: 400, headers: { "x-safe": "response-secret" } })])),
+      Effect.provide(fixedResponse("bad", { status: 400, headers: { "x-safe": "response-secret" } })),
      Effect.provideService(Headers.CurrentRedactedNames, ["x-safe"]),
    ),
  )
@@ -364,20 +384,18 @@ describe("RequestExecutor", () => {
      })
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response("rate limited", {
-            status: 429,
-            headers: {
-              "retry-after-ms": "0",
-              "x-ratelimit-limit-requests": "500",
-              "x-ratelimit-limit-tokens": "30000",
-              "x-ratelimit-remaining-requests": "499",
-              "x-ratelimit-remaining-tokens": "29900",
-              "x-ratelimit-reset-requests": "1s",
-              "x-ratelimit-reset-tokens": "10s",
-            },
-          }),
-        ]),
+        fixedResponse("rate limited", {
+          status: 429,
+          headers: {
+            "retry-after-ms": "0",
+            "x-ratelimit-limit-requests": "500",
+            "x-ratelimit-limit-tokens": "30000",
+            "x-ratelimit-remaining-requests": "499",
+            "x-ratelimit-remaining-tokens": "29900",
+            "x-ratelimit-reset-requests": "1s",
+            "x-ratelimit-reset-tokens": "10s",
+          },
+        }),
      ),
    ),
  )
@@ -388,8 +406,8 @@ describe("RequestExecutor", () => {
      const error = yield* executor.execute(request).pipe(Effect.flip)

      expectAIError(error)
-      expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
-      expect(errorHttp(error)?.rateLimit).toEqual({
+      expect(error.reason).toMatchObject({ _tag: "RateLimit" })
+      expect(error.reason._tag === "RateLimit" ? error.reason.rateLimit : undefined).toEqual({
        retryAfterMs: 0,
        limit: { requests: "100", "input-tokens": "10000" },
        remaining: { requests: "12", "input-tokens": "9000" },
@@ -397,20 +415,18 @@ describe("RequestExecutor", () => {
      })
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response("overloaded", {
-            status: 529,
-            headers: {
-              "retry-after-ms": "0",
-              "anthropic-ratelimit-requests-limit": "100",
-              "anthropic-ratelimit-requests-remaining": "12",
-              "anthropic-ratelimit-requests-reset": "2026-05-06T12:00:00Z",
-              "anthropic-ratelimit-input-tokens-limit": "10000",
-              "anthropic-ratelimit-input-tokens-remaining": "9000",
-              "anthropic-ratelimit-input-tokens-reset": "2026-05-06T12:00:10Z",
-            },
-          }),
-        ]),
+        fixedResponse("rate limited", {
+          status: 429,
+          headers: {
+            "retry-after-ms": "0",
+            "anthropic-ratelimit-requests-limit": "100",
+            "anthropic-ratelimit-requests-remaining": "12",
+            "anthropic-ratelimit-requests-reset": "2026-05-06T12:00:00Z",
+            "anthropic-ratelimit-input-tokens-limit": "10000",
+            "anthropic-ratelimit-input-tokens-remaining": "9000",
+            "anthropic-ratelimit-input-tokens-reset": "2026-05-06T12:00:10Z",
+          },
+        }),
      ),
    ),
  )
@@ -423,15 +439,20 @@ describe("RequestExecutor", () => {
        return yield* executor.execute(request).pipe(Effect.flip)
      }).pipe(
        Effect.provide(
-          countedResponsesLayer(attempts, [
-            new Response("busy", { status: 503, headers: { "retry-after-ms": "0" } }),
-            new Response("ok", { status: 200 }),
-          ]),
+          dynamicResponse((input) =>
+            Effect.gen(function* () {
+              const attempt = yield* Ref.getAndUpdate(attempts, (value) => value + 1)
+              return attempt === 0
+                ? input.respond("busy", { status: 503, headers: { "retry-after-ms": "0" } })
+                : input.respond("ok", { status: 200 })
+            }),
+          ),
        ),
      )

      expectAIError(error)
-      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", status: 503 })
+      expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
+      expect(error.reason.http?.status).toBe(503)
      expect(yield* Ref.get(attempts)).toBe(1)
    }),
  )
@@ -444,15 +465,14 @@ describe("RequestExecutor", () => {
          const error = yield* executor.execute(request).pipe(Effect.flip)

          expectAIError(error)
-          expect(error.reason).toMatchObject({ _tag: "ProviderInternal", status })
+          expect(error.reason).toMatchObject({ _tag: "ProviderInternal" })
+          expect(error.reason.http?.status).toBe(status)
        }).pipe(
          Effect.provide(
-            responsesLayer([
-              new Response("provider failure", {
-                status,
-                headers: { "retry-after-ms": "0" },
-              }),
-            ]),
+            fixedResponse("provider failure", {
+              status,
+              headers: { "retry-after-ms": "0" },
+            }),
          ),
        )

@@ -463,21 +483,28 @@ describe("RequestExecutor", () => {

  it.effect("preserves large authentication error bodies", () =>
    Effect.gen(function* () {
-      const executor = yield* RequestExecutor.Service
-      const error = yield* executor.execute(request).pipe(Effect.flip)
+      const attempts = yield* Ref.make(0)
+      const error = yield* Effect.gen(function* () {
+        const executor = yield* RequestExecutor.Service
+        return yield* executor.execute(request).pipe(Effect.flip)
+      }).pipe(
+        Effect.provide(
+          dynamicResponse((input) =>
+            Effect.gen(function* () {
+              const attempt = yield* Ref.getAndUpdate(attempts, (value) => value + 1)
+              return attempt === 0
+                ? input.respond("x".repeat(20_000), { status: 401 })
+                : input.respond("should not retry", { status: 200 })
+            }),
+          ),
+        ),
+      )

      expectAIError(error)
      expect(error.reason).toMatchObject({ _tag: "Authentication" })
-      expect(errorHttp(error)?.bodyTruncated).toBeUndefined()
-      expect(errorHttp(error)?.body).toHaveLength(20_000)
-    }).pipe(
-      Effect.provide(
-        responsesLayer([
-          new Response("x".repeat(20_000), { status: 401 }),
-          new Response("should not retry", { status: 200 }),
-        ]),
-      ),
-    ),
+      expect(error.reason.body).toHaveLength(20_000)
+      expect(yield* Ref.get(attempts)).toBe(1)
+    }),
  )

  it.effect("preserves response body fields", () =>
@@ -486,16 +513,12 @@ describe("RequestExecutor", () => {
      const error = yield* executor.execute(request).pipe(Effect.flip)

      expectAIError(error)
-      expect(errorHttp(error)?.body).toBe(
-        '{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}',
-      )
+      expect(error.reason.body).toBe('{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}')
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response('{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}', {
-            status: 400,
-          }),
-        ]),
+        fixedResponse('{"error":{"message":"bad","key":"body-secret","detail":"api_key=query-secret"}}', {
+          status: 400,
+        }),
      ),
    ),
  )
@@ -506,12 +529,10 @@ describe("RequestExecutor", () => {
      const error = yield* executor.execute(secretRequest).pipe(Effect.flip)

      expectAIError(error)
-      expect(errorHttp(error)?.body).toBe("provider echoed query-secret-123 and authorization header-secret-456")
+      expect(error.reason.body).toBe("provider echoed query-secret-123 and authorization header-secret-456")
    }).pipe(
      Effect.provide(
-        responsesLayer([
-          new Response("provider echoed query-secret-123 and authorization header-secret-456", { status: 400 }),
-        ]),
+        fixedResponse("provider echoed query-secret-123 and authorization header-secret-456", { status: 400 }),
      ),
    ),
  )
@@ -519,9 +540,7 @@ describe("RequestExecutor", () => {
  it.effect("does not re-execute after a successful response reaches stream parsing", () =>
    Effect.gen(function* () {
      const attempts = yield* Ref.make(0)
-      const model = OpenAIChat.route
-        .with({ endpoint: { baseURL: "https://api.openai.test/v1" } })
-        .model({ id: "gpt-4o-mini" })
+      const model = route.with({ endpoint: { baseURL: "https://api.openai.test/v1" } }).model({ id: "gpt-4o-mini" })
      const error = yield* LLMClient.generate(LLM.request({ model, prompt: "Say hello." })).pipe(
        Effect.provide(
          dynamicResponse((input) =>
@@ -543,19 +562,170 @@ describe("RequestExecutor", () => {

      expectAIError(error)
      expect(error.reason).toMatchObject({ _tag: "InvalidProviderOutput" })
+      expect(error.reason.body).toBe("not-json")
+      expect(error.reason.cause).toBeDefined()
+      expect(error.reason.http).toMatchObject({ status: 200, headers: { "content-type": "text/event-stream" } })
      expect(yield* Ref.get(attempts)).toBe(1)
    }),
  )
 })

 describe("WebSocket channel execution", () => {
-  const model = OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses("gpt-4.1-mini")
+  const model = configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses("gpt-4.1-mini")
  const request = LLM.request({ model, prompt: "Say hello." })
  const frames = [
    JSON.stringify({ type: "response.output_text.delta", item_id: "msg_1", delta: "Hi" }),
    JSON.stringify({ type: "response.completed", response: { id: "resp_1" } }),
  ]

+  it.effect("preserves close reasons and native event causes without fabricated HTTP metadata", () =>
+    Effect.gen(function* () {
+      class TestSocket extends EventTarget {
+        readyState = globalThis.WebSocket.OPEN
+        send() {}
+        close() {}
+      }
+      const socket = new TestSocket()
+      const connection = yield* WebSocketTransport.fromWebSocket(
+        // oxlint-disable-next-line typescript-eslint/no-unsafe-type-assertion
+        socket as unknown as globalThis.WebSocket,
+        { url: "wss://provider.test/responses", headers: Headers.empty },
+      )
+      const event = new CloseEvent("close", { code: 1011, reason: "upstream trace: req_close" })
+      socket.dispatchEvent(event)
+      const error = yield* connection.messages.pipe(Stream.runDrain, Effect.flip)
+
+      expect(error.reason).toMatchObject({ _tag: "Transport", code: "1011", phase: "close" })
+      expect(error.message).toBe("WebSocket closed with code 1011")
+      expect(error.reason.body).toBe(event.reason)
+      expect(error.reason.cause).toBe(event)
+      expect(error.reason.http).toBeUndefined()
+      yield* connection.close
+    }),
+  )
+
+  it.effect("preserves opening event errors and native send exceptions", () =>
+    Effect.gen(function* () {
+      const cause = new Error("native send failed")
+      class TestSocket extends EventTarget {
+        readyState = globalThis.WebSocket.CONNECTING
+        send() {
+          throw cause
+        }
+        close() {}
+      }
+      const socket = new TestSocket()
+      const open = WebSocketTransport.fromWebSocket(
+        // oxlint-disable-next-line typescript-eslint/no-unsafe-type-assertion
+        socket as unknown as globalThis.WebSocket,
+        { url: "wss://provider.test/responses", headers: Headers.empty },
+      )
+      const fiber = yield* open.pipe(Effect.flip, Effect.forkChild({ startImmediately: true }))
+      const event = new ErrorEvent("error", { message: "handshake rejected", error: cause })
+      socket.dispatchEvent(event)
+      const error = yield* Fiber.join(fiber)
+      expect(error.reason.cause).toBe(cause)
+      expect(error.message).toContain("handshake rejected")
+      expect(error.reason.http).toBeUndefined()
+
+      socket.readyState = globalThis.WebSocket.OPEN
+      const connection = yield* open
+      const sent = yield* connection.sendText("create").pipe(Effect.flip)
+      expect(sent.reason.cause).toBe(cause)
+      expect(sent.message).toBe(cause.message)
+      yield* connection.close
+    }),
+  )
+
+  it.effect("preserves raw driver failures and known upgrade metadata", () =>
+    Effect.gen(function* () {
+      const cause = new Error("driver validation failed")
+      const frame = '{ "error": "failed", "trace": "original" }'
+      const http = new HttpContext({
+        url: "https://provider.test/responses",
+        status: 101,
+        headers: { upgrade: "websocket" },
+      })
+      const executor = WebSocketTransport.makeDirect({
+        open: () =>
+          Effect.succeed({
+            http,
+            sendText: () => Effect.void,
+            messages: Stream.make(frame),
+            close: Effect.void,
+          }),
+      })
+      const execution = yield* executor.execute({
+        id: "exchange_error",
+        connect: { url: "wss://provider.test/responses", headers: Headers.empty },
+        fallback: () => Stream.empty,
+        driver: {
+          create: () => Effect.succeed({ message: "create", mode: "full" }),
+          observe: () =>
+            Effect.succeed({
+              type: "provider-failure",
+              error: new AIError({
+                reason: new InvalidProviderOutputError({
+                  message: "Driver failed",
+                  cause,
+                  body: "narrowed",
+                }),
+              }),
+            }),
+        },
+      })
+      const error = yield* execution.frames.pipe(Stream.runDrain, Effect.flip)
+
+      expect(error.message).toBe("Driver failed")
+      expect(error.reason.body).toBe(frame)
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason.http).toBe(http)
+      expect(execution.http).toBe(http)
+    }),
+  )
+
+  it.effect("retains diagnostic fields when annotating transport delivery", () =>
+    Effect.gen(function* () {
+      const cause = new Error("connection closed")
+      const executor = WebSocketTransport.makeDirect({
+        open: () =>
+          Effect.succeed({
+            sendText: () => Effect.void,
+            messages: Stream.fail(
+              new AIError({
+                reason: new TransportError({
+                  message: "Socket closed",
+                  transport: "websocket",
+                  operation: "read",
+                  phase: "close",
+                  recovery: "retry-full",
+                  body: "server close detail",
+                  cause,
+                }),
+              }),
+            ),
+            close: Effect.void,
+          }),
+      })
+      const execution = yield* executor.execute({
+        id: "exchange_closed",
+        connect: { url: "wss://provider.test/responses", headers: Headers.empty },
+        fallback: () => Stream.empty,
+        driver: {
+          create: () => Effect.succeed({ message: "create", mode: "full" }),
+          observe: (_create, frame) => Effect.succeed({ type: "frame", frame }),
+        },
+      })
+      const error = yield* execution.frames.pipe(Stream.runDrain, Effect.flip)
+
+      expect(error.message).toBe("Socket closed")
+      expect(error.reason.body).toBe("server close detail")
+      expect(error.reason.cause).toBe(cause)
+      expect(error.reason).toMatchObject({ phase: "close", delivery: "ambiguous", recovery: "retry-full" })
+      expect(error.reason.http).toBeUndefined()
+    }),
+  )
+
  it.effect("runs a channel driver through the direct executor", () =>
    Effect.gen(function* () {
      const sent = yield* Ref.make("")
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/chat-streams-text",
    "recordedAt": "2026-08-23T17:21:53.198Z"
  },
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/responses-calls-a-tool",
    "recordedAt": "2026-08-23T17:21:55.170Z"
  },
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/responses-continues-after-a-tool-result",
    "recordedAt": "2026-08-23T17:21:56.397Z"
  },
@@ -1,10 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
    "name": "azure/responses-streams-text",
    "recordedAt": "2026-08-23T17:21:54.158Z"
  },
@@ -2,11 +2,7 @@
  "version": 1,
  "metadata": {
    "model": "openai.gpt-oss-120b",
-    "tags": [
-      "prefix:bedrock-mantle",
-      "provider:amazon-bedrock",
-      "protocol:openai-responses"
-    ],
+    "tags": ["prefix:bedrock-mantle", "provider:amazon-bedrock", "protocol:openai-responses"],
    "name": "bedrock-mantle/streams-text",
    "recordedAt": "2026-08-25T03:29:02.968Z"
  },
@@ -7,13 +7,7 @@
    "route": "cloudflare-workers-ai",
    "transport": "http",
    "model": "@cf/openai/gpt-oss-20b",
-    "tags": [
-      "prefix:cloudflare-workers-ai",
-      "provider:cloudflare-workers-ai",
-      "tool",
-      "tool-call",
-      "golden"
-    ]
+    "tags": ["prefix:cloudflare-workers-ai", "provider:cloudflare-workers-ai", "tool", "tool-call", "golden"]
  },
  "interactions": [
    {
@@ -35,4 +29,4 @@
      }
    }
  ]
-}
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "text", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-text",
+    "recordedAt": "2026-08-26T00:34:03.019Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":40,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Hello\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"!\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":3.46e-6,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-call", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-call",
+    "recordedAt": "2026-08-26T00:34:04.173Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_SMfBjXa8eCmHLyjfeyARxe3a\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":0.000027480000000000005,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,50 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-loop", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-loop",
+    "recordedAt": "2026-08-26T00:34:05.656Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":0.000028780000000000002,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}],\"reasoning_content\":\"\"},{\"role\":\"tool\",\"tool_call_id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"service_tier\":null,\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Paris\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" is\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" sunny\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\".\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":0.0000292,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -1,11 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
    "name": "google-vertex/calls-a-tool",
    "recordedAt": "2026-08-23T17:21:51.036Z"
  },
@@ -1,11 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
    "name": "google-vertex/continues-after-a-tool-result",
    "recordedAt": "2026-08-23T17:21:51.853Z"
  },
@@ -1,11 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
    "name": "google-vertex/streams-text",
    "recordedAt": "2026-08-23T17:21:50.112Z"
  },
@@ -0,0 +1,29 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "openai/gpt-oss-20b",
+    "tags": ["prefix:groq-chat", "provider:groq", "protocol:groq-chat", "text", "usage"],
+    "name": "groq-chat/streams-text-with-usage",
+    "recordedAt": "2026-08-26T14:40:09.833Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.groq.com/openai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"openai/gpt-oss-20b\",\"messages\":[{\"role\":\"user\",\"content\":\"Reply with exactly one word: hello\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"reasoning_effort\":\"low\",\"max_completion_tokens\":512,\"include_reasoning\":false,\"service_tier\":\"on_demand\",\"user\":\"recorded-test\"}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"logprobs\":null,\"finish_reason\":null}],\"x_groq\":{\"id\":\"req_01m0z87915eep9bpf10gg7331e\",\"seed\":94036161}}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"hello\"},\"logprobs\":null,\"finish_reason\":null}]}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"x_groq\":{\"id\":\"req_01m0z87915eep9bpf10gg7331e\",\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}}},\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}}}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[],\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}},\"service_tier\":\"on_demand\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -1,12 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:openai-responses-cache",
-      "provider:openai",
-      "protocol:openai-responses",
-      "cache"
-    ],
+    "tags": ["prefix:openai-responses-cache", "provider:openai", "protocol:openai-responses", "cache"],
    "name": "openai-responses-cache/reports-cached-tokens-on-identical-second-call",
    "recordedAt": "2026-08-25T03:29:25.124Z"
  },
@@ -5,14 +5,7 @@
    "route": "openai-responses",
    "transport": "http",
    "model": "gpt-5.5",
-    "tags": [
-      "prefix:openai-responses",
-      "provider:openai",
-      "flagship",
-      "tool",
-      "tool-loop",
-      "golden"
-    ],
+    "tags": ["prefix:openai-responses", "provider:openai", "flagship", "tool", "tool-loop", "golden"],
    "name": "openai-responses/openai-responses-gpt-5-5-tool-loop",
    "recordedAt": "2026-08-20T06:30:22.262Z"
  },
@@ -1,14 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:openai",
-      "protocol:openai-responses",
-      "tool",
-      "tool-result"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:openai", "protocol:openai-responses", "tool", "tool-result"],
    "name": "pdf/openai-tool-result",
    "recordedAt": "2026-08-25T03:29:08.297Z"
  },
@@ -1,13 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:openai",
-      "protocol:openai-responses",
-      "user-input"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:openai", "protocol:openai-responses", "user-input"],
    "name": "pdf/openai-user-input",
    "recordedAt": "2026-08-25T03:29:05.645Z"
  },
@@ -1,14 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:xai",
-      "protocol:xai-responses",
-      "tool",
-      "tool-result"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:xai", "protocol:xai-responses", "tool", "tool-result"],
    "name": "pdf/xai-tool-result",
    "recordedAt": "2026-08-25T03:29:11.774Z"
  },
@@ -1,13 +1,7 @@
 {
  "version": 1,
  "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:xai",
-      "protocol:xai-responses",
-      "user-input"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:xai", "protocol:xai-responses", "user-input"],
    "name": "pdf/xai-user-input",
    "recordedAt": "2026-08-25T03:29:10.612Z"
  },
@@ -52,4 +52,4 @@
      }
    }
  ]
-}
+}
@@ -2,12 +2,7 @@
  "version": 1,
  "metadata": {
    "model": "anthropic/claude-sonnet-4.6",
-    "tags": [
-      "prefix:openai-compatible-chat",
-      "provider:vercel-ai-gateway",
-      "protocol:openai-chat",
-      "reasoning"
-    ],
+    "tags": ["prefix:openai-compatible-chat", "provider:vercel-ai-gateway", "protocol:openai-chat", "reasoning"],
    "name": "vercel-ai-gateway-reasoning",
    "recordedAt": "2026-07-18T11:28:42.077Z"
  },
@@ -31,4 +26,4 @@
      }
    }
  ]
-}
+}
@@ -44,21 +44,23 @@ describe("Tool.make (dynamic JSON Schema)", () => {
    expect(definition?.inputSchema).toEqual(jsonSchema)
  })

-  test("execute receives the raw input untouched", async () => {
-    const seen: unknown[] = []
-    const tool = Tool.make({
-      description: "echo",
-      jsonSchema: { type: "object" },
-      execute: (params) =>
-        Effect.sync(() => {
-          seen.push(params)
-          return { ok: true }
-        }),
-    })
-    const result = await Effect.runPromise(tool.execute({ hello: "world" }))
-    expect(seen).toEqual([{ hello: "world" }])
-    expect(result).toEqual({ ok: true })
-  })
+  it.effect("execute receives the raw input untouched", () =>
+    Effect.gen(function* () {
+      const seen: unknown[] = []
+      const tool = Tool.make({
+        description: "echo",
+        jsonSchema: { type: "object" },
+        execute: (params) =>
+          Effect.sync(() => {
+            seen.push(params)
+            return { ok: true }
+          }),
+      })
+      const result = yield* tool.execute({ hello: "world" })
+      expect(seen).toEqual([{ hello: "world" }])
+      expect(result).toEqual({ ok: true })
+    }),
+  )
 })

 describe("LLM.generateObject", () => {
@@ -7,6 +7,37 @@ import { it } from "./lib/effect.js"
 import { dynamicResponse } from "./lib/http.js"

 describe("Image", () => {
+  for (const provider of [OpenAI, Google, XAI, ZAI]) {
+    const model = provider.configure({ apiKey: "test", baseURL: "https://image.test" }).image("image-model")
+    for (const body of ['{"data":42,"candidates":42,"opaque":{"nested":[1,2]},"trace":"outer"}', '{"invalid":']) {
+      it.effect(`retains ${model.provider} image response body and decode cause: ${body}`, () =>
+        Effect.gen(function* () {
+          const error = yield* Image.generate({ model, prompt: "hello" }).pipe(Effect.flip)
+          expect(error.reason._tag).toBe("InvalidProviderOutput")
+          expect(error.message).toContain("invalid response")
+          expect(error.reason.body).toBe(body)
+          expect(error.reason.http).toMatchObject({ status: 200, headers: { "x-image-trace": "trace-1" } })
+          expect(error.reason.http?.url).toStartWith("https://image.test/")
+          expect(error.reason.cause).toBeInstanceOf(Error)
+        }).pipe(
+          Effect.provide(
+            ImageClient.layer.pipe(
+              Layer.provide(
+                dynamicResponse((input) =>
+                  Effect.succeed(
+                    input.respond(body, {
+                      headers: { "content-type": "application/json", "x-image-trace": "trace-1" },
+                    }),
+                  ),
+                ),
+              ),
+            ),
+          ),
+        ),
+      )
+    }
+  }
+
  it.effect("generates images through the OpenAI Images API", () =>
    Effect.gen(function* () {
      const response = yield* Image.generate({
@@ -530,22 +561,17 @@ describe("Image", () => {
        Effect.sync(() => {
          expect(error.reason._tag).toBe("InvalidProviderOutput")
          if (error.reason._tag !== "InvalidProviderOutput") return
-          expect(error.reason.message).toContain("finish reasons: IMAGE_SAFETY")
-          expect(error.reason.providerMetadata).toEqual({
-            google: {
-              promptFeedback: { blockReason: "SAFETY" },
-              candidates: [
-                {
-                  index: 0,
-                  finishReason: "IMAGE_SAFETY",
-                  finishMessage: "The generated image was blocked by safety filters.",
-                  safetyRatings: [{ category: "HARM_CATEGORY_DANGEROUS_CONTENT", blocked: true }],
-                  citationMetadata: undefined,
-                  groundingMetadata: undefined,
-                  parts: [{ type: "text", text: "blocked", thought: false, thoughtSignature: undefined }],
-                },
-              ],
-            },
+          expect(error.message).toContain("finish reasons: IMAGE_SAFETY")
+          expect(JSON.parse(error.reason.body ?? "")).toEqual({
+            promptFeedback: { blockReason: "SAFETY" },
+            candidates: [
+              {
+                finishReason: "IMAGE_SAFETY",
+                finishMessage: "The generated image was blocked by safety filters.",
+                safetyRatings: [{ category: "HARM_CATEGORY_DANGEROUS_CONTENT", blocked: true }],
+                content: { parts: [{ text: "blocked", thought: false }] },
+              },
+            ],
          })
        }),
      ),
@@ -11,6 +11,7 @@ describe("provider error classification", () => {
      "Input length 131393 exceeds the maximum allowed input length of 131040 tokens.",
      "The input (516368 tokens) is longer than the model's context length (262144 tokens).",
      "Prompt has 5,958,968 tokens, but the configured context size is 256,000 tokens",
+      "Range of input length should be [1, 129024]",
      "Too many tokens",
      "Token limit exceeded",
    ]
@@ -18,13 +19,19 @@ describe("provider error classification", () => {
    expect(messages.every(isContextOverflow)).toBe(true)
  })

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

@@ -33,7 +40,6 @@ describe("provider error classification", () => {
        expect.objectContaining({ _tag: "InvalidRequest", classification: "payload-too-large" }),
      ),
    )
-    expect(isContextOverflow("413 status code (no body)")).toBe(false)
  })

  test("does not classify rate limits as context overflow", () => {
@@ -82,12 +88,12 @@ describe("provider error classification", () => {
    ])
  })

-  test("classifies network error text as provider internal", () => {
+  test("classifies any remaining 4xx status as an invalid request", () => {
    expect(
-      ["network error", "network-error", "network_error"].map(
-        (message) => classifyProviderFailure({ message })._tag,
+      [400, 402, 404, 418, 422, 451].map(
+        (status) => classifyProviderFailure({ message: `HTTP ${status}`, status })._tag,
      ),
-    ).toEqual(["ProviderInternal", "ProviderInternal", "ProviderInternal"])
+    ).toEqual(Array(6).fill("InvalidRequest"))
  })

  test("classifies nested provider codes when a top-level code is also present", () => {
@@ -100,14 +106,49 @@ describe("provider error classification", () => {
    ).toEqual(["QuotaExceeded", "ProviderInternal", "InvalidRequest"])
  })

-  test("keeps unknown and malformed provider payloads non-retryable", () => {
+  test("leaves unrecognized failures unclassified for the retry default", () => {
    expect(classifyProviderFailure({ message: '{"error":{"message":"no_kv_space"}}' })._tag).toBe("UnknownProvider")
    expect(classifyProviderFailure({ message: '{"type":"error","error":{"code":123}}' })._tag).toBe("UnknownProvider")
    expect(classifyProviderFailure({ message: "not-json" })._tag).toBe("UnknownProvider")
+    expect(classifyProviderFailure({ message: "network error" })._tag).toBe("UnknownProvider")
+    expect(classifyProviderFailure({ message: "Provider returned error" })._tag).toBe("UnknownProvider")
  })
 })

 describe("provider error rawBody classification", () => {
+  test("classifies provider envelopes without separate code inputs", () => {
+    const cases = [
+      ['{"type":"error","error":{"type":"overloaded_error","message":"Try again"}}', "ProviderInternal"],
+      ['{"error":{"code":"insufficient_quota","message":"Request failed"}}', "QuotaExceeded"],
+      [
+        '{"type":"response.failed","response":{"error":{"code":"authentication_error","message":"Denied"}}}',
+        "Authentication",
+      ],
+      ['{"error":{"code":429,"status":"RESOURCE_EXHAUSTED","message":"Try again"}}', "ProviderInternal"],
+      ['{"exception":{"type":"throttlingException","details":{"message":"Try again"}}}', "RateLimit"],
+    ] as const
+    for (const [rawBody, expected] of cases) {
+      const reason = classifyProviderFailure({ message: "Request failed", rawBody })
+      expect(reason._tag).toBe(expected)
+      expect(reason.body).toBe(rawBody)
+      expect(reason).not.toHaveProperty("code")
+    }
+  })
+
+  test("classifies separately supplied SDK data without replacing the response body", () => {
+    const data = { error: { code: "authentication_error" } }
+    for (const value of [data, JSON.stringify(data)]) {
+      const reason = classifyProviderFailure({
+        message: "Request failed",
+        status: 400,
+        rawBody: '{"message":"Request failed"}',
+        data: value,
+      })
+      expect(reason._tag).toBe("Authentication")
+      expect(reason.body).toBe('{"message":"Request failed"}')
+    }
+  })
+
  test("classifies overflow signals buried in the raw payload when the summary is vague", () => {
    const reason = classifyProviderFailure({
      message: "Request failed",
@@ -28,6 +28,8 @@ describe("provider package entrypoints", () => {
      import("@opencode-ai/ai/providers/amazon-bedrock/mantle/responses"),
      import("@opencode-ai/ai/providers/togetherai"),
      import("@opencode-ai/ai/providers/cerebras"),
+      import("@opencode-ai/ai/providers/deepinfra"),
+      import("@opencode-ai/ai/providers/groq"),
    ])

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

+  test("maps DeepInfra package settings onto its native executable model", async () => {
+    const DeepInfra = await import("@opencode-ai/ai/providers/deepinfra")
+    const settings = {
+      apiKey: "fixture",
+      baseURL: "https://provider.example.test/v1/",
+      headers: { "x-application": "opencode" },
+      body: { service_tier: "priority" },
+      providerOptions: { reasoningEffort: "high" as const },
+    }
+    const deepinfra = DeepInfra.model("google/gemma-3-27b-it", settings)
+
+    expect(deepinfra.route.id).toBe("deepinfra-chat")
+    expect(deepinfra.route.endpoint.baseURL).toBe("https://provider.example.test/v1/openai")
+    expect(deepinfra.route.defaults.providerOptions).toEqual(settings.providerOptions)
+    expect(deepinfra.route.defaults.headers).toEqual(settings.headers)
+    expect(deepinfra.route.defaults.http?.body).toEqual(settings.body)
+  })
+
  test("maps OpenRouter and xAI package settings onto executable models", async () => {
    const OpenRouter = await import("@opencode-ai/ai/providers/openrouter")
    const XAI = await import("@opencode-ai/ai/providers/xai")
@@ -39,7 +39,7 @@ describe("Anthropic Messages sad-path recorded", () => {

      expect(error).toBeInstanceOf(AIError)
      expect(error.reason).toMatchObject({ _tag: "InvalidRequest" })
-      expect(error.reason.message).toContain("`tool_use` ids were found without `tool_result` blocks")
+      expect(error.message).toContain("`tool_use` ids were found without `tool_result` blocks")
    }),
  )
 })
@@ -2,7 +2,7 @@ import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { CacheHint, LLM, AIError, LLMRequest, Message, ToolCallPart, ToolDefinition, Usage } from "../../src/index.js"
-import { Auth, LLMClient } from "../../src/route.js"
+import { Auth, Endpoint, LLMClient, Route } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
 import * as AnthropicMessages from "../../src/protocols/anthropic-messages.js"
 import { GoogleVertexMessages } from "../../src/providers.js"
@@ -810,6 +810,99 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("round-trips compatible provider metadata in its own namespace", () =>
+    Effect.gen(function* () {
+      const compatible = Route.make({
+        id: "custom-anthropic-messages",
+        provider: "custom-anthropic",
+        protocol: AnthropicMessages.protocol,
+        endpoint: Endpoint.path("/messages", { baseURL: "https://compatible.test/v1" }),
+        auth: Auth.header("x-api-key", "test"),
+        framing: AnthropicMessages.framing,
+      }).model({ id: "custom-model" })
+      const result = [
+        {
+          type: "web_search_result",
+          url: "https://example.com",
+          citations: [{ type: "web_search_result_location", cited_text: "Example" }],
+        },
+      ]
+      const response = yield* LLMClient.generate(LLM.request({ model: compatible, prompt: "Search." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5, custom_start: true } } },
+              { type: "content_block_start", index: 0, content_block: { type: "thinking", thinking: "Thinking." } },
+              { type: "content_block_delta", index: 0, delta: { type: "signature_delta", signature: "custom_sig" } },
+              { type: "content_block_stop", index: 0 },
+              {
+                type: "content_block_start",
+                index: 1,
+                content_block: { type: "redacted_thinking", data: "custom_redacted" },
+              },
+              { type: "content_block_stop", index: 1 },
+              {
+                type: "content_block_start",
+                index: 2,
+                content_block: {
+                  type: "server_tool_use",
+                  id: "custom_tool",
+                  name: "web_search",
+                  input: { query: "example" },
+                },
+              },
+              { type: "content_block_stop", index: 2 },
+              {
+                type: "content_block_start",
+                index: 3,
+                content_block: { type: "web_search_tool_result", tool_use_id: "custom_tool", content: result },
+              },
+              { type: "content_block_stop", index: 3 },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "end_turn", stop_sequence: "custom_stop" },
+                usage: { output_tokens: 2, custom_terminal: true },
+              },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toMatchObject([
+        { type: "reasoning", text: "Thinking.", providerMetadata: { "custom-anthropic": { signature: "custom_sig" } } },
+        { type: "reasoning", text: "", providerMetadata: { "custom-anthropic": { redactedData: "custom_redacted" } } },
+        { type: "tool-call", id: "custom_tool", providerExecuted: true },
+        {
+          type: "tool-result",
+          providerExecuted: true,
+          providerMetadata: { "custom-anthropic": { blockType: "web_search_tool_result", result } },
+        },
+      ])
+      expect(response.usage?.providerMetadata).toEqual({
+        "custom-anthropic": { input_tokens: 5, custom_start: true, output_tokens: 2, custom_terminal: true },
+      })
+      expect(response.events.at(-1)).toMatchObject({
+        providerMetadata: { "custom-anthropic": { stopSequence: "custom_stop" } },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model: compatible, messages: [response.message], cache: "none" }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { type: "thinking", thinking: "Thinking.", signature: "custom_sig" },
+            { type: "redacted_thinking", data: "custom_redacted" },
+            { type: "server_tool_use", id: "custom_tool", name: "web_search", input: { query: "example" } },
+            { type: "web_search_tool_result", tool_use_id: "custom_tool", content: result },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("parses text, reasoning, and usage stream fixtures", () =>
    Effect.gen(function* () {
      const body = sseEvents(
@@ -873,9 +966,8 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({
-        _tag: "InvalidProviderOutput",
-        classification: "incomplete-stream",
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput", classification: "incomplete-stream" },
        message: "The provider response ended unexpectedly.",
      })
    }),
@@ -964,8 +1056,8 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({
-        _tag: "InvalidProviderOutput",
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput" },
        message: "Invalid anthropic/anthropic-messages stream event",
      })
    }),
@@ -986,8 +1078,8 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({
-        _tag: "InvalidProviderOutput",
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput" },
        message: "Invalid anthropic/anthropic-messages stream event",
      })
    }),
@@ -1013,8 +1105,8 @@ describe("Anthropic Messages route", () => {
            Effect.flip,
          )

-          expect(error.reason).toMatchObject({
-            _tag: "InvalidProviderOutput",
+          expect(error).toMatchObject({
+            reason: { _tag: "InvalidProviderOutput" },
            message: "Invalid anthropic/anthropic-messages stream event",
          })
        }),
@@ -1029,8 +1121,8 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({
-        _tag: "InvalidProviderOutput",
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput" },
        message: "Invalid anthropic/anthropic-messages stream event",
      })
    }),
@@ -1552,7 +1644,7 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", message: "overloaded_error: Overloaded" })
+      expect(error).toMatchObject({ reason: { _tag: "ProviderInternal" }, message: "overloaded_error: Overloaded" })
    }),
  )

@@ -1570,10 +1662,9 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({
-        _tag: "InvalidRequest",
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidRequest", classification: "context-overflow" },
        message: "invalid_request_error: prompt is too long: 210000 tokens",
-        classification: "context-overflow",
      })
    }),
  )
@@ -1585,7 +1676,7 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", message: "overloaded_error" })
+      expect(error).toMatchObject({ reason: { _tag: "ProviderInternal" }, message: "overloaded_error" })
    }),
  )

@@ -1596,7 +1687,7 @@ describe("Anthropic Messages route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({ _tag: "UnknownProvider", message: "Anthropic Messages stream error" })
+      expect(error).toMatchObject({ reason: { _tag: "UnknownProvider" }, message: "Anthropic Messages stream error" })
    }),
  )

@@ -1613,7 +1704,7 @@ describe("Anthropic Messages route", () => {
      )

      expect(error).toBeInstanceOf(AIError)
-      expect(error.reason).toMatchObject({ _tag: "InvalidRequest", message: "Bad request" })
+      expect(error).toMatchObject({ reason: { _tag: "InvalidRequest" }, message: "Bad request" })
    }),
  )

@@ -491,6 +491,59 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("ignores late tool deltas after contentBlockStop", () =>
+    Effect.gen(function* () {
+      const body = eventStreamBody(
+        [
+          "contentBlockStart",
+          {
+            contentBlockIndex: 0,
+            start: { toolUse: { toolUseId: "tool_1", name: "lookup" } },
+          },
+        ],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { toolUse: { input: '{"query":"weather"}' } } }],
+        ["contentBlockStop", { contentBlockIndex: 0 }],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { toolUse: { input: '{"late":true}' } } }],
+        ["messageStop", { stopReason: "tool_use" }],
+      )
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.toolCalls).toEqual([
+        { type: "tool-call", id: "tool_1", name: "lookup", input: { query: "weather" } },
+      ])
+      expect(response.events.filter((event) => event.type === "tool-input-delta")).toEqual([
+        {
+          type: "tool-input-delta",
+          id: "tool_1",
+          name: "lookup",
+          text: '{"query":"weather"}',
+          input: { query: "weather" },
+        },
+      ])
+    }),
+  )
+
+  it.effect("rejects tool deltas without contentBlockStart", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { toolUse: { input: "{}" } } }],
+              ["messageStop", { stopReason: "tool_use" }],
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput" },
+        message: "Bedrock Converse tool delta is missing its tool call",
+      })
+    }),
+  )
+
  it.effect("recovers incomplete tool input at finalization", () =>
    Effect.gen(function* () {
      const body = eventStreamBody(
@@ -531,6 +584,104 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  for (const signature of [undefined, "", "   "]) {
+    for (const cache of ["none", "auto"] as const) {
+      it.effect(`demotes unsigned reasoning to text (${JSON.stringify(signature)}, cache: ${cache})`, () =>
+        Effect.gen(function* () {
+          const prepared = yield* compileRequest(
+            LLM.request({
+              model,
+              messages: [
+                Message.user("Think"),
+                Message.assistant([
+                  {
+                    type: "reasoning",
+                    text: "Partial thought",
+                    providerMetadata: signature === undefined ? undefined : { bedrock: { signature } },
+                    cache: new CacheHint({ type: "ephemeral" }),
+                  },
+                ]),
+                Message.user("Continue"),
+              ],
+              cache,
+            }),
+          )
+          expect(prepared.body.messages[1]).toEqual({
+            role: "assistant",
+            content: [{ text: "Partial thought" }, { cachePoint: { type: "default" } }],
+          })
+        }),
+      )
+    }
+  }
+
+  it.effect("omits empty unsigned reasoning without leaving an empty or cache-only assistant", () =>
+    Effect.gen(function* () {
+      const cache = new CacheHint({ type: "ephemeral" })
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("Think"),
+            Message.assistant([
+              { type: "reasoning", text: "", cache },
+              { type: "reasoning", text: "  ", providerMetadata: { bedrock: { signature: "" } }, cache },
+            ]),
+            Message.user([{ type: "text", text: "Continue", cache }]),
+          ],
+          cache: "none",
+        }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "user",
+          content: [{ text: "Think" }, { text: "Continue" }, { cachePoint: { type: "default" } }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("demotes foreign reasoning while preserving signed, redacted, text, and tool blocks", () =>
+    Effect.gen(function* () {
+      const cache = new CacheHint({ type: "ephemeral" })
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              { type: "reasoning", text: "Foreign thought", providerMetadata: { anthropic: { signature: "old" } } },
+              {
+                type: "reasoning",
+                text: "Signed thought",
+                providerMetadata: { bedrock: { signature: "sig_1" } },
+                cache,
+              },
+              { type: "reasoning", text: "", encrypted: "sig_2", cache },
+              { type: "reasoning", text: "", providerMetadata: { bedrock: { redactedData: "cmVkYWN0ZWQ=" } }, cache },
+              { type: "text", text: "Checking" },
+              ToolCallPart.make({ id: "call_1", name: "lookup", input: {} }),
+            ]),
+          ],
+          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+          cache: "none",
+        }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { text: "Foreign thought" },
+            { reasoningContent: { reasoningText: { text: "Signed thought", signature: "sig_1" } } },
+            { reasoningContent: { reasoningText: { text: "", signature: "sig_2" } } },
+            { reasoningContent: { redactedContent: "cmVkYWN0ZWQ=" } },
+            { text: "Checking" },
+            { toolUse: { toolUseId: "call_1", name: "lookup", input: {} } },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("preserves streamed reasoning signatures for continuation lowering", () =>
    Effect.gen(function* () {
      const body = eventStreamBody(
@@ -569,6 +720,57 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("round-trips reassigned provider reasoning and usage metadata in its own namespace", () =>
+    Effect.gen(function* () {
+      const compatible = model.route.with({ provider: "custom-bedrock" }).model({ id: model.id })
+      const redactedData = "cmVkYWN0ZWQtdGhpbmtpbmc="
+      const response = yield* LLMClient.generate(LLMRequest.update(baseRequest, { model: compatible })).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["messageStart", { role: "assistant" }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { text: "Let me think." } } }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { signature: "custom_sig" } } }],
+              ["contentBlockStop", { contentBlockIndex: 0 }],
+              [
+                "contentBlockDelta",
+                { contentBlockIndex: 1, delta: { reasoningContent: { redactedContent: redactedData } } },
+              ],
+              ["contentBlockStop", { contentBlockIndex: 1 }],
+              ["messageStop", { stopReason: "end_turn" }],
+              ["metadata", { usage: { inputTokens: 5, outputTokens: 2, totalTokens: 7 } }],
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Let me think.",
+          providerMetadata: { "custom-bedrock": { signature: "custom_sig" } },
+        },
+        { type: "reasoning", text: "", providerMetadata: { "custom-bedrock": { redactedData } } },
+      ])
+      expect(response.usage?.providerMetadata).toEqual({
+        "custom-bedrock": { inputTokens: 5, outputTokens: 2, totalTokens: 7 },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model: compatible, messages: [response.message], cache: "none" }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { reasoningContent: { reasoningText: { text: "Let me think.", signature: "custom_sig" } } },
+            { reasoningContent: { redactedContent: redactedData } },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("preserves reasoning signatures when contentBlockStop is missing", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(baseRequest).pipe(
@@ -738,19 +940,32 @@ describe("Bedrock Converse route", () => {
      ])
      const error = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)), Effect.flip)

-      expect(error.reason).toMatchObject({ _tag: "UnknownProvider", message: "A future provider failure" })
+      expect(error).toMatchObject({ reason: { _tag: "UnknownProvider" }, message: "A future provider failure" })
    }),
  )

  it.effect("classifies throttlingException as a rate limit", () =>
    Effect.gen(function* () {
+      const payload = { message: "Slow down", details: { opaque: [1, 2] }, trace: "outer", p: "padding" }
      const body = concat([
        eventFrame("messageStart", { role: "assistant" }),
-        exceptionFrame("throttlingException", { message: "Slow down" }),
+        exceptionFrame("throttlingException", payload),
      ])
      const error = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)), Effect.flip)

-      expect(error.reason).toMatchObject({ _tag: "RateLimit", message: "Slow down" })
+      expect(error).toMatchObject({ reason: { _tag: "RateLimit" }, message: "Slow down" })
+      expect(JSON.parse(error.reason.body ?? "")).toEqual({
+        headers: {
+          ":message-type": { type: "string", value: "exception" },
+          ":exception-type": { type: "string", value: "throttlingException" },
+          ":content-type": { type: "string", value: "application/json" },
+        },
+        body: JSON.stringify(payload),
+      })
+      expect(error.reason.http).toMatchObject({
+        status: 200,
+        headers: { "content-type": "application/vnd.amazon.eventstream" },
+      })
    }),
  )

@@ -763,10 +978,9 @@ describe("Bedrock Converse route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({
-        _tag: "InvalidRequest",
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidRequest", classification: "context-overflow" },
        message: "Input is too long for requested model",
-        classification: "context-overflow",
      })
    }),
  )
@@ -785,7 +999,7 @@ describe("Bedrock Converse route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", message: "Upstream model failed" })
+      expect(error).toMatchObject({ reason: { _tag: "ProviderInternal" }, message: "Upstream model failed" })
    }),
  )

@@ -796,10 +1010,32 @@ describe("Bedrock Converse route", () => {
        Effect.flip,
      )

-      expect(error.reason).toMatchObject({
-        _tag: "InvalidProviderOutput",
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput" },
        message: "BadStream: Stream failed",
      })
+      expect(JSON.parse(error.reason.body ?? "")).toMatchObject({
+        headers: { ":error-code": { value: "BadStream" } },
+        body: "",
+      })
+    }),
+  )
+
+  it.effect("retains malformed AWS payloads with headers and decode cause", () =>
+    Effect.gen(function* () {
+      const headers = {
+        ":message-type": { type: "string" as const, value: "event" },
+        ":event-type": { type: "string" as const, value: "messageStart" },
+      }
+      const body = '{"malformed":'
+      const error = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(fixedBytes(codec.encode({ headers, body: utf8Encoder.encode(body) }))),
+        Effect.flip,
+      )
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      expect(JSON.parse(error.reason.body ?? "")).toEqual({ headers, body })
+      expect(error.reason.cause).toBeInstanceOf(Error)
+      expect(error.reason.http?.status).toBe(200)
    }),
  )

@@ -1,11 +1,13 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
-import { LLM } from "../../src/index.js"
+import { LLM, Message } from "../../src/index.js"
 import { AmazonBedrockMantle } from "../../src/providers.js"
+import { model } from "../../src/providers/amazon-bedrock/mantle.js"
+import { OpenAIResponses } from "../../src/protocols/openai-responses.js"
 import { compileRequest, LLMClient } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
-import { dynamicResponse } from "../lib/http.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
 import { sseEvents } from "../lib/sse.js"
 import { recordedTests } from "../recorded-test.js"

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

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

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

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

 const recorded = recordedTests({
@@ -1,4 +1,4 @@
-import { describe, expect } from "bun:test"
+import { describe, expect, test } from "bun:test"
 import { ConfigProvider, Effect, Schema } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, LLMEvent } from "../../src/index.js"
@@ -126,7 +126,7 @@ describe("Cloudflare", () => {
      expect(response.reasoning).toBe("Thinking")
      expect(response.events.filter(LLMEvent.is.reasoningDelta)).toHaveLength(2)
      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
-        openai: { reasoningField: "reasoning", reasoningDetails: merged },
+        "cloudflare-ai-gateway": { reasoningField: "reasoning", reasoningDetails: merged },
      })

      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
@@ -136,17 +136,15 @@ describe("Cloudflare", () => {
    }),
  )

-  it.effect("defaults AI Gateway id to default when omitted or blank", () =>
-    Effect.gen(function* () {
-      expect(
-        CloudflareAIGateway.configure({
-          accountId: "test-account",
-          gatewayId: "",
-          gatewayApiKey: "test-token",
-        }).model("workers-ai/@cf/meta/llama-3.3-70b-instruct").route.endpoint.baseURL,
-      ).toBe("https://gateway.ai.cloudflare.com/v1/test-account/default/compat")
-    }),
-  )
+  test("defaults AI Gateway id to default when omitted or blank", () => {
+    expect(
+      CloudflareAIGateway.configure({
+        accountId: "test-account",
+        gatewayId: "",
+        gatewayApiKey: "test-token",
+      }).model("workers-ai/@cf/meta/llama-3.3-70b-instruct").route.endpoint.baseURL,
+    ).toBe("https://gateway.ai.cloudflare.com/v1/test-account/default/compat")
+  })

  it.effect("supports authenticated AI Gateway plus upstream provider auth", () =>
    Effect.gen(function* () {
--- a/Show More
+++ b/Show More