test(core): isolate shell hook registry

Co-authored-by: Aiden Cline <rekram1-node@users.noreply.github.com>

.changeset/calm-services-start.md

@@ -1,7 +0,0 @@
----
-"@opencode-ai/client": patch
-"@opencode-ai/protocol": patch
-"@opencode-ai/cli": patch
----
-
-Expose background-service lifecycle status, preserve one process-held owner through startup and failure, reconnect TUIs without activating replacement, and stop exact service instances gracefully.

.changeset/calm-sessions-header.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/cli": patch
----
-
-Expose a TUI plugin slot at the top of the session view.

.changeset/clean-sessions-generate.md

@@ -1,7 +0,0 @@
----
-"@opencode-ai/client": patch
-"@opencode-ai/plugin": patch
-"@opencode-ai/protocol": patch
----
-
-Expose transient, read-only session generation through the HTTP API, generated clients, and V2 plugin session context.

.changeset/fresh-composers-slot.md

@@ -1,5 +0,0 @@
----
-"@opencode-ai/cli": patch
----
-
-Expose a TUI plugin slot above the session composer.

.github/workflows/deploy-www.yml

@@ -0,0 +1,37 @@
+name: deploy-www
+
+on:
+  push:
+    branches:
+      - dev
+      - v2
+  workflow_dispatch:
+
+concurrency:
+  group: deploy-www-${{ github.ref_name }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+
+jobs:
+  deploy:
+    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'dev' || github.ref_name == 'v2')
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+
+      - name: Build
+        working-directory: packages/www
+        run: bun run build
+        env:
+          BLUME_ENV: ${{ github.ref_name == 'v2' && 'production' || 'dev' }}
+          CLOUDFLARE_ENV: ${{ github.ref_name == 'v2' && 'production' || 'dev' }}
+
+      - name: Deploy
+        working-directory: packages/www
+        run: bun run deploy
+        env:
+          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}

.github/workflows/test.yml

@@ -99,7 +99,7 @@ jobs:
 
       - name: Check generated documentation
         if: runner.os == 'Linux'
-        working-directory: packages/docs
+        working-directory: packages/www
         run: bun run check:generated
 
   e2e:

.opencode/plugins/tui/discovery-smoke.ts

@@ -0,0 +1,13 @@
+import type { Context } from "../../../packages/plugin/src/tui/context"
+
+export default {
+  id: "test.tui-discovery-smoke",
+  setup(_context: Context) {
+    // context.ui.toast.show({
+    //   title: "TUI plugin discovery works",
+    //   message: "Loaded .opencode/plugins/tui/discovery-smoke.ts",
+    //   variant: "success",
+    //   duration: 30_000,
+    // })
+  },
+}

.opencode/skills/ideal-pseudocode/SKILL.md

@@ -0,0 +1,68 @@
+---
+name: ideal-pseudocode
+description: Function-by-function refactoring loop driven by ideal pseudocode. Use when the user says "ideal pseudocode", asks to make a function read like its pseudocode, or wants a dense module cleaned up one function at a time.
+---
+
+# Ideal Pseudocode
+
+Clean up one function at a time by writing the pseudocode it _should_ read as, naming every delta between that and the real code, and closing only the gaps the user approves.
+
+## Loop
+
+One function per round. Never touch code before the user picks a direction.
+
+1. **Pick the target** with the user — usually the next function up or down the call chain from the last round.
+2. **Read the current code** fresh from disk. It may have unsaved or parallel edits; ask before overwriting anything unexpected.
+3. **Distill.** Write the function's ideal pseudocode in a `ts`-fenced code block — TypeScript-flavored for syntax highlighting, but pseudocode: comments over mechanics, one line per idea, every arm of a loop visible as an arm. For a dense or unfamiliar function, first show the _current_ structure as pseudocode, then the ideal.
+4. **Name the deltas.** A numbered list; each delta is one concrete gap: control flow smuggled through combinators, a flag-typed parameter, colliding names, side-effect plumbing drowning a loop body, twin functions differing only by a filter. For each, give an honest cost/benefit and a decisive recommendation — including "leave it" when the density has reasons (state the reasons).
+5. **Verify before proposing.** Any API the proposal builds on (a combinator, schedule, matcher) gets checked against the installed version's source, not memory.
+6. **Get buy-in, then apply.** Minimal edits. Quarantine irreducible ugliness inside a small helper whose name is the pseudocode line it implements. Collapse shallow sibling functions into one deep function only when the distinguishing argument is a domain noun, not a behavior switch.
+7. **Validate.** Typecheck and the focused test suites from the package directory; rerun timing-sensitive tests repeatedly. Flag any semantic change the restructure introduced, even when tests pass.
+8. **Offer the next target** and stop.
+
+## Style
+
+- Keep each round tight: pseudocode block, deltas, recommendation, one question.
+- State mutation belongs in one visible place (usually the loop tail), never spread across error channels or callback closures.
+- Prefer early return and `if` chains for control flow; matchers are for producing values.
+- Compare names against the question each caller is actually asking; rename toward the domain noun.
+
+## Example round
+
+Every round should read like this (abridged from a real one, on an Effect step-runner):
+
+> `runStep` it is. The ideal pseudocode:
+>
+> ```ts
+> // One logical step: one settled model call, however many attempts it takes.
+> // An attempt can end without settling in two ways:
+> //   transient provider failure -> retry the same call, same assistant message
+> //   compaction rewrote history -> rebuild the request and call again
+> function runStep(session, promotable, step) {
+>   while (true) {
+>     const result = callModel(session, promotable, step)
+>
+>     if (result.completed) return result
+>
+>     if (result.retryable) {
+>       wait(retrySchedule) // exhausted -> publish Step.Failed, fail drain
+>       promotable = none // never re-promote on a second attempt
+>       continue
+>     }
+>
+>     // compaction restarted the step: fresh request from rewritten history
+>     promotable = none
+>     step = result.step
+>   }
+> }
+> ```
+>
+> Comparing against the real thing, three deltas:
+>
+> 1. **The retry arm doesn't live in the loop.** In the pseudocode, retry is just another loop arm. In the real code it's smuggled through the error channel: `Effect.tapError` mutating loop variables via `Effect.sync` closures, then `Effect.retryOrElse` re-running a suspended call. Four mutable variables mutated from three different places — the state flow is invisible unless you already know the trick.
+> 2. **`recoverOverflow: typeof compaction.compact | undefined`** — a function-typed parameter used as a one-shot armed/disarmed flag. The callee already closes over `compaction`; a boolean says what it is.
+> 3. **`assistantMessageID`** exists solely to resume the same durable message across retries — fine, but only legible once the retry arm is visible in the loop.
+>
+> The unifying move for (1): a schedule driver, so a retryable failure becomes a caught value handled in the loop body exactly like the compaction arm. One honest trade: it's more manual than the blessed combinator — but the manual version is the loop the combinator was hiding, and the loop already exists for restarts. Two restart mechanisms, one control structure.
+>
+> Want me to apply it — unified loop, simplified schedule input, boolean `recoverOverflow`?

AGENTS.md

@@ -1,4 +1,3 @@
-- To regenerate the legacy JavaScript SDK, run `./packages/sdk/js/script/build.ts`.
 - After changing the public Protocol or Server `HttpApi`, run `bun run generate` from `packages/client`. Do not edit `src/generated` or `src/generated-effect` directly.
 - Keep runtime dependencies directed from Schema to Core and Protocol, then from Core and Protocol to Server. Client runtime code may depend on Schema and Protocol but never Core or Server; `sdk-next` composes Client, Core, and Server.
 - Do not modify `packages/opencode` unless the user explicitly asks for V1 work. `packages/opencode` is the V1 implementation and is present for reference only. New implementation changes should land in the V2 package set: `packages/core`, `packages/cli`, `packages/server`, `packages/protocol`, `packages/schema`, and related generated client surfaces when required.
@@ -61,6 +60,7 @@ const { a, b } = obj
 ### Imports
 
 - Never alias imports. Do not use `import { foo as bar } from "..."` or renamed imports like `resolve as pathResolve`.
+- Never use type-position `import("...")` references such as `Schema.declare<import("@opencode-ai/plugin/effect/plugin").Plugin["effect"]>`. Only when two imports genuinely collide on a name and no other option exists, an aliased type import (`import type { Plugin as PluginDefinition } from "..."`) is permitted as a last resort — still strongly preferred not to.
 - Never use star imports. Do not use `import * as Foo from "..."` or `import type * as Foo from "..."`.
 - If a namespace-style value is needed, import the module's own exported namespace by name, for example `import { Project } from "@opencode-ai/core/project"`, then reference `Project.ID`.
 - Prefer dynamic imports for heavy modules that are only needed in selected code paths, especially in startup-sensitive entrypoints. Destructure dynamic import bindings near the top of the narrowest scope that needs them so they read like normal imports. Avoid inline chains such as `await import("./module").then((mod) => mod.value())` or `(await import("./module")).value()`. Keep branch-specific imports inside the branch that needs them to preserve lazy loading.

bunfig.toml

@@ -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", "@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"]
+minimumReleaseAgeExcludes = ["@ai-sdk/amazon-bedrock", "@ai-sdk/anthropic", "@opencode-ai/sdk", "@opentui/core", "@opentui/core-darwin-arm64", "@opentui/core-darwin-x64", "@opentui/core-linux-arm64", "@opentui/core-linux-arm64-musl", "@opentui/core-linux-x64", "@opentui/core-linux-x64-musl", "@opentui/core-win32-arm64", "@opentui/core-win32-x64", "@opentui/keymap", "@opentui/solid", "opentui-spinner", "gitlab-ai-provider", "opencode-gitlab-auth", "@ff-labs/fff-node", "@ff-labs/fff-bun", "@ff-labs/fff-bin-darwin-arm64", "@ff-labs/fff-bin-darwin-x64", "@ff-labs/fff-bin-linux-arm64-gnu", "@ff-labs/fff-bin-linux-arm64-musl", "@ff-labs/fff-bin-linux-x64-gnu", "@ff-labs/fff-bin-linux-x64-musl", "@ff-labs/fff-bin-win32-arm64", "@ff-labs/fff-bin-win32-x64", "@pierre/diffs", "@pierre/theming", "app-builder-lib", "dmg-builder", "electron-builder", "electron-publish"]
 
 [test]
 root = "./do-not-run-tests-from-root"

docs/design/v1-v2-database-migration.md

@@ -0,0 +1,118 @@
+# V1 to V2 Database Migration
+
+## Approach
+
+- Use the `dev` branch database schema and migration registry as the V1 baseline.
+- Remove migrations that exist only on the V2 branch.
+- Generate one canonical migration from the `dev` schema to the final V2 schema.
+- Add explicit data operations to that migration where generated DDL is insufficient.
+- Test the migration against a populated database at the exact `dev` schema.
+
+## Preserve
+
+The canonical V1 data remains in its existing tables. In particular, preserve `session`, `message`, and `part` rows.
+
+Preserve `workspace` rows and existing `session.workspace_id` values unchanged. The migration must not clear or rebuild
+workspace relationships.
+
+Keep the `todo` table and its data unchanged. V2 does not currently migrate todos into another representation, and the
+generated migration must not drop the table.
+
+## Truncate
+
+Truncate these pre-launch V2 tables before applying schema changes:
+
+- `event`
+- `event_sequence`
+- `session_message`
+
+These rows are not canonical V1 data. Truncating `event` before adding the required `event.created` column means the
+column needs neither a backfill nor a default. After truncation, rebuild `session_message` from canonical V1 `message`
+and `part` rows rather than retaining its pre-launch V2 contents.
+
+## Message Backfill
+
+Backfill canonical V1 history from `message` and `part` into `session_message`. This is the main data transformation in
+the migration. Preserving the V1 tables alone keeps the data safe but does not make existing history visible through the
+V2 session APIs, which read `session_message`.
+
+Reuse each V1 `message.id` as the corresponding `session_message.id`. Stable IDs keep the migration deterministic and
+avoid rewriting other persisted state that may refer to a message.
+
+Within each session, order V1 messages by `time_created` and then `id`, matching the existing V1 message index. Assign
+contiguous `session_message.seq` values starting at `0`.
+
+Map ordinary V1 messages one-to-one by role. Each ordinary V1 user message becomes one V2 `user` row, and each ordinary
+V1 assistant message becomes one V2 `assistant` row. Fold the source message's ordered V1 parts into that row's V2
+payload.
+
+Handle semantic marker parts before applying the ordinary mapping. In particular, a V1 user message containing a
+`compaction` part and its paired assistant summary represent one compaction operation, not two ordinary messages. Special
+part mappings must be decided explicitly before implementing the backfill.
+
+V1 synthetic content is represented by user text parts with `synthetic: true`, not by a separate message role. A V1 user
+message whose visible text parts are all synthetic should become a V2 `synthetic` message. If a V1 user message mixes
+ordinary and synthetic content, preserve the ordinary content in the V2 `user` row and emit the synthetic content as an
+adjacent V2 `synthetic` row. Ignore text parts marked `ignored`, matching V1 model-history behavior.
+
+Use the V1 compaction user message ID as the ID of the collapsed V2 compaction message. This matches V2's use of the
+admitted compaction input ID and preserves references to the initiating message.
+
+For a completed compaction, create one V2 `compaction` row with `status: "completed"`. Set `reason` from the V1
+compaction part's `auto` flag, join the paired summary assistant's nonempty text parts with blank lines for `summary`, and
+serialize the retained V1 tail beginning at `tail_start_id` for `recent`. Use an empty `recent` value when no tail was
+retained, and use the compaction user message creation time. Do not emit the paired summary assistant as a separate V2
+assistant row.
+
+After rebuilding `session_message`, seed `event_sequence` with one row per migrated session. Set its watermark to that
+session's maximum backfilled `session_message.seq`. This prevents new V2 events from reusing sequence numbers or sorting
+before migrated history. The `event` table remains empty.
+
+## Drop
+
+Drop these pre-launch V2 tables without preserving or transforming their rows:
+
+- `session_input`
+- `session_context_epoch`
+
+Do not transfer `session_input` rows into `session_pending`.
+
+## Create Empty
+
+Let the generated migration create these tables empty:
+
+- `instruction_blob`
+- `instruction_entry`
+- `instruction_state`
+- `session_pending`
+- `kv`
+
+V1 has no canonical data to backfill into these tables. V2 initializes their state as it runs.
+
+## Fork Storage
+
+V1 has no fork-boundary state to backfill. New V2 forks use a required message boundary and persist it in
+`session.fork_boundary`. The durable fork event contains no parent sequence. Its resolved boundary is one of:
+
+- `before`: copy messages before the identified message.
+- `through`: copy messages through the identified message.
+
+Forking an empty session is not supported. `session.fork_seq` and `session.fork_message_id` are not part of the final V2
+schema.
+
+New nullable session columns, including `fork_session_id`, `fork_boundary`, and `time_suspended`, require no explicit
+backfill. Existing rows naturally receive `NULL` when the generated migration adds the columns.
+
+## Verification
+
+The canonical migration test should seed representative V1 sessions, messages, parts, todos, projects, accounts,
+credentials, permissions, shares, and workspaces. After migration, it should verify:
+
+- Preserved rows and encoded values remain unchanged.
+- Todo rows remain available in the unchanged `todo` table.
+- `event` is empty, and stale pre-launch rows are absent from the rebuilt projections.
+- Backfilled `session_message` rows represent the canonical V1 `message` and `part` history.
+- Each migrated session's `event_sequence` watermark matches its maximum backfilled message sequence.
+- Dropped tables no longer exist.
+- New tables exist and are empty.
+- The final schema has no ungenerated changes.

github/package.json

@@ -15,6 +15,6 @@
     "@actions/github": "6.0.1",
     "@octokit/graphql": "9.0.1",
     "@octokit/rest": "catalog:",
-    "@opencode-ai/sdk": "workspace:*"
+    "@opencode-ai/sdk": "1.18.5"
   }
 }

nix/hashes.json

@@ -1,8 +1,8 @@
 {
   "nodeModules": {
-    "x86_64-linux": "sha256-qt11SKmOjq0KU542QFbs+u7YyJicn4drCcwCdg325yk=",
-    "aarch64-linux": "sha256-z68doReXTrWS7HeiAjc0btIjAsvzeZZ7hXAlHr0c77Q=",
-    "aarch64-darwin": "sha256-PILYH1Pi8XBvSkuZ+1sNnUTao5kba+m5Z8iJKx6YXPo=",
-    "x86_64-darwin": "sha256-KpcJzP4m0SUavu/WaSffgzOxrHq8ljdy0GOzs9p16lo="
+    "x86_64-linux": "sha256-RFek0QoEEjsgbqmTE/SxQAmPtYyzs0IPR2ugFn5Okrs=",
+    "aarch64-linux": "sha256-BmAxapY1YrAFn7mVq3/6A9+6Au5UIvSqBboHMkyJH3I=",
+    "aarch64-darwin": "sha256-Sx3bGWQqLlgoa/RudJxanjSzhFRNklckT2ffnO2I5F4=",
+    "x86_64-darwin": "sha256-CMOhiisHNowg06qadvgg4K+60zrynglwiT0qKYQ4NiA="
   }
 }

package.json

@@ -33,13 +33,12 @@
       "packages/*",
       "packages/console/*",
       "packages/stats/*",
-      "packages/sdk/js",
       "packages/slack"
     ],
     "catalog": {
-      "@effect/opentelemetry": "4.0.0-beta.98",
-      "@effect/platform-node": "4.0.0-beta.98",
-      "@effect/sql-sqlite-bun": "4.0.0-beta.98",
+      "@effect/opentelemetry": "4.0.0-beta.101",
+      "@effect/platform-node": "4.0.0-beta.101",
+      "@effect/sql-sqlite-bun": "4.0.0-beta.101",
       "@npmcli/arborist": "9.4.0",
       "@types/bun": "1.3.13",
       "@types/cross-spawn": "6.0.6",
@@ -51,6 +50,7 @@
       "@opentui/solid": "0.4.5",
       "@tanstack/solid-virtual": "3.13.32",
       "@shikijs/stream": "4.2.0",
+      "@standard-schema/spec": "1.1.0",
       "ulid": "3.0.1",
       "@kobalte/core": "0.13.11",
       "@corvu/drawer": "0.2.4",
@@ -69,7 +69,7 @@
       "dompurify": "3.3.1",
       "drizzle-kit": "1.0.0-rc.2",
       "drizzle-orm": "1.0.0-rc.2",
-      "effect": "4.0.0-beta.98",
+      "effect": "4.0.0-beta.101",
       "ai": "6.0.168",
       "cross-spawn": "7.0.6",
       "hono": "4.10.7",
@@ -124,7 +124,7 @@
     "@aws-sdk/client-s3": "3.933.0",
     "@opencode-ai/plugin": "workspace:*",
     "@opencode-ai/script": "workspace:*",
-    "@opencode-ai/sdk": "workspace:*",
+    "@opencode-ai/sdk": "1.18.5",
     "heap-snapshot-toolkit": "1.1.3",
     "typescript": "catalog:"
   },
@@ -152,21 +152,22 @@
     "@opentui/keymap": "catalog:",
     "@opentui/solid": "catalog:",
     "@types/bun": "catalog:",
-    "@types/node": "catalog:"
+    "@types/node": "catalog:",
+    "effect": "catalog:"
   },
   "patchedDependencies": {
-    "@ff-labs/fff-bun@0.9.3": "patches/@ff-labs%2Ffff-bun@0.9.3.patch",
     "@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",
     "@ai-sdk/xai@3.0.102": "patches/@ai-sdk%2Fxai@3.0.102.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",
     "@ai-sdk/google@3.0.73": "patches/@ai-sdk%2Fgoogle@3.0.73.patch",
     "@pierre/trees@1.0.0-beta.4": "patches/@pierre%2Ftrees@1.0.0-beta.4.patch",
     "@modelcontextprotocol/sdk@1.29.0": "patches/@modelcontextprotocol%2Fsdk@1.29.0.patch",
-    "effect@4.0.0-beta.98": "patches/effect@4.0.0-beta.98.patch",
+    "effect@4.0.0-beta.101": "patches/effect@4.0.0-beta.101.patch",
     "@tanstack/virtual-core@3.17.3": "patches/@tanstack%2Fvirtual-core@3.17.3.patch"
   }
 }

packages/ai/AGENTS.md

@@ -46,7 +46,7 @@ const response = yield * LLMClient.generate(request)
 
 `LLM.request(...)` builds an `LLMRequest`. `LLMClient.generate(...)` reads the executable route carried by `request.model.route`, builds the provider-native body, asks the route's transport for a real `HttpClientRequest.HttpClientRequest`, sends it through `RequestExecutor.Service`, parses the provider stream into common `LLMEvent`s, and finally returns an `LLMResponse`.
 
-Use `LLMClient.stream(request)` when callers want incremental `LLMEvent`s. Use `LLMClient.generate(request)` when callers want those same events collected into an `LLMResponse`. Use `LLMClient.prepare<Body>(request)` to compile a request through the route pipeline without sending it — the optional `Body` type argument narrows `.body` to the route's native shape (e.g. `prepare<OpenAIChatBody>(...)` returns a `PreparedRequestOf<OpenAIChatBody>`). The runtime body is identical; the generic is a type-level assertion.
+Use `LLMClient.stream(request)` when callers want incremental `LLMEvent`s. Use `LLMClient.generate(request)` when callers want those same events collected into an `LLMResponse`.
 
 Filter or narrow `LLMEvent` streams with `LLMEvent.is.*` (camelCase guards, e.g. `events.filter(LLMEvent.is.toolCall)`). The kebab-case `LLMEvent.guards["tool-call"]` form also works but prefer `is.*` in new code.
 
@@ -54,7 +54,7 @@ Filter or narrow `LLMEvent` streams with `LLMEvent.is.*` (camelCase guards, e.g.
 
 A route is the registered, runnable composition of four orthogonal pieces:
 
-- **`Protocol`** (`src/route/protocol.ts`) — semantic API contract. Owns request body construction (`body.from`), the body schema (`body.schema`), the streaming-event schema (`stream.event`), and the event-to-`LLMEvent` state machine (`stream.step`). `Route.make(...)` validates and JSON-encodes the body from `body.schema` and decodes frames with `stream.event`. Examples: `OpenAIChat.protocol`, `OpenAIResponses.protocol`, `AnthropicMessages.protocol`, `Gemini.protocol`, `BedrockConverse.protocol`.
+- **`Protocol`** (`src/route/protocol.ts`) — semantic API contract. Owns request body construction (`body.from`), the body schema (`body.schema`), the streaming-event schema (`stream.event`), and the event-to-`LLMEvent` state machine (`stream.step`). `Route.make(...)` validates and JSON-encodes the body from `body.schema` and decodes frames with `stream.event`. Examples: `OpenAIChat.protocol`, `OpenResponses.protocol`, `OpenAIResponses.protocol`, `AnthropicMessages.protocol`, `Gemini.protocol`, `BedrockConverse.protocol`.
 - **`Endpoint`** (`src/route/endpoint.ts`) — URL construction. The host, path, and route query live on the endpoint. `Endpoint.path("/chat/completions", { baseURL })` is the common case; pass a function for paths that embed the model id or a body field (e.g. `Endpoint.path(({ body }) => `/model/${body.modelId}/converse-stream`)`).
 - **`Auth`** (`src/route/auth.ts`) — per-request transport authentication. Provider facades configure credentials onto the route before model selection, usually via `Auth.bearer(apiKey)` or `Auth.header(name, apiKey)`. Routes that need per-request signing (Bedrock SigV4, future Vertex IAM, Azure AAD) implement `Auth` as a function that signs the body and merges signed headers into the result.
 - **`Framing`** (`src/route/framing.ts`) — bytes → frames. SSE (`Framing.sse`) is shared; Bedrock keeps its AWS event-stream framing as a typed `Framing<object>` value alongside its protocol.
@@ -138,13 +138,13 @@ packages/ai/src/
     ids.ts                  branded IDs, literal types, ProviderMetadata
     options.ts              Generation/Provider/Http options, Limits, Model, cache policy
     messages.ts             content parts, Message, ToolDefinition, LLMRequest
-    events.ts               Usage, individual events, LLMEvent, PreparedRequest, LLMResponse
+    events.ts               Usage, individual events, LLMEvent, LLMResponse
     errors.ts               error reasons, LLMError, ToolFailure
     index.ts                barrel
   llm.ts                    request constructors and convenience helpers
   route/
     index.ts                @opencode-ai/ai/route advanced barrel
-    client.ts               Route.make + LLMClient.prepare/stream/generate
+    client.ts               Route.make + LLMClient.stream/generate
     executor.ts             RequestExecutor service + transport error mapping
     protocol.ts             Protocol type + Protocol.make
     endpoint.ts             Endpoint type + Endpoint.path
@@ -158,13 +158,14 @@ packages/ai/src/
   protocols/
     shared.ts               ProviderShared toolkit used inside protocol impls
     openai-chat.ts          protocol + route (compose OpenAIChat.protocol)
-    openai-responses.ts
+    open-responses.ts         provider-neutral Responses protocol baseline
+    openai-responses.ts       OpenAI tools/events/transports composed over OpenResponses
     anthropic-messages.ts
     gemini.ts
     bedrock-converse.ts
     bedrock-event-stream.ts framing for AWS event-stream binary frames
     openai-compatible-chat.ts route that reuses OpenAIChat.protocol, no canonical URL
-    openai-compatible-responses.ts route that reuses OpenAIResponses.protocol, no canonical URL
+    openai-compatible-responses.ts deployment adapter that reuses OpenResponses.protocol, no canonical URL
     utils/                  per-protocol helpers (auth, cache, media, tool-stream, ...)
   providers/
     openai-compatible.ts    generic Chat helper + family model helpers
@@ -175,7 +176,7 @@ packages/ai/src/
   tool-runtime.ts           narrow one-call typed tool dispatcher
 ```
 
-The dependency arrow points down: `providers/*.ts` files import protocol routes and auth-option utilities; protocol modules import `endpoint`, `auth`, `framing`, and transport pieces. Protocols do not import provider facades. Lower-level modules know nothing about provider catalog metadata.
+The dependency arrow points down: `providers/*.ts` files import protocol routes and auth-option utilities; protocol modules import `endpoint`, `auth`, `framing`, and transport pieces. Protocols do not import provider facades. Lower-level modules know nothing about provider catalog metadata. `OpenAIResponses` composes the provider-neutral `OpenResponses` protocol; the baseline never imports the OpenAI extension.
 
 ### Shared protocol helpers
 
@@ -240,7 +241,7 @@ const get_weather = tool({
 
 const tools = { get_weather, get_time, ... }
 const events = yield* LLM.stream(
-  LLM.updateRequest(request, { tools: Tool.toDefinitions(tools) }),
+  LLMRequest.update(request, { tools: Tool.toDefinitions(tools) }),
 ).pipe(Stream.runCollect)
 
 const call = Array.from(events).find(LLMEvent.is.toolCall)

packages/ai/DESIGN.md

@@ -315,7 +315,8 @@ const longer = {
 }
 ```
 
-There is no `LLM.updateRequest(...)` helper and no request Schema class.
+There is no `LLM.updateRequest(...)` helper. The current Schema-backed implementation
+uses `LLMRequest.update(...)` when canonical request data must be derived.
 
 ### Conversation history
 
@@ -436,7 +437,7 @@ const call = Array.from(events).find(LLMEvent.is.toolCall)
 
 if (call && !call.providerExecuted) {
   const dispatched = yield * ToolRuntime.dispatch(tools, call)
-  const followUp = LLM.updateRequest(request, {
+  const followUp = LLMRequest.update(request, {
     messages: [...request.messages, Message.assistant([call]), Message.tool({ ...call, result: dispatched.result })],
   })
   // Caller must invoke the provider again and repeat the loop.

packages/ai/README.md

@@ -196,7 +196,6 @@ The hosted result is represented as a provider-executed tool call and tool resul
 - **`LLM.generate` / `LLM.stream`** — re-exported from `LLMClient` for one-import use.
 - **`Message.user(...)` / `Message.assistant(...)` / `Message.tool(...)`** — message constructors from the canonical schema model.
 - **`Model.make(...)` / `ToolCallPart.make(...)` / `ToolResultPart.make(...)` / `ToolDefinition.make(...)`** — model and tool-related constructors from the canonical schema model.
-- **`LLMClient.prepare(request)`** — compile a request through protocol body construction, validation, and HTTP preparation without sending. Useful for inspection and testing.
 - **`LLMEvent.is.*`** — typed guards (`is.textDelta`, `is.toolCall`, `is.finish`, …) for filtering streams.
 - **`Image.generate({...})`** — generate images through a provider-neutral image request and response model.
 - **`ImageClient`** — Effect service and layer for image execution, parallel to `LLMClient`.
@@ -207,7 +206,9 @@ Prompt caching is **on by default**. Every `LLMRequest` resolves to `cache: "aut
 
 ### Auto placement
 
-`"auto"` places three breakpoints — last tool definition, last system part, latest user message. The last-user-message boundary is the load-bearing detail: in a tool-use loop, a single user turn expands into many assistant/tool round-trips, all sharing that prefix. Caching at that boundary lets every intra-turn API call hit.
+`"auto"` places up to four breakpoints — the last tool definition, the first system part, the last system part when distinct, and the final message boundary. These expose successively larger reusable prefixes for tools, the base agent, project instructions, and the active conversation. The rolling final-message boundary is the load-bearing detail in tool loops: it advances on every request so the previous cache entry stays within Anthropic's 20-block lookback.
+
+Tools precede every system and conversation block in the provider prefix, so tool definitions must remain byte-stable and deterministically ordered for downstream breakpoints to remain reusable.
 
 The math justifies the default: Anthropic's 5-minute cache write is 1.25× base, read is 0.1×, so a single reuse within 5 minutes already wins. One-shot completions below the per-model minimum-cacheable-token threshold silently no-op on the wire, so the worst case is harmless.
 
@@ -235,7 +236,7 @@ cache: {
 
 ### Manual hints
 
-Inline `CacheHint` on any text / system / tool / tool-result part overrides automatic placement. The auto policy preserves manual hints; it only fills gaps.
+Inline `CacheHint` on any text / system / tool / tool-result part overrides automatic placement. The auto policy preserves manual hints, counts them against Anthropic and Bedrock's four-breakpoint limit, and only fills the remaining slots.
 
 ```ts
 LLM.request({
@@ -251,8 +252,8 @@ LLM.request({
 
 | Protocol                | `cache: "auto"`                                                           |
 | ----------------------- | ------------------------------------------------------------------------- |
-| Anthropic Messages      | emits up to 3 `cache_control` markers (4-breakpoint cap enforced)         |
-| Bedrock Converse        | emits up to 3 `cachePoint` blocks (4-breakpoint cap enforced)             |
+| Anthropic Messages      | emits up to 4 `cache_control` markers (4-breakpoint cap enforced)         |
+| Bedrock Converse        | emits up to 4 `cachePoint` blocks (4-breakpoint cap enforced)             |
 | OpenAI Chat / Responses | no-op (implicit caching above 1024 tokens)                                |
 | Gemini                  | no-op (implicit caching on 2.5+; explicit `CachedContent` is out-of-band) |
 
@@ -300,7 +301,7 @@ OpenAI Chat and OpenAI Responses are separate semantic entrypoints:
 - `@opencode-ai/ai/providers/google-vertex/responses`
 - `@opencode-ai/ai/providers/google-vertex/messages`
 
-Responses HTTP versus WebSocket is a scoped `transport` setting on the OpenAI Responses entrypoint, not another entrypoint. Azure follows the same Chat/Responses split at `providers/azure/chat` and `providers/azure/responses`. Generic OpenAI-compatible Chat remains at `providers/openai-compatible`; compatible Responses is separate at `providers/openai-compatible/responses`. Generic Anthropic Messages-compatible providers use `providers/anthropic-compatible`, which the named Anthropic provider composes. Google Gemini and Amazon Bedrock expose their single native API through their existing provider paths.
+Responses HTTP versus WebSocket is a scoped `transport` setting on the OpenAI Responses entrypoint, not another entrypoint. Azure follows the same Chat/Responses split at `providers/azure/chat` and `providers/azure/responses`. Generic OpenAI-compatible Chat remains at `providers/openai-compatible`; the Responses adapter at `providers/openai-compatible/responses` uses the provider-neutral Open Responses protocol. OpenAI Responses extends that baseline with OpenAI tools, event variants, metadata, defaults, and transports. Generic Anthropic Messages-compatible providers use `providers/anthropic-compatible`, which the named Anthropic provider composes. Google Gemini and Amazon Bedrock expose their single native API through their existing provider paths.
 
 Vertex Gemini, Vertex Chat, Vertex Responses, and Vertex Messages are separate API entrypoints. All accept `project`, `location`, and an optional `accessToken`; when no explicit token or auth override is supplied they lazily use Google Application Default Credentials. Vertex Gemini instead selects express mode when `apiKey` or `GOOGLE_VERTEX_API_KEY` is present. Vertex Chat targets MaaS models through the OpenAI-compatible Chat Completions endpoint, while Vertex Responses targets Grok models and defaults `store` to `false` as required by Vertex. `providers/google-vertex` remains the default alias for `providers/google-vertex/gemini`.
 

packages/ai/STATUS.md

@@ -1,6 +1,6 @@
 # LLM Provider Parity Status
 
-Last reviewed: 2026-07-17
+Last reviewed: 2026-07-24
 
 This file tracks the gap between the native `@opencode-ai/ai` package and the AI SDK provider packages that opencode still depends on for many catalog/runtime paths.
 
@@ -13,26 +13,26 @@ This file tracks the gap between the native `@opencode-ai/ai` package and the AI
 
 ## Current Implementation Snapshot
 
-| Native slice                       | Source                                                                                         | Current state                                                                                                                                                            | Main gaps                                                                                                                                                                                                                                                 |
-| ---------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| OpenAI Chat                        | `src/protocols/openai-chat.ts`, `src/providers/openai.ts`                                      | Usable. Streams text, reasoning deltas, tool calls, usage, images, and common generation controls.                                                                       | No typed structured-output / `response_format` path. Limited typed OpenAI option surface compared with SDK escape hatches.                                                                                                                                |
-| OpenAI Responses HTTP              | `src/protocols/openai-responses.ts`, `src/providers/openai.ts`                                 | Usable. Supports hosted-tool event surfacing, reasoning replay metadata, GPT-5 defaults, and cache usage.                                                                | No explicit `previous_response_id` path. Typed options cover only a subset of Responses fields. Structured output is still mostly synthetic-tool based.                                                                                                   |
-| OpenAI Responses WebSocket         | `src/protocols/openai-responses.ts`, `src/route/transport/websocket.ts`                        | Present as `OpenAI.responsesWebSocket(...)`.                                                                                                                             | Runner/catalog support explicitly must not downgrade WebSocket routes; broader runtime selection is not complete.                                                                                                                                         |
-| OpenAI-compatible Chat             | `src/protocols/openai-compatible-chat.ts`, `src/providers/openai-compatible.ts`                | Usable for generic Chat and several profiles: Baseten, Cerebras, DeepInfra, DeepSeek, Fireworks, Groq, TogetherAI.                                                       | Family quirks are mostly endpoint defaults, not full typed behavior.                                                                                                                                                                                      |
-| OpenAI-compatible Responses        | `src/protocols/openai-compatible-responses.ts`, `src/providers/openai-compatible-responses.ts` | Usable for deployments that implement the OpenAI Responses wire protocol.                                                                                                | No named family profiles or recorded deployment coverage yet.                                                                                                                                                                                             |
-| Anthropic-compatible Messages      | `src/protocols/anthropic-messages.ts`, `src/providers/anthropic-compatible.ts`                 | Usable for deployments that implement the Anthropic Messages wire protocol. Named Anthropic composes this base; MiniMax M3 has recorded text and tool-loop coverage.       | No named compatible family profiles yet.                                                                                                                                                                                                                  |
-| Anthropic Messages                 | `src/protocols/anthropic-messages.ts`, `src/providers/anthropic.ts`                            | Usable. Supports tools, thinking, cache control, images, server-hosted tool events, and usage.                                                                           | Provider option surface is small. Beta/header handling, metadata, and newer Messages fields need a typed parity pass.                                                                                                                                     |
-| Gemini Developer API               | `src/protocols/gemini.ts`, `src/providers/google.ts`                                           | Usable for Google API key flow. Supports text, images, tools, thinking signatures, and cache usage.                                                                      | This is not Vertex. Typed provider options are narrow; many Gemini request fields currently require raw `http.body` overlays.                                                                                                                             |
-| Vertex Gemini                      | `src/protocols/gemini.ts`, `src/providers/google-vertex.ts`                                    | Usable through API-key express mode, explicit OAuth tokens, or ADC with project/location endpoint derivation, including tuned `endpoints/...` deployments.               | Core runner/catalog mapping and recorded provider coverage are missing.                                                                                                                                                                                   |
-| Vertex Chat                        | `src/protocols/openai-chat.ts`, `src/providers/google-vertex-chat.ts`                          | Usable for MaaS models through OpenAI-compatible Chat Completions with explicit OAuth tokens or ADC and project/location endpoint derivation.                            | Core runner/catalog mapping and recorded provider coverage are missing; MaaS family-specific request parity needs review.                                                                                                                                 |
-| Vertex Responses                   | `src/protocols/openai-responses.ts`, `src/providers/google-vertex-responses.ts`                | Usable for Grok models through OpenAI-compatible Responses with explicit OAuth tokens or ADC, project/location endpoint derivation, and storage disabled by default.     | Core runner/catalog mapping and recorded provider coverage are missing; stateful continuation is not supported by Vertex.                                                                                                                                 |
-| Vertex Messages                    | `src/protocols/anthropic-messages.ts`, `src/providers/google-vertex-messages.ts`               | Usable through explicit OAuth tokens or ADC, including global, regional, and `eu`/`us` multi-region endpoints.                                                           | Core runner/catalog mapping and recorded provider coverage are missing; Vertex-specific hosted-tool parity needs review.                                                                                                                                  |
-| Bedrock Converse                   | `src/protocols/bedrock-converse.ts`, `src/providers/amazon-bedrock.ts`                         | Partial but real. Supports AWS event-stream framing, SigV4 with supplied credentials, bearer auth, tools, reasoning signatures, media, cache points, and recorded tests. | Native facade does not mirror the AI SDK plugin's default AWS credential chain/profile behavior. Runner/catalog mapping is missing. Guardrails, inference profiles, region-specific model ID fixes, and model-specific request fields need a parity pass. |
-| Azure OpenAI                       | `src/providers/azure.ts` using OpenAI Chat/Responses protocols                                 | Partial. Supports resource/base URL setup, API key auth, API version query, Chat, and Responses selectors.                                                               | Core runner does not map `@ai-sdk/azure` to this native facade. AAD/token auth and Azure-specific endpoint variants need review.                                                                                                                          |
-| Cloudflare AI Gateway / Workers AI | `src/providers/cloudflare.ts`                                                                  | Present via OpenAI-compatible Chat routes.                                                                                                                               | Useful but not part of the critical AI SDK replacement set yet. Needs per-product recorded coverage before relying on it broadly.                                                                                                                         |
-| OpenRouter                         | `src/providers/openrouter.ts`                                                                  | Present with OpenRouter-specific usage/reasoning/prompt-cache options over Chat.                                                                                         | Responses-style OpenRouter support is absent.                                                                                                                                                                                                             |
-| xAI                                | `src/providers/xai.ts`                                                                         | Present with Responses and Chat selectors.                                                                                                                               | Needs package-parity review against the AI SDK xAI provider.                                                                                                                                                                                              |
-| GitHub Copilot                     | `src/providers/github-copilot.ts`                                                              | Present as explicit-base-URL OpenAI Chat/Responses facade.                                                                                                               | Runtime/catalog integration remains specialized and should stay separate from public OpenAI-compatible defaults.                                                                                                                                          |
+| Native slice                       | Source                                                                                                                            | Current state                                                                                                                                                            | Main gaps                                                                                                                                                                                                                                                 |
+| ---------------------------------- | --------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| OpenAI Chat                        | `src/protocols/openai-chat.ts`, `src/providers/openai.ts`                                                                         | Usable. Streams text, reasoning deltas, tool calls, usage, images, and common generation controls.                                                                       | No typed structured-output / `response_format` path. Limited typed OpenAI option surface compared with SDK escape hatches.                                                                                                                                |
+| OpenAI Responses HTTP              | `src/protocols/open-responses.ts`, `src/protocols/openai-responses.ts`, `src/providers/openai.ts`                                 | Usable. Extends the Open Responses baseline with hosted-tool event surfacing, reasoning replay metadata, GPT-5 defaults, and cache usage.                                | No explicit `previous_response_id` path. Typed options cover only a subset of Responses fields. Structured output is still mostly synthetic-tool based.                                                                                                   |
+| OpenAI Responses WebSocket         | `src/protocols/openai-responses.ts`, `src/route/transport/websocket.ts`                                                           | Present as `OpenAI.responsesWebSocket(...)`.                                                                                                                             | Runner/catalog support explicitly must not downgrade WebSocket routes; broader runtime selection is not complete.                                                                                                                                         |
+| OpenAI-compatible Chat             | `src/protocols/openai-compatible-chat.ts`, `src/providers/openai-compatible.ts`                                                   | Usable for generic Chat and several profiles: Baseten, Cerebras, DeepInfra, DeepSeek, Fireworks, Groq, TogetherAI.                                                       | Family quirks are mostly endpoint defaults, not full typed behavior.                                                                                                                                                                                      |
+| Open Responses-compatible          | `src/protocols/open-responses.ts`, `src/protocols/openai-compatible-responses.ts`, `src/providers/openai-compatible-responses.ts` | Usable for deployments that implement the provider-neutral Open Responses protocol. The deployment adapter does not inherit OpenAI tools, events, metadata, or defaults. | No named family profiles or recorded deployment coverage yet.                                                                                                                                                                                             |
+| Anthropic-compatible Messages      | `src/protocols/anthropic-messages.ts`, `src/providers/anthropic-compatible.ts`                                                    | Usable for deployments that implement the Anthropic Messages wire protocol. Named Anthropic composes this base; MiniMax M3 has recorded text and tool-loop coverage.     | No named compatible family profiles yet.                                                                                                                                                                                                                  |
+| Anthropic Messages                 | `src/protocols/anthropic-messages.ts`, `src/providers/anthropic.ts`                                                               | Usable. Supports tools, thinking, cache control, images, server-hosted tool events, and usage.                                                                           | Provider option surface is small. Beta/header handling, metadata, and newer Messages fields need a typed parity pass.                                                                                                                                     |
+| Gemini Developer API               | `src/protocols/gemini.ts`, `src/providers/google.ts`                                                                              | Usable for Google API key flow. Supports text, images, tools, thinking signatures, and cache usage.                                                                      | This is not Vertex. Typed provider options are narrow; many Gemini request fields currently require raw `http.body` overlays.                                                                                                                             |
+| Vertex Gemini                      | `src/protocols/gemini.ts`, `src/providers/google-vertex.ts`                                                                       | Usable through API-key express mode, explicit OAuth tokens, or ADC with project/location endpoint derivation, including tuned `endpoints/...` deployments.               | Core runner/catalog mapping and recorded provider coverage are missing.                                                                                                                                                                                   |
+| Vertex Chat                        | `src/protocols/openai-chat.ts`, `src/providers/google-vertex-chat.ts`                                                             | Usable for MaaS models through OpenAI-compatible Chat Completions with explicit OAuth tokens or ADC and project/location endpoint derivation.                            | Core runner/catalog mapping and recorded provider coverage are missing; MaaS family-specific request parity needs review.                                                                                                                                 |
+| Vertex Responses                   | `src/protocols/open-responses.ts`, `src/providers/google-vertex-responses.ts`                                                     | Usable for Grok models through Open Responses with explicit OAuth tokens or ADC, project/location endpoint derivation, and an explicit `store: false` Vertex default.    | Core runner/catalog mapping and recorded provider coverage are missing; stateful continuation is not supported by Vertex.                                                                                                                                 |
+| Vertex Messages                    | `src/protocols/anthropic-messages.ts`, `src/providers/google-vertex-messages.ts`                                                  | Usable through explicit OAuth tokens or ADC, including global, regional, and `eu`/`us` multi-region endpoints.                                                           | Core runner/catalog mapping and recorded provider coverage are missing; Vertex-specific hosted-tool parity needs review.                                                                                                                                  |
+| Bedrock Converse                   | `src/protocols/bedrock-converse.ts`, `src/providers/amazon-bedrock.ts`                                                            | Partial but real. Supports AWS event-stream framing, SigV4 with supplied credentials, bearer auth, tools, reasoning signatures, media, cache points, and recorded tests. | Native facade does not mirror the AI SDK plugin's default AWS credential chain/profile behavior. Runner/catalog mapping is missing. Guardrails, inference profiles, region-specific model ID fixes, and model-specific request fields need a parity pass. |
+| Azure OpenAI                       | `src/providers/azure.ts` using OpenAI Chat/Responses protocols                                                                    | Partial. Supports resource/base URL setup, API key auth, API version query, Chat, and Responses selectors.                                                               | Core runner does not map `@ai-sdk/azure` to this native facade. AAD/token auth and Azure-specific endpoint variants need review.                                                                                                                          |
+| Cloudflare AI Gateway / Workers AI | `src/providers/cloudflare.ts`                                                                                                     | Present via OpenAI-compatible Chat routes.                                                                                                                               | Useful but not part of the critical AI SDK replacement set yet. Needs per-product recorded coverage before relying on it broadly.                                                                                                                         |
+| OpenRouter                         | `src/providers/openrouter.ts`                                                                                                     | Present with OpenRouter-specific usage/reasoning/prompt-cache options over Chat.                                                                                         | Responses-style OpenRouter support is absent.                                                                                                                                                                                                             |
+| xAI                                | `src/providers/xai.ts`                                                                                                            | Present with Responses and Chat selectors.                                                                                                                               | Needs package-parity review against the AI SDK xAI provider.                                                                                                                                                                                              |
+| GitHub Copilot                     | `src/providers/github-copilot.ts`                                                                                                 | Present as explicit-base-URL OpenAI Chat/Responses facade.                                                                                                               | Runtime/catalog integration remains specialized and should stay separate from public OpenAI-compatible defaults.                                                                                                                                          |
 
 ## V2 Runner Status
 
@@ -65,7 +65,7 @@ Other `aisdk:` packages, including Google Vertex, Azure, and Bedrock, currently
 ## Highest-Risk Gaps
 
 1. Runner support is narrower than the LLM package. The package has native provider facades for Google, Azure, and Bedrock, but the V2 Session runner only maps OpenAI, Anthropic, and explicit OpenAI-compatible Chat from `aisdk` catalog metadata.
-2. OpenAI-compatible Responses is available as a separate package entrypoint, but the V2 runner still maps `@ai-sdk/openai-compatible` to Chat only. Catalog selection must become API-aware before Responses deployments can use it.
+2. The Open Responses adapter is available through a separate package entrypoint, but the V2 runner still maps `@ai-sdk/openai-compatible` to Chat only. Catalog selection must become API-aware before Responses deployments can use it.
 3. Bedrock native auth is not AI SDK parity. The AI SDK plugin uses the default AWS provider chain, profile, container credentials, and Bedrock bearer token env behavior. Native Bedrock currently expects explicit credentials or bearer auth on the facade.
 4. Vertex Gemini, Vertex Chat, Vertex Responses, and Vertex Messages now have native package entrypoints, but the core runner does not map catalog metadata to them yet and recorded provider coverage is still missing.
 5. Azure is only a provider facade, not a full runtime replacement. Native Azure exists, but the catalog runner does not select it, and token auth/resource variants need review.
@@ -83,13 +83,13 @@ These are implementation/API slices, not separate npm packages.
 | OpenAI Chat                   | `@opencode-ai/ai/providers/openai/chat`                 | OpenAI `/chat/completions` semantics.                                        |
 | OpenAI Responses              | `@opencode-ai/ai/providers/openai/responses`            | OpenAI `/responses` semantics with HTTP/WebSocket selected through settings. |
 | OpenAI-compatible Chat        | `@opencode-ai/ai/providers/openai-compatible`           | Generic OpenAI-compatible `/chat/completions`.                               |
-| OpenAI-compatible Responses   | `@opencode-ai/ai/providers/openai-compatible/responses` | Generic OpenAI-compatible `/responses`.                                      |
+| Open Responses-compatible     | `@opencode-ai/ai/providers/openai-compatible/responses` | Generic provider-neutral `/responses`.                                       |
 | Anthropic-compatible Messages | `@opencode-ai/ai/providers/anthropic-compatible`        | Generic Anthropic-compatible `/messages`.                                    |
 | Anthropic Messages            | `@opencode-ai/ai/providers/anthropic`                   | Anthropic Messages API.                                                      |
 | Gemini Developer API          | `@opencode-ai/ai/providers/google`                      | Google AI Studio Gemini API.                                                 |
 | Vertex Gemini                 | `@opencode-ai/ai/providers/google-vertex/gemini`        | Vertex Gemini API; `providers/google-vertex` is the default alias.           |
 | Vertex Chat                   | `@opencode-ai/ai/providers/google-vertex/chat`          | Vertex OpenAI-compatible Chat Completions for MaaS models.                   |
-| Vertex Responses              | `@opencode-ai/ai/providers/google-vertex/responses`     | Vertex OpenAI-compatible Responses for Grok models.                          |
+| Vertex Responses              | `@opencode-ai/ai/providers/google-vertex/responses`     | Vertex Open Responses for Grok models.                                       |
 | Vertex Messages               | `@opencode-ai/ai/providers/google-vertex/messages`      | Vertex-hosted Anthropic Messages API.                                        |
 | Bedrock Converse              | `@opencode-ai/ai/providers/amazon-bedrock`              | AWS Bedrock Converse API.                                                    |
 | Bedrock Mantle                | Missing                                                 | AWS Bedrock Mantle OpenAI-compatible APIs.                                   |

packages/ai/example/call-sites.md

@@ -568,7 +568,7 @@ App boundary = explicit durable-config -> typed-provider call
       calling `.model(...)`.
 - [x] Remove request-shaping defaults from `Model`; selected models now carry only
       id, provider, and configured route while defaults live on routes or requests.
-- [x] Rework `LLMClient.prepare` / `stream` / `generate` to read
+- [x] Rework `LLMClient.stream` / `generate` to read
       `request.model.route` directly instead of calling `registeredRoute(...)`.
 - [x] Remove `Route.make(...)` global registration from the normal execution
       path; keep route ids only as diagnostics/provider API labels.

packages/ai/example/tutorial.ts

@@ -1,5 +1,5 @@
 import { Config, Effect, Formatter, Layer, Schema, Stream } from "effect"
-import { LLM, LLMClient, Message, ProviderID, Tool, ToolRuntime } from "@opencode-ai/ai"
+import { LLM, LLMClient, LLMRequest, Message, ProviderID, Tool, ToolRuntime } from "@opencode-ai/ai"
 import { Route, Auth, Endpoint, Framing, Protocol, RequestExecutor, WebSocketExecutor } from "@opencode-ai/ai/route"
 import { OpenAI } from "@opencode-ai/ai/providers"
 
@@ -50,18 +50,6 @@ const request = LLM.request({
   },
 })
 
-// `http` is intentionally not needed for normal calls. This shows the shape for
-// newly released provider fields before they deserve a typed provider option.
-const rawOverlayExample = LLM.request({
-  model,
-  prompt: "Show the final HTTP overlay shape.",
-  http: {
-    body: { metadata: { example: "tutorial" } },
-    headers: { "x-opencode-tutorial": "1" },
-    query: { debug: "1" },
-  },
-})
-
 // 3. `generate` sends the request and collects the event stream into one
 // response object. `response.text` is the collected text output.
 const generateOnce = Effect.gen(function* () {
@@ -78,7 +66,10 @@ const streamText = LLM.stream(request).pipe(
   Stream.tap((event) =>
     Effect.sync(() => {
       if (event.type === "text-delta") process.stdout.write(`\ntext: ${event.text}`)
-      if (event.type === "finish") process.stdout.write(`\nfinish: ${event.reason}\n`)
+      if (event.type === "finish")
+        process.stdout.write(
+          `\nfinish: ${event.reason.normalized}${event.reason.raw ? ` (${event.reason.raw})` : ""}\n`,
+        )
     }),
   ),
   Stream.runDrain,
@@ -113,7 +104,7 @@ const streamWithTools = Effect.gen(function* () {
 
     // A durable agent would persist these messages before starting another
     // raw model turn. This tutorial keeps the boundary visible instead.
-    const followUp = LLM.updateRequest(request, {
+    const followUp = LLMRequest.update(request, {
       messages: [
         ...request.messages,
         Message.assistant([event]),
@@ -194,7 +185,7 @@ const FakeProtocol = Protocol.make<FakeBody, string, string, void>({
     event: Schema.String,
     initial: () => undefined,
     step: (_, frame) => Effect.succeed([undefined, [{ type: "text-delta", id: "text-0", text: frame }]] as const),
-    onHalt: () => [{ type: "finish", reason: "stop" }],
+    onHalt: () => [{ type: "finish", reason: { normalized: "stop" } }],
   },
 })
 
@@ -219,33 +210,15 @@ const FakeEcho = {
   }),
 }
 
-// `LLMClient.prepare` is the lower-level inspection hook: it compiles through
-// body conversion, validation, endpoint, auth, and HTTP construction without
-// sending anything over the network.
-const inspectFakeProvider = Effect.gen(function* () {
-  const prepared = yield* LLMClient.prepare(
-    LLM.request({
-      model: FakeEcho.configure().model("tiny-echo"),
-      prompt: "Show me the provider pipeline.",
-    }),
-  )
-
-  console.log("\n== fake provider prepare ==")
-  console.log("route:", prepared.route)
-  console.log("body:", Formatter.formatJson(prepared.body, { space: 2 }))
-})
-
 // Provide the LLM runtime and the HTTP request executor once. Keep one path
-// enabled at a time so the tutorial can demonstrate generate, prepare, stream,
-// or tool-loop behavior without spending tokens on every example.
+// enabled at a time so the tutorial can demonstrate generate, stream, or
+// tool-loop behavior without spending tokens on every example.
 const requestExecutorLayer = RequestExecutor.fetchLayer
 const llmDeps = Layer.mergeAll(requestExecutorLayer, WebSocketExecutor.layer)
 const llmClientLayer = LLMClient.layer.pipe(Layer.provide(llmDeps))
 
 const program = Effect.gen(function* () {
   // yield* generateOnce
-  // yield* inspectFakeProvider
-  // yield* LLMClient.prepare(rawOverlayExample).pipe(Effect.andThen((prepared) => Effect.sync(() => console.log(prepared.body))))
   // yield* streamText
   // yield* generateStructuredObject
   // yield* generateDynamicObject.pipe(Effect.andThen((response) => Effect.sync(() => console.log(response.object))))

packages/ai/package.json

@@ -15,6 +15,7 @@
   ],
   "exports": {
     ".": "./src/index.ts",
+    "./testing": "./src/testing.ts",
     "./*": "./src/*.ts"
   },
   "devDependencies": {

packages/ai/src/cache-policy.ts

@@ -2,32 +2,31 @@
 // the policy designates. Runs once at compile time, before the per-protocol
 // body builder, so the existing inline-hint lowering path handles the rest.
 //
-// The default `"auto"` shape places one breakpoint at the last tool definition,
-// one at the last system part, and one at the latest user message. This
-// matches what production agent harnesses (LangChain's caching middleware,
-// kern-ai's 10x cost-reduction playbook) converge on for tool-use loops: the
-// latest user message stays put while a single turn explodes into many
-// assistant/tool round-trips, so caching at that boundary lets every
-// intra-turn API call hit the prefix.
+// The default `"auto"` shape places breakpoints at the last tool definition,
+// the first and last distinct system parts, and the conversation tail. This
+// exposes reusable tool, base-agent, project, and session prefixes while
+// advancing the tail after each tool result keeps the previous cache entry
+// within Anthropic's 20-block lookback during long agent turns.
 //
-// Manual `cache: CacheHint` placements on individual parts are preserved —
-// this function only fills gaps the caller left empty.
+// Manual `cache: CacheHint` placements on individual parts are preserved and
+// count against the four-breakpoint budget; auto only fills remaining slots.
 import { CacheHint, type CachePolicy, type CachePolicyObject } from "./schema/options"
 import { LLMRequest, Message, ToolDefinition, type ContentPart } from "./schema/messages"
 
 const AUTO: CachePolicyObject = {
   tools: true,
   system: true,
-  messages: "latest-user-message",
+  messages: { tail: 1 },
 }
 
 const NONE: CachePolicyObject = {}
+const BREAKPOINT_CAP = 4
 
 // Resolution rules:
 //   - undefined   → "auto" — caching is on by default. The math favors it:
 //                   Anthropic 5m-cache write is 1.25x base, read is 0.1x,
 //                   so a single reuse within 5 minutes already wins.
-//   - "auto"      → tools + system + latest user msg.
+//   - "auto"      → tools + first/last system + final message boundary.
 //   - "none"      → no auto placement; manual `CacheHint`s still flow.
 //   - object form → exactly what the caller asked for.
 const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
@@ -44,18 +43,32 @@ const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "bedrock-converse"]
 const makeHint = (ttlSeconds: number | undefined): CacheHint =>
   ttlSeconds !== undefined ? new CacheHint({ type: "ephemeral", ttlSeconds }) : new CacheHint({ type: "ephemeral" })
 
-const markLastTool = (tools: ReadonlyArray<ToolDefinition>, hint: CacheHint): ReadonlyArray<ToolDefinition> => {
+interface Budget {
+  remaining: number
+}
+
+const markLastTool = (
+  tools: ReadonlyArray<ToolDefinition>,
+  hint: CacheHint,
+  budget: Budget,
+): ReadonlyArray<ToolDefinition> => {
   if (tools.length === 0) return tools
   const last = tools.length - 1
-  if (tools[last]!.cache) return tools
+  if (tools[last]!.cache || budget.remaining === 0) return tools
+  budget.remaining -= 1
   return tools.map((tool, i) => (i === last ? new ToolDefinition({ ...tool, cache: hint }) : tool))
 }
 
-const markLastSystem = (system: LLMRequest["system"], hint: CacheHint): LLMRequest["system"] => {
+const markSystemBoundaries = (system: LLMRequest["system"], hint: CacheHint, budget: Budget): LLMRequest["system"] => {
   if (system.length === 0) return system
-  const last = system.length - 1
-  if (system[last]!.cache) return system
-  return system.map((part, i) => (i === last ? { ...part, cache: hint } : part))
+  let changed = false
+  const next = system.map((part, index) => {
+    if ((index !== 0 && index !== system.length - 1) || part.cache || budget.remaining === 0) return part
+    budget.remaining -= 1
+    changed = true
+    return { ...part, cache: hint }
+  })
+  return changed ? next : system
 }
 
 const lastIndexOfRole = (messages: ReadonlyArray<Message>, role: Message["role"]): number =>
@@ -64,14 +77,20 @@ const lastIndexOfRole = (messages: ReadonlyArray<Message>, role: Message["role"]
 // Mark the last text part of `messages[index]`. If no text part exists, mark
 // the last content part regardless of type — that's the breakpoint position
 // in tool-result-only messages too.
-const markMessageAt = (messages: ReadonlyArray<Message>, index: number, hint: CacheHint): ReadonlyArray<Message> => {
+const markMessageAt = (
+  messages: ReadonlyArray<Message>,
+  index: number,
+  hint: CacheHint,
+  budget: Budget,
+): ReadonlyArray<Message> => {
   if (index < 0 || index >= messages.length) return messages
   const target = messages[index]!
   if (target.content.length === 0) return messages
   const lastTextIndex = target.content.findLastIndex((part) => part.type === "text")
   const markAt = lastTextIndex >= 0 ? lastTextIndex : target.content.length - 1
   const existing = target.content[markAt]!
-  if ("cache" in existing && existing.cache) return messages
+  if (("cache" in existing && existing.cache) || budget.remaining === 0) return messages
+  budget.remaining -= 1
   const nextContent = target.content.map((part, i) => (i === markAt ? ({ ...part, cache: hint } as ContentPart) : part))
   const next = new Message({ ...target, content: nextContent })
   // Single pass over `messages`, substituting the one updated entry. Long
@@ -86,25 +105,42 @@ const markMessages = (
   messages: ReadonlyArray<Message>,
   strategy: NonNullable<CachePolicyObject["messages"]>,
   hint: CacheHint,
+  budget: Budget,
 ): ReadonlyArray<Message> => {
   if (messages.length === 0) return messages
-  if (strategy === "latest-user-message") return markMessageAt(messages, lastIndexOfRole(messages, "user"), hint)
-  if (strategy === "latest-assistant") return markMessageAt(messages, lastIndexOfRole(messages, "assistant"), hint)
+  if (strategy === "latest-user-message")
+    return markMessageAt(messages, lastIndexOfRole(messages, "user"), hint, budget)
+  if (strategy === "latest-assistant")
+    return markMessageAt(messages, lastIndexOfRole(messages, "assistant"), hint, budget)
   const start = Math.max(0, messages.length - strategy.tail)
   let next = messages
-  for (let i = start; i < messages.length; i++) next = markMessageAt(next, i, hint)
+  for (let i = start; i < messages.length; i++) next = markMessageAt(next, i, hint, budget)
   return next
 }
 
+const countHints = (request: LLMRequest) =>
+  request.tools.reduce((count, tool) => count + (tool.cache === undefined ? 0 : 1), 0) +
+  request.system.reduce((count, part) => count + (part.cache === undefined ? 0 : 1), 0) +
+  request.messages.reduce(
+    (count, message) =>
+      count +
+      message.content.reduce(
+        (contentCount, part) => contentCount + ("cache" in part && part.cache !== undefined ? 1 : 0),
+        0,
+      ),
+    0,
+  )
+
 export const applyCachePolicy = (request: LLMRequest): LLMRequest => {
   if (!RESPECTS_INLINE_HINTS.has(request.model.route.id)) return request
   const policy = resolve(request.cache)
   if (!policy.tools && !policy.system && !policy.messages) return request
 
   const hint = makeHint(policy.ttlSeconds)
-  const tools = policy.tools ? markLastTool(request.tools, hint) : request.tools
-  const system = policy.system ? markLastSystem(request.system, hint) : request.system
-  const messages = policy.messages ? markMessages(request.messages, policy.messages, hint) : request.messages
+  const budget = { remaining: Math.max(0, BREAKPOINT_CAP - countHints(request)) }
+  const tools = policy.tools ? markLastTool(request.tools, hint, budget) : request.tools
+  const system = policy.system ? markSystemBoundaries(request.system, hint, budget) : request.system
+  const messages = policy.messages ? markMessages(request.messages, policy.messages, hint, budget) : request.messages
 
   if (tools === request.tools && system === request.system && messages === request.messages) return request
   return LLMRequest.update(request, { tools, system, messages })

packages/ai/src/llm.ts

@@ -9,36 +9,28 @@ import {
   LLMRequest,
   LLMResponse,
   Message,
-  type ModelInput as SchemaModelInput,
+  Model,
   SystemPart,
   ToolChoice,
   ToolDefinition,
   type ContentPart,
-  ToolResultPart,
+  type ModelProviderOptions,
 } from "./schema"
 import { make as makeTool, toDefinitions, type ToolSchema } from "./tool"
 
-export type ModelInput = SchemaModelInput
-
-export type MessageInput = Message.Input
-
-export type ToolChoiceInput = ToolChoice.Input
-export type ToolChoiceMode = ToolChoice.Mode
-
-export type ToolResultInput = Parameters<typeof ToolResultPart.make>[0]
-
 /** Input accepted by `LLM.request`, normalized into the canonical `LLMRequest` class. */
-export type RequestInput = Omit<
+export type RequestInput<SelectedModel extends Model = Model> = Omit<
   ConstructorParameters<typeof LLMRequest>[0],
-  "system" | "messages" | "tools" | "toolChoice" | "generation" | "http" | "providerOptions"
+  "model" | "system" | "messages" | "tools" | "toolChoice" | "generation" | "http" | "providerOptions"
 > & {
+  readonly model: SelectedModel
   readonly system?: string | SystemPart | ReadonlyArray<SystemPart>
   readonly prompt?: string | ContentPart | ReadonlyArray<ContentPart>
-  readonly messages?: ReadonlyArray<Message | MessageInput>
+  readonly messages?: ReadonlyArray<Message | Message.Input>
   readonly tools?: ReadonlyArray<ToolDefinition.Input>
-  readonly toolChoice?: ToolChoiceInput
+  readonly toolChoice?: ToolChoice.Input
   readonly generation?: GenerationOptions.Input
-  readonly providerOptions?: ConstructorParameters<typeof LLMRequest>[0]["providerOptions"]
+  readonly providerOptions?: NoInfer<ModelProviderOptions<SelectedModel>>
   readonly http?: HttpOptions.Input
 }
 
@@ -46,11 +38,7 @@ export const generate = LLMClient.generate
 
 export const stream = LLMClient.stream
 
-export const requestInput = (input: LLMRequest): RequestInput => ({
-  ...LLMRequest.input(input),
-})
-
-export const request = (input: RequestInput) => {
+export const request = <const SelectedModel extends Model>(input: RequestInput<SelectedModel>) => {
   const {
     system: requestSystem,
     prompt,
@@ -74,14 +62,11 @@ export const request = (input: RequestInput) => {
   })
 }
 
-export const updateRequest = (input: LLMRequest, patch: Partial<RequestInput>) =>
-  request({ ...requestInput(input), ...patch })
-
 const GENERATE_OBJECT_TOOL_NAME = "generate_object"
 
 const GENERATE_OBJECT_TOOL_DESCRIPTION = "Return the structured result by calling this tool."
 
-type GenerateObjectBase = Omit<RequestInput, "tools" | "toolChoice" | "responseFormat">
+type GenerateObjectBase<SelectedModel extends Model = Model> = Omit<RequestInput<SelectedModel>, "tools" | "toolChoice">
 
 export class GenerateObjectResponse<T> {
   constructor(
@@ -98,11 +83,13 @@ export class GenerateObjectResponse<T> {
   }
 }
 
-export interface GenerateObjectOptions<S extends ToolSchema<any>> extends GenerateObjectBase {
+export interface GenerateObjectOptions<S extends ToolSchema<any>, SelectedModel extends Model = Model>
+  extends GenerateObjectBase<SelectedModel> {
   readonly schema: S
 }
 
-export interface GenerateObjectDynamicOptions extends GenerateObjectBase {
+export interface GenerateObjectDynamicOptions<SelectedModel extends Model = Model>
+  extends GenerateObjectBase<SelectedModel> {
   /** Raw JSON Schema object describing the expected output shape. */
   readonly jsonSchema: JsonSchema.JsonSchema
 }
@@ -155,11 +142,11 @@ const runGenerateObject = Effect.fn("LLM.generateObject")(function* (
  * 2. `jsonSchema: JsonSchema.JsonSchema` — `.object` is `unknown`. Use when
  *    the schema is only available at runtime (MCP, plugin manifests). Caller validates.
  */
-export function generateObject<S extends ToolSchema<any>>(
-  options: GenerateObjectOptions<S>,
+export function generateObject<const SelectedModel extends Model, S extends ToolSchema<any>>(
+  options: GenerateObjectOptions<S, SelectedModel>,
 ): Effect.Effect<GenerateObjectResponse<Schema.Schema.Type<S>>, LLMError>
-export function generateObject(
-  options: GenerateObjectDynamicOptions,
+export function generateObject<const SelectedModel extends Model>(
+  options: GenerateObjectDynamicOptions<SelectedModel>,
 ): Effect.Effect<GenerateObjectResponse<unknown>, LLMError>
 export function generateObject(options: GenerateObjectOptions<ToolSchema<any>> | GenerateObjectDynamicOptions) {
   if ("schema" in options) {

packages/ai/src/protocols/anthropic-messages.ts

@@ -1,4 +1,5 @@
 import { Effect, Schema } from "effect"
+import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client"
 import { Auth } from "../route/auth"
 import { Endpoint } from "../route/endpoint"
@@ -7,16 +8,18 @@ import { Protocol } from "../route/protocol"
 import {
   LLMError,
   LLMEvent,
+  mergeJsonRecords,
   Usage,
   type CacheHint,
+  type FinishReasonDetails,
   type FinishReason,
   type JsonSchema,
   type LLMRequest,
   type MediaPart,
+  type ProviderOptions,
   type ProviderMetadata,
   type ToolCallPart,
   type ToolDefinition,
-  type ToolContent,
   type ToolResultPart,
 } from "../schema"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared"
@@ -31,6 +34,29 @@ const MEDIA_MIMES = new Set<string>([...ProviderShared.IMAGE_MIMES, ...ProviderS
 export const DEFAULT_BASE_URL = "https://api.anthropic.com/v1"
 export const PATH = "/messages"
 
+export type ThinkingInput =
+  | {
+      readonly type: "adaptive"
+      readonly display?: "summarized" | "omitted"
+    }
+  | {
+      readonly type: "disabled"
+    }
+  | ({ readonly type: "enabled" } & (
+      | { readonly budgetTokens: number; readonly budget_tokens?: number }
+      | { readonly budgetTokens?: number; readonly budget_tokens: number }
+    ))
+
+export interface OptionsInput {
+  readonly [key: string]: unknown
+  readonly thinking?: ThinkingInput
+  readonly effort?: string
+}
+
+export type ProviderOptionsInput = ProviderOptions & {
+  readonly anthropic?: OptionsInput
+}
+
 // =============================================================================
 // Request Body Schema
 // =============================================================================
@@ -75,6 +101,15 @@ const AnthropicThinkingBlock = Schema.Struct({
   cache_control: Schema.optional(AnthropicCacheControl),
 })
 
+// Safety-filtered thinking arrives as an opaque encrypted `data` payload with
+// no visible text. It must round-trip verbatim so multi-turn thinking + tool
+// use conversations keep their reasoning continuity.
+const AnthropicRedactedThinkingBlock = Schema.Struct({
+  type: Schema.tag("redacted_thinking"),
+  data: Schema.String,
+  cache_control: Schema.optional(AnthropicCacheControl),
+})
+
 const AnthropicToolUseBlock = Schema.Struct({
   type: Schema.tag("tool_use"),
   id: Schema.String,
@@ -136,6 +171,7 @@ type AnthropicUserBlock = Schema.Schema.Type<typeof AnthropicUserBlock>
 const AnthropicAssistantBlock = Schema.Union([
   AnthropicTextBlock,
   AnthropicThinkingBlock,
+  AnthropicRedactedThinkingBlock,
   AnthropicToolUseBlock,
   AnthropicServerToolUseBlock,
   AnthropicServerToolResultBlock,
@@ -159,7 +195,7 @@ const AnthropicTool = Schema.Struct({
 type AnthropicTool = Schema.Schema.Type<typeof AnthropicTool>
 
 const AnthropicToolChoice = Schema.Union([
-  Schema.Struct({ type: Schema.Literals(["auto", "any"]) }),
+  Schema.Struct({ type: Schema.Literals(["auto", "any", "none"]) }),
   Schema.Struct({ type: Schema.tag("tool"), name: Schema.String }),
 ])
 
@@ -199,12 +235,27 @@ const AnthropicBodyFields = {
 export const AnthropicMessagesBody = Schema.Struct(AnthropicBodyFields)
 export type AnthropicMessagesBody = Schema.Schema.Type<typeof AnthropicMessagesBody>
 
-const AnthropicUsage = Schema.Struct({
-  input_tokens: Schema.optional(Schema.Number),
-  output_tokens: Schema.optional(Schema.Number),
-  cache_creation_input_tokens: optionalNull(Schema.Number),
-  cache_read_input_tokens: optionalNull(Schema.Number),
-})
+const AnthropicUsage = Schema.StructWithRest(
+  Schema.Struct({
+    input_tokens: Schema.optional(Schema.Number),
+    output_tokens: Schema.optional(Schema.Number),
+    cache_creation_input_tokens: optionalNull(Schema.Number),
+    cache_read_input_tokens: optionalNull(Schema.Number),
+    server_tool_use: optionalNull(
+      Schema.StructWithRest(
+        Schema.Struct({ web_search_requests: Schema.optional(Schema.Number) }),
+        [Schema.Record(Schema.String, Schema.Unknown)],
+      ),
+    ),
+    output_tokens_details: optionalNull(
+      Schema.StructWithRest(
+        Schema.Struct({ thinking_tokens: Schema.optional(Schema.Number) }),
+        [Schema.Record(Schema.String, Schema.Unknown)],
+      ),
+    ),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
 type AnthropicUsage = Schema.Schema.Type<typeof AnthropicUsage>
 
 const AnthropicStreamBlock = Schema.Struct({
@@ -214,6 +265,9 @@ const AnthropicStreamBlock = Schema.Struct({
   text: Schema.optional(Schema.String),
   thinking: Schema.optional(Schema.String),
   signature: Schema.optional(Schema.String),
+  // redacted_thinking blocks arrive whole in content_block_start with the
+  // encrypted payload in `data`; there is no streaming delta sequence.
+  data: Schema.optional(Schema.String),
   input: Schema.optional(Schema.Unknown),
   // *_tool_result blocks arrive whole as content_block_start (no streaming
   // delta) with the structured payload in `content` and the originating
@@ -251,7 +305,12 @@ type AnthropicEvent = Schema.Schema.Type<typeof AnthropicEvent>
 
 interface ParserState {
   readonly tools: ToolStream.State<number>
+  readonly reasoningSignatures: Readonly<Record<number, string>>
   readonly usage?: Usage
+  readonly pendingFinish?: {
+    readonly reason: FinishReasonDetails
+    readonly providerMetadata?: ProviderMetadata
+  }
   readonly lifecycle: Lifecycle.State
 }
 
@@ -287,6 +346,12 @@ const signatureFromMetadata = (metadata: ProviderMetadata | undefined): string |
   return typeof anthropic.signature === "string" ? anthropic.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 lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition, inputSchema: JsonSchema): AnthropicTool => ({
   name: tool.name,
   description: tool.description,
@@ -297,7 +362,7 @@ const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition, inputSc
 const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
   ProviderShared.matchToolChoice("Anthropic Messages", toolChoice, {
     auto: () => ({ type: "auto" as const }),
-    none: () => undefined,
+    none: () => ({ type: "none" as const }),
     required: () => ({ type: "any" as const }),
     tool: (name) => ({ type: "tool" as const, name }),
   })
@@ -330,7 +395,10 @@ const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult
   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}`)
-  return { type: wireType, tool_use_id: part.id, content: part.result.value } satisfies AnthropicServerToolResultBlock
+  // Prefer the provider-owned replay payload; fall back to the result value for
+  // histories constructed directly from provider events.
+  const payload = part.providerMetadata?.anthropic?.["result"] ?? part.result.value
+  return { type: wireType, tool_use_id: part.id, content: payload } satisfies AnthropicServerToolResultBlock
 })
 
 const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (part: MediaPart) {
@@ -357,7 +425,7 @@ const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (part: Me
 // Tool results may carry structured text, images, and documents. Keep media as provider-native
 // content instead of JSON-stringifying base64 into a prompt string.
 const lowerToolResultContentItem = Effect.fn("AnthropicMessages.lowerToolResultContentItem")(function* (
-  item: ToolContent,
+  item: Tool.Content,
 ) {
   if (item.type === "text") return { type: "text" as const, text: item.text } satisfies AnthropicTextBlock
   return yield* lowerMedia({ type: "media", mediaType: item.mime, data: item.uri, filename: item.name })
@@ -368,7 +436,7 @@ const lowerToolResultContent = Effect.fn("AnthropicMessages.lowerToolResultConte
   // with existing cassettes and provider expectations.
   if (part.result.type !== "content") return ProviderShared.toolResultText(part)
   // Preserve the narrowed array element type when compiled through a consumer package.
-  const content: ReadonlyArray<ToolContent> = part.result.value
+  const content: ReadonlyArray<Tool.Content> = part.result.value
   return yield* Effect.forEach(content, lowerToolResultContentItem)
 })
 
@@ -469,11 +537,16 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
           continue
         }
         if (part.type === "reasoning") {
-          content.push({
-            type: "thinking",
-            thinking: part.text,
-            signature: part.encrypted ?? signatureFromMetadata(part.providerMetadata),
-          })
+          // Mirrors Vercel's @ai-sdk/anthropic: a signature marks visible
+          // thinking; only signature-less parts carrying redactedData
+          // round-trip as opaque redacted_thinking blocks.
+          const signature = part.encrypted ?? signatureFromMetadata(part.providerMetadata)
+          const redactedData = redactedDataFromMetadata(part.providerMetadata)
+          if (signature === undefined && redactedData !== undefined) {
+            content.push({ type: "redacted_thinking", data: redactedData })
+            continue
+          }
+          content.push({ type: "thinking", thinking: part.text, signature })
           continue
         }
         if (part.type === "tool-call") {
@@ -510,39 +583,39 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
   return messages
 })
 
-const anthropicOptions = (request: LLMRequest) => request.providerOptions?.anthropic
+const resolveOptions = Effect.fn("AnthropicMessages.resolveOptions")(function* (request: LLMRequest) {
+  const input = request.providerOptions?.anthropic
+  return {
+    thinking: yield* resolveThinking(input?.thinking),
+    effort: typeof input?.effort === "string" ? input.effort : undefined,
+  }
+})
 
-const lowerThinking = Effect.fn("AnthropicMessages.lowerThinking")(function* (request: LLMRequest) {
-  const thinking = anthropicOptions(request)?.thinking
-  if (!ProviderShared.isRecord(thinking)) return undefined
-  if (thinking.type === "adaptive") {
+const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function* (input: unknown) {
+  if (!ProviderShared.isRecord(input)) return undefined
+  if (input.type === "adaptive") {
     const display =
-      thinking.display === "summarized"
+      input.display === "summarized"
         ? ("summarized" as const)
-        : thinking.display === "omitted"
+        : input.display === "omitted"
           ? ("omitted" as const)
           : undefined
     return { type: "adaptive" as const, ...(display === undefined ? {} : { display }) }
   }
-  if (thinking.type === "disabled") return { type: "disabled" as const }
-  if (thinking.type !== "enabled") return undefined
+  if (input.type === "disabled") return { type: "disabled" as const }
+  if (input.type !== "enabled") return undefined
   const budget =
-    typeof thinking.budgetTokens === "number"
-      ? thinking.budgetTokens
-      : typeof thinking.budget_tokens === "number"
-        ? thinking.budget_tokens
+    typeof input.budgetTokens === "number"
+      ? input.budgetTokens
+      : typeof input.budget_tokens === "number"
+        ? input.budget_tokens
         : undefined
-  if (budget === undefined) return yield* invalid("Anthropic thinking provider option requires budgetTokens")
+  if (budget === undefined)
+    return yield* ProviderShared.invalidRequest("Anthropic thinking provider option requires budgetTokens")
   return { type: "enabled" as const, budget_tokens: budget }
 })
 
-const outputConfig = (request: LLMRequest) => {
-  const effort = anthropicOptions(request)?.effort
-  return typeof effort === "string" ? { effort } : undefined
-}
-
 const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (request: LLMRequest) {
-  const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
   const generation = request.generation
   const toolSchemaCompatibility = request.model.compatibility?.toolSchema
   const outputLimit = request.model.defaults?.limits?.output ?? request.model.route.defaults.limits?.output ?? 4096
@@ -551,7 +624,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
   // over-mark we keep their tool hints and shed the message-tail ones first.
   const breakpoints = Cache.newBreakpoints(ANTHROPIC_BREAKPOINT_CAP)
   const tools =
-    request.tools.length === 0 || request.toolChoice?.type === "none"
+    request.tools.length === 0
       ? undefined
       : request.tools.map((tool) =>
           lowerTool(
@@ -560,6 +633,8 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
             ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility),
           ),
         )
+  // Anthropic rejects tool_choice when tools are absent; "none" is only meaningful with tools present.
+  const toolChoice = tools === undefined || !request.toolChoice ? undefined : yield* lowerToolChoice(request.toolChoice)
   const system =
     request.system.length === 0
       ? undefined
@@ -574,6 +649,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
       `Anthropic Messages: dropped ${breakpoints.dropped} cache breakpoint(s); the API allows at most ${ANTHROPIC_BREAKPOINT_CAP} per request.`,
     )
   }
+  const options = yield* resolveOptions(request)
   return {
     model: request.model.id,
     system,
@@ -586,8 +662,8 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
     top_p: generation?.topP,
     top_k: generation?.topK,
     stop_sequences: generation?.stop,
-    thinking: yield* lowerThinking(request),
-    output_config: outputConfig(request),
+    thinking: options.thinking,
+    output_config: options.effort === undefined ? undefined : { effort: options.effort },
   }
 })
 
@@ -596,7 +672,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
 // =============================================================================
 const mapFinishReason = (reason: string | null | undefined): FinishReason => {
   if (reason === "end_turn" || reason === "stop_sequence" || reason === "pause_turn") return "stop"
-  if (reason === "max_tokens") return "length"
+  if (reason === "max_tokens" || reason === "model_context_window_exceeded") return "length"
   if (reason === "tool_use") return "tool-calls"
   if (reason === "refusal") return "content-filter"
   return "unknown"
@@ -606,9 +682,8 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // `input_tokens` is the *non-cached* count per the Messages API docs, with
 // cache reads and writes as separate fields. We sum them to derive the
 // inclusive `inputTokens` the rest of the contract expects. Extended
-// thinking tokens are *not* broken out by Anthropic — they're billed as
-// part of `output_tokens`, so `reasoningTokens` stays `undefined` and
-// `outputTokens` carries the combined total.
+// 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 => {
   if (!usage) return undefined
   const nonCached = usage.input_tokens
@@ -621,6 +696,7 @@ const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
     nonCachedInputTokens: nonCached,
     cacheReadInputTokens: cacheRead,
     cacheWriteInputTokens: cacheWrite,
+    reasoningTokens: usage.output_tokens_details?.thinking_tokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, usage.output_tokens, undefined),
     providerMetadata: { anthropic: usage },
   })
@@ -639,18 +715,18 @@ const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
   const cacheWriteInputTokens = right.cacheWriteInputTokens ?? left.cacheWriteInputTokens
   const inputTokens = ProviderShared.sumTokens(nonCachedInputTokens, cacheReadInputTokens, cacheWriteInputTokens)
   const outputTokens = right.outputTokens ?? left.outputTokens
+  const reasoningTokens = right.reasoningTokens ?? left.reasoningTokens
   return new Usage({
     inputTokens,
     outputTokens,
     nonCachedInputTokens,
     cacheReadInputTokens,
     cacheWriteInputTokens,
+    reasoningTokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, outputTokens, undefined),
     providerMetadata: {
-      anthropic: {
-        ...left.providerMetadata?.["anthropic"],
-        ...right.providerMetadata?.["anthropic"],
-      },
+      anthropic:
+        mergeJsonRecords(left.providerMetadata?.["anthropic"], right.providerMetadata?.["anthropic"]) ?? {},
     },
   })
 }
@@ -680,7 +756,9 @@ const serverToolResultEvent = (block: NonNullable<AnthropicEvent["content_block"
     name: SERVER_TOOL_RESULT_NAMES[block.type],
     result: isError ? { type: "error", value: block.content } : { type: "json", value: block.content },
     providerExecuted: true,
-    providerMetadata: anthropicMetadata({ blockType: block.type }),
+    // 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 }),
   })
 }
 
@@ -707,6 +785,10 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
         tools: ToolStream.start(state.tools, event.index, {
           id: block.id ?? String(event.index),
           name: block.name ?? "",
+          input:
+            block.input !== undefined && (!ProviderShared.isRecord(block.input) || Object.keys(block.input).length > 0)
+              ? ProviderShared.encodeJson(block.input)
+              : undefined,
           providerExecuted: block.type === "server_tool_use",
         }),
       },
@@ -721,20 +803,50 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
     ]
   }
 
-  if (block.type === "text" && block.text) {
+  if (block.type === "text" && block.text !== undefined) {
     const events: LLMEvent[] = []
+    const id = `text-${event.index ?? 0}`
+    const lifecycle = Lifecycle.textStart(state.lifecycle, events, id)
     return [
-      { ...state, lifecycle: Lifecycle.textDelta(state.lifecycle, events, `text-${event.index ?? 0}`, block.text) },
+      { ...state, lifecycle: block.text ? Lifecycle.textDelta(lifecycle, events, id, block.text) : lifecycle },
       events,
     ]
   }
 
-  if (block.type === "thinking" && block.thinking) {
+  if (block.type === "thinking" && block.thinking !== undefined) {
+    const events: LLMEvent[] = []
+    const id = `reasoning-${event.index ?? 0}`
+    const providerMetadata = block.signature === undefined ? undefined : anthropicMetadata({ signature: block.signature })
+    const lifecycle = Lifecycle.reasoningStart(state.lifecycle, events, id, providerMetadata)
+    return [
+      {
+        ...state,
+        lifecycle: block.thinking
+          ? Lifecycle.reasoningDelta(lifecycle, events, id, block.thinking, providerMetadata)
+          : lifecycle,
+        reasoningSignatures:
+          event.index === undefined || block.signature === undefined
+            ? state.reasoningSignatures
+            : { ...state.reasoningSignatures, [event.index]: block.signature },
+      },
+      events,
+    ]
+  }
+
+  // Redacted thinking surfaces as an empty reasoning part carrying the opaque
+  // payload as `redactedData` metadata (same model as Vercel's
+  // @ai-sdk/anthropic). The existing content_block_stop closes the part.
+  if (block.type === "redacted_thinking" && block.data !== undefined) {
     const events: LLMEvent[] = []
     return [
       {
         ...state,
-        lifecycle: Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${event.index ?? 0}`, block.thinking),
+        lifecycle: Lifecycle.reasoningStart(
+          state.lifecycle,
+          events,
+          `reasoning-${event.index ?? 0}`,
+          anthropicMetadata({ redactedData: block.data }),
+        ),
       },
       events,
     ]
@@ -772,18 +884,13 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
   }
 
   if (delta?.type === "signature_delta" && delta.signature) {
-    const events: LLMEvent[] = []
+    const index = event.index ?? 0
     return [
       {
         ...state,
-        lifecycle: Lifecycle.reasoningEnd(
-          state.lifecycle,
-          events,
-          `reasoning-${event.index ?? 0}`,
-          anthropicMetadata({ signature: delta.signature }),
-        ),
+        reasoningSignatures: { ...state.reasoningSignatures, [index]: delta.signature },
       },
-      events,
+      NO_EVENTS,
     ] satisfies StepResult
   }
 
@@ -814,28 +921,53 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
   const result = yield* ToolStream.finish(ADAPTER, state.tools, event.index)
   const events: LLMEvent[] = []
   const resultEvents = result.events ?? []
+  const signature = state.reasoningSignatures[event.index]
   const lifecycle = resultEvents.length
     ? Lifecycle.stepStart(state.lifecycle, events)
     : Lifecycle.reasoningEnd(
         Lifecycle.textEnd(state.lifecycle, events, `text-${event.index}`),
         events,
         `reasoning-${event.index}`,
+        signature === undefined ? undefined : anthropicMetadata({ signature }),
       )
   events.push(...resultEvents)
-  return [{ ...state, lifecycle, tools: result.tools }, events] satisfies StepResult
+  const reasoningSignatures = { ...state.reasoningSignatures }
+  delete reasoningSignatures[event.index]
+  return [{ ...state, lifecycle, tools: result.tools, reasoningSignatures }, events] satisfies StepResult
 })
 
 const onMessageDelta = (state: ParserState, event: AnthropicEvent): StepResult => {
   const usage = mergeUsage(state.usage, mapUsage(event.usage))
+  return [
+    {
+      ...state,
+      usage,
+      pendingFinish: {
+        reason: {
+          normalized: mapFinishReason(event.delta?.stop_reason),
+          raw: event.delta?.stop_reason ?? undefined,
+        },
+        providerMetadata:
+          event.delta?.stop_sequence === null || event.delta?.stop_sequence === undefined
+            ? undefined
+            : anthropicMetadata({ stopSequence: event.delta.stop_sequence }),
+      },
+    },
+    NO_EVENTS,
+  ]
+}
+
+const onMessageStop = (state: ParserState): StepResult => {
   const events: LLMEvent[] = []
   const lifecycle = Lifecycle.finish(state.lifecycle, events, {
-    reason: mapFinishReason(event.delta?.stop_reason),
-    usage,
-    providerMetadata: event.delta?.stop_sequence
-      ? anthropicMetadata({ stopSequence: event.delta.stop_sequence })
-      : undefined,
+    reason: state.pendingFinish?.reason ?? {
+      normalized: "unknown",
+      raw: undefined,
+    },
+    usage: state.usage,
+    providerMetadata: state.pendingFinish?.providerMetadata,
   })
-  return [{ ...state, lifecycle, usage }, events]
+  return [{ ...state, lifecycle }, events]
 }
 
 // Prefix `error.type` so overloads, rate limits, and quota errors are visible
@@ -860,6 +992,7 @@ const step = (state: ParserState, event: AnthropicEvent) => {
   if (event.type === "content_block_delta") return onContentBlockDelta(state, event)
   if (event.type === "content_block_stop") return onContentBlockStop(state, event)
   if (event.type === "message_delta") return Effect.succeed(onMessageDelta(state, event))
+  if (event.type === "message_stop") return Effect.succeed(onMessageStop(state))
   if (event.type === "error") return onError(event)
   return Effect.succeed<StepResult>([state, NO_EVENTS])
 }
@@ -880,7 +1013,11 @@ export const protocol = Protocol.make({
   },
   stream: {
     event: Protocol.jsonEvent(AnthropicEvent),
-    initial: () => ({ tools: ToolStream.empty<number>(), lifecycle: Lifecycle.initial() }),
+    initial: () => ({
+      tools: ToolStream.empty<number>(),
+      reasoningSignatures: {},
+      lifecycle: Lifecycle.initial(),
+    }),
     step,
   },
 })

packages/ai/src/protocols/bedrock-converse.ts

@@ -8,6 +8,7 @@ import {
   Usage,
   type CacheHint,
   type FinishReason,
+  type FinishReasonDetails,
   type JsonSchema,
   type LLMRequest,
   type ModelToolSchemaCompatibility,
@@ -65,14 +66,15 @@ const BedrockToolResultBlock = Schema.Struct({
 type BedrockToolResultBlock = Schema.Schema.Type<typeof BedrockToolResultBlock>
 
 const BedrockReasoningBlock = Schema.Struct({
-  reasoningContent: Schema.Struct({
-    reasoningText: Schema.optional(
-      Schema.Struct({
+  reasoningContent: Schema.Union([
+    Schema.Struct({
+      reasoningText: Schema.Struct({
         text: Schema.String,
         signature: Schema.optional(Schema.String),
       }),
-    ),
-  }),
+    }),
+    Schema.Struct({ redactedContent: Schema.String }),
+  ]),
 })
 
 const BedrockUserBlock = Schema.Union([
@@ -153,6 +155,12 @@ const BedrockUsageSchema = Schema.Struct({
 })
 type BedrockUsageSchema = Schema.Schema.Type<typeof BedrockUsageSchema>
 
+const BedrockStreamException = Schema.Struct({
+  message: Schema.optional(Schema.String),
+  originalMessage: Schema.optional(Schema.String),
+  originalStatusCode: Schema.optional(Schema.Number),
+})
+
 // Streaming event shape — the AWS event stream wraps each JSON payload by its
 // `:event-type` header (e.g. `messageStart`, `contentBlockDelta`). We
 // reconstruct that wrapping in `decodeFrames` below so the event schema can
@@ -180,6 +188,11 @@ const BedrockEvent = Schema.Struct({
             Schema.Struct({
               text: Schema.optional(Schema.String),
               signature: Schema.optional(Schema.String),
+              // Blob fields in Bedrock's JSON event stream are base64 strings.
+              redactedContent: Schema.optional(Schema.String),
+              // Vercel's Bedrock provider exposes the same delta under
+              // Anthropic's shorter `data` spelling.
+              data: Schema.optional(Schema.String),
             }),
           ),
         }),
@@ -199,11 +212,11 @@ const BedrockEvent = Schema.Struct({
       metrics: Schema.optional(Schema.Unknown),
     }),
   ),
-  internalServerException: Schema.optional(Schema.Struct({ message: Schema.String })),
-  modelStreamErrorException: Schema.optional(Schema.Struct({ message: Schema.String })),
-  validationException: Schema.optional(Schema.Struct({ message: Schema.String })),
-  throttlingException: Schema.optional(Schema.Struct({ message: Schema.String })),
-  serviceUnavailableException: Schema.optional(Schema.Struct({ message: Schema.String })),
+  internalServerException: Schema.optional(BedrockStreamException),
+  modelStreamErrorException: Schema.optional(BedrockStreamException),
+  validationException: Schema.optional(BedrockStreamException),
+  throttlingException: Schema.optional(BedrockStreamException),
+  serviceUnavailableException: Schema.optional(BedrockStreamException),
 })
 type BedrockEvent = Schema.Schema.Type<typeof BedrockEvent>
 
@@ -259,6 +272,13 @@ const reasoningSignature = (part: ReasoningPart) => {
   )
 }
 
+const reasoningRedactedData = (part: ReasoningPart) => {
+  const bedrock = part.providerMetadata?.bedrock
+  return ProviderShared.isRecord(bedrock) && typeof bedrock.redactedData === "string"
+    ? bedrock.redactedData
+    : undefined
+}
+
 const lowerToolCall = (part: ToolCallPart): BedrockToolUseBlock => ({
   toolUse: {
     toolUseId: part.id,
@@ -348,11 +368,13 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
           continue
         }
         if (part.type === "reasoning") {
-          content.push({
-            reasoningContent: {
-              reasoningText: { text: part.text, signature: reasoningSignature(part) },
-            },
-          })
+          const signature = reasoningSignature(part)
+          const redactedData = reasoningRedactedData(part)
+          if (signature === undefined && redactedData !== undefined) {
+            content.push({ reasoningContent: { redactedContent: redactedData } })
+            continue
+          }
+          content.push({ reasoningContent: { reasoningText: { text: part.text, signature } } })
           continue
         }
         if (part.type === "tool-call") {
@@ -392,8 +414,13 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
   // tools → system → messages order to favour the highest-impact prefixes.
   const breakpoints = BedrockCache.breakpoints()
   const toolConfig =
-    request.tools.length > 0 && request.toolChoice?.type !== "none"
-      ? { tools: lowerTools(request.model.compatibility?.toolSchema, breakpoints, request.tools), toolChoice }
+    request.tools.length > 0
+      ? {
+          tools: lowerTools(request.model.compatibility?.toolSchema, breakpoints, request.tools),
+          // Converse has no native "none". Keep definitions stable for prompt
+          // caching and omit only the unsupported choice.
+          toolChoice,
+        }
       : undefined
   const system = request.system.length === 0 ? undefined : lowerSystem(breakpoints, request.system)
   const messages = yield* lowerMessages(request, breakpoints)
@@ -430,9 +457,10 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
 // =============================================================================
 const mapFinishReason = (reason: string): FinishReason => {
   if (reason === "end_turn" || reason === "stop_sequence") return "stop"
-  if (reason === "max_tokens") return "length"
+  if (reason === "max_tokens" || reason === "model_context_window_exceeded") return "length"
   if (reason === "tool_use") return "tool-calls"
   if (reason === "content_filtered" || reason === "guardrail_intervened") return "content-filter"
+  if (reason === "malformed_model_output" || reason === "malformed_tool_use") return "error"
   return "unknown"
 }
 
@@ -461,7 +489,7 @@ interface ParserState {
   // 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.
-  readonly pendingFinish: { readonly reason: FinishReason; readonly usage?: Usage } | undefined
+  readonly pendingFinish: { readonly reason: FinishReasonDetails; readonly usage?: Usage } | undefined
   readonly hasToolCalls: boolean
   readonly lifecycle: Lifecycle.State
   readonly reasoningSignatures: Readonly<Record<number, string>>
@@ -512,12 +540,26 @@ const step = (state: ParserState, event: BedrockEvent) =>
       const index = event.contentBlockDelta.contentBlockIndex
       const reasoning = event.contentBlockDelta.delta.reasoningContent
       const events: LLMEvent[] = []
+      const redactedData = reasoning.redactedContent ?? reasoning.data
+      const providerMetadata = reasoning.signature
+        ? bedrockMetadata({ signature: reasoning.signature })
+        : redactedData !== undefined
+          ? bedrockMetadata({ redactedData })
+          : undefined
+      const lifecycle =
+        reasoning.text !== undefined || providerMetadata !== undefined
+          ? Lifecycle.reasoningDelta(
+              state.lifecycle,
+              events,
+              `reasoning-${index}`,
+              reasoning.text ?? "",
+              providerMetadata,
+            )
+          : state.lifecycle
       return [
         {
           ...state,
-          lifecycle: reasoning.text
-            ? Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text)
-            : state.lifecycle,
+          lifecycle,
           reasoningSignatures: reasoning.signature
             ? { ...state.reasoningSignatures, [index]: reasoning.signature }
             : state.reasoningSignatures,
@@ -578,15 +620,30 @@ const step = (state: ParserState, event: BedrockEvent) =>
       return [
         {
           ...state,
-          pendingFinish: { reason: mapFinishReason(event.messageStop.stopReason), usage: state.pendingFinish?.usage },
+          pendingFinish: {
+            reason: {
+              normalized: mapFinishReason(event.messageStop.stopReason),
+              raw: event.messageStop.stopReason,
+            },
+            usage: state.pendingFinish?.usage,
+          },
         },
         [],
       ] as const
     }
 
     if (event.metadata) {
-      const usage = mapUsage(event.metadata.usage)
-      return [{ ...state, pendingFinish: { reason: state.pendingFinish?.reason ?? "stop", usage } }, []] as const
+      const usage = mapUsage(event.metadata.usage) ?? state.pendingFinish?.usage
+      return [
+        {
+          ...state,
+          pendingFinish: {
+            reason: state.pendingFinish?.reason ?? { normalized: "stop" },
+            usage,
+          },
+        },
+        [],
+      ] as const
     }
 
     const exception = (
@@ -603,7 +660,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
         module: ADAPTER,
         method: "stream",
         reason: classifyProviderFailure({
-          message: exception[1]?.message ?? "Bedrock Converse stream error",
+          message: exception[1]?.message ?? exception[1]?.originalMessage ?? "Bedrock Converse stream error",
           code: exception[0],
         }),
       })
@@ -619,8 +676,13 @@ const onHalt = (state: ParserState): ReadonlyArray<LLMEvent> =>
     ? (() => {
         const events: LLMEvent[] = []
         Lifecycle.finish(state.lifecycle, events, {
-          reason:
-            state.pendingFinish.reason === "stop" && state.hasToolCalls ? "tool-calls" : state.pendingFinish.reason,
+          reason: {
+            ...state.pendingFinish.reason,
+            normalized:
+              state.pendingFinish.reason.normalized === "stop" && state.hasToolCalls
+                ? "tool-calls"
+                : state.pendingFinish.reason.normalized,
+          },
           usage: state.pendingFinish.usage,
         })
         return events

packages/ai/src/protocols/bedrock-event-stream.ts

@@ -53,8 +53,22 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
       })
       cursor = { buffer: cursor.buffer, offset: cursor.offset + totalLength }
 
-      if (decoded.headers[":message-type"]?.value !== "event") continue
-      const eventType = decoded.headers[":event-type"]?.value
+      const messageType = decoded.headers[":message-type"]?.value
+      if (messageType === "error") {
+        const code = decoded.headers[":error-code"]?.value
+        const message = decoded.headers[":error-message"]?.value
+        return yield* ProviderShared.eventError(
+          route,
+          [code, message].filter((value): value is string => typeof value === "string").join(": ") ||
+            "Bedrock Converse event-stream error",
+        )
+      }
+      const eventType =
+        messageType === "event"
+          ? decoded.headers[":event-type"]?.value
+          : messageType === "exception"
+            ? decoded.headers[":exception-type"]?.value
+            : undefined
       if (typeof eventType !== "string") continue
       const payload = utf8.decode(decoded.body)
       if (!payload) continue

packages/ai/src/protocols/gemini.ts

@@ -1,4 +1,5 @@
 import { Effect, Schema } from "effect"
+import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client"
 import { Auth } from "../route/auth"
 import { Endpoint } from "../route/endpoint"
@@ -11,11 +12,11 @@ import {
   type JsonSchema,
   type LLMRequest,
   type MediaPart,
+  type ProviderOptions,
   type ProviderMetadata,
   type TextPart,
   type ToolCallPart,
   type ToolDefinition,
-  type ToolContent,
 } from "../schema"
 import { JsonObject, optionalArray, ProviderShared } from "./shared"
 import { GeminiToolSchema } from "./utils/gemini-tool-schema"
@@ -26,6 +27,18 @@ const ADAPTER = "gemini"
 const MEDIA_MIMES = new Set<string>(ProviderShared.MEDIA_MIMES)
 export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
 
+export interface OptionsInput {
+  readonly [key: string]: unknown
+  readonly thinkingConfig?: {
+    readonly thinkingBudget?: number
+    readonly includeThoughts?: boolean
+  }
+}
+
+export type ProviderOptionsInput = ProviderOptions & {
+  readonly gemini?: OptionsInput
+}
+
 // =============================================================================
 // Request Body Schema
 // =============================================================================
@@ -203,7 +216,9 @@ const thoughtSignature = (providerMetadata: ProviderMetadata | undefined) => {
 
 const functionCallId = (providerMetadata: ProviderMetadata | undefined) => {
   const google = providerMetadata?.google
-  return ProviderShared.isRecord(google) && typeof google.functionCallId === "string" ? google.functionCallId : undefined
+  return ProviderShared.isRecord(google) && typeof google.functionCallId === "string"
+    ? google.functionCallId
+    : undefined
 }
 
 const lowerToolCall = (part: ToolCallPart) => ({
@@ -274,7 +289,7 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
         })
         continue
       }
-      const content: ReadonlyArray<ToolContent> = part.result.value
+      const content: ReadonlyArray<Tool.Content> = part.result.value
       const text = content.filter((item) => item.type === "text").map((item) => item.text)
       const media: GeminiInlineDataPart[] = []
       for (const item of content) {
@@ -300,21 +315,22 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
   return contents
 })
 
-const geminiOptions = (request: LLMRequest) => request.providerOptions?.gemini
-
-const thinkingConfig = (request: LLMRequest) => {
-  const value = geminiOptions(request)?.thinkingConfig
-  if (!ProviderShared.isRecord(value)) return undefined
-  const result = {
+const resolveOptions = (request: LLMRequest) => {
+  const value = request.providerOptions?.gemini?.thinkingConfig
+  if (!ProviderShared.isRecord(value)) return {}
+  const thinkingConfig = {
     thinkingBudget: typeof value.thinkingBudget === "number" ? value.thinkingBudget : undefined,
     includeThoughts: typeof value.includeThoughts === "boolean" ? value.includeThoughts : undefined,
   }
-  return Object.values(result).some((item) => item !== undefined) ? result : undefined
+  return {
+    thinkingConfig: Object.values(thinkingConfig).some((item) => item !== undefined) ? thinkingConfig : undefined,
+  }
 }
 
 const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMRequest) {
-  const toolsEnabled = request.tools.length > 0 && request.toolChoice?.type !== "none"
+  const hasTools = request.tools.length > 0
   const generation = request.generation
+  const options = resolveOptions(request)
   const toolSchemaCompatibility = request.model.compatibility?.toolSchema
   const generationConfig = {
     maxOutputTokens: generation?.maxTokens,
@@ -322,14 +338,14 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
     topP: generation?.topP,
     topK: generation?.topK,
     stopSequences: generation?.stop,
-    thinkingConfig: thinkingConfig(request),
+    thinkingConfig: options.thinkingConfig,
   }
 
   return {
     contents: yield* lowerMessages(request),
     systemInstruction:
       request.system.length === 0 ? undefined : { parts: [{ text: ProviderShared.joinText(request.system) }] },
-    tools: toolsEnabled
+    tools: hasTools
       ? [
           {
             functionDeclarations: request.tools.map((tool) =>
@@ -338,7 +354,7 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
           },
         ]
       : undefined,
-    toolConfig: toolsEnabled && request.toolChoice ? yield* lowerToolConfig(request.toolChoice) : undefined,
+    toolConfig: hasTools && request.toolChoice ? yield* lowerToolConfig(request.toolChoice) : undefined,
     generationConfig: Object.values(generationConfig).some((value) => value !== undefined)
       ? generationConfig
       : undefined,
@@ -382,10 +398,22 @@ const mapFinishReason = (finishReason: string | undefined, hasToolCalls: boolean
     finishReason === "SAFETY" ||
     finishReason === "BLOCKLIST" ||
     finishReason === "PROHIBITED_CONTENT" ||
-    finishReason === "SPII"
+    finishReason === "SPII" ||
+    finishReason === "MODEL_ARMOR" ||
+    finishReason === "IMAGE_PROHIBITED_CONTENT" ||
+    finishReason === "IMAGE_RECITATION" ||
+    finishReason === "LANGUAGE"
   )
     return "content-filter"
-  if (finishReason === "MALFORMED_FUNCTION_CALL") return "error"
+  if (
+    finishReason === "MALFORMED_FUNCTION_CALL" ||
+    finishReason === "UNEXPECTED_TOOL_CALL" ||
+    finishReason === "NO_IMAGE" ||
+    finishReason === "TOO_MANY_TOOL_CALLS" ||
+    finishReason === "MISSING_THOUGHT_SIGNATURE" ||
+    finishReason === "MALFORMED_RESPONSE"
+  )
+    return "error"
   return "unknown"
 }
 
@@ -402,7 +430,10 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> =>
             )
           : state.lifecycle
         Lifecycle.finish(lifecycle, events, {
-          reason: mapFinishReason(state.finishReason, state.hasToolCalls),
+          reason: {
+            normalized: mapFinishReason(state.finishReason, state.hasToolCalls),
+            raw: state.finishReason,
+          },
           usage: state.usage,
         })
         return events

packages/ai/src/protocols/index.ts

@@ -6,3 +6,4 @@ export * as OpenAIImages from "./openai-images"
 export * as OpenAICompatibleChat from "./openai-compatible-chat"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses"
 export * as OpenAIResponses from "./openai-responses"
+export * as OpenResponses from "./open-responses"

packages/ai/src/protocols/openai-chat.ts

@@ -1,13 +1,16 @@
 import { Effect, Schema } from "effect"
+import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client"
 import { Auth } from "../route/auth"
 import { Endpoint } from "../route/endpoint"
 import { HttpTransport } from "../route/transport"
 import { Protocol } from "../route/protocol"
 import {
+  LLMError,
   LLMEvent,
   Usage,
   type FinishReason,
+  type FinishReasonDetails,
   type JsonSchema,
   type LLMRequest,
   type MediaPart,
@@ -15,8 +18,8 @@ import {
   type TextPart,
   type ToolCallPart,
   type ToolDefinition,
-  type ToolContent,
 } from "../schema"
+import { classifyProviderFailure } from "../provider-error"
 import { isRecord, JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared"
 import { OpenAIOptions } from "./utils/openai-options"
 import { Lifecycle } from "./utils/lifecycle"
@@ -128,6 +131,7 @@ const OpenAIChatUsage = Schema.Struct({
   prompt_tokens_details: optionalNull(
     Schema.Struct({
       cached_tokens: Schema.optional(Schema.Number),
+      cache_write_tokens: Schema.optional(Schema.Number),
     }),
   ),
   completion_tokens_details: optionalNull(
@@ -164,11 +168,18 @@ const OpenAIChatDelta = Schema.StructWithRest(
 const OpenAIChatChoice = Schema.Struct({
   delta: optionalNull(OpenAIChatDelta),
   finish_reason: optionalNull(Schema.String),
+  native_finish_reason: optionalNull(Schema.String),
+})
+
+const OpenAIChatError = Schema.Struct({
+  code: optionalNull(Schema.Union([Schema.String, Schema.Number])),
+  message: Schema.String,
 })
 
 export const OpenAIChatEvent = Schema.Struct({
-  choices: Schema.Array(OpenAIChatChoice),
+  choices: optionalNull(Schema.Array(OpenAIChatChoice)),
   usage: optionalNull(OpenAIChatUsage),
+  error: optionalNull(OpenAIChatError),
 })
 export type OpenAIChatEvent = Schema.Schema.Type<typeof OpenAIChatEvent>
 type OpenAIChatRequestMessage = LLMRequest["messages"][number]
@@ -184,7 +195,7 @@ export interface ParserState {
   readonly pendingTools: Partial<Record<number, PendingToolDelta>>
   readonly toolCallEvents: ReadonlyArray<LLMEvent>
   readonly usage?: Usage
-  readonly finishReason?: FinishReason
+  readonly finishReason?: FinishReasonDetails
   readonly lifecycle: Lifecycle.State
   readonly reasoningField?: string
   readonly reasoningDetails: Array<unknown>
@@ -324,7 +335,7 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (m
       messages.push({ role: "tool", tool_call_id: part.id, content: ProviderShared.toolResultText(part) })
       continue
     }
-    const content: ReadonlyArray<ToolContent> = part.result.value
+    const content: ReadonlyArray<Tool.Content> = part.result.value
     const text = content.filter((item) => item.type === "text").map((item) => item.text)
     messages.push({ role: "tool", tool_call_id: part.id, content: text.join("\n") })
     const files = content.filter((item) => item.type === "file")
@@ -386,14 +397,13 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
   return messages
 })
 
-const lowerOptions = Effect.fn("OpenAIChat.lowerOptions")(function* (request: LLMRequest) {
-  const store = OpenAIOptions.store(request)
-  const reasoningEffort = OpenAIOptions.reasoningEffort(request)
+const lowerOptions = (request: LLMRequest) => {
+  const options = OpenAIOptions.resolve(request)
   return {
-    ...(store !== undefined ? { store } : {}),
-    ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),
+    ...(options.store !== undefined ? { store: options.store } : {}),
+    ...(options.reasoningEffort ? { reasoning_effort: options.reasoningEffort } : {}),
   }
-})
+}
 
 const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (request: LLMRequest) {
   // `fromRequest` returns the provider body only. Endpoint, auth, framing,
@@ -424,7 +434,7 @@ const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (request: LLMR
     presence_penalty: generation?.presencePenalty,
     seed: generation?.seed,
     stop: generation?.stop,
-    ...(yield* lowerOptions(request)),
+    ...lowerOptions(request),
   }
 })
 
@@ -439,24 +449,27 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
   if (reason === "length") return "length"
   if (reason === "content_filter") return "content-filter"
   if (reason === "function_call" || reason === "tool_calls") return "tool-calls"
+  if (reason === "error") return "error"
   return "unknown"
 }
 
 // OpenAI Chat reports `prompt_tokens` (inclusive total) with a
-// `cached_tokens` subset, and `completion_tokens` (inclusive total) with
-// a `reasoning_tokens` subset. We pass the inclusive totals through and
-// derive the non-cached breakdown so the `LLM.Usage` contract is
+// cached-read and cache-write subsets, and `completion_tokens` (inclusive
+// total) with a `reasoning_tokens` subset. We pass the inclusive totals
+// through and derive the non-cached breakdown so the `LLM.Usage` contract is
 // satisfied on both sides.
 const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
   if (!usage) return undefined
   const cached = usage.prompt_tokens_details?.cached_tokens
+  const cacheWrite = usage.prompt_tokens_details?.cache_write_tokens
   const reasoning = usage.completion_tokens_details?.reasoning_tokens
-  const nonCached = ProviderShared.subtractTokens(usage.prompt_tokens, cached)
+  const nonCached = ProviderShared.subtractTokens(usage.prompt_tokens, ProviderShared.sumTokens(cached, cacheWrite))
   return new Usage({
     inputTokens: usage.prompt_tokens,
     outputTokens: usage.completion_tokens,
     nonCachedInputTokens: nonCached,
     cacheReadInputTokens: cached,
+    cacheWriteInputTokens: cacheWrite,
     reasoningTokens: reasoning,
     totalTokens: ProviderShared.totalTokens(usage.prompt_tokens, usage.completion_tokens, usage.total_tokens),
     providerMetadata: { openai: usage },
@@ -532,10 +545,22 @@ const reasoningMetadata = (field: ParserState["reasoningField"], details?: Reado
 
 const step = (state: ParserState, event: OpenAIChatEvent) =>
   Effect.gen(function* () {
+    if (event.error)
+      return yield* new LLMError({
+        module: ADAPTER,
+        method: "stream",
+        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,
+        }),
+      })
     const events: LLMEvent[] = []
     const usage = mapUsage(event.usage) ?? state.usage
-    const choice = event.choices[0]
-    const finishReason = choice?.finish_reason ? mapFinishReason(choice.finish_reason) : state.finishReason
+    const choice = event.choices?.[0]
+    const finishReason = choice?.finish_reason
+      ? { normalized: mapFinishReason(choice.finish_reason), raw: choice.native_finish_reason ?? choice.finish_reason }
+      : state.finishReason
     const delta = choice?.delta
     const toolDeltas = delta?.tool_calls ?? []
     let tools = state.tools
@@ -627,7 +652,13 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
 const finishEvents = (state: ParserState): ReadonlyArray<LLMEvent> => {
   const events: LLMEvent[] = []
   const hasToolCalls = state.toolCallEvents.length > 0
-  const reason = state.finishReason === "stop" && hasToolCalls ? "tool-calls" : state.finishReason
+  const reason = state.finishReason
+    ? {
+        ...state.finishReason,
+        normalized:
+          state.finishReason.normalized === "stop" && hasToolCalls ? "tool-calls" : state.finishReason.normalized,
+      }
+    : undefined
   const metadata = reasoningMetadata(
     state.reasoningField,
     state.reasoningDetailsObserved ? state.reasoningDetails : undefined,

packages/ai/src/protocols/openai-compatible-responses.ts

@@ -1,23 +1,22 @@
 import { Route, type RouteRoutedModelInput } from "../route/client"
 import { Endpoint } from "../route/endpoint"
-import { OpenAIResponses } from "./openai-responses"
+import { OpenResponses } from "./open-responses"
 
 const ADAPTER = "openai-compatible-responses"
 
 export type OpenAICompatibleResponsesModelInput = RouteRoutedModelInput
 
 /**
- * Route for providers that expose an OpenAI Responses-compatible `/responses`
- * endpoint. Provider helpers configure identity, endpoint, and auth before
- * model selection while this route reuses the OpenAI Responses protocol.
+ * Deployment adapter for providers that expose an Open Responses-compatible
+ * `/responses` endpoint. Provider helpers configure identity, endpoint, and
+ * auth while the semantic protocol remains provider-neutral.
  */
 export const route = Route.make({
   id: ADAPTER,
-  providerMetadataKey: "openai",
-  protocol: OpenAIResponses.protocol,
-  endpoint: Endpoint.path(OpenAIResponses.PATH),
-  transport: OpenAIResponses.httpTransport,
-  defaults: { providerOptions: { openai: { store: false } } },
+  providerMetadataKey: "openresponses",
+  protocol: OpenResponses.protocol,
+  endpoint: Endpoint.path(OpenResponses.PATH),
+  transport: OpenResponses.httpTransport,
 })
 
 export * as OpenAICompatibleResponses from "./openai-compatible-responses"

packages/ai/src/protocols/shared.ts

@@ -1,4 +1,5 @@
 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"
@@ -9,7 +10,6 @@ import {
   type ContentPart,
   type LLMRequest,
   type MediaPart,
-  type ToolFileContent,
   type TextPart,
   type ToolResultPart,
 } from "../schema"
@@ -206,7 +206,7 @@ export const validateMedia = Effect.fn("ProviderShared.validateMedia")(function*
   return { mime, base64, dataUrl: `data:${mime};base64,${base64}`, bytes } satisfies ValidatedMedia
 })
 
-export const validateToolFile = (route: string, part: ToolFileContent, supportedMimes: ReadonlySet<string>) =>
+export const validateToolFile = (route: string, part: Tool.FileContent, supportedMimes: ReadonlySet<string>) =>
   validateMedia(route, { type: "media", mediaType: part.mime, data: part.uri, filename: part.name }, supportedMimes)
 
 export const trimBaseUrl = (value: string) => value.replace(/\/+$/, "")

packages/ai/src/protocols/utils/lifecycle.ts

@@ -1,4 +1,4 @@
-import { LLMEvent, type FinishReason, type ProviderMetadata, type Usage } from "../../schema"
+import { LLMEvent, type FinishReasonDetails, type ProviderMetadata, type Usage } from "../../schema"
 
 export interface State {
   readonly stepStarted: boolean
@@ -14,16 +14,19 @@ export const stepStart = (state: State, events: LLMEvent[]): State => {
   return { ...state, stepStarted: true }
 }
 
-export const textDelta = (state: State, events: LLMEvent[], id: string, text: string): State => {
+export const textStart = (state: State, events: LLMEvent[], id: string, providerMetadata?: ProviderMetadata): State => {
+  if (state.text.has(id)) return state
   const stepped = stepStart(state, events)
-  if (stepped.text.has(id)) {
-    events.push(LLMEvent.textDelta({ id, text }))
-    return stepped
-  }
-  events.push(LLMEvent.textStart({ id }), LLMEvent.textDelta({ id, text }))
+  events.push(LLMEvent.textStart({ id, providerMetadata }))
   return { ...stepped, text: new Set([...stepped.text, id]) }
 }
 
+export const textDelta = (state: State, events: LLMEvent[], id: string, text: string): State => {
+  const started = textStart(state, events, id)
+  events.push(LLMEvent.textDelta({ id, text }))
+  return started
+}
+
 export const reasoningStart = (
   state: State,
   events: LLMEvent[],
@@ -81,7 +84,7 @@ export const finish = (
   state: State,
   events: LLMEvent[],
   input: {
-    readonly reason: FinishReason
+    readonly reason: FinishReasonDetails
     readonly usage?: Usage
     readonly providerMetadata?: ProviderMetadata
   },

packages/ai/src/protocols/utils/open-responses-options.ts

@@ -0,0 +1,65 @@
+import { Schema } from "effect"
+import { TextVerbosity, type LLMRequest } from "../../schema"
+
+export const ResponseIncludables = [
+  "file_search_call.results",
+  "web_search_call.results",
+  "web_search_call.action.sources",
+  "message.input_image.image_url",
+  "computer_call_output.output.image_url",
+  "code_interpreter_call.outputs",
+  "reasoning.encrypted_content",
+  "message.output_text.logprobs",
+] as const
+export type ResponseIncludable = (typeof ResponseIncludables)[number]
+
+export const ServiceTiers = ["auto", "default", "flex", "priority"] as const
+export type ServiceTier = (typeof ServiceTiers)[number]
+
+const TEXT_VERBOSITY = new Set<string>(["low", "medium", "high"])
+const INCLUDABLES = new Set<string>(ResponseIncludables)
+const SERVICE_TIERS = new Set<string>(ServiceTiers)
+
+const isTextVerbosity = (value: unknown): value is Schema.Schema.Type<typeof TextVerbosity> =>
+  typeof value === "string" && TEXT_VERBOSITY.has(value)
+
+const isServiceTier = (value: unknown): value is ServiceTier => typeof value === "string" && SERVICE_TIERS.has(value)
+
+export const ReasoningEffort = Schema.String
+export const TextVerbositySchema = TextVerbosity
+export const ResponseIncludableSchema = Schema.Literals(ResponseIncludables)
+export const ServiceTierSchema = Schema.Literals(ServiceTiers)
+
+export interface Resolved {
+  readonly instructions?: string
+  readonly store?: boolean
+  readonly promptCacheKey?: string
+  readonly reasoningEffort?: string
+  readonly reasoningSummary?: "auto" | "concise" | "detailed"
+  readonly include?: ReadonlyArray<ResponseIncludable>
+  readonly textVerbosity?: Schema.Schema.Type<typeof TextVerbosity>
+  readonly serviceTier?: ServiceTier
+}
+
+export const resolve = (request: LLMRequest): Resolved => {
+  const input = request.providerOptions?.[request.model.route.providerMetadataKey ?? "openresponses"]
+  const include = Array.isArray(input?.include)
+    ? input.include.filter((entry): entry is ResponseIncludable => INCLUDABLES.has(entry))
+    : []
+  const reasoningSummary = input?.reasoningSummary
+  return {
+    instructions: typeof input?.instructions === "string" ? input.instructions : undefined,
+    store: typeof input?.store === "boolean" ? input.store : undefined,
+    promptCacheKey: typeof input?.promptCacheKey === "string" ? input.promptCacheKey : undefined,
+    reasoningEffort: typeof input?.reasoningEffort === "string" ? input.reasoningEffort : undefined,
+    reasoningSummary:
+      reasoningSummary === "auto" || reasoningSummary === "concise" || reasoningSummary === "detailed"
+        ? reasoningSummary
+        : undefined,
+    include: include.length > 0 ? include : undefined,
+    textVerbosity: isTextVerbosity(input?.textVerbosity) ? input.textVerbosity : undefined,
+    serviceTier: isServiceTier(input?.serviceTier) ? input.serviceTier : undefined,
+  }
+}
+
+export * as OpenResponsesOptions from "./open-responses-options"

packages/ai/src/protocols/utils/openai-options.ts

@@ -1,85 +1,23 @@
-import { Schema } from "effect"
-import type { LLMRequest, TextVerbosity as TextVerbosityValue } from "../../schema"
-import { ReasoningEfforts, TextVerbosity } from "../../schema"
+import { ReasoningEfforts } from "../../schema"
+import { OpenResponsesOptions } from "./open-responses-options"
 
 export const OpenAIReasoningEfforts = ReasoningEfforts
 export type OpenAIReasoningEffort = string
 
 // Mirrors OpenAI's `ResponseIncludable` union from the official SDK. Keep this
 // in lockstep with `openai-node/src/resources/responses/responses.ts`.
-export const OpenAIResponseIncludables = [
-  "file_search_call.results",
-  "web_search_call.results",
-  "web_search_call.action.sources",
-  "message.input_image.image_url",
-  "computer_call_output.output.image_url",
-  "code_interpreter_call.outputs",
-  "reasoning.encrypted_content",
-  "message.output_text.logprobs",
-] as const
-export type OpenAIResponseIncludable = (typeof OpenAIResponseIncludables)[number]
-export const OpenAIServiceTiers = ["auto", "default", "flex", "priority"] as const
-export type OpenAIServiceTier = (typeof OpenAIServiceTiers)[number]
+export const OpenAIResponseIncludables = OpenResponsesOptions.ResponseIncludables
+export type OpenAIResponseIncludable = OpenResponsesOptions.ResponseIncludable
+export const OpenAIServiceTiers = OpenResponsesOptions.ServiceTiers
+export type OpenAIServiceTier = OpenResponsesOptions.ServiceTier
 
-const TEXT_VERBOSITY = new Set<string>(["low", "medium", "high"])
-const INCLUDABLES = new Set<string>(OpenAIResponseIncludables)
-const SERVICE_TIERS = new Set<string>(OpenAIServiceTiers)
-
-export const OpenAIReasoningEffort = Schema.String
-export const OpenAITextVerbosity = TextVerbosity
-export const OpenAIResponseIncludable = Schema.Literals(OpenAIResponseIncludables)
-export const OpenAIServiceTier = Schema.Literals(OpenAIServiceTiers)
+export const OpenAIReasoningEffort = OpenResponsesOptions.ReasoningEffort
+export const OpenAITextVerbosity = OpenResponsesOptions.TextVerbositySchema
+export const OpenAIResponseIncludable = OpenResponsesOptions.ResponseIncludableSchema
+export const OpenAIServiceTier = OpenResponsesOptions.ServiceTierSchema
 
 export const isReasoningEffort = (effort: unknown): effort is OpenAIReasoningEffort => typeof effort === "string"
 
-const isTextVerbosity = (value: unknown): value is TextVerbosityValue =>
-  typeof value === "string" && TEXT_VERBOSITY.has(value)
-
-const options = (request: LLMRequest) => request.providerOptions?.openai
-
-export const store = (request: LLMRequest): boolean | undefined => {
-  const value = options(request)?.store
-  return typeof value === "boolean" ? value : undefined
-}
-
-export const reasoningEffort = (request: LLMRequest): string | undefined => {
-  const value = options(request)?.reasoningEffort
-  return typeof value === "string" ? value : undefined
-}
-
-export const reasoningSummary = (request: LLMRequest): "auto" | undefined =>
-  options(request)?.reasoningSummary === "auto" ? "auto" : undefined
-
-// Resolve the OpenAI Responses `include` field. Filters out unknown
-// includable values defensively so a typo in upstream config drops the
-// invalid entry instead of poisoning the wire body. An empty array (either
-// passed directly or produced by filtering) is treated as "no include" and
-// returns undefined so the request body omits the field entirely.
-export const include = (request: LLMRequest): ReadonlyArray<OpenAIResponseIncludable> | undefined => {
-  const value = options(request)?.include
-  if (!Array.isArray(value)) return undefined
-  const filtered = value.filter((entry): entry is OpenAIResponseIncludable => INCLUDABLES.has(entry))
-  return filtered.length > 0 ? filtered : undefined
-}
-
-export const promptCacheKey = (request: LLMRequest) => {
-  const value = options(request)?.promptCacheKey
-  return typeof value === "string" ? value : undefined
-}
-
-export const textVerbosity = (request: LLMRequest) => {
-  const value = options(request)?.textVerbosity
-  return isTextVerbosity(value) ? value : undefined
-}
-
-export const serviceTier = (request: LLMRequest) => {
-  const value = options(request)?.serviceTier
-  return typeof value === "string" && SERVICE_TIERS.has(value) ? (value as OpenAIServiceTier) : undefined
-}
-
-export const instructions = (request: LLMRequest) => {
-  const value = options(request)?.instructions
-  return typeof value === "string" ? value : undefined
-}
+export const resolve = OpenResponsesOptions.resolve
 
 export * as OpenAIOptions from "./openai-options"

packages/ai/src/protocols/utils/tool-schema.ts

@@ -63,6 +63,8 @@ const openAI = (schema: JsonSchema): JsonSchema => {
   return isRecord(normalized) ? normalized : { type: "object" }
 }
 
+const responses = openAI
+
 const gemini = (schema: JsonSchema): JsonSchema => GeminiToolSchema.convert(schema) ?? {}
 
 const modelCompatibility = (
@@ -83,4 +85,5 @@ export const ToolSchemaProjection = {
   modelCompatibility,
   moonshot,
   openAI,
+  responses,
 } as const

packages/ai/src/provider-error.ts

@@ -135,7 +135,7 @@ export function classifyProviderFailure(input: ProviderFailure): LLMError["reaso
       rateLimit: input.rateLimit,
     })
   }
-  if (input.status !== undefined && input.status >= 500)
+  if (input.status === 408 || input.status === 409 || (input.status !== undefined && input.status >= 500))
     return new ProviderInternalReason({
       ...common,
       status: input.status,
@@ -145,7 +145,6 @@ export function classifyProviderFailure(input: ProviderFailure): LLMError["reaso
   if (
     input.status === 400 ||
     input.status === 404 ||
-    input.status === 409 ||
     input.status === 413 ||
     input.status === 422
   )

packages/ai/src/provider-package.ts

@@ -1,4 +1,4 @@
-import type { Model } from "./schema"
+import type { Model, ProviderOptions } from "./schema"
 
 export interface Settings extends Readonly<Record<string, unknown>> {
   readonly headers?: Readonly<Record<string, string>>
@@ -9,8 +9,11 @@ export interface Settings extends Readonly<Record<string, unknown>> {
   }
 }
 
-export interface Definition<ProviderSettings extends Settings = Settings> {
-  readonly model: (modelID: string, settings: ProviderSettings) => Model
+export interface Definition<
+  ProviderSettings extends Settings = Settings,
+  Options extends ProviderOptions = ProviderOptions,
+> {
+  readonly model: (modelID: string, settings: ProviderSettings) => Model<Options>
 }
 
 export * as ProviderPackage from "./provider-package"

packages/ai/src/providers/anthropic-compatible.ts

@@ -5,12 +5,17 @@ import type { ProviderAuthOption } from "../route/auth-options"
 import type { RouteDefaultsInput } from "../route/client"
 import { ProviderID, type ModelID } from "../schema"
 
+export type AnthropicOptionsInput = AnthropicMessages.OptionsInput
+export type AnthropicProviderOptionsInput = AnthropicMessages.ProviderOptionsInput
+export type AnthropicThinkingInput = AnthropicMessages.ThinkingInput
+
 export const id = ProviderID.make("anthropic-compatible")
 
 export type Config = RouteDefaultsInput &
   ProviderAuthOption<"optional"> & {
     readonly provider?: string
     readonly baseURL: string
+    readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
   }
 
 export type Settings = ProviderPackage.Settings &
@@ -20,6 +25,7 @@ export type Settings = ProviderPackage.Settings &
   ) & {
     readonly baseURL: string
     readonly provider?: string
+    readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
   }
 
 export const routes = [AnthropicMessages.route]
@@ -41,7 +47,7 @@ export const configure = (input: Config) => {
   })
   return {
     id: ProviderID.make(provider),
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<AnthropicMessages.ProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -51,7 +57,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, AnthropicMessages.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined && settings.authToken !== undefined)
     throw new Error("Anthropic-compatible apiKey cannot be combined with authToken")
   return configure({
@@ -61,6 +70,7 @@ export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, se
     http: settings.body === undefined ? undefined : { body: { ...settings.body } },
     limits: settings.limits,
     provider: settings.provider,
+    providerOptions: settings.providerOptions,
   }).model(modelID)
 }
 

packages/ai/src/providers/anthropic.ts

@@ -6,11 +6,19 @@ import { ProviderID, type ModelID } from "../schema"
 import { AnthropicMessages } from "../protocols/anthropic-messages"
 import { AnthropicCompatible } from "./anthropic-compatible"
 
+export type AnthropicOptionsInput = AnthropicMessages.OptionsInput
+export type AnthropicProviderOptionsInput = AnthropicMessages.ProviderOptionsInput
+export type AnthropicThinkingInput = AnthropicMessages.ThinkingInput
+
 export const id = ProviderID.make("anthropic")
 
 export const routes = [AnthropicMessages.route]
 
-export type Config = RouteDefaultsInput & ProviderAuthOption<"optional"> & { readonly baseURL?: string }
+export type Config = RouteDefaultsInput &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
+  }
 
 export type Settings = ProviderPackage.Settings &
   (
@@ -18,6 +26,7 @@ export type Settings = ProviderPackage.Settings &
     | { readonly apiKey?: never; readonly authToken?: string }
   ) & {
     readonly baseURL?: string
+    readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
   }
 
 const auth = (options: ProviderAuthOption<"optional">) => {
@@ -43,7 +52,10 @@ export const configure = (input: Config = {}) => {
 }
 
 export const provider = configure()
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, AnthropicMessages.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined && settings.authToken !== undefined)
     throw new Error("Anthropic apiKey cannot be combined with authToken")
   return configure({
@@ -52,5 +64,6 @@ export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, se
     headers: settings.headers === undefined ? undefined : { ...settings.headers },
     http: settings.body === undefined ? undefined : { body: { ...settings.body } },
     limits: settings.limits,
+    providerOptions: settings.providerOptions,
   }).model(modelID)
 }

packages/ai/src/providers/azure.ts

@@ -99,10 +99,14 @@ export const configure = (input: Config) => {
   const modelDefaults = defaults(input)
 
   const responses = (modelID: string | ModelID) =>
-    configuredResponsesRoute.with(withOpenAIOptions(modelID, modelDefaults)).model({ id: modelID })
+    configuredResponsesRoute
+      .with(withOpenAIOptions(modelID, modelDefaults))
+      .model<OpenAIProviderOptionsInput>({ id: modelID })
 
   const chat = (modelID: string | ModelID) =>
-    configuredChatRoute.with(withOpenAIOptions(modelID, modelDefaults)).model({ id: modelID })
+    configuredChatRoute
+      .with(withOpenAIOptions(modelID, modelDefaults))
+      .model<OpenAIProviderOptionsInput>({ id: modelID })
 
   return {
     id,
@@ -133,8 +137,12 @@ const config = (settings: Settings): Config => {
   throw new Error("Azure requires resourceName or baseURL")
 }
 
-export const responsesModel: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
-  configure(config(settings)).responses(modelID)
-export const chatModel: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
-  configure(config(settings)).chat(modelID)
+export const responsesModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => configure(config(settings)).responses(modelID)
+export const chatModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => configure(config(settings)).chat(modelID)
 export const model = responsesModel

packages/ai/src/providers/cloudflare.ts

@@ -4,6 +4,7 @@ import { Auth } from "../route/auth"
 import { AuthOptions, type AtLeastOne, type ProviderAuthOption } from "../route/auth-options"
 import type { RouteDefaultsInput } from "../route/client"
 import { ProviderID, type ModelID } from "../schema"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const aiGatewayID = ProviderID.make("cloudflare-ai-gateway")
 export const workersAIID = ProviderID.make("cloudflare-workers-ai")
@@ -20,10 +21,11 @@ type GatewayURL = AtLeastOne<{
 }
 
 export type AIGatewayOptions = GatewayURL &
-  RouteDefaultsInput &
+  Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     /** Cloudflare AI Gateway authentication token. Sent as `cf-aig-authorization`. */
     readonly gatewayApiKey?: CloudflareSecret
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 type WorkersAIURL = AtLeastOne<{
@@ -31,7 +33,11 @@ type WorkersAIURL = AtLeastOne<{
   readonly baseURL: string
 }>
 
-export type WorkersAIOptions = WorkersAIURL & RouteDefaultsInput & ProviderAuthOption<"optional">
+export type WorkersAIOptions = WorkersAIURL &
+  Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
 
 export const aiGatewayBaseURL = (input: GatewayURL) => {
   if (input.baseURL) return input.baseURL
@@ -98,7 +104,7 @@ const configureAIGateway = (options: AIGatewayOptions) => {
   })
   return {
     id: aiGatewayID,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenAIProviderOptionsInput>({ id: modelID }),
     configure: configureAIGateway,
   }
 }
@@ -111,7 +117,7 @@ const configureWorkersAI = (options: WorkersAIOptions) => {
   })
   return {
     id: workersAIID,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenAIProviderOptionsInput>({ id: modelID }),
     configure: configureWorkersAI,
   }
 }

packages/ai/src/providers/github-copilot.ts

@@ -50,9 +50,11 @@ export const configure = (options: ModelOptions) => {
   const responsesRoute = configuredResponsesRoute(options)
   const chatRoute = configuredChatRoute(options)
   const responses = (modelID: string | ModelID) =>
-    responsesRoute.with(withOpenAIOptions(modelID, defaults(options))).model({ id: modelID })
+    responsesRoute
+      .with(withOpenAIOptions(modelID, defaults(options)))
+      .model<OpenAIProviderOptionsInput>({ id: modelID })
   const chat = (modelID: string | ModelID) =>
-    chatRoute.with(withOpenAIOptions(modelID, defaults(options))).model({ id: modelID })
+    chatRoute.with(withOpenAIOptions(modelID, defaults(options))).model<OpenAIProviderOptionsInput>({ id: modelID })
   return {
     id,
     model: (modelID: string | ModelID) =>

packages/ai/src/providers/google-vertex-chat.ts

@@ -1,8 +1,9 @@
 import type { ProviderPackage } from "../provider-package"
 import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat"
 import type { RouteDefaultsInput } from "../route/client"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { ProviderID, type ModelID } from "../schema"
 import { GoogleVertexShared } from "./google-vertex-shared"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const id = ProviderID.make("google-vertex")
 
@@ -11,6 +12,7 @@ export type Config = RouteDefaultsInput &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -19,7 +21,7 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly location?: string
   readonly project?: string
-  readonly providerOptions?: ProviderOptions
+  readonly providerOptions?: OpenAIProviderOptionsInput
 }
 
 const route = OpenAICompatibleChat.route.with({
@@ -56,7 +58,7 @@ export const configure = (input: Config = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenAIProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -66,7 +68,7 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) => {
   if (settings.apiKey !== undefined) throw new Error("Google Vertex Chat does not support API keys")
   return configure({
     accessToken: settings.accessToken,

packages/ai/src/providers/google-vertex-messages.ts

@@ -6,9 +6,13 @@ import { Route, type RouteDefaultsInput } from "../route/client"
 import { Endpoint } from "../route/endpoint"
 import { Framing } from "../route/framing"
 import { Protocol } from "../route/protocol"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { ProviderID, type ModelID } from "../schema"
 import { GoogleVertexShared } from "./google-vertex-shared"
 
+export type AnthropicOptionsInput = AnthropicMessages.OptionsInput
+export type AnthropicProviderOptionsInput = AnthropicMessages.ProviderOptionsInput
+export type AnthropicThinkingInput = AnthropicMessages.ThinkingInput
+
 const VERSION = "vertex-2023-10-16" as const
 
 // models.dev uses this provider id even though the API contract is Anthropic Messages.
@@ -19,6 +23,7 @@ export type Config = RouteDefaultsInput &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -27,7 +32,7 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly location?: string
   readonly project?: string
-  readonly providerOptions?: ProviderOptions
+  readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
 }
 
 const route = Route.make({
@@ -86,7 +91,7 @@ export const configure = (input: Config = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<AnthropicMessages.ProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -96,7 +101,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, AnthropicMessages.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined) throw new Error("Google Vertex Messages does not support API keys")
   return configure({
     accessToken: settings.accessToken,

packages/ai/src/providers/google-vertex-responses.ts

@@ -1,8 +1,9 @@
 import type { ProviderPackage } from "../provider-package"
 import { OpenAICompatibleResponses } from "../protocols/openai-compatible-responses"
 import type { RouteDefaultsInput } from "../route/client"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { ProviderID, type ModelID } from "../schema"
 import { GoogleVertexShared } from "./google-vertex-shared"
+import type { OpenResponsesProviderOptionsInput } from "./open-responses-options"
 
 export const id = ProviderID.make("google-vertex")
 
@@ -11,6 +12,7 @@ export type Config = RouteDefaultsInput &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: OpenResponsesProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -19,12 +21,13 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly location?: string
   readonly project?: string
-  readonly providerOptions?: ProviderOptions
+  readonly providerOptions?: OpenResponsesProviderOptionsInput
 }
 
 const route = OpenAICompatibleResponses.route.with({
   id: "google-vertex-responses",
   provider: id,
+  providerOptions: { openresponses: { store: false } },
 })
 
 export const routes = [route]
@@ -57,7 +60,7 @@ export const configure = (input: Config = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenResponsesProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -67,7 +70,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, OpenResponsesProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined) throw new Error("Google Vertex Responses does not support API keys")
   return configure({
     accessToken: settings.accessToken,

packages/ai/src/providers/google-vertex.ts

@@ -4,9 +4,12 @@ import { Auth } from "../route/auth"
 import { Route, type RouteDefaultsInput } from "../route/client"
 import { Endpoint } from "../route/endpoint"
 import { Framing } from "../route/framing"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { ProviderID, type ModelID } from "../schema"
 import { GoogleVertexShared } from "./google-vertex-shared"
 
+export type GeminiOptionsInput = Gemini.OptionsInput
+export type GeminiProviderOptionsInput = Gemini.ProviderOptionsInput
+
 export const id = ProviderID.make("google-vertex")
 
 export type Config = RouteDefaultsInput &
@@ -14,6 +17,7 @@ export type Config = RouteDefaultsInput &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: Gemini.ProviderOptionsInput
   }
 
 export type Settings = ProviderPackage.Settings &
@@ -24,7 +28,7 @@ export type Settings = ProviderPackage.Settings &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
-    readonly providerOptions?: ProviderOptions
+    readonly providerOptions?: Gemini.ProviderOptionsInput
   }
 
 const route = Route.make({
@@ -73,7 +77,8 @@ const configuredRoute = (input: Config, modelID: string | ModelID) => {
 export const configure = (input: Config = {}) => {
   return {
     id,
-    model: (modelID: string | ModelID) => configuredRoute(input, modelID).model({ id: modelID }),
+    model: (modelID: string | ModelID) =>
+      configuredRoute(input, modelID).model<Gemini.ProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -82,7 +87,10 @@ export const provider = {
   id,
   configure,
 }
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, Gemini.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined && settings.accessToken !== undefined)
     throw new Error("Google Vertex apiKey cannot be combined with accessToken or auth")
   return configure({

packages/ai/src/providers/google.ts

@@ -2,11 +2,13 @@ import type { RouteDefaultsInput } from "../route/client"
 import { Auth } from "../route/auth"
 import type { ProviderAuthOption } from "../route/auth-options"
 import type { ProviderPackage } from "../provider-package"
-import { HttpOptions, ProviderID, mergeHttpOptions, type ModelID, type ProviderOptions } from "../schema"
+import { HttpOptions, ProviderID, mergeHttpOptions, type ModelID } from "../schema"
 import { Gemini } from "../protocols/gemini"
 import { GoogleImages } from "../protocols/google-images"
 
 export type { GoogleImageOptions } from "../protocols/google-images"
+export type GeminiOptionsInput = Gemini.OptionsInput
+export type GeminiProviderOptionsInput = Gemini.ProviderOptionsInput
 
 export const id = ProviderID.make("google")
 
@@ -15,12 +17,13 @@ export const routes = [Gemini.route]
 export type Config = RouteDefaultsInput &
   ProviderAuthOption<"optional"> & {
     readonly baseURL?: string
+    readonly providerOptions?: Gemini.ProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
   readonly apiKey?: string
   readonly baseURL?: string
-  readonly providerOptions?: ProviderOptions
+  readonly providerOptions?: Gemini.ProviderOptionsInput
 }
 
 const auth = (options: ProviderAuthOption<"optional">) => {
@@ -47,14 +50,14 @@ export const configure = (input: Config = {}) => {
     })
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<Gemini.ProviderOptionsInput>({ id: modelID }),
     image,
     configure,
   }
 }
 
 export const provider = configure()
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+export const model: ProviderPackage.Definition<Settings, Gemini.ProviderOptionsInput>["model"] = (modelID, settings) =>
   configure({
     apiKey: settings.apiKey,
     baseURL: settings.baseURL,

packages/ai/src/providers/open-responses-options.ts

@@ -0,0 +1,20 @@
+import type { ResponseIncludable, ServiceTier } from "../protocols/utils/open-responses-options"
+import type { ProviderOptions, ReasoningEffort, TextVerbosity } from "../schema"
+
+export interface OpenResponsesOptionsInput {
+  readonly [key: string]: unknown
+  readonly instructions?: string
+  readonly store?: boolean
+  readonly promptCacheKey?: string
+  readonly reasoningEffort?: ReasoningEffort
+  readonly reasoningSummary?: "auto" | "concise" | "detailed"
+  readonly include?: ReadonlyArray<ResponseIncludable>
+  readonly textVerbosity?: TextVerbosity
+  readonly serviceTier?: ServiceTier
+}
+
+export type OpenResponsesProviderOptionsInput = ProviderOptions & {
+  readonly openresponses?: OpenResponsesOptionsInput
+}
+
+export * as OpenResponsesProviderOptions from "./open-responses-options"

packages/ai/src/providers/openai-compatible-responses.ts

@@ -3,7 +3,9 @@ import { OpenAICompatibleResponses } from "../protocols/openai-compatible-respon
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options"
 import type { RouteDefaultsInput } from "../route/client"
 import { ProviderID, type ModelID } from "../schema"
-import type { OpenAIProviderOptionsInput } from "./openai-options"
+import type { OpenResponsesProviderOptionsInput } from "./open-responses-options"
+
+export type { OpenResponsesOptionsInput, OpenResponsesProviderOptionsInput } from "./open-responses-options"
 
 export const id = ProviderID.make("openai-compatible")
 
@@ -11,13 +13,14 @@ export type Config = RouteDefaultsInput &
   ProviderAuthOption<"optional"> & {
     readonly provider?: string
     readonly baseURL: string
+    readonly providerOptions?: OpenResponsesProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
   readonly apiKey?: string
   readonly baseURL: string
   readonly provider?: string
-  readonly providerOptions?: OpenAIProviderOptionsInput
+  readonly providerOptions?: OpenResponsesProviderOptionsInput
 }
 
 export const routes = [OpenAICompatibleResponses.route]
@@ -33,7 +36,7 @@ export const configure = (input: Config) => {
   })
   return {
     id: ProviderID.make(provider),
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenResponsesProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -43,7 +46,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+export const model: ProviderPackage.Definition<Settings, OpenResponsesProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) =>
   configure({
     apiKey: settings.apiKey,
     baseURL: settings.baseURL,

packages/ai/src/providers/openai-compatible.ts

@@ -4,13 +4,15 @@ import type { RouteDefaultsInput } from "../route/client"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options"
 import type { ProviderPackage } from "../provider-package"
 import { profiles, type OpenAICompatibleProfile } from "./openai-compatible-profile"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const id = ProviderID.make("openai-compatible")
 
-type GenericModelOptions = RouteDefaultsInput &
+type GenericModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     readonly provider?: string
     readonly baseURL: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -19,9 +21,10 @@ export interface Settings extends ProviderPackage.Settings {
   readonly provider?: string
 }
 
-export type FamilyModelOptions = RouteDefaultsInput &
+export type FamilyModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export const routes = [OpenAICompatibleChat.route]
@@ -37,7 +40,8 @@ export const configure = (input: GenericModelOptions) => {
   })
   return {
     id: ProviderID.make(provider),
-    model: (modelID: string | ModelID) => route.model({ id: modelID, provider: ProviderID.make(provider) }),
+    model: (modelID: string | ModelID) =>
+      route.model<OpenAIProviderOptionsInput>({ id: modelID, provider: ProviderID.make(provider) }),
     configure,
   }
 }
@@ -63,7 +67,7 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
   configure({
     apiKey: settings.apiKey,
     baseURL: settings.baseURL,

packages/ai/src/providers/openai-options.ts

@@ -1,22 +1,10 @@
-import type { ProviderOptions, ReasoningEffort, TextVerbosity } from "../schema"
+import type { ProviderOptions } from "../schema"
 import { mergeProviderOptions } from "../schema"
-import type { OpenAIResponseIncludable, OpenAIServiceTier } from "../protocols/utils/openai-options"
+import type { OpenResponsesOptionsInput } from "./open-responses-options"
 
 export type { OpenAIResponseIncludable, OpenAIServiceTier } from "../protocols/utils/openai-options"
 
-export interface OpenAIOptionsInput {
-  readonly [key: string]: unknown
-  readonly store?: boolean
-  readonly promptCacheKey?: string
-  readonly reasoningEffort?: ReasoningEffort
-  readonly reasoningSummary?: "auto"
-  // OpenAI Responses `include` wire field. Mirrors the official SDK's
-  // `ResponseIncludable[]` union exactly so AI SDK callers and direct
-  // native-SDK callers share one shape and no translation is required.
-  readonly include?: ReadonlyArray<OpenAIResponseIncludable>
-  readonly textVerbosity?: TextVerbosity
-  readonly serviceTier?: OpenAIServiceTier
-}
+export type OpenAIOptionsInput = OpenResponsesOptionsInput
 
 export type OpenAIProviderOptionsInput = ProviderOptions & {
   readonly openai?: OpenAIOptionsInput

packages/ai/src/providers/openai.ts

@@ -86,10 +86,15 @@ export const configure = (input: Config = {}) => {
   const chatRoute = configuredRoute(OpenAIChat.route, input)
   const modelDefaults = defaults(input)
   const responses = (id: string | ModelID) =>
-    responsesRoute.with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true })).model({ id })
+    responsesRoute
+      .with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true }))
+      .model<OpenAIProviderOptionsInput>({ id })
   const responsesWebSocket = (id: string | ModelID) =>
-    responsesWebSocketRoute.with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true })).model({ id })
-  const chat = (id: string | ModelID) => chatRoute.with(withOpenAIOptions(id, modelDefaults)).model({ id })
+    responsesWebSocketRoute
+      .with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true }))
+      .model<OpenAIProviderOptionsInput>({ id })
+  const chat = (id: string | ModelID) =>
+    chatRoute.with(withOpenAIOptions(id, modelDefaults)).model<OpenAIProviderOptionsInput>({ id })
   const image = (modelID: string | ModelID) =>
     OpenAIImages.model({
       id: modelID,
@@ -132,15 +137,17 @@ const config = (settings: Settings): Config => {
   }
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) => {
   const configured = configure(config(settings))
   if (settings.transport === undefined || settings.transport === "http") return configured.responses(modelID)
   if (settings.transport === "websocket") return configured.responsesWebSocket(modelID)
   throw new Error(`Unsupported OpenAI Responses transport: ${String(settings.transport)}`)
 }
 
-export const chatModel: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
-  configure(config(settings)).chat(modelID)
+export const chatModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => configure(config(settings)).chat(modelID)
 export const responses = provider.responses
 export const responsesWebSocket = provider.responsesWebSocket
 export const chat = provider.chat

packages/ai/src/providers/openrouter.ts

@@ -107,7 +107,7 @@ export const configure = (input: ModelOptions = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenRouterProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }

packages/ai/src/providers/xai.ts

@@ -5,12 +5,14 @@ import * as OpenAICompatibleProfiles from "./openai-compatible-profile"
 import * as OpenAICompatibleChat from "../protocols/openai-compatible-chat"
 import * as OpenAIResponses from "../protocols/openai-responses"
 import { XAIImages } from "../protocols/xai-images"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const id = ProviderID.make("xai")
 
-export type ModelOptions = RouteDefaultsInput &
+export type ModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export type { XAIImageOptions } from "../protocols/xai-images"
@@ -42,8 +44,8 @@ const configuredChatRoute = (input: ModelOptions) => {
 export const configure = (input: ModelOptions = {}) => {
   const responsesRoute = configuredResponsesRoute(input)
   const chatRoute = configuredChatRoute(input)
-  const responses = (modelID: string | ModelID) => responsesRoute.model({ id: modelID })
-  const chat = (modelID: string | ModelID) => chatRoute.model({ id: modelID })
+  const responses = (modelID: string | ModelID) => responsesRoute.model<OpenAIProviderOptionsInput>({ id: modelID })
+  const chat = (modelID: string | ModelID) => chatRoute.model<OpenAIProviderOptionsInput>({ id: modelID })
   const image = (modelID: string | ModelID) =>
     XAIImages.model({
       id: modelID,

packages/ai/src/route/client.ts

@@ -10,7 +10,7 @@ import { WebSocketExecutor } from "./transport"
 import type { Protocol } from "./protocol"
 import { applyCachePolicy } from "../cache-policy"
 import * as ProviderShared from "../protocols/shared"
-import type { LLMError, PreparedRequestOf, ProtocolID, ProviderOptions } from "../schema"
+import type { LLMError, ProtocolID, ProviderOptions } from "../schema"
 import {
   GenerationOptions,
   HttpOptions,
@@ -20,7 +20,6 @@ import {
   ModelLimits,
   LLMError as LLMErrorClass,
   LLMEvent,
-  PreparedRequest,
   ProviderID,
   mergeGenerationOptions,
   mergeHttpOptions,
@@ -46,7 +45,7 @@ export interface Route<Body, Prepared = unknown> {
   readonly defaults: RouteDefaults
   readonly body: RouteBody<Body>
   readonly with: (patch: RoutePatch<Body, Prepared>) => Route<Body, Prepared>
-  readonly model: (input: RouteMappedModelInput) => Model
+  readonly model: <Options extends ProviderOptions = ProviderOptions>(input: RouteMappedModelInput) => Model<Options>
   readonly prepareTransport: (body: Body, request: LLMRequest) => Effect.Effect<Prepared, LLMError>
   readonly streamPrepared: (
     prepared: Prepared,
@@ -93,12 +92,12 @@ export interface RoutePatch<Body, Prepared> extends RouteDefaultsInput {
 
 type RouteMappedModelInput = RouteModelInput | RouteRoutedModelInput
 
-const makeRouteModel = (route: AnyRoute, mapped: RouteMappedModelInput) => {
+const makeRouteModel = <Options extends ProviderOptions = ProviderOptions>(route: AnyRoute, mapped: RouteMappedModelInput) => {
   const provider = route.provider ?? ("provider" in mapped ? mapped.provider : undefined)
   if (!provider) throw new Error(`Route.model(${route.id}) requires a provider`)
   if (!endpointBaseURL(route.endpoint))
     throw new Error(`Route.model(${route.id}) requires an endpoint baseURL — configure it on the route first`)
-  return Model.make({
+  return Model.make<Options>({
     ...mapped,
     provider,
     route,
@@ -142,17 +141,6 @@ export const httpOptions = (input: HttpOptionsInput | undefined) => {
 }
 
 export interface Interface {
-  /**
-   * Compile a request through protocol body construction, validation, and HTTP
-   * preparation without sending it. Returns the prepared request including the
-   * provider-native body.
-   *
-   * Pass a `Body` type argument to statically expose the route's body
-   * shape (e.g. `prepare<OpenAIChatBody>(...)`) — the runtime body is
-   * identical, so this is a type-level assertion the caller makes about which
-   * route the request will resolve to.
-   */
-  readonly prepare: <Body = unknown>(request: LLMRequest) => Effect.Effect<PreparedRequestOf<Body>, LLMError>
   readonly stream: StreamMethod
   readonly generate: GenerateMethod
 }
@@ -296,7 +284,8 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
           defaults: mergeRouteDefaults(route.defaults, defaults),
         })
       },
-      model: (input) => makeRouteModel(route, input),
+      model: <Options extends ProviderOptions = ProviderOptions>(input: RouteMappedModelInput) =>
+        makeRouteModel<Options>(route, input),
       prepareTransport: (body, request) =>
         routeInput.transport.prepare({
           body,
@@ -370,9 +359,6 @@ export function make<Body, Prepared, Frame, Event, State>(
   })
 }
 
-// `compile` is the important boundary: it turns a common `LLMRequest` into a
-// validated provider body plus transport-private prepared data, but does not
-// execute transport.
 const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest) {
   const resolved = applyCachePolicy(resolveRequestOptions(request))
   const route = resolved.model.route
@@ -390,17 +376,17 @@ const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest) {
   }
 })
 
-const prepareWith = Effect.fn("LLMClient.prepare")(function* (request: LLMRequest) {
+/** @internal Test-only projection of the execution compiler; not exported from package barrels. */
+export const compileRequest = Effect.fn("LLM.compileRequest")(function* (request: LLMRequest) {
   const compiled = yield* compile(request)
-
-  return new PreparedRequest({
+  return {
     id: compiled.request.id ?? "request",
     route: compiled.route.id,
     protocol: compiled.route.protocol,
     model: compiled.request.model,
     body: compiled.body,
     metadata: { transport: compiled.route.transport.id },
-  })
+  }
 })
 
 const streamRequestWith = (runtime: TransportRuntime) => (request: LLMRequest) =>
@@ -422,9 +408,6 @@ const generateWith = (stream: Interface["stream"]) =>
     )
   })
 
-export const prepare = <Body = unknown>(request: LLMRequest) =>
-  prepareWith(request) as Effect.Effect<PreparedRequestOf<Body>, LLMError>
-
 export function stream(request: LLMRequest): Stream.Stream<LLMEvent, LLMError> {
   return Stream.unwrap(
     Effect.gen(function* () {
@@ -453,7 +436,7 @@ export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer
       http: yield* RequestExecutor.Service,
       webSocket: Option.getOrUndefined(yield* Effect.serviceOption(WebSocketExecutor.Service)),
     })
-    return Service.of({ prepare: prepareWith as Interface["prepare"], stream, generate: generateWith(stream) })
+    return Service.of({ stream, generate: generateWith(stream) })
   }),
 )
 
@@ -462,7 +445,6 @@ export const Route = { make } as const
 export const LLMClient = {
   Service,
   layer,
-  prepare,
   stream,
   generate,
 } as const

packages/ai/src/route/protocol.ts

@@ -12,7 +12,8 @@ import type { LLMError, LLMEvent, LLMRequest, ProtocolID } from "../schema"
  * Examples:
  *
  * - `OpenAIChat.protocol` — chat completions style
- * - `OpenAIResponses.protocol` — responses API
+ * - `OpenResponses.protocol` — provider-neutral Responses API baseline
+ * - `OpenAIResponses.protocol` — OpenAI extensions to that baseline
  * - `AnthropicMessages.protocol` — messages API with content blocks
  * - `Gemini.protocol` — generateContent
  * - `BedrockConverse.protocol` — Converse with binary event-stream framing

packages/ai/src/schema/errors.ts

@@ -1,4 +1,5 @@
 import { Schema } from "effect"
+import { Tool } from "@opencode-ai/schema/tool"
 import { ModelID, ProviderID, ProviderMetadata, RouteID } from "./ids"
 
 export const ProviderFailureClassification = Schema.Literal("context-overflow")
@@ -152,8 +153,4 @@ export class LLMError extends Schema.TaggedErrorClass<LLMError>()("LLM.Error", {
  * Anything thrown or yielded by a handler that is not a `ToolFailure` is
  * treated as a defect and fails the stream.
  */
-export class ToolFailure extends Schema.TaggedErrorClass<ToolFailure>()("LLM.ToolFailure", {
-  message: Schema.String,
-  error: Schema.optional(Schema.Defect()),
-  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
-}) {}
+export class ToolFailure extends Tool.Error {}

packages/ai/src/schema/events.ts

@@ -1,6 +1,5 @@
 import { Schema } from "effect"
-import { ContentBlockID, FinishReason, ProtocolID, ProviderMetadata, RouteID, ToolCallID } from "./ids"
-import { ModelSchema } from "./options"
+import { ContentBlockID, FinishReason, ProviderMetadata, ToolCallID } from "./ids"
 import { Message, ToolCallPart, ToolOutput, ToolResultPart, ToolResultValue, type ContentPart } from "./messages"
 import { ProviderFailureClassification } from "./errors"
 
@@ -40,9 +39,9 @@ import { ProviderFailureClassification } from "./errors"
  * - Anthropic and Bedrock report the input breakdown natively: Anthropic's
  *   `input_tokens` and Bedrock's `inputTokens` are non-cached only. Their
  *   mappers sum the breakdown to derive the inclusive `inputTokens`.
- *   Anthropic does *not* break extended-thinking out of `output_tokens`, so
- *   `reasoningTokens` is `undefined` and `outputTokens` carries the
- *   combined total — a documented limitation of the Anthropic API.
+ *   Anthropic's `outputTokens` includes extended thinking. Newer responses
+ *   expose that subset as `output_tokens_details.thinking_tokens`, which maps
+ *   to `reasoningTokens`; older responses leave it undefined.
  *
  * `providerMetadata` always carries the provider's raw usage payload —
  * keyed by provider name (`{ openai: ... }`, `{ anthropic: ... }`, etc.)
@@ -191,10 +190,16 @@ export const ToolError = Schema.Struct({
 }).annotate({ identifier: "LLM.Event.ToolError" })
 export type ToolError = Schema.Schema.Type<typeof ToolError>
 
+export const FinishReasonDetails = Schema.Struct({
+  normalized: FinishReason,
+  raw: Schema.optional(Schema.String),
+}).annotate({ identifier: "LLM.FinishReasonDetails" })
+export type FinishReasonDetails = Schema.Schema.Type<typeof FinishReasonDetails>
+
 export const StepFinish = Schema.Struct({
   type: Schema.tag("step-finish"),
   index: Schema.Number,
-  reason: FinishReason,
+  reason: FinishReasonDetails,
   usage: Schema.optional(Usage),
   providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.StepFinish" })
@@ -202,7 +207,7 @@ export type StepFinish = Schema.Schema.Type<typeof StepFinish>
 
 export const Finish = Schema.Struct({
   type: Schema.tag("finish"),
-  reason: FinishReason,
+  reason: FinishReasonDetails,
   usage: Schema.optional(Usage),
   providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.Finish" })
@@ -308,29 +313,6 @@ export const LLMEvent = Object.assign(llmEventTagged, {
 })
 export type LLMEvent = Schema.Schema.Type<typeof llmEventTagged>
 
-export class PreparedRequest extends Schema.Class<PreparedRequest>("LLM.PreparedRequest")({
-  id: Schema.String,
-  route: RouteID,
-  protocol: ProtocolID,
-  model: ModelSchema,
-  body: Schema.Unknown,
-  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
-}) {}
-
-/**
- * A `PreparedRequest` whose `body` is typed as `Body`. Use with the generic
- * on `LLMClient.prepare<Body>(...)` when the caller knows which route their
- * request will resolve to and wants its native shape statically exposed
- * (debug UIs, request previews, plan rendering).
- *
- * The runtime body is identical — the route still emits `body: unknown` — so
- * this is a type-level assertion the caller makes about what they expect to
- * find. The prepare runtime does not validate the assertion.
- */
-export type PreparedRequestOf<Body> = Omit<PreparedRequest, "body"> & {
-  readonly body: Body
-}
-
 const responseText = (events: ReadonlyArray<LLMEvent>) =>
   events
     .filter(LLMEvent.is.textDelta)
@@ -365,7 +347,7 @@ interface ResponseState {
   readonly events: ReadonlyArray<LLMEvent>
   readonly message: Message
   readonly usage?: Usage
-  readonly finishReason?: FinishReason
+  readonly finishReason?: FinishReasonDetails
   readonly textParts: Readonly<Record<string, ContentAssembly>>
   readonly reasoningParts: Readonly<Record<string, ContentAssembly>>
   readonly toolInputs: Readonly<Record<string, ToolInputAssembly>>
@@ -393,7 +375,7 @@ const appendEvent = (state: ResponseState, event: LLMEvent): ResponseState => {
     return {
       ...state,
       events,
-      finishReason: state.finishReason ?? "error",
+      finishReason: state.finishReason ?? { normalized: "error" },
     }
   }
   return {
@@ -580,7 +562,7 @@ export class LLMResponse extends Schema.Class<LLMResponse>("LLM.Response")({
   message: Message,
   events: Schema.Array(LLMEvent),
   usage: Schema.optional(Usage),
-  finishReason: FinishReason,
+  finishReason: FinishReasonDetails,
 }) {
   /** Concatenated assistant text assembled from streamed `text-delta` events. */
   get text() {

packages/ai/src/schema/messages.ts

@@ -1,5 +1,5 @@
 import { Schema } from "effect"
-import { ToolContent, ToolFileContent, ToolTextContent } from "@opencode-ai/schema/llm"
+import { Tool } from "@opencode-ai/schema/tool"
 import { JsonSchema, MessageRole, ProviderMetadata } from "./ids"
 import { CacheHint, CachePolicy, GenerationOptions, HttpOptions, ModelSchema, ProviderOptions } from "./options"
 import { isRecord } from "../utils/record"
@@ -40,8 +40,6 @@ export const MediaPart = Schema.Struct({
 }).annotate({ identifier: "LLM.Content.Media" })
 export type MediaPart = Schema.Schema.Type<typeof MediaPart>
 
-export { ToolContent, ToolFileContent, ToolTextContent }
-
 const isToolResultValue = (value: unknown): value is ToolResultValue =>
   isRecord(value) &&
   (value.type === "text" || value.type === "json" || value.type === "error" || value.type === "content") &&
@@ -63,7 +61,7 @@ export const ToolResultValue = Object.assign(
     }),
     Schema.Struct({
       type: Schema.Literal("content"),
-      value: Schema.Array(ToolContent),
+      value: Schema.Array(Tool.Content),
     }),
   ]).annotate({ identifier: "LLM.ToolResult" }),
   {
@@ -79,16 +77,16 @@ export type ToolResultValue = Schema.Schema.Type<typeof ToolResultValue>
 
 export interface ToolOutput {
   readonly structured: unknown
-  readonly content: ReadonlyArray<ToolContent>
+  readonly content: ReadonlyArray<Tool.Content>
 }
 
 export const ToolOutput = Object.assign(
   Schema.Struct({
     structured: Schema.Unknown,
-    content: Schema.Array(ToolContent),
+    content: Schema.Array(Tool.Content),
   }).annotate({ identifier: "LLM.ToolOutput" }),
   {
-    make: (structured: unknown, content: ReadonlyArray<ToolContent> = []): ToolOutput => ({ structured, content }),
+    make: (structured: unknown, content: ReadonlyArray<Tool.Content> = []): ToolOutput => ({ structured, content }),
     fromResultValue: (result: ToolResultValue): ToolOutput | undefined => {
       switch (result.type) {
         case "json":
@@ -261,13 +259,6 @@ export namespace ToolChoice {
   }
 }
 
-export const ResponseFormat = Schema.Union([
-  Schema.Struct({ type: Schema.Literal("text") }),
-  Schema.Struct({ type: Schema.Literal("json"), schema: JsonSchema }),
-  Schema.Struct({ type: Schema.Literal("tool"), tool: ToolDefinition }),
-]).pipe(Schema.toTaggedUnion("type"))
-export type ResponseFormat = Schema.Schema.Type<typeof ResponseFormat>
-
 export class LLMRequest extends Schema.Class<LLMRequest>("LLM.Request")({
   id: Schema.optional(Schema.String),
   model: ModelSchema,
@@ -278,7 +269,6 @@ export class LLMRequest extends Schema.Class<LLMRequest>("LLM.Request")({
   generation: Schema.optional(GenerationOptions),
   providerOptions: Schema.optional(ProviderOptions),
   http: Schema.optional(HttpOptions),
-  responseFormat: Schema.optional(ResponseFormat),
   cache: Schema.optional(CachePolicy),
   metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
 }) {}
@@ -296,7 +286,6 @@ export namespace LLMRequest {
     generation: request.generation,
     providerOptions: request.providerOptions,
     http: request.http,
-    responseFormat: request.responseFormat,
     cache: request.cache,
     metadata: request.metadata,
   })

packages/ai/src/schema/options.ts

@@ -178,7 +178,8 @@ export namespace ModelCompatibility {
   export const make = (input: Input) => (input instanceof ModelCompatibility ? input : new ModelCompatibility(input))
 }
 
-export class Model {
+export class Model<Options extends ProviderOptions = ProviderOptions> {
+  declare protected readonly _ProviderOptions: Options
   readonly id: ModelID
   readonly provider: ProviderID
   readonly route: AnyRoute
@@ -193,8 +194,8 @@ export class Model {
     this.compatibility = input.compatibility
   }
 
-  static make(input: Model.Input) {
-    return new Model({
+  static make<Options extends ProviderOptions = ProviderOptions>(input: Model.Input) {
+    return new Model<Options>({
       id: ModelID.make(input.id),
       provider: ProviderID.make(input.provider),
       route: input.route,
@@ -203,7 +204,7 @@ export class Model {
     })
   }
 
-  static input(model: Model): Model.ConstructorInput {
+  static input<Options extends ProviderOptions>(model: Model<Options>): Model.ConstructorInput {
     return {
       id: model.id,
       provider: model.provider,
@@ -213,9 +214,9 @@ export class Model {
     }
   }
 
-  static update(model: Model, patch: Partial<Model.Input>) {
+  static update<Options extends ProviderOptions>(model: Model<Options>, patch: Partial<Model.Input>) {
     if (Object.keys(patch).length === 0) return model
-    return Model.make({
+    return Model.make<Options>({
       ...Model.input(model),
       ...patch,
     })
@@ -241,6 +242,8 @@ export namespace Model {
 
 export type ModelInput = Model.Input
 
+export type ModelProviderOptions<SelectedModel> = SelectedModel extends Model<infer Options> ? Options : never
+
 export const ModelSchema = Schema.declare((value): value is Model => value instanceof Model, { expected: "LLM.Model" })
 
 export class CacheHint extends Schema.Class<CacheHint>("LLM.CacheHint")({
@@ -251,11 +254,11 @@ export class CacheHint extends Schema.Class<CacheHint>("LLM.CacheHint")({
 // Auto-placement policy for prompt caching. The protocol-neutral lowering step
 // reads this and injects `CacheHint`s at the configured boundaries; the
 // per-protocol body builders then translate those hints into wire markers as
-// usual. `"auto"` is the recommended default for agent loops — it places one
-// breakpoint at the last tool definition, one at the last system part, and one
-// at the latest user message. The combination of provider invalidation
-// hierarchy (tools → system → messages) and Anthropic/Bedrock's 20-block
-// lookback means three trailing breakpoints reliably cover the static prefix.
+// usual. `"auto"` is the recommended default for agent loops — it places
+// breakpoints at the last tool definition, the first and last distinct system
+// parts, and the conversation tail. The rolling message breakpoint keeps a
+// prior cache entry within Anthropic/Bedrock's 20-block lookback during long
+// tool loops.
 //
 // Pass `"none"` to opt out entirely (the legacy behavior). Pass the granular
 // object form to override individual choices.

packages/ai/src/testing.ts

@@ -0,0 +1,156 @@
+export * as TestLLM from "./testing"
+
+import { LLMClient, type Interface as LLMClientShape } from "./route/client"
+import {
+  LLMEvent,
+  LLMResponse,
+  type FinishReasonDetails,
+  type LLMError,
+  type LLMRequest,
+  type UsageInput,
+} from "./schema"
+import { Context, Deferred, Effect, Latch, Layer, Queue, Scope, Stream } from "effect"
+
+export type Response = readonly LLMEvent[] | Stream.Stream<LLMEvent, LLMError>
+
+export type Gate = Readonly<{ started: Effect.Effect<void>; release: Effect.Effect<void> }>
+
+export interface Interface {
+  readonly requests: LLMRequest[]
+  readonly push: (...responses: readonly Response[]) => Effect.Effect<void>
+  readonly always: (response: Response) => Effect.Effect<void>
+  readonly wait: (count: number) => Effect.Effect<void>
+  readonly gate: Effect.Effect<Gate, never, Scope.Scope>
+  readonly client: LLMClientShape
+}
+
+export interface LayerOptions {
+  readonly transformRequest?: (request: LLMRequest) => LLMRequest
+  /** Used after the one-shot response queue is exhausted. Omit to defect on unexpected requests. */
+  readonly fallback?: Response
+}
+
+export class Service extends Context.Service<Service, Interface>()("@opencode/ai/TestLLM") {}
+
+export const complete = (
+  options: { readonly reason: FinishReasonDetails; readonly usage?: UsageInput },
+  ...events: readonly LLMEvent[]
+) => [
+  LLMEvent.stepStart({ index: 0 }),
+  ...events,
+  LLMEvent.stepFinish({ index: 0, reason: options.reason, usage: options.usage }),
+  LLMEvent.finish({ reason: options.reason }),
+]
+
+export const stop = (...events: readonly LLMEvent[]) => complete({ reason: { normalized: "stop" } }, ...events)
+
+export const toolCalls = (...events: readonly LLMEvent[]) =>
+  complete({ reason: { normalized: "tool-calls" } }, ...events)
+
+const textEvents = (value: string, id: string) => [
+  LLMEvent.textStart({ id }),
+  LLMEvent.textDelta({ id, text: value }),
+  LLMEvent.textEnd({ id }),
+]
+
+export const text = (value: string, id: string) => stop(...textEvents(value, id))
+
+export const textWithUsage = (value: string, id: string, inputTokens: number) =>
+  complete(
+    { reason: { normalized: "stop" }, usage: { inputTokens, nonCachedInputTokens: inputTokens } },
+    ...textEvents(value, id),
+  )
+
+export const tool = (id: string, name: string, input: unknown) => toolCalls(LLMEvent.toolCall({ id, name, input }))
+
+export const failAfter = (error: LLMError, ...events: readonly LLMEvent[]) =>
+  Stream.fromIterable(events).pipe(Stream.concat(Stream.fail(error)))
+
+export const hangAfter = (...events: readonly LLMEvent[]) => Stream.concat(Stream.fromIterable(events), Stream.never)
+
+const toStream = (response: Response) => (Stream.isStream(response) ? response : Stream.fromIterable(response))
+
+export const layer = (options: LayerOptions = {}) =>
+  Layer.effect(
+    Service,
+    Effect.gen(function* () {
+      const requests: LLMRequest[] = []
+      const responses: Response[] = []
+      let started = Deferred.makeUnsafe<void>()
+      let fallback = options.fallback
+      let activeGate: { readonly started: Queue.Queue<void>; readonly release: Latch.Latch } | undefined
+      const wait = (count: number): Effect.Effect<void> =>
+        Effect.suspend(() =>
+          requests.length >= count ? Effect.void : Deferred.await(started).pipe(Effect.andThen(wait(count))),
+        )
+
+      const stream = ((request: LLMRequest) => {
+        requests.push(options.transformRequest?.(request) ?? request)
+        const waiting = started
+        started = Deferred.makeUnsafe()
+        Deferred.doneUnsafe(waiting, Effect.void)
+        const response = responses.shift() ?? fallback
+        if (!response) return Stream.die(new Error(`TestLLM has no response for request ${requests.length}`))
+        const streamed = toStream(response)
+        const gate = activeGate
+        if (!gate) return streamed
+        return Stream.unwrap(
+          Queue.offer(gate.started, undefined).pipe(Effect.andThen(gate.release.await), Effect.as(streamed)),
+        )
+      }) as LLMClientShape["stream"]
+      const client = LLMClient.Service.of({
+        stream,
+        generate: (request) =>
+          stream(request).pipe(
+            Stream.runFold(LLMResponse.empty, LLMResponse.reduce),
+            Effect.flatMap((state) => {
+              const response = LLMResponse.complete(state)
+              if (response) return Effect.succeed(response)
+              return Effect.die("TestLLM response ended without a terminal finish event")
+            }),
+          ),
+      })
+
+      return Service.of({
+        requests,
+        push: (...input) =>
+          Effect.sync(() => {
+            responses.push(...input)
+          }),
+        always: (response) =>
+          Effect.sync(() => {
+            fallback = response
+          }),
+        wait,
+        gate: Effect.gen(function* () {
+          const gate = {
+            started: yield* Effect.acquireRelease(Queue.unbounded<void>(), Queue.shutdown),
+            release: yield* Latch.make(),
+          }
+          activeGate = gate
+          const release = Effect.sync(() => {
+            if (activeGate === gate) activeGate = undefined
+          }).pipe(Effect.andThen(gate.release.open), Effect.asVoid)
+          yield* Effect.addFinalizer(() => release)
+          return {
+            started: Queue.take(gate.started),
+            release,
+          }
+        }),
+        client,
+      })
+    }),
+  )
+
+export const clientLayer = Layer.effect(
+  LLMClient.Service,
+  Effect.map(Service, (service) => service.client),
+)
+
+export const push = (...responses: readonly Response[]) => Service.use((service) => service.push(...responses))
+
+export const always = (response: Response) => Service.use((service) => service.always(response))
+
+export const wait = (count: number) => Service.use((service) => service.wait(count))
+
+export const gate = Service.use((service) => service.gate)

packages/ai/src/tool-runtime.ts

@@ -28,7 +28,7 @@ export const dispatch = (tools: Tools, call: ToolCallPart): Effect.Effect<Dispat
 
   return decodeAndExecute(tool, call).pipe(
     Effect.map((value) => result(call, value)),
-    Effect.catchTag("LLM.ToolFailure", (failure) =>
+    Effect.catchTag("Tool.Error", (failure) =>
       Effect.succeed(result(call, { type: "error", value: failure.message }, failure.error)),
     ),
   )

packages/ai/src/tool.ts

@@ -1,7 +1,7 @@
 import { Effect, JsonSchema, Schema } from "effect"
+import { Tool } from "@opencode-ai/schema/tool"
 import type {
   ToolCallPart,
-  ToolContent,
   ToolDefinition as ToolDefinitionClass,
   ToolOutput as ToolOutputType,
 } from "./schema"
@@ -31,7 +31,7 @@ export interface ToolModelOutputInput<Parameters, Output> {
 
 export type ToolToModelOutput<Parameters extends ToolSchema<any>, Success extends ToolSchema<any>> = (
   input: ToolModelOutputInput<Schema.Schema.Type<Parameters>, Success["Encoded"]>,
-) => ReadonlyArray<ToolContent>
+) => ReadonlyArray<Tool.Content>
 
 /**
  * A type-safe LLM tool. Each tool bundles its own description, parameter
@@ -95,7 +95,7 @@ type DynamicToolConfig = {
   readonly jsonSchema: JsonSchema.JsonSchema
   readonly outputSchema?: JsonSchema.JsonSchema
   readonly execute?: (params: unknown, context?: ToolExecuteContext) => Effect.Effect<unknown, ToolFailure>
-  readonly toModelOutput?: (input: ToolModelOutputInput<unknown, unknown>) => ReadonlyArray<ToolContent>
+  readonly toModelOutput?: (input: ToolModelOutputInput<unknown, unknown>) => ReadonlyArray<Tool.Content>
   readonly toStructuredOutput?: (output: unknown) => unknown
 }
 
@@ -151,7 +151,7 @@ export function make(config: {
   readonly jsonSchema: JsonSchema.JsonSchema
   readonly outputSchema?: JsonSchema.JsonSchema
   readonly execute: (params: unknown, context?: ToolExecuteContext) => Effect.Effect<unknown, ToolFailure>
-  readonly toModelOutput?: (input: ToolModelOutputInput<unknown, unknown>) => ReadonlyArray<ToolContent>
+  readonly toModelOutput?: (input: ToolModelOutputInput<unknown, unknown>) => ReadonlyArray<Tool.Content>
   readonly toStructuredOutput?: (output: unknown) => unknown
 }): AnyExecutableTool
 export function make(config: {
@@ -159,7 +159,7 @@ export function make(config: {
   readonly jsonSchema: JsonSchema.JsonSchema
   readonly outputSchema?: JsonSchema.JsonSchema
   readonly execute?: undefined
-  readonly toModelOutput?: (input: ToolModelOutputInput<unknown, unknown>) => ReadonlyArray<ToolContent>
+  readonly toModelOutput?: (input: ToolModelOutputInput<unknown, unknown>) => ReadonlyArray<Tool.Content>
   readonly toStructuredOutput?: (output: unknown) => unknown
 }): AnyTool
 export function make(config: TypedToolConfig | DynamicToolConfig): AnyTool {
@@ -236,7 +236,7 @@ const toJsonSchema = (schema: Schema.Top): JsonSchema.JsonSchema => {
 }
 
 const project = (
-  toModelOutput: ((input: ToolModelOutputInput<any, any>) => ReadonlyArray<ToolContent>) | undefined,
+  toModelOutput: ((input: ToolModelOutputInput<any, any>) => ReadonlyArray<Tool.Content>) | undefined,
   toStructuredOutput: ((output: unknown) => unknown) | undefined,
   parameters: unknown,
   callID: ToolCallPart["id"],

packages/ai/test/adapter.test.ts

@@ -1,7 +1,8 @@
 import { describe, expect } from "bun:test"
 import { Effect, Schema, Stream } from "effect"
-import { LLM, LLMResponse } from "../src"
+import { LLM, LLMRequest, LLMResponse } from "../src"
 import { Route, Endpoint, LLMClient, Protocol, type FramingDef } from "../src/route"
+import { compileRequest } from "../src/route/client"
 import { Model } from "../src/schema"
 import { testEffect } from "./lib/effect"
 import { dynamicResponse } from "./lib/http"
@@ -40,7 +41,7 @@ const fakeFraming: FramingDef<FakeEvent> = {
 
 const raiseEvent = (event: FakeEvent): import("../src/schema").LLMEvent =>
   event.type === "finish"
-    ? { type: "finish", reason: event.reason }
+    ? { type: "finish", reason: { normalized: event.reason } }
     : { type: "text-delta", id: "text-0", text: event.text }
 
 const fakeProtocol = Protocol.make<FakeBody, FakeEvent, FakeEvent, void>({
@@ -139,9 +140,8 @@ describe("llm route", () => {
 
   it.effect("selects routes by model route value", () =>
     Effect.gen(function* () {
-      const llm = yield* LLMClient.Service
-      const prepared = yield* llm.prepare(
-        LLM.updateRequest(request, { model: updateModel(request.model, { route: configuredGemini }) }),
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, { model: updateModel(request.model, { route: configuredGemini }) }),
       )
 
       expect(prepared.route).toBe("gemini-fake")
@@ -173,8 +173,8 @@ describe("llm route", () => {
         framing: fakeFraming,
       })
 
-      const prepared = yield* (yield* LLMClient.Service).prepare(
-        LLM.updateRequest(request, { model: updateModel(request.model, { route: duplicate }) }),
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, { model: updateModel(request.model, { route: duplicate }) }),
       )
 
       expect(prepared.body).toEqual({ body: "late-default" })

packages/ai/test/auth-options.types.ts

@@ -137,15 +137,26 @@ Azure.configure({ apiKey: "azure-key", resourceName: "resource" }).chat("deploym
 Azure.configure({ resourceName: "resource", apiKey: "azure-key", auth: Auth.header("api-key", "override") })
 
 Anthropic.configure({ apiKey: "anthropic-key" }).model("claude-haiku")
+Anthropic.configure({
+  apiKey: "anthropic-key",
+  providerOptions: {
+    anthropic: { thinking: { type: "enabled", budgetTokens: 1_024 }, effort: "high" },
+  },
+}).model("claude-haiku")
 // @ts-expect-error Anthropic model selectors only accept model ids.
 Anthropic.configure({ apiKey: "anthropic-key" }).model("claude-haiku", {})
 // @ts-expect-error Anthropic package settings accept only one auth source.
 Anthropic.model("claude-sonnet-4-6", { apiKey: "anthropic-key", authToken: "anthropic-token" })
+// @ts-expect-error Enabled Anthropic thinking requires a token budget.
+Anthropic.configure({ providerOptions: { anthropic: { thinking: { type: "enabled" } } } })
+// @ts-expect-error Anthropic thinking budgets must be numbers.
+Anthropic.configure({ providerOptions: { anthropic: { thinking: { type: "enabled", budgetTokens: "large" } } } })
 
 AnthropicCompatible.configure({
   apiKey: "messages-key",
   baseURL: "https://messages.example.com/v1",
   provider: "example",
+  providerOptions: { anthropic: { thinking: { type: "disabled" } } },
 }).model("compatible-model")
 // @ts-expect-error Anthropic-compatible providers require a base URL.
 AnthropicCompatible.configure({ apiKey: "messages-key" })
@@ -159,10 +170,19 @@ AnthropicCompatible.model("compatible-model", {
 })
 
 Google.configure({ apiKey: "google-key" }).model("gemini-2.5-flash")
+Google.configure({
+  apiKey: "google-key",
+  providerOptions: { gemini: { thinkingConfig: { thinkingBudget: 0, includeThoughts: false } } },
+}).model("gemini-2.5-flash")
 // @ts-expect-error Google model selectors only accept model ids.
 Google.configure({ apiKey: "google-key" }).model("gemini-2.5-flash", {})
+// @ts-expect-error Gemini thinking budgets must be numbers.
+Google.configure({ providerOptions: { gemini: { thinkingConfig: { thinkingBudget: "large" } } } })
 
-GoogleVertex.configure({ apiKey: "vertex-key" }).model("gemini-3.5-flash")
+GoogleVertex.configure({
+  apiKey: "vertex-key",
+  providerOptions: { gemini: { thinkingConfig: { thinkingBudget: 1_024 } } },
+}).model("gemini-3.5-flash")
 GoogleVertex.configure({ accessToken: "vertex-token", project: "project" }).model("gemini-3.5-flash")
 GoogleVertex.configure({ auth: Auth.bearer("vertex-token"), project: "project" }).model("gemini-3.5-flash")
 // @ts-expect-error Vertex Gemini model selectors only accept model ids.
@@ -208,7 +228,11 @@ GoogleVertexResponses.configure({
   project: "project",
 })
 
-GoogleVertexMessages.configure({ accessToken: "vertex-token", project: "project" }).model("claude-sonnet-4-6")
+GoogleVertexMessages.configure({
+  accessToken: "vertex-token",
+  project: "project",
+  providerOptions: { anthropic: { thinking: { type: "adaptive", display: "omitted" }, effort: "low" } },
+}).model("claude-sonnet-4-6")
 // @ts-expect-error Vertex Messages package settings do not accept API keys.
 GoogleVertexMessages.model("claude-sonnet-4-6", { apiKey: "vertex-key", project: "project" })
 GoogleVertexMessages.configure({ auth: Auth.bearer("vertex-token"), project: "project" }).model("claude-sonnet-4-6")

packages/ai/test/cache-policy.test.ts

@@ -1,7 +1,8 @@
 import { describe, expect, test } from "bun:test"
 import { Effect } from "effect"
 import { CacheHint, LLM, Message } from "../src"
-import { Auth, LLMClient } from "../src/route"
+import { Auth } from "../src/route"
+import { compileRequest } from "../src/route/client"
 import { AmazonBedrock } from "../src/providers"
 import * as AnthropicMessages from "../src/protocols/anthropic-messages"
 import * as Gemini from "../src/protocols/gemini"
@@ -31,7 +32,7 @@ const geminiModel = Gemini.route
 describe("applyCachePolicy", () => {
   it.effect("undefined cache resolves to 'auto' (the recommended default)", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "You are concise.",
@@ -39,8 +40,8 @@ describe("applyCachePolicy", () => {
         }),
       )
 
-      // No explicit cache field → auto policy fires → last system part + latest
-      // user message both get cache_control markers.
+      // A single system block is both the first and last boundary, so the auto
+      // policy deduplicates it and still marks the conversation tail.
       expect(prepared.body).toMatchObject({
         system: [{ type: "text", text: "You are concise.", cache_control: { type: "ephemeral" } }],
         messages: [{ role: "user", content: [{ type: "text", text: "hi", cache_control: { type: "ephemeral" } }] }],
@@ -48,12 +49,15 @@ describe("applyCachePolicy", () => {
     }),
   )
 
-  it.effect("'auto' marks the last tool, last system part, and latest user message on Anthropic", () =>
+  it.effect("'auto' marks the last tool, first and last system parts, and final message boundary on Anthropic", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
-          system: "Sys A",
+          system: [
+            { type: "text", text: "Base agent" },
+            { type: "text", text: "Project instructions" },
+          ],
           tools: [{ name: "t1", description: "t1", inputSchema: { type: "object", properties: {} } }],
           messages: [
             Message.user("first user"),
@@ -66,7 +70,10 @@ describe("applyCachePolicy", () => {
 
       expect(prepared.body).toMatchObject({
         tools: [{ name: "t1", cache_control: { type: "ephemeral" } }],
-        system: [{ type: "text", text: "Sys A", cache_control: { type: "ephemeral" } }],
+        system: [
+          { type: "text", text: "Base agent", cache_control: { type: "ephemeral" } },
+          { type: "text", text: "Project instructions", cache_control: { type: "ephemeral" } },
+        ],
         messages: [
           { role: "user", content: [{ type: "text", text: "first user" }] },
           { role: "assistant", content: [{ type: "text", text: "assistant reply" }] },
@@ -81,7 +88,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'auto' is a no-op on OpenAI (implicit caching protocol)", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: openaiModel,
           system: "Sys",
@@ -100,7 +107,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'auto' is a no-op on Gemini (out-of-band caching protocol)", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: geminiModel,
           system: "Sys",
@@ -117,10 +124,13 @@ describe("applyCachePolicy", () => {
 
   it.effect("'auto' on Bedrock emits cachePoint markers in the right places", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: bedrockModel,
-          system: "Sys",
+          system: [
+            { type: "text", text: "Base agent" },
+            { type: "text", text: "Project instructions" },
+          ],
           tools: [{ name: "t1", description: "t1", inputSchema: { type: "object", properties: {} } }],
           messages: [Message.user("first user"), Message.assistant("reply"), Message.user("latest user")],
           cache: "auto",
@@ -131,7 +141,12 @@ describe("applyCachePolicy", () => {
         toolConfig: {
           tools: [{ toolSpec: { name: "t1" } }, { cachePoint: { type: "default" } }],
         },
-        system: [{ text: "Sys" }, { cachePoint: { type: "default" } }],
+        system: [
+          { text: "Base agent" },
+          { cachePoint: { type: "default" } },
+          { text: "Project instructions" },
+          { cachePoint: { type: "default" } },
+        ],
         messages: [
           { role: "user", content: [{ text: "first user" }] },
           { role: "assistant", content: [{ text: "reply" }] },
@@ -143,7 +158,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'none' disables auto placement even when manual hints exist", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "Sys",
@@ -162,7 +177,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("granular object form: tools-only marks just tools", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "Sys",
@@ -181,7 +196,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("auto policy preserves manual CacheHints on other parts", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: [
@@ -193,15 +208,61 @@ describe("applyCachePolicy", () => {
         }),
       )
 
-      const body = prepared.body as { system: Array<{ text: string; cache_control?: unknown }> }
+      const body = prepared.body as {
+        system: Array<{ text: string; cache_control?: unknown }>
+        messages: Array<{ content: Array<{ cache_control?: unknown }> }>
+      }
       expect(body.system[0]?.cache_control).toEqual({ type: "ephemeral", ttl: "1h" })
       expect(body.system[1]?.cache_control).toEqual({ type: "ephemeral" })
+      expect(body.messages[0]?.content[0]?.cache_control).toEqual({ type: "ephemeral" })
+    }),
+  )
+
+  it.effect("auto policy stays within the four-breakpoint cap when preserving manual hints", () =>
+    Effect.gen(function* () {
+      const request = LLM.request({
+        model: anthropicModel,
+        system: [
+          { type: "text", text: "Base agent" },
+          {
+            type: "text",
+            text: "Manual context",
+            cache: new CacheHint({ type: "ephemeral", ttlSeconds: 3600 }),
+          },
+          { type: "text", text: "Project instructions" },
+        ],
+        tools: [{ name: "t1", description: "t1", inputSchema: { type: "object", properties: {} } }],
+        prompt: "hi",
+        cache: "auto",
+      })
+      const applied = applyCachePolicy(request)
+      expect(applied.tools[0]?.cache).toBeDefined()
+      expect(applied.system.map((part) => part.cache !== undefined)).toEqual([true, true, true])
+      const tail = applied.messages[0]!.content[0]!
+      expect("cache" in tail ? tail.cache : undefined).toBeUndefined()
+      expect(applyCachePolicy(applied)).toBe(applied)
+
+      const prepared = yield* compileRequest(request)
+
+      const body = prepared.body as {
+        tools: Array<{ cache_control?: unknown }>
+        system: Array<{ cache_control?: unknown }>
+        messages: Array<{ content: Array<{ cache_control?: unknown }> }>
+      }
+      const marked = [
+        ...body.tools.map((tool) => tool.cache_control),
+        ...body.system.map((part) => part.cache_control),
+        ...body.messages.flatMap((message) => message.content.map((part) => part.cache_control)),
+      ].filter((cache) => cache !== undefined)
+      expect(marked).toHaveLength(4)
+      expect(body.system[1]?.cache_control).toEqual({ type: "ephemeral", ttl: "1h" })
+      expect(body.messages[0]?.content[0]?.cache_control).toBeUndefined()
     }),
   )
 
   it.effect("ttlSeconds in the policy flows through to wire markers", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "Sys",
@@ -218,7 +279,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("messages: { tail: 2 } marks the last 2 message boundaries", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           messages: [Message.user("u1"), Message.assistant("a1"), Message.user("u2"), Message.assistant("a2")],
@@ -236,7 +297,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'latest-assistant' marks the last assistant message", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           messages: [Message.user("u1"), Message.assistant("a1"), Message.user("u2")],

packages/ai/test/compile.test.ts

@@ -4,6 +4,7 @@ import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, mergeProviderOptions } from "../src"
 import { AnthropicMessages, OpenAIChat } from "../src/protocols"
 import { Auth, LLMClient } from "../src/route"
+import { compileRequest } from "../src/route/client"
 import { it } from "./lib/effect"
 import { dynamicResponse } from "./lib/http"
 import { deltaChunk } from "./lib/openai-chunks"
@@ -44,7 +45,7 @@ describe("request option precedence", () => {
     })
   })
 
-  it.effect("prepares bodies with route defaults, model defaults, and call options in order", () =>
+  it.effect("compiles bodies with route defaults, model defaults, and call options in order", () =>
     Effect.gen(function* () {
       const route = OpenAIChat.route.with({
         endpoint: { baseURL: "https://api.openai.test/v1/" },
@@ -59,7 +60,7 @@ describe("request option precedence", () => {
           providerOptions: { openai: { reasoningEffort: "medium" } },
         },
       })
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "Say hello.",
@@ -141,7 +142,7 @@ describe("request option precedence", () => {
       const model = OpenAIChat.route
         .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
         .model({ id: "gpt-4o-mini" })
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           prompt: "Say hello.",
@@ -164,10 +165,8 @@ describe("request option precedence", () => {
         limits: { output: 128 },
       })
       const model = route.model({ id: "claude-sonnet-4-5", defaults: { limits: { output: 64 } } })
-      const withoutMaxTokens = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
-        LLM.request({ model, prompt: "Say hello.", cache: "none" }),
-      )
-      const withMaxTokens = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const withoutMaxTokens = yield* compileRequest(LLM.request({ model, prompt: "Say hello.", cache: "none" }))
+      const withMaxTokens = yield* compileRequest(
         LLM.request({ model, prompt: "Say hello.", cache: "none", generation: { maxTokens: 32 } }),
       )
 

packages/ai/test/exports.test.ts

@@ -11,8 +11,15 @@ import {
   XAI,
 } from "@opencode-ai/ai/providers"
 import * as GitHubCopilot from "@opencode-ai/ai/providers/github-copilot"
-import { OpenAIChat, OpenAICompatibleChat, OpenAICompatibleResponses, OpenAIResponses } from "@opencode-ai/ai/protocols"
+import {
+  OpenAIChat,
+  OpenAICompatibleChat,
+  OpenAICompatibleResponses,
+  OpenAIResponses,
+  OpenResponses,
+} from "@opencode-ai/ai/protocols"
 import * as AnthropicMessages from "@opencode-ai/ai/protocols/anthropic-messages"
+import { TestLLM } from "@opencode-ai/ai/testing"
 
 describe("public exports", () => {
   test("root exposes app-facing runtime APIs", () => {
@@ -22,6 +29,7 @@ describe("public exports", () => {
     expect(ImageInput.bytes).toBeFunction()
     expect(Provider.make).toBeFunction()
     expect(ProviderSubpath.make).toBe(Provider.make)
+    expect(TestLLM.layer).toBeFunction()
   })
 
   test("route barrel exposes route-authoring APIs", () => {
@@ -74,7 +82,9 @@ describe("public exports", () => {
   test("protocol barrels expose supported low-level routes", () => {
     expect(OpenAIChat.route.id).toBe("openai-chat")
     expect(OpenAICompatibleChat.route.id).toBe("openai-compatible-chat")
+    expect(OpenResponses.protocol.id).toBe("open-responses")
     expect(OpenAICompatibleResponses.route.id).toBe("openai-compatible-responses")
+    expect(OpenAICompatibleResponses.route.protocol).toBe("open-responses")
     expect(OpenAIResponses.route.id).toBe("openai-responses")
     expect(OpenAIResponses.webSocketRoute.id).toBe("openai-responses-websocket")
     expect(AnthropicMessages.route.id).toBe("anthropic-messages")

packages/ai/test/lib/tool-runtime.ts

@@ -83,7 +83,7 @@ const indexStep = (event: LLMEvent, index: number): LLMEvent => {
 const stepState = (events: ReadonlyArray<LLMEvent>) => {
   const assistantContent: ContentPart[] = []
   const toolCalls: ToolCallPart[] = []
-  let reason: Extract<LLMEvent, { type: "finish" }>["reason"] = "unknown"
+  let reason: Extract<LLMEvent, { type: "finish" }>["reason"] = { normalized: "unknown" }
   let usage: Usage | undefined
   let providerMetadata: ProviderMetadata | undefined
 

packages/ai/test/llm-option-types.types.ts

@@ -0,0 +1,47 @@
+import { Schema } from "effect"
+import { LLM, type Model, type ModelProviderOptions, type ProviderOptions } from "../src"
+import { OpenAIChat } from "../src/protocols"
+
+interface ExampleOptions {
+  readonly [key: string]: unknown
+  readonly mode?: "fast" | "thorough"
+}
+
+type ExampleProviderOptions = ProviderOptions & {
+  readonly example?: ExampleOptions
+}
+
+const model = OpenAIChat.route
+  .with({ endpoint: { baseURL: "https://example.com/v1" } })
+  .model<ExampleProviderOptions>({ id: "example" })
+
+LLM.request({ model, prompt: "Hello", providerOptions: { example: { mode: "fast" } } })
+LLM.request({ model, prompt: "Hello", providerOptions: { future: { option: true } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Known provider options preserve their value types.
+  providerOptions: { example: { mode: "slow" } },
+})
+
+LLM.generateObject({
+  model,
+  prompt: "Hello",
+  schema: Schema.Struct({ answer: Schema.String }),
+  providerOptions: { example: { mode: "thorough" } },
+})
+
+LLM.generateObject({
+  model,
+  prompt: "Hello",
+  jsonSchema: { type: "object" },
+  // @ts-expect-error Dynamic object generation uses the selected model's provider options.
+  providerOptions: { example: { mode: false } },
+})
+
+declare const generic: Model
+LLM.request({ model: generic, prompt: "Hello", providerOptions: { arbitrary: { option: true } } })
+
+const options: ModelProviderOptions<typeof model> = { example: { mode: "fast" } }
+void options

packages/ai/test/llm.test.ts

@@ -2,7 +2,16 @@ import { describe, expect, test } from "bun:test"
 import { CacheHint, LLM, LLMResponse } from "../src"
 import * as OpenAIChat from "../src/protocols/openai-chat"
 import * as OpenAIResponses from "../src/protocols/openai-responses"
-import { LLMRequest, Message, Model, ToolCallPart, ToolChoice, ToolDefinition, ToolResultPart } from "../src/schema"
+import {
+  GenerationOptions,
+  LLMRequest,
+  Message,
+  Model,
+  ToolCallPart,
+  ToolChoice,
+  ToolDefinition,
+  ToolResultPart,
+} from "../src/schema"
 
 const chatRoute = OpenAIChat.route
 const responsesRoute = OpenAIResponses.route
@@ -31,8 +40,8 @@ describe("llm constructors", () => {
       model: Model.make({ id: "fake-model", provider: "fake", route: chatRoute }),
       prompt: "Say hello.",
     })
-    const updated = LLM.updateRequest(base, {
-      generation: { maxTokens: 20 },
+    const updated = LLMRequest.update(base, {
+      generation: GenerationOptions.make({ maxTokens: 20 }),
       messages: [...base.messages, Message.assistant("Hi.")],
     })
 
@@ -191,7 +200,7 @@ describe("llm constructors", () => {
       LLMResponse.text({
         events: [
           { type: "text-delta", id: "text-0", text: "hi" },
-          { type: "finish", reason: "stop" },
+          { type: "finish", reason: { normalized: "stop" } },
         ],
       }),
     ).toBe("hi")

packages/ai/test/provider-error.test.ts

@@ -58,6 +58,12 @@ describe("provider error classification", () => {
     ).toEqual(["ProviderInternal", "ProviderInternal"])
   })
 
+  test("classifies transient client statuses as provider internal", () => {
+    expect(
+      [408, 409].map((status) => classifyProviderFailure({ message: `HTTP ${status}`, status })._tag),
+    ).toEqual(["ProviderInternal", "ProviderInternal"])
+  })
+
   test("classifies nested provider codes when a top-level code is also present", () => {
     expect(
       [

packages/ai/test/provider-options/anthropic-compatible.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { AnthropicCompatible } from "../../src/providers"
+
+const model = AnthropicCompatible.configure({ baseURL: "https://example.com" }).model("claude")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { anthropic: { effort: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Anthropic effort must be a string.
+  providerOptions: { anthropic: { effort: 1 } },
+})

packages/ai/test/provider-options/anthropic.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { Anthropic } from "../../src/providers"
+
+const model = Anthropic.provider.model("claude-sonnet-4-5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { anthropic: { thinking: { type: "adaptive" } } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Anthropic thinking modes are a fixed union.
+  providerOptions: { anthropic: { thinking: { type: "automatic" } } },
+})

packages/ai/test/provider-options/azure.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { Azure } from "../../src/providers"
+
+const model = Azure.configure({ resourceName: "example" }).responses("deployment")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { store: false } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Azure OpenAI store must be boolean.
+  providerOptions: { openai: { store: "false" } },
+})

packages/ai/test/provider-options/cloudflare.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { CloudflareWorkersAI } from "../../src/providers"
+
+const model = CloudflareWorkersAI.configure({ accountId: "account", apiKey: "test" }).model("model")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { promptCacheKey: "cache" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Cloudflare's OpenAI-compatible prompt cache key must be a string.
+  providerOptions: { openai: { promptCacheKey: 1 } },
+})

packages/ai/test/provider-options/github-copilot.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GitHubCopilot } from "../../src/providers"
+
+const model = GitHubCopilot.configure({ baseURL: "https://example.com" }).model("gpt-5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { reasoningSummary: "auto" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Copilot reasoning summaries use the OpenAI union.
+  providerOptions: { openai: { reasoningSummary: "full" } },
+})

packages/ai/test/provider-options/google-vertex-chat.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GoogleVertexChat } from "../../src/providers"
+
+const model = GoogleVertexChat.configure({ accessToken: "test", project: "project" }).model("gemini")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { serviceTier: "priority" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex OpenAI-compatible service tiers use the OpenAI union.
+  providerOptions: { openai: { serviceTier: "premium" } },
+})

packages/ai/test/provider-options/google-vertex-messages.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GoogleVertexMessages } from "../../src/providers"
+
+const model = GoogleVertexMessages.configure({ accessToken: "test", project: "project" }).model("claude")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { anthropic: { effort: "medium" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex Anthropic effort must be a string.
+  providerOptions: { anthropic: { effort: false } },
+})

packages/ai/test/provider-options/google-vertex-responses.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GoogleVertexResponses } from "../../src/providers"
+
+const model = GoogleVertexResponses.configure({ accessToken: "test", project: "project" }).model("gemini")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openresponses: { textVerbosity: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex Responses verbosity uses the Open Responses union.
+  providerOptions: { openresponses: { textVerbosity: "verbose" } },
+})

packages/ai/test/provider-options/google-vertex.types.ts

@@ -0,0 +1,17 @@
+import { LLM } from "../../src"
+import { GoogleVertex } from "../../src/providers"
+
+const model = GoogleVertex.provider.configure({ apiKey: "test" }).model("gemini-2.5-pro")
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  providerOptions: { gemini: { thinkingConfig: { includeThoughts: true } } },
+})
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex Gemini includeThoughts must be boolean.
+  providerOptions: { gemini: { thinkingConfig: { includeThoughts: "yes" } } },
+})

packages/ai/test/provider-options/google.types.ts

@@ -0,0 +1,17 @@
+import { LLM } from "../../src"
+import { Google } from "../../src/providers"
+
+const model = Google.provider.model("gemini-2.5-pro")
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  providerOptions: { gemini: { thinkingConfig: { thinkingBudget: 1024 } } },
+})
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Gemini thinking budgets must be numeric.
+  providerOptions: { gemini: { thinkingConfig: { thinkingBudget: "large" } } },
+})

packages/ai/test/provider-options/openai-compatible-responses.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenAICompatibleResponses } from "../../src/providers"
+
+const model = OpenAICompatibleResponses.configure({ baseURL: "https://example.com" }).model("model")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openresponses: { reasoningSummary: "detailed" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Open Responses reasoning summaries use a fixed union.
+  providerOptions: { openresponses: { reasoningSummary: "full" } },
+})

packages/ai/test/provider-options/openai-compatible.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenAICompatible } from "../../src/providers"
+
+const model = OpenAICompatible.deepseek.model("deepseek-chat")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { store: false } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error OpenAI-compatible store must be boolean.
+  providerOptions: { openai: { store: "false" } },
+})

packages/ai/test/provider-options/openai.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenAI } from "../../src/providers"
+
+const model = OpenAI.responses("gpt-5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { reasoningEffort: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error OpenAI reasoning effort must be a string.
+  providerOptions: { openai: { reasoningEffort: 1 } },
+})

packages/ai/test/provider-options/openrouter.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenRouter } from "../../src/providers"
+
+const model = OpenRouter.provider.model("anthropic/claude-sonnet-4.5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openrouter: { usage: true } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error OpenRouter usage must be boolean or an option record.
+  providerOptions: { openrouter: { usage: "yes" } },
+})

packages/ai/test/provider-options/xai.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { XAI } from "../../src/providers"
+
+const model = XAI.provider.model("grok-4")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { reasoningEffort: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error xAI's OpenAI-compatible reasoning effort must be a string.
+  providerOptions: { openai: { reasoningEffort: true } },
+})

packages/ai/test/provider-package.test.ts

@@ -59,7 +59,7 @@ describe("provider package entrypoints", () => {
       headers: { "x-application": "opencode" },
       body: { service_tier: "priority" },
       limits: { context: 200_000, output: 64_000 },
-      providerOptions: { openai: { reasoningEffort: "low", store: true } },
+      providerOptions: { openresponses: { reasoningEffort: "low", store: true } },
     })
 
     expect(String(selected.provider)).toBe("example")
@@ -72,7 +72,7 @@ describe("provider package entrypoints", () => {
     expect(selected.route.defaults.http?.body).toEqual({ service_tier: "priority" })
     expect(selected.route.defaults.limits).toEqual({ context: 200_000, output: 64_000 })
     expect(selected.route.defaults.providerOptions).toEqual({
-      openai: { reasoningEffort: "low", store: true },
+      openresponses: { reasoningEffort: "low", store: true },
     })
   })
 
@@ -85,6 +85,7 @@ describe("provider package entrypoints", () => {
       headers: { "x-application": "opencode" },
       body: { metadata: { user_id: "user_1" } },
       limits: { context: 200_000, output: 64_000 },
+      providerOptions: { anthropic: { effort: "low" } },
     })
 
     expect(String(selected.provider)).toBe("example")
@@ -96,6 +97,19 @@ describe("provider package entrypoints", () => {
     expect(selected.route.defaults.headers).toEqual({ "x-application": "opencode" })
     expect(selected.route.defaults.http?.body).toEqual({ metadata: { user_id: "user_1" } })
     expect(selected.route.defaults.limits).toEqual({ context: 200_000, output: 64_000 })
+    expect(selected.route.defaults.providerOptions).toEqual({ anthropic: { effort: "low" } })
+  })
+
+  test("maps Anthropic provider options onto the executable model", async () => {
+    const Anthropic = await import("@opencode-ai/ai/providers/anthropic")
+    const selected = Anthropic.model("claude-sonnet-4-6", {
+      apiKey: "fixture",
+      providerOptions: { anthropic: { thinking: { type: "adaptive" } } },
+    })
+
+    expect(selected.route.defaults.providerOptions).toEqual({
+      anthropic: { thinking: { type: "adaptive" } },
+    })
   })
 
   test("requires an Anthropic-compatible base URL at runtime", async () => {
@@ -235,12 +249,12 @@ describe("provider package entrypoints", () => {
       path: "/chat/completions",
     })
     expect(responses.route.id).toBe("google-vertex-responses")
-    expect(responses.route.protocol).toBe("openai-responses")
+    expect(responses.route.protocol).toBe("open-responses")
     expect(responses.route.endpoint).toMatchObject({
       baseURL: "https://aiplatform.googleapis.com/v1/projects/vertex-project/locations/global/endpoints/openapi",
       path: "/responses",
     })
-    expect(responses.route.defaults.providerOptions).toEqual({ openai: { store: false } })
+    expect(responses.route.defaults.providerOptions).toEqual({ openresponses: { store: false } })
   })
 
   test("rejects conflicting Vertex auth settings at runtime", async () => {

packages/ai/test/provider/anthropic-messages-cache.recorded.test.ts

@@ -1,6 +1,6 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
-import { CacheHint, LLM } from "../../src"
+import { CacheHint, LLM, LLMRequest, Message, ToolCallPart, ToolDefinition } from "../../src"
 import { LLMClient } from "../../src/route"
 import * as Anthropic from "../../src/providers/anthropic"
 import { LARGE_CACHEABLE_SYSTEM } from "../recorded-scenarios"
@@ -24,6 +24,39 @@ const cacheRequest = LLM.request({
   generation: { maxTokens: 16, temperature: 0 },
 })
 
+const lookup = ToolDefinition.make({
+  name: "lookup",
+  description: "Look up a fixture value.",
+  inputSchema: {
+    type: "object",
+    properties: { index: { type: "number" } },
+    required: ["index"],
+    additionalProperties: false,
+  },
+})
+const longToolTurn = [
+  Message.user("Run the fixture lookups."),
+  ...Array.from({ length: 11 }, (_, index) => {
+    const id = `lookup_${index}`
+    return [
+      Message.assistant(ToolCallPart.make({ id, name: lookup.name, input: { index } })),
+      Message.tool({
+        id,
+        name: lookup.name,
+        result: `Fixture result ${index}. `.repeat(80),
+      }),
+    ]
+  }).flat(),
+]
+const longToolTurnRequest = LLM.request({
+  id: "recorded_anthropic_cache_long_tool_turn",
+  model,
+  system: LARGE_CACHEABLE_SYSTEM,
+  messages: longToolTurn,
+  tools: [lookup],
+  generation: { maxTokens: 16, temperature: 0 },
+})
+
 const recorded = recordedTests({
   prefix: "anthropic-messages-cache",
   provider: "anthropic",
@@ -50,4 +83,28 @@ describe("Anthropic Messages cache recorded", () => {
       expect(second.usage?.cacheReadInputTokens ?? 0).toBeGreaterThan(0)
     }),
   )
+
+  recorded.effect.with("keeps a long tool turn inside the cache lookback", { tags: ["cache", "tool"] }, () =>
+    Effect.gen(function* () {
+      const first = yield* LLMClient.generate(longToolTurnRequest)
+      const firstRead = first.usage?.cacheReadInputTokens ?? 0
+      const firstWrite = first.usage?.cacheWriteInputTokens ?? 0
+      const firstCached = firstRead + firstWrite
+      // The prefix may already be warm when recording, so either a read or a
+      // write establishes that Anthropic recognized the cache boundary.
+      expect(firstCached).toBeGreaterThan(0)
+
+      const second = yield* LLMClient.generate(
+        LLMRequest.update(longToolTurnRequest, {
+          messages: [
+            ...longToolTurn,
+            Message.assistant("The fixture lookups are complete."),
+            Message.user("Reply exactly: OK"),
+          ],
+        }),
+      )
+      expect(second.usage?.cacheReadInputTokens ?? 0).toBeGreaterThanOrEqual(firstCached)
+      expect(second.usage?.cacheWriteInputTokens ?? 0).toBeLessThan(firstCached)
+    }),
+  )
 })

packages/ai/test/provider/anthropic-messages.test.ts

@@ -1,8 +1,9 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
-import { CacheHint, LLM, LLMError, Message, ToolCallPart, Usage } from "../../src"
+import { CacheHint, LLM, LLMError, LLMRequest, Message, ToolCallPart, ToolDefinition, Usage } from "../../src"
 import { Auth, LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import * as AnthropicMessages from "../../src/protocols/anthropic-messages"
 import { continuationRequest, nativeAnthropicMessagesContinuation } from "../continuation-scenarios"
 import { it } from "../lib/effect"
@@ -44,7 +45,7 @@ const expectToolResult = (body: AnthropicMessages.AnthropicMessagesBody): Anthro
 describe("Anthropic Messages route", () => {
   it.effect("prepares Anthropic Messages target", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(request)
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         model: "claude-sonnet-4-5",
@@ -59,8 +60,8 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers adaptive thinking settings with effort", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
-        LLM.updateRequest(request, {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, {
           providerOptions: {
             anthropic: { thinking: { type: "adaptive", display: "summarized" }, effort: "low" },
           },
@@ -74,9 +75,45 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("normalizes enabled and disabled thinking settings", () =>
+    Effect.gen(function* () {
+      const enabled = yield* compileRequest(
+        LLMRequest.update(request, {
+          providerOptions: { anthropic: { thinking: { type: "enabled", budgetTokens: 1_024 } } },
+        }),
+      )
+      const legacy = yield* compileRequest(
+        LLMRequest.update(request, {
+          providerOptions: { anthropic: { thinking: { type: "enabled", budget_tokens: 2_048 } } },
+        }),
+      )
+      const disabled = yield* compileRequest(
+        LLMRequest.update(request, {
+          providerOptions: { anthropic: { thinking: { type: "disabled" } } },
+        }),
+      )
+
+      expect(enabled.body.thinking).toEqual({ type: "enabled", budget_tokens: 1_024 })
+      expect(legacy.body.thinking).toEqual({ type: "enabled", budget_tokens: 2_048 })
+      expect(disabled.body.thinking).toEqual({ type: "disabled" })
+    }),
+  )
+
+  it.effect("rejects enabled thinking without a budget", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, {
+          providerOptions: { anthropic: { thinking: { type: "enabled" } } },
+        }),
+      ).pipe(Effect.flip)
+
+      expect(error.message).toContain("Anthropic thinking provider option requires budgetTokens")
+    }),
+  )
+
   it.effect("lowers chronological system updates natively for Claude Opus 4.8 with cache hints", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: opus48,
           messages: [
@@ -101,7 +138,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers chronological system updates to wrapped user text for unsupported Anthropic models", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -128,7 +165,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects non-text chronological system update content before send", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: opus48,
           messages: [
@@ -145,7 +182,7 @@ describe("Anthropic Messages route", () => {
   it.effect("falls back for unsupported native chronological system update placement", () =>
     Effect.gen(function* () {
       expect(
-        (yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+        (yield* compileRequest(
           LLM.request({
             model: opus48,
             messages: [Message.assistant("Plain."), Message.system("After plain assistant.")],
@@ -160,12 +197,11 @@ describe("Anthropic Messages route", () => {
         },
       ])
       expect(
-        (yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
-          LLM.request({ model: opus48, messages: [Message.system("First.")], cache: "none" }),
-        )).body.messages,
+        (yield* compileRequest(LLM.request({ model: opus48, messages: [Message.system("First.")], cache: "none" })))
+          .body.messages,
       ).toEqual([{ role: "user", content: [{ type: "text", text: "<system-update>\nFirst.\n</system-update>" }] }])
       expect(
-        (yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+        (yield* compileRequest(
           LLM.request({
             model: opus48,
             messages: [Message.user("Before."), Message.system("One."), Message.system("Two.")],
@@ -187,7 +223,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects a system update between a local tool call and its result", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: opus48,
           messages: [
@@ -206,7 +242,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("prepares tool call and tool result messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result",
           model,
@@ -235,11 +271,39 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("keeps tools and sends tool_choice none", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          id: "req_tool_choice_none",
+          model,
+          tools: [{ name: "lookup", description: "Look things up", inputSchema: { type: "object", properties: {} } }],
+          messages: [
+            Message.user("What is the weather?"),
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } })]),
+            Message.tool({ id: "call_1", name: "lookup", result: { forecast: "sunny" } }),
+          ],
+          toolChoice: "none",
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.tools).toEqual([
+        {
+          name: "lookup",
+          description: "Look things up",
+          input_schema: { type: "object", properties: {} },
+        },
+      ])
+      expect(prepared.body.tool_choice).toEqual({ type: "none" })
+    }),
+  )
+
   // Regression: read tool results must stay structured so base64 media data is
   // not JSON-stringified into `tool_result.content`.
   it.effect("lowers media tool-result content as structured blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_image",
           model,
@@ -271,7 +335,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers single-image tool-result content as a structured image block", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_image_only",
           model,
@@ -296,7 +360,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects unsupported media in tool-result content with a clear error", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_unsupported_media",
           model,
@@ -320,7 +384,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("prepares the composed native continuation request", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         continuationRequest({
           id: "req_native_continuation_anthropic",
           model,
@@ -364,7 +428,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers preserved Anthropic reasoning signature metadata", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -381,6 +445,34 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("round-trips redacted thinking as redacted_thinking blocks", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              { type: "reasoning", text: "", providerMetadata: { anthropic: { redactedData: "opaque_1" } } },
+              { type: "reasoning", text: "visible", providerMetadata: { anthropic: { signature: "sig_1" } } },
+            ]),
+          ],
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        messages: [
+          {
+            role: "assistant",
+            content: [
+              { type: "redacted_thinking", data: "opaque_1" },
+              { type: "thinking", thinking: "visible", signature: "sig_1" },
+            ],
+          },
+        ],
+      })
+    }),
+  )
+
   it.effect("parses text, reasoning, and usage stream fixtures", () =>
     Effect.gen(function* () {
       const body = sseEvents(
@@ -414,18 +506,349 @@ describe("Anthropic Messages route", () => {
       expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
         providerMetadata: { anthropic: { signature: "sig_1" } },
       })
+      expect(response.events.find((event) => event.type === "reasoning-delta" && event.text === "")).toBeUndefined()
       expect(response.message.content).toEqual([
         { type: "text", text: "Hello!" },
         { type: "reasoning", text: "thinking", providerMetadata: { anthropic: { signature: "sig_1" } } },
       ])
       expect(response.events.at(-1)).toMatchObject({
         type: "finish",
-        reason: "stop",
+        reason: { normalized: "stop", raw: "end_turn" },
         providerMetadata: { anthropic: { stopSequence: "\n\nHuman:" } },
       })
     }),
   )
 
+  it.effect("requires message_stop before completing a streamed message", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              { type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
+              { type: "content_block_delta", index: 0, delta: { type: "text_delta", text: "Hello" } },
+              { type: "content_block_stop", index: 0 },
+              { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 1 } },
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidProviderOutput",
+        message: "Provider stream ended without a terminal finish event",
+      })
+    }),
+  )
+
+  it.effect("maps thinking tokens and preserves unknown Anthropic usage fields", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "message_start",
+                message: {
+                  usage: {
+                    input_tokens: 5,
+                    cache_read_input_tokens: 2,
+                    service_tier: "standard",
+                    cache_creation: { ephemeral_5m_input_tokens: 1 },
+                    server_tool_use: { web_search_requests: 1, start_counter: 2 },
+                    output_tokens_details: { thinking_tokens: 3, start_detail: "preserved" },
+                  },
+                },
+              },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "end_turn" },
+                usage: {
+                  output_tokens: 8,
+                  server_tool_use: { web_search_requests: 2, terminal_counter: 3 },
+                  output_tokens_details: { terminal_detail: "preserved" },
+                  future_terminal: { requests: 4 },
+                },
+              },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.usage).toMatchObject({
+        inputTokens: 7,
+        outputTokens: 8,
+        reasoningTokens: 3,
+        totalTokens: 15,
+        providerMetadata: {
+          anthropic: {
+            input_tokens: 5,
+            cache_read_input_tokens: 2,
+            service_tier: "standard",
+            cache_creation: { ephemeral_5m_input_tokens: 1 },
+            server_tool_use: { web_search_requests: 2, start_counter: 2, terminal_counter: 3 },
+            output_tokens: 8,
+            output_tokens_details: {
+              thinking_tokens: 3,
+              start_detail: "preserved",
+              terminal_detail: "preserved",
+            },
+            future_terminal: { requests: 4 },
+          },
+        },
+      })
+    }),
+  )
+
+  it.effect("round-trips omitted thinking carried only by a signature delta", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              {
+                type: "content_block_start",
+                index: 0,
+                content_block: { type: "thinking", thinking: "", signature: "" },
+              },
+              { type: "content_block_delta", index: 0, delta: { type: "signature_delta", signature: "sig_1" } },
+              { type: "content_block_stop", index: 0 },
+              { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 1 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        { type: "reasoning", text: "", providerMetadata: { anthropic: { signature: "sig_1" } } },
+      ])
+
+      const prepared = yield* compileRequest(LLM.request({ model, messages: [response.message], cache: "none" }))
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: [{ type: "thinking", thinking: "", signature: "sig_1" }] },
+      ])
+    }),
+  )
+
+  it.effect("retains a thinking signature supplied in content_block_start", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              {
+                type: "content_block_start",
+                index: 0,
+                content_block: { type: "thinking", thinking: "", signature: "sig_1" },
+              },
+              { type: "content_block_stop", index: 0 },
+              { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 1 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        { type: "reasoning", text: "", providerMetadata: { anthropic: { signature: "sig_1" } } },
+      ])
+      expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
+        providerMetadata: { anthropic: { signature: "sig_1" } },
+      })
+    }),
+  )
+
+  it.effect("retains complete tool input from content_block_start", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              {
+                type: "content_block_start",
+                index: 0,
+                content_block: { type: "tool_use", id: "call_1", name: "lookup", input: { query: "weather" } },
+              },
+              { type: "content_block_stop", index: 0 },
+              { type: "message_delta", delta: { stop_reason: "tool_use" }, usage: { output_tokens: 1 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.toolCalls).toMatchObject([
+        { id: "call_1", name: "lookup", input: { query: "weather" } },
+      ])
+    }),
+  )
+
+  it.effect("retains empty text blocks", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              { type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
+              { type: "content_block_stop", index: 0 },
+              { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 1 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([{ type: "text", text: "" }])
+    }),
+  )
+
+  it.effect("parses redacted thinking into empty reasoning with redactedData metadata", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        { type: "message_start", message: { usage: { input_tokens: 5 } } },
+        { type: "content_block_start", index: 0, content_block: { type: "redacted_thinking", data: "opaque_1" } },
+        { type: "content_block_stop", index: 0 },
+        { type: "content_block_start", index: 1, content_block: { type: "text", text: "" } },
+        { type: "content_block_delta", index: 1, delta: { type: "text_delta", text: "Hello" } },
+        { type: "content_block_stop", index: 1 },
+        { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 2 } },
+        { type: "message_stop" },
+      )
+      const response = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events.find((event) => event.type === "reasoning-start")).toMatchObject({
+        providerMetadata: { anthropic: { redactedData: "opaque_1" } },
+      })
+      expect(response.message.content).toEqual([
+        { type: "reasoning", text: "", providerMetadata: { anthropic: { redactedData: "opaque_1" } } },
+        { type: "text", text: "Hello" },
+      ])
+    }),
+  )
+
+  it.effect("round-trips streamed redacted thinking with tool use into a continuation request", () =>
+    Effect.gen(function* () {
+      // Anthropic types `redacted_thinking.data` as an opaque string. Its
+      // contents are provider-owned and must be replayed without inspection.
+      const redactedData = "cmVkYWN0ZWQtdGhpbmtpbmc="
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              {
+                type: "content_block_start",
+                index: 0,
+                content_block: { type: "redacted_thinking", data: redactedData },
+              },
+              { type: "content_block_stop", index: 0 },
+              {
+                type: "content_block_start",
+                index: 1,
+                content_block: { type: "tool_use", id: "call_1", name: "lookup" },
+              },
+              {
+                type: "content_block_delta",
+                index: 1,
+                delta: { type: "input_json_delta", partial_json: '{"query":"weather"}' },
+              },
+              { type: "content_block_stop", index: 1 },
+              { type: "message_delta", delta: { stop_reason: "tool_use" }, usage: { output_tokens: 1 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("Say hello."),
+            response.message,
+            Message.tool({ id: "call_1", name: "lookup", result: "sunny", resultType: "text" }),
+          ],
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: [{ type: "text", text: "Say hello." }] },
+        {
+          role: "assistant",
+          content: [
+            { type: "redacted_thinking", data: redactedData },
+            { type: "tool_use", id: "call_1", name: "lookup", input: { query: "weather" } },
+          ],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: "sunny",
+              is_error: undefined,
+              cache_control: undefined,
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("maps context-window truncation to length", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "model_context_window_exceeded" },
+                usage: { output_tokens: 1 },
+              },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.finishReason).toEqual({ normalized: "length", raw: "model_context_window_exceeded" })
+    }),
+  )
+
+  it.effect("preserves pause_turn while normalizing it to stop", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              { type: "message_delta", delta: { stop_reason: "pause_turn" }, usage: { output_tokens: 1 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "pause_turn" })
+    }),
+  )
+
   it.effect("assembles streamed tool call input", () =>
     Effect.gen(function* () {
       const body = sseEvents(
@@ -435,10 +858,11 @@ describe("Anthropic Messages route", () => {
         { type: "content_block_delta", index: 0, delta: { type: "input_json_delta", partial_json: ':"weather"}' } },
         { type: "content_block_stop", index: 0 },
         { type: "message_delta", delta: { stop_reason: "tool_use" }, usage: { output_tokens: 1 } },
+        { type: "message_stop" },
       )
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
       const usage = new Usage({
@@ -475,10 +899,16 @@ describe("Anthropic Messages route", () => {
           providerExecuted: undefined,
           providerMetadata: undefined,
         },
-        { type: "step-finish", index: 0, reason: "tool-calls", usage, providerMetadata: undefined },
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "tool-calls", raw: "tool_use" },
+          usage,
+          providerMetadata: undefined,
+        },
         {
           type: "finish",
-          reason: "tool-calls",
+          reason: { normalized: "tool-calls", raw: "tool_use" },
           providerMetadata: undefined,
           usage,
         },
@@ -614,10 +1044,13 @@ describe("Anthropic Messages route", () => {
         { type: "content_block_delta", index: 2, delta: { type: "text_delta", text: "Found it." } },
         { type: "content_block_stop", index: 2 },
         { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 8 } },
+        { type: "message_stop" },
       )
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "web_search", description: "Web search", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [
+            ToolDefinition.make({ name: "web_search", description: "Web search", inputSchema: { type: "object" } }),
+          ],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
 
@@ -636,10 +1069,20 @@ describe("Anthropic Messages route", () => {
         name: "web_search",
         result: { type: "json", value: [{ type: "web_search_result", url: "https://example.com", title: "Example" }] },
         providerExecuted: true,
-        providerMetadata: { anthropic: { blockType: "web_search_tool_result" } },
+        // The complete payload rides in provider metadata as irreducible replay
+        // state for later stateless requests.
+        providerMetadata: {
+          anthropic: {
+            blockType: "web_search_tool_result",
+            result: [{ type: "web_search_result", url: "https://example.com", title: "Example" }],
+          },
+        },
       })
       expect(response.text).toBe("Found it.")
-      expect(response.events.at(-1)).toMatchObject({ type: "finish", reason: "stop" })
+      expect(response.events.at(-1)).toMatchObject({
+        type: "finish",
+        reason: { normalized: "stop", raw: "end_turn" },
+      })
     }),
   )
 
@@ -665,10 +1108,13 @@ describe("Anthropic Messages route", () => {
         },
         { type: "content_block_stop", index: 1 },
         { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 1 } },
+        { type: "message_stop" },
       )
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "web_search", description: "Web search", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [
+            ToolDefinition.make({ name: "web_search", description: "Web search", inputSchema: { type: "object" } }),
+          ],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
 
@@ -685,7 +1131,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("round-trips provider-executed assistant content into server tool blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_round_trip",
           model,
@@ -736,7 +1182,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects round-trip for unknown server tool names", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_unknown_server_tool",
           model,
@@ -784,7 +1230,10 @@ describe("Anthropic Messages route", () => {
                     content: [
                       { type: "text", text: "What is in this image?" },
                       { type: "image", source: { type: "base64", media_type: "image/png", data: "AAECAw==" } },
-                      { type: "document", source: { type: "base64", media_type: "application/pdf", data: "JVBERi0xLjQ=" } },
+                      {
+                        type: "document",
+                        source: { type: "base64", media_type: "application/pdf", data: "JVBERi0xLjQ=" },
+                      },
                     ],
                   },
                 ],
@@ -810,7 +1259,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("maps ttlSeconds >= 3600 to cache_control ttl: '1h'", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: { type: "text", text: "system", cache: new CacheHint({ type: "ephemeral", ttlSeconds: 3600 }) },
@@ -826,7 +1275,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("emits cache_control on tool definitions and tool-result blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           tools: [
@@ -867,7 +1316,7 @@ describe("Anthropic Messages route", () => {
   it.effect("drops cache_control breakpoints past the 4-per-request cap", () =>
     Effect.gen(function* () {
       const hint = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: [
@@ -893,7 +1342,7 @@ describe("Anthropic Messages route", () => {
   it.effect("spends breakpoint budget on tools before system before messages", () =>
     Effect.gen(function* () {
       const hint = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           tools: [

packages/ai/test/provider/bedrock-converse.test.ts

@@ -2,8 +2,18 @@ import { EventStreamCodec } from "@smithy/eventstream-codec"
 import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
-import { CacheHint, LLM, Message, ToolCallPart, ToolChoice } from "../../src"
+import {
+  CacheHint,
+  GenerationOptions,
+  LLM,
+  LLMRequest,
+  Message,
+  ToolCallPart,
+  ToolChoice,
+  ToolDefinition,
+} from "../../src"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { AmazonBedrock } from "../../src/providers"
 import * as BedrockConverse from "../../src/protocols/bedrock-converse"
 import { it } from "../lib/effect"
@@ -34,6 +44,26 @@ const eventFrame = (type: string, payload: object) =>
     body: utf8Encoder.encode(JSON.stringify(payload)),
   })
 
+const exceptionFrame = (type: string, payload: object) =>
+  codec.encode({
+    headers: {
+      ":message-type": { type: "string", value: "exception" },
+      ":exception-type": { type: "string", value: type },
+      ":content-type": { type: "string", value: "application/json" },
+    },
+    body: utf8Encoder.encode(JSON.stringify(payload)),
+  })
+
+const errorFrame = (code: string, message: string) =>
+  codec.encode({
+    headers: {
+      ":message-type": { type: "string", value: "error" },
+      ":error-code": { type: "string", value: code },
+      ":error-message": { type: "string", value: message },
+    },
+    body: new Uint8Array(),
+  })
+
 const concat = (frames: ReadonlyArray<Uint8Array>) => {
   const total = frames.reduce((sum, frame) => sum + frame.length, 0)
   const out = new Uint8Array(total)
@@ -72,7 +102,7 @@ const baseRequest = LLM.request({
 describe("Bedrock Converse route", () => {
   it.effect("prepares Converse target with system, inference config, and messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(baseRequest)
+      const prepared = yield* compileRequest(baseRequest)
 
       expect(prepared.body).toEqual({
         modelId: "anthropic.claude-3-5-sonnet-20240620-v1:0",
@@ -85,8 +115,10 @@ describe("Bedrock Converse route", () => {
 
   it.effect("passes topK through additionalModelRequestFields as top_k", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
-        LLM.updateRequest(baseRequest, { generation: { maxTokens: 64, temperature: 0, topK: 40 } }),
+      const prepared = yield* compileRequest(
+        LLMRequest.update(baseRequest, {
+          generation: GenerationOptions.make({ maxTokens: 64, temperature: 0, topK: 40 }),
+        }),
       )
 
       // Converse's inferenceConfig has no topK; Anthropic/Nova read it from
@@ -98,14 +130,14 @@ describe("Bedrock Converse route", () => {
 
   it.effect("omits additionalModelRequestFields when topK is unset", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(baseRequest)
+      const prepared = yield* compileRequest(baseRequest)
       expect(prepared.body.additionalModelRequestFields).toBeUndefined()
     }),
   )
 
   it.effect("lowers chronological system updates to wrapped user text in order", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [Message.user("Before."), Message.system("Update."), Message.assistant("After.")],
@@ -122,14 +154,14 @@ describe("Bedrock Converse route", () => {
 
   it.effect("prepares tool config with toolSpec and toolChoice", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
-        LLM.updateRequest(baseRequest, {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(baseRequest, {
           tools: [
-            {
+            ToolDefinition.make({
               name: "lookup",
               description: "Lookup data",
               inputSchema: { type: "object", properties: { query: { type: "string" } }, required: ["query"] },
-            },
+            }),
           ],
           toolChoice: ToolChoice.make({ type: "required" }),
         }),
@@ -154,9 +186,39 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("keeps tools and omits the unsupported choice when tool choice is none", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(baseRequest, {
+          tools: [
+            ToolDefinition.make({
+              name: "lookup",
+              description: "Lookup data",
+              inputSchema: { type: "object", properties: { query: { type: "string" } } },
+            }),
+          ],
+          toolChoice: ToolChoice.make({ type: "none" }),
+        }),
+      )
+
+      expect(prepared.body.toolConfig).toMatchObject({
+        tools: [
+          {
+            toolSpec: {
+              name: "lookup",
+              description: "Lookup data",
+              inputSchema: { json: { type: "object", properties: { query: { type: "string" } } } },
+            },
+          },
+        ],
+      })
+      expect(prepared.body.toolConfig?.toolChoice).toBeUndefined()
+    }),
+  )
+
   it.effect("lowers assistant tool-call + tool-result message history", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_history",
           model,
@@ -195,7 +257,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers image content in tool-result messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_image",
           model,
@@ -260,7 +322,10 @@ describe("Bedrock Converse route", () => {
       // `metadata` (carries usage). We consolidate them into a single
       // terminal `finish` event with both.
       expect(finishes).toHaveLength(1)
-      expect(finishes[0]).toMatchObject({ type: "finish", reason: "stop" })
+      expect(finishes[0]).toMatchObject({
+        type: "finish",
+        reason: { normalized: "stop", raw: "end_turn" },
+      })
       expect(response.usage).toMatchObject({
         inputTokens: 5,
         outputTokens: 2,
@@ -269,6 +334,23 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("maps truncation and malformed output stop reasons", () =>
+    Effect.gen(function* () {
+      const reasons = [
+        ["model_context_window_exceeded", "length"],
+        ["malformed_model_output", "error"],
+        ["malformed_tool_use", "error"],
+      ] as const
+
+      for (const [raw, normalized] of reasons) {
+        const response = yield* LLMClient.generate(baseRequest).pipe(
+          Effect.provide(fixedBytes(eventStreamBody(["messageStop", { stopReason: raw }]))),
+        )
+        expect(response.finishReason).toEqual({ normalized, raw })
+      }
+    }),
+  )
+
   it.effect("adds cache reads and writes to Bedrock input usage", () =>
     Effect.gen(function* () {
       const body = eventStreamBody(
@@ -302,6 +384,19 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("preserves usage across later metadata events without usage", () =>
+    Effect.gen(function* () {
+      const body = eventStreamBody(
+        ["messageStop", { stopReason: "end_turn" }],
+        ["metadata", { usage: { inputTokens: 5, outputTokens: 2, totalTokens: 7 } }],
+        ["metadata", { metrics: { latencyMs: 100 } }],
+      )
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.usage).toMatchObject({ inputTokens: 5, outputTokens: 2, totalTokens: 7 })
+    }),
+  )
+
   it.effect("assembles streamed tool call input", () =>
     Effect.gen(function* () {
       const body = eventStreamBody(
@@ -319,8 +414,8 @@ describe("Bedrock Converse route", () => {
         ["messageStop", { stopReason: "tool_use" }],
       )
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(baseRequest, {
-          tools: [{ name: "lookup", description: "Lookup", inputSchema: { type: "object" } }],
+        LLMRequest.update(baseRequest, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedBytes(body)))
 
@@ -332,7 +427,10 @@ describe("Bedrock Converse route", () => {
         { type: "tool-input-delta", id: "tool_1", name: "lookup", text: '{"query"' },
         { type: "tool-input-delta", id: "tool_1", name: "lookup", text: ':"weather"}' },
       ])
-      expect(response.events.at(-1)).toMatchObject({ type: "finish", reason: "tool-calls" })
+      expect(response.events.at(-1)).toMatchObject({
+        type: "finish",
+        reason: { normalized: "tool-calls", raw: "tool_use" },
+      })
     }),
   )
 
@@ -358,7 +456,7 @@ describe("Bedrock Converse route", () => {
         name: "lookup",
         raw: '{"query":"partial',
       })
-      expect(response.finishReason).toBe("tool-calls")
+      expect(response.finishReason).toEqual({ normalized: "tool-calls", raw: "end_turn" })
     }),
   )
 
@@ -394,7 +492,7 @@ describe("Bedrock Converse route", () => {
         providerMetadata: { bedrock: { signature: "sig_1" } },
       })
 
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -414,12 +512,168 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
-  it.effect("classifies throttlingException as a rate limit", () =>
+  it.effect("preserves reasoning signatures when contentBlockStop is missing", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["messageStart", { role: "assistant" }],
+              [
+                "contentBlockDelta",
+                { contentBlockIndex: 0, delta: { reasoningContent: { text: "Let me think." } } },
+              ],
+              [
+                "contentBlockDelta",
+                { contentBlockIndex: 0, delta: { reasoningContent: { signature: "sig_1" } } },
+              ],
+              ["messageStop", { stopReason: "end_turn" }],
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.find((event) => event.type === "reasoning-delta" && event.text === "")).toEqual({
+        type: "reasoning-delta",
+        id: "reasoning-0",
+        text: "",
+        providerMetadata: { bedrock: { signature: "sig_1" } },
+      })
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Let me think.",
+          providerMetadata: { bedrock: { signature: "sig_1" } },
+        },
+      ])
+
+      const prepared = yield* compileRequest(LLM.request({ model, messages: [response.message], cache: "none" }))
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ reasoningContent: { reasoningText: { text: "Let me think.", signature: "sig_1" } } }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("preserves signature-only reasoning blocks", () =>
     Effect.gen(function* () {
       const body = eventStreamBody(
         ["messageStart", { role: "assistant" }],
-        ["throttlingException", { message: "Slow down" }],
+        [
+          "contentBlockDelta",
+          { contentBlockIndex: 0, delta: { reasoningContent: { signature: "sig_1" } } },
+        ],
+        ["contentBlockStop", { contentBlockIndex: 0 }],
+        ["messageStop", { stopReason: "end_turn" }],
       )
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.message.content).toEqual([
+        { type: "reasoning", text: "", providerMetadata: { bedrock: { signature: "sig_1" } } },
+      ])
+    }),
+  )
+
+  it.effect("accepts Vercel-compatible redacted reasoning data deltas", () =>
+    Effect.gen(function* () {
+      const redactedData = "cmVkYWN0ZWQtdGhpbmtpbmc="
+      const body = eventStreamBody(
+        ["messageStart", { role: "assistant" }],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { data: redactedData } } }],
+        ["contentBlockStop", { contentBlockIndex: 0 }],
+        ["messageStop", { stopReason: "end_turn" }],
+      )
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.events.find((event) => event.type === "reasoning-delta" && event.text === "")).toEqual({
+        type: "reasoning-delta",
+        id: "reasoning-0",
+        text: "",
+        providerMetadata: { bedrock: { redactedData } },
+      })
+      expect(response.message.content).toEqual([
+        { type: "reasoning", text: "", providerMetadata: { bedrock: { redactedData } } },
+      ])
+    }),
+  )
+
+  it.effect("round-trips streamed redacted reasoning with tool use into a continuation request", () =>
+    Effect.gen(function* () {
+      // Bedrock represents redactedContent blobs as base64 strings on its JSON
+      // wire. The provider owns the payload and requires byte-exact replay.
+      const redactedData = "cmVkYWN0ZWQtdGhpbmtpbmc="
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(baseRequest, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["messageStart", { role: "assistant" }],
+              [
+                "contentBlockDelta",
+                { contentBlockIndex: 0, delta: { reasoningContent: { redactedContent: redactedData } } },
+              ],
+              ["contentBlockStop", { contentBlockIndex: 0 }],
+              [
+                "contentBlockStart",
+                {
+                  contentBlockIndex: 1,
+                  start: { toolUse: { toolUseId: "tool_1", name: "lookup" } },
+                },
+              ],
+              ["contentBlockDelta", { contentBlockIndex: 1, delta: { toolUse: { input: '{"query":"weather"}' } } }],
+              ["contentBlockStop", { contentBlockIndex: 1 }],
+              ["messageStop", { stopReason: "tool_use" }],
+            ),
+          ),
+        ),
+      )
+      expect(response.events.find((event) => event.type === "reasoning-delta" && event.text === "")).toEqual({
+        type: "reasoning-delta",
+        id: "reasoning-0",
+        text: "",
+        providerMetadata: { bedrock: { redactedData } },
+      })
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("Say hello."),
+            response.message,
+            Message.tool({ id: "tool_1", name: "lookup", result: "sunny", resultType: "text" }),
+          ],
+          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: [{ text: "Say hello." }] },
+        {
+          role: "assistant",
+          content: [
+            { reasoningContent: { redactedContent: redactedData } },
+            { toolUse: { toolUseId: "tool_1", name: "lookup", input: { query: "weather" } } },
+          ],
+        },
+        {
+          role: "user",
+          content: [{ toolResult: { toolUseId: "tool_1", content: [{ text: "sunny" }], status: "success" } }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("classifies throttlingException as a rate limit", () =>
+    Effect.gen(function* () {
+      const body = concat([
+        eventFrame("messageStart", { role: "assistant" }),
+        exceptionFrame("throttlingException", { message: "Slow down" }),
+      ])
       const error = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)), Effect.flip)
 
       expect(error.reason).toMatchObject({ _tag: "RateLimit", message: "Slow down" })
@@ -430,7 +684,7 @@ describe("Bedrock Converse route", () => {
     Effect.gen(function* () {
       const error = yield* LLMClient.generate(baseRequest).pipe(
         Effect.provide(
-          fixedBytes(eventStreamBody(["validationException", { message: "Input is too long for requested model" }])),
+          fixedBytes(exceptionFrame("validationException", { message: "Input is too long for requested model" })),
         ),
         Effect.flip,
       )
@@ -443,12 +697,44 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("uses originalMessage from model stream exception frames", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(
+          fixedBytes(
+            exceptionFrame("modelStreamErrorException", {
+              originalMessage: "Upstream model failed",
+              originalStatusCode: 500,
+            }),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", message: "Upstream model failed" })
+    }),
+  )
+
+  it.effect("fails unmodeled AWS event-stream errors", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(fixedBytes(errorFrame("BadStream", "Stream failed"))),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "InvalidProviderOutput",
+        message: "BadStream: Stream failed",
+      })
+    }),
+  )
+
   it.effect("rejects requests with no auth path", () =>
     Effect.gen(function* () {
       const unsignedModel = AmazonBedrock.configure({
         baseURL: "https://bedrock-runtime.test",
       }).model("anthropic.claude-3-5-sonnet-20240620-v1:0")
-      const error = yield* LLMClient.generate(LLM.updateRequest(baseRequest, { model: unsignedModel })).pipe(
+      const error = yield* LLMClient.generate(LLMRequest.update(baseRequest, { model: unsignedModel })).pipe(
         Effect.provide(fixedBytes(eventStreamBody(["messageStop", { stopReason: "end_turn" }]))),
         Effect.flip,
       )
@@ -467,7 +753,7 @@ describe("Bedrock Converse route", () => {
           secretAccessKey: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY",
         },
       }).model("anthropic.claude-3-5-sonnet-20240620-v1:0")
-      const prepared = yield* LLMClient.prepare(LLM.updateRequest(baseRequest, { model: signed }))
+      const prepared = yield* compileRequest(LLMRequest.update(baseRequest, { model: signed }))
 
       expect(prepared.route).toBe("bedrock-converse")
       expect(prepared.model).toBe(signed)
@@ -477,7 +763,7 @@ describe("Bedrock Converse route", () => {
   it.effect("emits cachePoint markers after system, user-text, and assistant-text with cache hints", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_cache",
           model,
@@ -509,7 +795,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("does not emit cachePoint when no cache hint is set", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(baseRequest)
+      const prepared = yield* compileRequest(baseRequest)
       expect(prepared.body).toMatchObject({
         system: [{ text: "You are concise." }],
         messages: [{ role: "user", content: [{ text: "Say hello." }] }],
@@ -519,7 +805,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers image media into Bedrock image blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_image",
           model,
@@ -556,7 +842,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("base64-encodes Uint8Array image bytes", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_image_bytes",
           model,
@@ -578,7 +864,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers document media into Bedrock document blocks with format and name", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_doc",
           model,
@@ -610,7 +896,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("requires names for document media", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           messages: [Message.user({ type: "media", mediaType: "application/pdf", data: "UERGREFUQQ==" })],
@@ -623,7 +909,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("passes named document-only messages through for provider validation", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           cache: "none",
@@ -649,9 +935,10 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers document media in tool results", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
+          cache: "none",
           messages: [
             Message.assistant([ToolCallPart.make({ id: "call_1", name: "read", input: { path: "report.pdf" } })]),
             Message.tool({
@@ -700,7 +987,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("rejects unsupported image media types", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_bad_image",
           model,
@@ -714,7 +1001,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("rejects unsupported document media types", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_bad_doc",
           model,
@@ -729,7 +1016,7 @@ describe("Bedrock Converse route", () => {
   it.effect("maps ttlSeconds >= 3600 to cachePoint ttl: '1h'", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral", ttlSeconds: 3600 })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: [{ type: "text", text: "system", cache }],
@@ -746,7 +1033,7 @@ describe("Bedrock Converse route", () => {
   it.effect("appends cachePoint after marked tool definitions and tool-result blocks", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           tools: [{ name: "lookup", description: "lookup", inputSchema: { type: "object", properties: {} }, cache }],
@@ -778,7 +1065,7 @@ describe("Bedrock Converse route", () => {
   it.effect("drops cachePoint markers past the 4-per-request cap", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: [

packages/ai/test/provider/cloudflare.test.ts

@@ -3,7 +3,7 @@ import { ConfigProvider, Effect, Schema } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, LLMEvent } from "../../src"
 import { CloudflareAIGateway, CloudflareWorkersAI } from "../../src/providers/cloudflare"
-import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { it } from "../lib/effect"
 import { dynamicResponse } from "../lib/http"
 import { sseEvents } from "../lib/sse"
@@ -34,7 +34,7 @@ describe("Cloudflare", () => {
       })
       expect(model.route.endpoint.baseURL).toBe("https://gateway.ai.cloudflare.com/v1/test-account/test-gateway/compat")
 
-      const prepared = yield* LLMClient.prepare(LLM.request({ model, prompt: "Say hello." }))
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Say hello." }))
 
       expect(prepared.route).toBe("cloudflare-ai-gateway")
       expect(prepared.body).toMatchObject({
@@ -129,7 +129,7 @@ describe("Cloudflare", () => {
         openai: { reasoningField: "reasoning", reasoningDetails: merged },
       })
 
-      const replay = yield* LLMClient.prepare(LLM.request({ model, messages: [response.message] }))
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         { role: "assistant", content: "Hello", reasoning: "Thinking", reasoning_details: merged },
       ])
@@ -180,7 +180,7 @@ describe("Cloudflare", () => {
 
   it.effect("allows a fully configured baseURL override", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: CloudflareAIGateway.configure({
             baseURL: "https://gateway.proxy.test/v1/custom/compat",
@@ -208,7 +208,7 @@ describe("Cloudflare", () => {
       })
       expect(model.route.endpoint.baseURL).toBe("https://api.cloudflare.com/client/v4/accounts/test-account/ai/v1")
 
-      const prepared = yield* LLMClient.prepare(LLM.request({ model, prompt: "Say hello." }))
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Say hello." }))
 
       expect(prepared.route).toBe("cloudflare-workers-ai")
       expect(prepared.body).toMatchObject({

packages/ai/test/provider/gemini.test.ts

@@ -1,7 +1,8 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
-import { LLM, LLMError, Message, ToolCallPart, Usage } from "../../src"
+import { LLM, LLMError, LLMRequest, Message, ToolCallPart, ToolDefinition, Usage } from "../../src"
 import { Auth, LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import * as Gemini from "../../src/protocols/gemini"
 import { ProviderShared } from "../../src/protocols/shared"
 import { it } from "../lib/effect"
@@ -26,7 +27,7 @@ const request = LLM.request({
 describe("Gemini route", () => {
   it.effect("prepares Gemini target", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(request)
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         contents: [{ role: "user", parts: [{ text: "Say hello." }] }],
@@ -36,9 +37,30 @@ describe("Gemini route", () => {
     }),
   )
 
+  it.effect("normalizes Gemini thinking options", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, {
+          providerOptions: { gemini: { thinkingConfig: { thinkingBudget: 0, includeThoughts: false } } },
+        }),
+      )
+      const filtered = yield* compileRequest(
+        LLMRequest.update(request, {
+          providerOptions: { gemini: { thinkingConfig: { thinkingBudget: "invalid", includeThoughts: false } } },
+        }),
+      )
+
+      expect(prepared.body.generationConfig?.thinkingConfig).toEqual({
+        thinkingBudget: 0,
+        includeThoughts: false,
+      })
+      expect(filtered.body.generationConfig?.thinkingConfig).toEqual({ includeThoughts: false })
+    }),
+  )
+
   it.effect("lowers chronological system updates to wrapped user text in order", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [Message.user("Before."), Message.system("Update."), Message.assistant("After.")],
@@ -54,7 +76,7 @@ describe("Gemini route", () => {
 
   it.effect("prepares multimodal user input and tool history", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result",
           model,
@@ -122,7 +144,7 @@ describe("Gemini route", () => {
 
   it.effect("continues media tool results as inline model input without base64 text", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -167,7 +189,7 @@ describe("Gemini route", () => {
 
   it.effect("strips matching data URLs to raw base64 inlineData", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -208,7 +230,7 @@ describe("Gemini route", () => {
   ] as const)
     it.effect(`rejects ${name}`, () =>
       Effect.gen(function* () {
-        const error = yield* LLMClient.prepare(
+        const error = yield* compileRequest(
           LLM.request({ model, messages: [Message.user({ type: "media", ...media })] }),
         ).pipe(Effect.flip)
         expect(error.message).toMatch(/does not support|does not match|valid base64/)
@@ -217,7 +239,7 @@ describe("Gemini route", () => {
 
   it.effect("rejects oversized image input", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -233,27 +255,29 @@ describe("Gemini route", () => {
     }),
   )
 
-  it.effect("omits tools when tool choice is none", () =>
+  it.effect("keeps tools and sends function calling mode NONE", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
-          id: "req_no_tools",
+          id: "req_tool_choice_none",
           model,
           prompt: "Say hello.",
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
           toolChoice: { type: "none" },
         }),
       )
 
-      expect(prepared.body).toEqual({
+      expect(prepared.body).toMatchObject({
         contents: [{ role: "user", parts: [{ text: "Say hello." }] }],
+        tools: [{ functionDeclarations: [{ name: "lookup", description: "Lookup data" }] }],
+        toolConfig: { functionCallingConfig: { mode: "NONE" } },
       })
     }),
   )
 
   it.effect("sanitizes integer enums, dangling required, untyped arrays, and scalar object keys", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_schema_patch",
           model,
@@ -371,10 +395,16 @@ describe("Gemini route", () => {
         { type: "text-delta", id: "text-0", text: "Hello" },
         { type: "text-delta", id: "text-0", text: "!" },
         { type: "text-end", id: "text-0" },
-        { type: "step-finish", index: 0, reason: "stop", usage, providerMetadata: undefined },
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "stop", raw: "STOP" },
+          usage,
+          providerMetadata: undefined,
+        },
         {
           type: "finish",
-          reason: "stop",
+          reason: { normalized: "stop", raw: "STOP" },
           usage,
         },
       ])
@@ -402,8 +432,8 @@ describe("Gemini route", () => {
         ],
       })
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
       const reasoning = response.events.find((event) => event.type === "reasoning-start")
@@ -428,7 +458,7 @@ describe("Gemini route", () => {
         response.events.findIndex((event) => event.type === "tool-call"),
       )
 
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -493,8 +523,8 @@ describe("Gemini route", () => {
         usageMetadata: { promptTokenCount: 5, candidatesTokenCount: 1 },
       })
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
       const usage = new Usage({
@@ -527,10 +557,16 @@ describe("Gemini route", () => {
           providerExecuted: undefined,
           providerMetadata: undefined,
         },
-        { type: "step-finish", index: 0, reason: "tool-calls", usage, providerMetadata: undefined },
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "tool-calls", raw: "STOP" },
+          usage,
+          providerMetadata: undefined,
+        },
         {
           type: "finish",
-          reason: "tool-calls",
+          reason: { normalized: "tool-calls", raw: "STOP" },
           usage,
         },
       ])
@@ -554,8 +590,8 @@ describe("Gemini route", () => {
         ],
       })
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
 
@@ -569,7 +605,10 @@ describe("Gemini route", () => {
         },
         { type: "tool-call", id: "tool_1", name: "lookup", input: { query: "news" } },
       ])
-      expect(response.events.at(-1)).toMatchObject({ type: "finish", reason: "tool-calls" })
+      expect(response.events.at(-1)).toMatchObject({
+        type: "finish",
+        reason: { normalized: "tool-calls", raw: "STOP" },
+      })
     }),
   )
 
@@ -589,9 +628,41 @@ describe("Gemini route", () => {
       )
 
       expect(length.events.map((event) => event.type)).toEqual(["step-start", "step-finish", "finish"])
-      expect(length.events.at(-1)).toMatchObject({ type: "finish", reason: "length" })
+      expect(length.events.at(-1)).toMatchObject({
+        type: "finish",
+        reason: { normalized: "length", raw: "MAX_TOKENS" },
+      })
       expect(filtered.events.map((event) => event.type)).toEqual(["step-start", "step-finish", "finish"])
-      expect(filtered.events.at(-1)).toMatchObject({ type: "finish", reason: "content-filter" })
+      expect(filtered.events.at(-1)).toMatchObject({
+        type: "finish",
+        reason: { normalized: "content-filter", raw: "SAFETY" },
+      })
+    }),
+  )
+
+  it.effect("maps current blocking and invalid-output finish reasons", () =>
+    Effect.gen(function* () {
+      const reasons = [
+        ["MODEL_ARMOR", "content-filter"],
+        ["IMAGE_PROHIBITED_CONTENT", "content-filter"],
+        ["IMAGE_RECITATION", "content-filter"],
+        ["LANGUAGE", "content-filter"],
+        ["UNEXPECTED_TOOL_CALL", "error"],
+        ["NO_IMAGE", "error"],
+        ["IMAGE_OTHER", "unknown"],
+        ["TOO_MANY_TOOL_CALLS", "error"],
+        ["MISSING_THOUGHT_SIGNATURE", "error"],
+        ["MALFORMED_RESPONSE", "error"],
+      ] as const
+
+      for (const [raw, normalized] of reasons) {
+        const response = yield* LLMClient.generate(request).pipe(
+          Effect.provide(
+            fixedResponse(sseEvents({ candidates: [{ content: { role: "model", parts: [] }, finishReason: raw }] })),
+          ),
+        )
+        expect(response.finishReason).toEqual({ normalized, raw })
+      }
     }),
   )
 
@@ -621,7 +692,7 @@ describe("Gemini route", () => {
 
   it.effect("rejects unsupported assistant media content", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_media",
           model,

packages/ai/test/provider/google-vertex.test.ts

@@ -4,6 +4,7 @@ import { HttpClientRequest } from "effect/unstable/http"
 import { LLM } from "../../src"
 import { GoogleVertex, GoogleVertexChat, GoogleVertexMessages, GoogleVertexResponses } from "../../src/providers"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { it } from "../lib/effect"
 import { dynamicResponse } from "../lib/http"
 import { deltaChunk, finishChunk } from "../lib/openai-chunks"
@@ -182,7 +183,7 @@ describe("Google Vertex providers", () => {
 
   it.effect("protects the Vertex Messages API version from body overlays", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: GoogleVertexMessages.configure({
             accessToken: "vertex-token",

packages/ai/test/provider/openai-chat-reasoning.recorded.test.ts

@@ -5,6 +5,7 @@ import { OpenAIChat } from "../../src/protocols/openai-chat"
 import * as OpenAICompatible from "../../src/providers/openai-compatible"
 import * as OpenRouter from "../../src/providers/openrouter"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { recordedTests } from "../recorded-test"
 import { expectWeatherToolLoop, goldenWeatherToolLoopRequest, runWeatherToolLoop } from "../recorded-scenarios"
 
@@ -84,9 +85,7 @@ for (const item of cases) {
             ),
           ).toBe(true)
 
-          const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-            LLM.request({ model: item.model, messages: [response.message] }),
-          )
+          const replay = yield* compileRequest(LLM.request({ model: item.model, messages: [response.message] }))
           expect(replay.body.messages).toMatchObject([
             { role: "assistant", content: response.text, reasoning: response.reasoning },
           ])

packages/ai/test/provider/openai-chat.test.ts

@@ -1,12 +1,24 @@
 import { describe, expect } from "bun:test"
 import { Effect, Schema, Stream } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
-import { LLM, LLMError, LLMEvent, Message, Model, ToolCallPart, Usage } from "../../src"
+import {
+  HttpOptions,
+  LLM,
+  LLMError,
+  LLMEvent,
+  LLMRequest,
+  Message,
+  Model,
+  ToolCallPart,
+  ToolDefinition,
+  Usage,
+} from "../../src"
 import * as Azure from "../../src/providers/azure"
 import * as OpenAI from "../../src/providers/openai"
 import * as OpenAIChat from "../../src/protocols/openai-chat"
 import { ProviderShared } from "../../src/protocols/shared"
 import { Auth, LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { it } from "../lib/effect"
 import { dynamicResponse, fixedResponse, truncatedStream } from "../lib/http"
 import { deltaChunk, usageChunk } from "../lib/openai-chunks"
@@ -31,11 +43,7 @@ const request = LLM.request({
 describe("OpenAI Chat route", () => {
   it.effect("prepares OpenAI Chat payload", () =>
     Effect.gen(function* () {
-      // Pass the OpenAIChat payload type so `prepared.body` is statically
-      // typed to the route's native shape — the assertions below read field
-      // names without `unknown` casts.
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(request)
-      const _typed: { readonly model: string; readonly stream: true } = prepared.body
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         model: "gpt-4o-mini",
@@ -53,7 +61,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("lowers chronological system updates to escaped user wrappers in order", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -76,7 +84,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("replays canonical reasoning as OpenAI-compatible reasoning_content", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -94,7 +102,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("writes reasoning to a configured custom field on every assistant message", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: Model.update(model, { compatibility: { reasoningField: "vendor_reasoning" } }),
           messages: [
@@ -120,7 +128,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("rejects reasoning fields that conflict with assistant message fields", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: Model.update(model, { compatibility: { reasoningField: "content" } }),
           messages: [Message.assistant([{ type: "reasoning", text: "thinking" }])],
@@ -133,7 +141,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("maps OpenAI provider options to Chat options", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).chat("gpt-4o-mini"),
           prompt: "think",
@@ -148,7 +156,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("passes through custom OpenAI-compatible reasoning effort strings", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "think",
@@ -162,7 +170,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("adds native query params to the Chat Completions URL", () =>
     LLMClient.generate(
-      LLM.updateRequest(request, {
+      LLMRequest.update(request, {
         model: Model.update(model, { route: model.route.with({ endpoint: { query: { "api-version": "v1" } } }) }),
       }),
     ).pipe(
@@ -182,7 +190,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("uses Azure api-key header for static OpenAI Chat keys", () =>
     LLMClient.generate(
-      LLM.updateRequest(request, {
+      LLMRequest.update(request, {
         model: Azure.configure({
           baseURL: "https://opencode-test.openai.azure.com/openai/v1/",
           apiKey: "azure-key",
@@ -208,15 +216,15 @@ describe("OpenAI Chat route", () => {
 
   it.effect("applies serializable HTTP overlays after payload lowering", () =>
     LLMClient.generate(
-      LLM.updateRequest(request, {
+      LLMRequest.update(request, {
         model: model.route
           .with({ auth: Auth.bearer("fresh-key"), headers: { authorization: "Bearer stale" } })
           .model({ id: model.id }),
-        http: {
+        http: HttpOptions.make({
           body: { metadata: { source: "test" } },
           headers: { authorization: "Bearer request", "x-custom": "yes" },
           query: { debug: "1" },
-        },
+        }),
       }),
     ).pipe(
       Effect.provide(
@@ -242,7 +250,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("prepares assistant tool-call and tool-result messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result",
           model,
@@ -280,7 +288,7 @@ describe("OpenAI Chat route", () => {
   it.effect("preserves structured tool errors for the model", () =>
     Effect.gen(function* () {
       const error = { error: { type: "unknown", message: "Tool execution interrupted" } }
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -300,7 +308,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("continues image tool results as vision input without base64 text", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -344,7 +352,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("orders parallel tool responses before one aggregated vision message", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -394,7 +402,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("aggregates consecutive tool images with a following system update", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -435,7 +443,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("appends system updates without replacing multipart user content", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -463,7 +471,7 @@ describe("OpenAI Chat route", () => {
   ] as const)
     it.effect(`rejects ${name}`, () =>
       Effect.gen(function* () {
-        const error = yield* LLMClient.prepare(
+        const error = yield* compileRequest(
           LLM.request({ model, messages: [Message.user({ type: "media", ...media })] }),
         ).pipe(Effect.flip)
         expect(error.message).toMatch(/does not support|does not match|valid base64/)
@@ -472,7 +480,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("rejects oversized image input", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -490,7 +498,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("prepares raw and data URL image media as vision input", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_media",
           model,
@@ -517,7 +525,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("lowers reasoning-only assistant history", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_reasoning",
           model,
@@ -539,7 +547,7 @@ describe("OpenAI Chat route", () => {
           prompt_tokens: 5,
           completion_tokens: 2,
           total_tokens: 7,
-          prompt_tokens_details: { cached_tokens: 1 },
+          prompt_tokens_details: { cached_tokens: 1, cache_write_tokens: 2 },
           completion_tokens_details: { reasoning_tokens: 0 },
         }),
       )
@@ -547,8 +555,9 @@ describe("OpenAI Chat route", () => {
       const usage = new Usage({
         inputTokens: 5,
         outputTokens: 2,
-        nonCachedInputTokens: 4,
+        nonCachedInputTokens: 2,
         cacheReadInputTokens: 1,
+        cacheWriteInputTokens: 2,
         reasoningTokens: 0,
         totalTokens: 7,
         providerMetadata: {
@@ -556,7 +565,7 @@ describe("OpenAI Chat route", () => {
             prompt_tokens: 5,
             completion_tokens: 2,
             total_tokens: 7,
-            prompt_tokens_details: { cached_tokens: 1 },
+            prompt_tokens_details: { cached_tokens: 1, cache_write_tokens: 2 },
             completion_tokens_details: { reasoning_tokens: 0 },
           },
         },
@@ -569,10 +578,16 @@ describe("OpenAI Chat route", () => {
         { type: "text-delta", id: "text-0", text: "Hello" },
         { type: "text-delta", id: "text-0", text: "!" },
         { type: "text-end", id: "text-0" },
-        { type: "step-finish", index: 0, reason: "stop", usage, providerMetadata: undefined },
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "stop", raw: "stop" },
+          usage,
+          providerMetadata: undefined,
+        },
         {
           type: "finish",
-          reason: "stop",
+          reason: { normalized: "stop", raw: "stop" },
           usage,
         },
       ])
@@ -601,9 +616,7 @@ describe("OpenAI Chat route", () => {
           openai: { reasoningField: field },
         })
 
-        const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-          LLM.request({ model, messages: [response.message] }),
-        )
+        const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
         expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", [field]: "thinking" }])
       }
     }),
@@ -612,7 +625,7 @@ describe("OpenAI Chat route", () => {
   it.effect("parses and replays a configured custom reasoning field", () =>
     Effect.gen(function* () {
       const custom = Model.update(model, { compatibility: { reasoningField: "vendor_reasoning" } })
-      const response = yield* LLMClient.generate(LLM.updateRequest(request, { model: custom })).pipe(
+      const response = yield* LLMClient.generate(LLMRequest.update(request, { model: custom })).pipe(
         Effect.provide(
           fixedResponse(
             sseEvents(
@@ -629,12 +642,8 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningField: "vendor_reasoning" },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model: custom, messages: [response.message] }),
-      )
-      expect(replay.body.messages).toEqual([
-        { role: "assistant", content: "Hello", vendor_reasoning: "thinking" },
-      ])
+      const replay = yield* compileRequest(LLM.request({ model: custom, messages: [response.message] }))
+      expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", vendor_reasoning: "thinking" }])
     }),
   )
 
@@ -645,8 +654,8 @@ describe("OpenAI Chat route", () => {
         { type: "reasoning.encrypted", data: "opaque", format: "anthropic-claude-v1", index: 1 },
       ]
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(
         Effect.provide(
@@ -676,9 +685,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningField: "reasoning", reasoningDetails: details },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         {
           role: "assistant",
@@ -721,9 +728,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningDetails: details },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_details: details }])
     }),
   )
@@ -748,9 +753,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningField: "reasoning", reasoningDetails: details },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         { role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: details },
       ])
@@ -823,9 +826,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningDetails: [] },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_details: [] }])
     }),
   )
@@ -873,9 +874,7 @@ describe("OpenAI Chat route", () => {
         response.events.findIndex(LLMEvent.is.textStart),
       )
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         { role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: merged },
       ])
@@ -902,9 +901,7 @@ describe("OpenAI Chat route", () => {
       expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
       expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([{ role: "assistant", content: null, reasoning_details: details }])
     }),
   )
@@ -934,7 +931,7 @@ describe("OpenAI Chat route", () => {
     Effect.gen(function* () {
       const first = { type: "reasoning.text", text: "first", signature: "signed-0", index: 0 }
       const second = { type: "reasoning.text", text: "second", signature: "signed-1", index: 1 }
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const replay = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -963,7 +960,7 @@ describe("OpenAI Chat route", () => {
   it.effect("retains scalar replay for mixed structured reasoning parts", () =>
     Effect.gen(function* () {
       const detail = { type: "reasoning.encrypted", data: "opaque", index: 0 }
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const replay = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -988,7 +985,7 @@ describe("OpenAI Chat route", () => {
   it.effect("replays native scalar reasoning alongside native details", () =>
     Effect.gen(function* () {
       const details = [{ type: "reasoning.encrypted", data: "opaque", index: 0 }]
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const replay = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -1018,8 +1015,8 @@ describe("OpenAI Chat route", () => {
         deltaChunk({}, "tool_calls"),
       )
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
 
@@ -1037,8 +1034,14 @@ describe("OpenAI Chat route", () => {
           providerExecuted: undefined,
           providerMetadata: undefined,
         },
-        { type: "step-finish", index: 0, reason: "tool-calls", usage: undefined, providerMetadata: undefined },
-        { type: "finish", reason: "tool-calls", usage: undefined },
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "tool-calls", raw: "tool_calls" },
+          usage: undefined,
+          providerMetadata: undefined,
+        },
+        { type: "finish", reason: { normalized: "tool-calls", raw: "tool_calls" }, usage: undefined },
       ])
     }),
   )
@@ -1055,8 +1058,8 @@ describe("OpenAI Chat route", () => {
         deltaChunk({}, "tool_calls"),
       )
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
 
@@ -1077,8 +1080,8 @@ describe("OpenAI Chat route", () => {
         deltaChunk({}, "tool_calls"),
       )
       const response = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)))
 
@@ -1095,8 +1098,8 @@ describe("OpenAI Chat route", () => {
         deltaChunk({}, "tool_calls"),
       )
       const error = yield* LLMClient.generate(
-        LLM.updateRequest(request, {
-          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
         }),
       ).pipe(Effect.provide(fixedResponse(body)), Effect.flip)
 
@@ -1113,8 +1116,8 @@ describe("OpenAI Chat route", () => {
         }),
         deltaChunk({ tool_calls: [{ index: 0, function: { arguments: ':"weather"}' } }] }),
       )
-      const input = LLM.updateRequest(request, {
-        tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+      const input = LLMRequest.update(request, {
+        tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
       })
       const events: LLMEvent[] = []
       const streamError = yield* LLMClient.stream(input).pipe(