refactor(tui): remove unavailable sharing commands

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.

.gitattributes

@@ -1,2 +1,3 @@
 packages/core/migration/**/snapshot.json linguist-generated
 packages/core/src/database/migration.gen.ts linguist-generated
+packages/core/src/**/*.txt text eol=lf

.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/publish.yml

@@ -90,11 +90,18 @@ jobs:
           opencode-app-id: ${{ vars.OPENCODE_APP_ID }}
           opencode-app-secret: ${{ secrets.OPENCODE_APP_SECRET }}
 
-      - name: Build
+      - name: Build legacy CLI
+        if: github.ref_name != 'v2'
+        run: ./packages/opencode/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
+        env:
+          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
+          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
+          GH_REPO: ${{ needs.version.outputs.repo }}
+          GH_TOKEN: ${{ steps.committer.outputs.token }}
+
+      - name: Build preview CLI
         id: build
-        run: |
+        run: ./packages/cli/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
-          ./packages/opencode/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
-          ./packages/cli/script/build.ts ${{ (github.ref_name == 'beta' && '--sourcemaps') || '' }}
         env:
           OPENCODE_VERSION: ${{ needs.version.outputs.version }}
           OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
@@ -102,6 +109,7 @@ jobs:
           GH_TOKEN: ${{ steps.committer.outputs.token }}
 
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        if: github.ref_name != 'v2'
         with:
           name: opencode-cli
           path: |
@@ -109,6 +117,7 @@ jobs:
             packages/opencode/dist/opencode-linux*
 
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        if: github.ref_name != 'v2'
         with:
           name: opencode-cli-windows
           path: packages/opencode/dist/opencode-windows*
@@ -121,6 +130,55 @@ jobs:
     outputs:
       version: ${{ needs.version.outputs.version }}
 
+  build-node-cli:
+    needs: version
+    if: github.repository == 'anomalyco/opencode'
+    strategy:
+      fail-fast: false
+      matrix:
+        settings:
+          - target: linux-arm64
+            host: blacksmith-4vcpu-ubuntu-2404-arm
+          - target: linux-x64
+            host: blacksmith-4vcpu-ubuntu-2404
+          - target: darwin-arm64
+            host: macos-26
+          - target: windows-arm64
+            host: blacksmith-4vcpu-windows-2025
+          - target: windows-x64
+            host: blacksmith-4vcpu-windows-2025
+    runs-on: ${{ matrix.settings.host }}
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # v3.6.0
+
+      - uses: ./.github/actions/setup-bun
+        with:
+          install-flags: --os=* --cpu=*
+
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
+        with:
+          node-version: "26.4.0"
+
+      - name: Build
+        run: bun packages/cli/script/build-node.ts --target=${{ matrix.settings.target }} --skip-install --outdir=dist/node
+        env:
+          OPENCODE_VERSION: ${{ needs.version.outputs.version }}
+          OPENCODE_RELEASE: ${{ needs.version.outputs.release }}
+
+      - name: Verify service lifecycle
+        if: matrix.settings.target != 'windows-arm64'
+        working-directory: packages/cli
+        run: bun run script/service-smoke.ts --node
+
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: opencode-node-cli-${{ matrix.settings.target }}
+          path: packages/cli/dist/node/cli-node-*
+          if-no-files-found: error
+
   sign-cli-windows:
     needs:
       - build-cli
@@ -413,6 +471,7 @@ jobs:
     needs:
       - version
       - build-cli
+      - build-node-cli
       - sign-cli-windows
       - build-electron
     if: always() && !failure() && !cancelled()
@@ -441,11 +500,13 @@ jobs:
           registry-url: "https://registry.npmjs.org"
 
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        if: github.ref_name != 'v2'
         with:
           name: opencode-cli
           path: packages/opencode/dist
 
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        if: github.ref_name != 'v2'
         with:
           name: opencode-cli-windows
           path: packages/opencode/dist
@@ -461,6 +522,12 @@ jobs:
           name: opencode-preview-cli
           path: packages/cli/dist
 
+      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        with:
+          pattern: opencode-node-cli-*
+          path: packages/cli/dist/node
+          merge-multiple: true
+
       - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
         if: needs.version.outputs.release
         with:

.github/workflows/test.yml

@@ -78,11 +78,30 @@ jobs:
           bun run script/build.ts --single --skip-install
           bun run script/service-smoke.ts
 
+      - name: Setup Node build runtime
+        if: always()
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
+        with:
+          node-version: "26.4.0"
+
+      - name: Verify Node build
+        if: always()
+        timeout-minutes: 15
+        working-directory: packages/cli
+        run: |
+          bun run script/build-node.ts --single --skip-install --outdir=dist/node
+          bun run script/service-smoke.ts --node
+
       - name: Check generated client
         if: runner.os == 'Linux'
         working-directory: packages/client
         run: bun run check:generated
 
+      - name: Check generated documentation
+        if: runner.os == 'Linux'
+        working-directory: packages/www
+        run: bun run check:generated
+
   e2e:
     name: e2e (${{ matrix.settings.name }})
     if: github.ref_name != 'v2' && github.head_ref != 'v2'

.gitignore

@@ -11,6 +11,7 @@ node_modules
 playground
 tmp
 dist
+dist-node
 ts-dist
 .turbo
 .typecheck-profiles
@@ -25,6 +26,7 @@ Session.vim
 a.out
 target
 .scripts
+.cache
 .direnv/
 
 # Local dev files

.opencode/command/translate.md

@@ -1,6 +1,6 @@
 ---
 description: translate English to other languages
-model: opencode/claude-opus-4-8
+model: opencode/gpt-5.6-sol
 ---
 
 run git diff and translate changed english doc and UI copy files to other international languages. Translate all languages in parallel to save time.

.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.
@@ -19,8 +18,6 @@ Valid types are `feat`, `fix`, `docs`, `chore`, `refactor`, and `test`. Scopes a
 
 Examples: `fix(tui): simplify thinking toggle styling`, `docs: update contributing guide`, `chore(sdk): regenerate types`.
 
-Never bypass Git hooks. Do not use `--no-verify` or otherwise disable, skip, or circumvent commit or push hooks. If a hook fails, fix the failure or stop and report it to the user.
-
 ## Style Guide
 
 ### General Principles
@@ -63,6 +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"
   }
 }

infra/stage.ts

@@ -8,6 +8,25 @@ export const zoneID = "430ba34c138cfb5360826c4909f99be8"
 export const awsStage = $app.stage === "production" ? "production" : "dev"
 export const deployAws = $app.stage === awsStage
 
+if ($app.stage === "production") {
+  new cloudflare.DnsRecord("TrustCenter", {
+    zoneId: zoneID,
+    name: "trust.opencode.ai",
+    type: "CNAME",
+    content: "3a69a5bb27875189.vercel-dns-016.com",
+    proxied: false,
+    ttl: 60,
+  })
+
+  new cloudflare.DnsRecord("TrustCenterVerification", {
+    zoneId: zoneID,
+    name: "opencode.ai",
+    type: "TXT",
+    content: "compai-domain-verification=org_6993a99c6200a2d642bb115d",
+    ttl: 60,
+  })
+}
+
 new cloudflare.RegionalHostname("RegionalHostname", {
   hostname: domain,
   regionKey: "us",

nix/desktop.nix

@@ -8,6 +8,8 @@
   makeWrapper,
   writableTmpDirAsHomeHook,
   autoPatchelfHook,
+  copyDesktopItems,
+  makeDesktopItem,
   opencode,
 }:
 let
@@ -27,9 +29,12 @@ stdenv.mkDerivation (finalAttrs: {
     nodejs
     makeWrapper
     writableTmpDirAsHomeHook
-  ] ++ lib.optionals stdenv.hostPlatform.isLinux [
+  ]
+  ++ lib.optionals stdenv.hostPlatform.isLinux [
     autoPatchelfHook
-  ] ++ lib.optionals stdenv.hostPlatform.isDarwin [
+    copyDesktopItems
+  ]
+  ++ lib.optionals stdenv.hostPlatform.isDarwin [
     # Ad-hoc sign the .app: --config.mac.identity=null below skips signing.
     darwin.autoSignDarwinBinariesHook
   ];
@@ -38,20 +43,37 @@ stdenv.mkDerivation (finalAttrs: {
     (lib.getLib stdenv.cc.cc)
   ];
 
+  desktopItems = lib.optional stdenv.hostPlatform.isLinux (makeDesktopItem {
+    name = "ai.opencode.desktop";
+    desktopName = "OpenCode";
+    exec = "opencode-desktop %U";
+    icon = "ai.opencode.desktop";
+    # Electron 41 derives X11 WM_CLASS from app.name.
+    startupWMClass = "OpenCode";
+    categories = [ "Development" ];
+  });
+
   env = opencode.env // {
     ELECTRON_SKIP_BINARY_DOWNLOAD = "1";
   };
 
-  # https://github.com/electron/electron/issues/31121
+  postPatch =
-  # mac builds use a .app bundle which doesnt have this issue
+    # NOTE: Relax Bun version check to be a warning instead of an error
-  postPatch = lib.optionalString stdenv.isLinux ''
+    ''
-    BASE_PATH=packages/desktop
+      substituteInPlace packages/script/src/index.ts \
-    FILES=(src/main/windows.ts)
+        --replace-fail 'throw new Error(`This script requires bun@''${expectedBunVersionRange}' \
-    for file in "''${FILES[@]}"; do
+                       'console.warn(`Warning: This script requires bun@''${expectedBunVersionRange}'
-      substituteInPlace $BASE_PATH/$file \
+    ''
-        --replace-fail "process.resourcesPath" "'$out/opt/opencode-desktop/resources'"
+    # https://github.com/electron/electron/issues/31121
-    done
+    # mac builds use a .app bundle which doesnt have this issue
-  '';
+    + lib.optionalString stdenv.isLinux ''
+      BASE_PATH=packages/desktop
+      FILES=(src/main/windows.ts)
+      for file in "''${FILES[@]}"; do
+        substituteInPlace $BASE_PATH/$file \
+          --replace-fail "process.resourcesPath" "'$out/opt/opencode-desktop/resources'"
+      done
+    '';
 
   preBuild = ''
     cp -r "${electron.dist}" $HOME/.electron-dist
@@ -76,27 +98,38 @@ stdenv.mkDerivation (finalAttrs: {
     runHook postBuild
   '';
 
-  installPhase =
+  installPhase = ''
-    ''
+    runHook preInstall
-      runHook preInstall
+  ''
-    ''
+  + lib.optionalString stdenv.hostPlatform.isDarwin ''
-    + lib.optionalString stdenv.hostPlatform.isDarwin ''
+    mkdir -p $out/Applications
-      mkdir -p $out/Applications
+    mv dist/mac*/*.app $out/Applications
-      mv dist/mac*/*.app $out/Applications
+    makeWrapper "$out/Applications/OpenCode.app/Contents/MacOS/OpenCode" $out/bin/opencode-desktop
-      makeWrapper "$out/Applications/OpenCode.app/Contents/MacOS/OpenCode" $out/bin/opencode-desktop
+  ''
-    ''
+  + lib.optionalString stdenv.hostPlatform.isLinux ''
-    + lib.optionalString stdenv.hostPlatform.isLinux ''
+    mkdir -p $out/opt/opencode-desktop
-      mkdir -p $out/opt/opencode-desktop
+    cp -r dist/linux*-unpacked/{resources,LICENSE*} $out/opt/opencode-desktop
-      cp -r dist/linux*-unpacked/{resources,LICENSE*} $out/opt/opencode-desktop
+    install -Dm644 resources/icons/32x32.png \
-      makeWrapper ${lib.getExe electron} $out/bin/opencode-desktop \
+      "$out/share/icons/hicolor/32x32/apps/ai.opencode.desktop.png"
-        --inherit-argv0 \
+    install -Dm644 resources/icons/64x64.png \
-        --set ELECTRON_FORCE_IS_PACKAGED 1 \
+      "$out/share/icons/hicolor/64x64/apps/ai.opencode.desktop.png"
-        --add-flags $out/opt/opencode-desktop/resources/app.asar \
+    install -Dm644 resources/icons/128x128.png \
-        --add-flags "\''${NIXOS_OZONE_WL:+\''${WAYLAND_DISPLAY:+--ozone-platform-hint=auto --enable-features=WaylandWindowDecorations --enable-wayland-ime=true}}"
+      "$out/share/icons/hicolor/128x128/apps/ai.opencode.desktop.png"
-    ''
+    install -Dm644 resources/icons/128x128@2x.png \
-    + ''
+      "$out/share/icons/hicolor/256x256/apps/ai.opencode.desktop.png"
-      runHook postInstall
+    install -Dm644 resources/icons/icon.png \
-    '';
+      "$out/share/icons/hicolor/512x512/apps/ai.opencode.desktop.png"
+    install -Dm644 resources/ai.opencode.desktop.metainfo.xml \
+      "$out/share/metainfo/ai.opencode.desktop.metainfo.xml"
+    makeWrapper ${lib.getExe electron} $out/bin/opencode-desktop \
+     --inherit-argv0 \
+     --set ELECTRON_FORCE_IS_PACKAGED 1 \
+     --add-flags $out/opt/opencode-desktop/resources/app.asar \
+     --add-flags "\''${NIXOS_OZONE_WL:+\''${WAYLAND_DISPLAY:+--ozone-platform-hint=auto --enable-features=WaylandWindowDecorations --enable-wayland-ime=true}}"
+  ''
+  + ''
+    runHook postInstall
+  '';
 
   autoPatchelfIgnoreMissingDeps = [
     "libc.musl-x86_64.so.1"

nix/hashes.json

@@ -1,8 +1,8 @@
 {
   "nodeModules": {
-    "x86_64-linux": "sha256-F1luclnqCPQk9yxfmeSYGaM/nScf28yBu9K3Fv+Xd24=",
+    "x86_64-linux": "sha256-RFek0QoEEjsgbqmTE/SxQAmPtYyzs0IPR2ugFn5Okrs=",
-    "aarch64-linux": "sha256-XW0XZnsCRkU3MFJH9TjMRYZHffzVy3cQyiNCkec2gl4=",
+    "aarch64-linux": "sha256-BmAxapY1YrAFn7mVq3/6A9+6Au5UIvSqBboHMkyJH3I=",
-    "aarch64-darwin": "sha256-bf8kvORs3Fs2UYLp3PekF+AJR7NKOcHb+fIQA79RtMk=",
+    "aarch64-darwin": "sha256-Sx3bGWQqLlgoa/RudJxanjSzhFRNklckT2ffnO2I5F4=",
-    "x86_64-darwin": "sha256-sBdQPkzd7JXNW6Lbi9JHiAsfHwdLwTKWY+uPeXAv2Nw="
+    "x86_64-darwin": "sha256-CMOhiisHNowg06qadvgg4K+60zrynglwiT0qKYQ4NiA="
   }
 }

package.json

@@ -15,7 +15,7 @@
     "dev:www": "bun run --cwd packages/www dev",
     "dev:storybook": "bun --cwd packages/storybook storybook",
     "lint": "oxlint",
-    "lint:effect-patterns": "ast-grep scan -c script/ast-grep/sgconfig.yml packages/core/src packages/server/src packages/protocol/src packages/cli/src",
+    "lint:effect-patterns": "ast-grep scan -c script/ast-grep/sgconfig.yml packages/util/src packages/core/src packages/server/src packages/protocol/src packages/cli/src",
     "test:lint-rules": "ast-grep test -c script/ast-grep/sgconfig.yml",
     "typecheck": "bun turbo typecheck --concurrency=3",
     "typecheck:profile": "bun script/profile-typecheck.ts",
@@ -33,24 +33,24 @@
       "packages/*",
       "packages/console/*",
       "packages/stats/*",
-      "packages/sdk/js",
       "packages/slack"
     ],
     "catalog": {
-      "@effect/opentelemetry": "4.0.0-beta.83",
+      "@effect/opentelemetry": "4.0.0-beta.101",
-      "@effect/platform-node": "4.0.0-beta.83",
+      "@effect/platform-node": "4.0.0-beta.101",
-      "@effect/sql-sqlite-bun": "4.0.0-beta.83",
+      "@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",
       "@octokit/rest": "22.0.0",
       "@hono/standard-validator": "0.2.0",
       "@hono/zod-validator": "0.4.2",
-      "@opentui/core": "0.4.3",
+      "@opentui/core": "0.4.5",
-      "@opentui/keymap": "0.4.3",
+      "@opentui/keymap": "0.4.5",
-      "@opentui/solid": "0.4.3",
+      "@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",
@@ -69,12 +69,13 @@
       "dompurify": "3.3.1",
       "drizzle-kit": "1.0.0-rc.2",
       "drizzle-orm": "1.0.0-rc.2",
-      "effect": "4.0.0-beta.83",
+      "effect": "4.0.0-beta.101",
       "ai": "6.0.168",
       "cross-spawn": "7.0.6",
       "hono": "4.10.7",
       "hono-openapi": "1.1.2",
       "fuzzysort": "3.1.0",
+      "get-east-asian-width": "1.6.0",
       "luxon": "3.6.1",
       "marked": "17.0.6",
       "marked-shiki": "1.2.1",
@@ -85,9 +86,11 @@
       "@typescript/native-preview": "7.0.0-dev.20251207.1",
       "zod": "4.1.8",
       "remeda": "2.26.0",
+      "resolve.exports": "2.0.3",
       "sst": "4.13.1",
       "shiki": "4.2.0",
       "solid-list": "0.3.0",
+      "string-width": "7.2.0",
       "tailwindcss": "4.1.11",
       "vite": "7.1.4",
       "@solidjs/meta": "0.29.4",
@@ -121,7 +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:"
   },
@@ -149,21 +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.83": "patches/effect@4.0.0-beta.83.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

@@ -12,6 +12,8 @@
 
 Per-type constructors live on the type, not as top-level re-exports. Use `Message.system(...)`, `Message.user(...)`, `Message.assistant(...)`, `Message.tool(...)`, `Model.make(...)`, `ToolDefinition.make(...)`, `ToolCallPart.make(...)`, `ToolResultPart.make(...)`, `ToolChoice.make(...)`, `ToolChoice.named(...)`, `SystemPart.make(...)`, and `GenerationOptions.make(...)` directly. The top-level `LLM` namespace is reserved for request-shaped call APIs: `LLM.request`, `LLM.generate`, `LLM.stream`, `LLM.updateRequest`, and `LLM.generateObject`. Two ways to construct the same thing is one too many.
 
+- Keep provider-defined string enums forward-compatible. Expose known values for autocomplete while accepting future values with `Known | (string & {})`; use `Schema.String` at runtime unless rejecting unknown values is required for correctness.
+
 ## Tests
 
 - Use `testEffect(...)` from `test/lib/effect.ts` for tests requiring Effect layers.
@@ -46,7 +48,7 @@ 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 +56,7 @@ 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 +140,13 @@ 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 +160,14 @@ 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 +178,7 @@ 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 +243,7 @@ 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

@@ -1,6 +1,6 @@
 # @opencode-ai/ai
 
-Schema-first LLM core for opencode. One typed request, response, event, and tool language; provider quirks live in adapters, not in calling code.
+Schema-first AI primitives for opencode. Provider quirks live in adapters, not in calling code.
 
 ```ts
 import { Effect } from "effect"
@@ -24,14 +24,181 @@ const program = Effect.gen(function* () {
 
 Run `LLMClient.stream(request)` instead of `generate` when you want incremental `LLMEvent`s. The event stream is provider-neutral — same shape across OpenAI Chat, OpenAI Responses, Anthropic Messages, Gemini, Bedrock Converse, and any OpenAI-compatible deployment.
 
+## Image generation
+
+Use `Image.generate` with an image model for direct asset generation:
+
+```ts
+import { Image, ImageInput } from "@opencode-ai/ai"
+import { OpenAI } from "@opencode-ai/ai/providers"
+
+const program = Effect.gen(function* () {
+  const response = yield* Image.generate({
+    model: OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY }).image("gpt-image-2"),
+    prompt: "A robot tending a rooftop garden",
+    options: {
+      n: 2,
+      size: "1024x1024",
+      quality: "high", // inferred from the OpenAI image model
+      outputFormat: "webp",
+      future_option: true, // unknown native options pass through unchanged
+    },
+  })
+
+  return response.images // GeneratedImage[] with owned bytes or a provider URL
+})
+```
+
+Pass ordered image inputs to the same method for editing, composition, or image-conditioned generation:
+
+```ts
+const response =
+  yield *
+  Image.generate({
+    model,
+    prompt: "Combine these product photos into one studio scene",
+    images: [
+      ImageInput.bytes(firstBytes, "image/png"),
+      ImageInput.url("https://example.com/second.webp"),
+      ImageInput.file("file_123"),
+    ],
+    options,
+    http,
+  })
+```
+
+`ImageInput.fileUri(uri, mediaType)` represents provider file URIs such as Gemini Files. Raw strings are not
+accepted as image inputs, avoiding ambiguity between base64, URLs, and provider IDs. Empty or omitted `images`
+uses text-to-image generation; a non-empty array selects the provider's edit behavior without enforcing provider
+image-count limits locally. `images` is the only common image-editing field. OpenAI uses multipart for byte/data-URL
+edits and its JSON reference body for URL or file-ID edits. Its provider-specific `options.mask` accepts an
+`ImageInput` for inpainting:
+
+```ts
+yield *
+  Image.generate({
+    model: OpenAI.configure({ apiKey }).image("gpt-image-2"),
+    prompt,
+    images: [ImageInput.bytes(sourceBytes, "image/png")],
+    options: { mask: ImageInput.bytes(maskBytes, "image/png") },
+  })
+```
+
+The OpenAI adapter extracts this helper value into the edit request's native `mask` field rather than passing the
+tagged `ImageInput` object through as an ordinary option. On multipart requests, `http.body` can override option
+fields but not structural `model`, `prompt`, `image[]`, or `mask` fields, and the transport owns the multipart
+`Content-Type` boundary. For JSON requests, `http.body` remains the final raw-native overlay. Gemini does not fetch
+public HTTP URLs, and hosted Z.ai image generation does not accept image inputs. These cases fail with
+`InvalidRequest` before network I/O.
+
+Provider-native image options belong to each request. Raw `http.body` fields have final precedence over them:
+
+```ts
+const model = OpenAI.configure({ apiKey }).image("gpt-image-2")
+
+yield *
+  Image.generate({
+    model,
+    prompt,
+    options: { quality: "medium" },
+    http,
+  })
+```
+
+xAI image models use the same request API with xAI-native controls:
+
+```ts
+yield *
+  Image.generate({
+    model: XAI.configure({ apiKey }).image("any-model-id"),
+    prompt,
+    options: {
+      n: 2,
+      aspectRatio: "16:9",
+      resolution: "1k",
+      responseFormat: "b64_json",
+      future_option: true,
+    },
+    http,
+  })
+```
+
+Google's current Gemini image models use the same direct API:
+
+```ts
+import { Google } from "@opencode-ai/ai/providers"
+
+const googleProgram = Effect.gen(function* () {
+  const response = yield* Image.generate({
+    model: Google.configure({ apiKey }).image("any-model-id"),
+    prompt: "A robot tending a rooftop garden",
+    options: {
+      aspectRatio: "16:9",
+      imageSize: "2K",
+      seed: 42,
+      thinkingLevel: "HIGH",
+      includeThoughts: true,
+      futureOption: true,
+    },
+    http,
+  })
+
+  return response.images
+})
+```
+
+Google image options are request-scoped and inferred from the selected model. Known fields autocomplete while
+future string values and arbitrary native Gemini `generationConfig` fields remain available. Native fields override
+their mapped aliases, and `http.body` is the final deep overlay. The selected model ID is sent to Gemini
+`generateContent` without a local allowlist.
+
+Z.ai image models infer open Z.ai-native options from the selected model:
+
+```ts
+yield *
+  Image.generate({
+    model: ZAI.configure({ apiKey }).image("any-model-id"),
+    prompt,
+    options: {
+      quality: "hd",
+      userID: "user-123",
+      future_option: true,
+    },
+    http,
+  })
+```
+
+Z.ai does not include trustworthy MIME metadata for output URLs, so generated images use
+`application/octet-stream`. Output URLs expire after 30 days; download and persist them promptly if they must
+remain available.
+
+Conversational image generation remains part of the LLM interaction. OpenAI Responses exposes it through its hosted image tool:
+
+```ts
+const program = Effect.gen(function* () {
+  const response = yield* LLM.generate(
+    LLM.request({
+      model: OpenAI.configure({ apiKey }).responses("gpt-5"),
+      prompt: "Design a solarpunk rooftop garden, then show me.",
+      tools: [OpenAI.imageGeneration({ quality: "high" })],
+    }),
+  )
+
+  return response.message
+})
+```
+
+The hosted result is represented as a provider-executed tool call and tool result. Its image is a `file` content item with a data URI, so retaining `response.message` preserves the generated image for continuation.
+
 ## Public API
 
 - **`LLM.request({...})`** — build a provider-neutral `LLMRequest`. Accepts ergonomic inputs (`system: string`, `prompt: string`) that normalize into the canonical Schema classes.
 - **`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`.
 
 ## Caching
 
@@ -39,7 +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.
 
@@ -67,7 +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({
@@ -83,8 +252,8 @@ LLM.request({
 
 | Protocol                | `cache: "auto"`                                                           |
 | ----------------------- | ------------------------------------------------------------------------- |
-| Anthropic Messages      | emits up to 3 `cache_control` markers (4-breakpoint cap enforced)         |
+| Anthropic Messages      | emits up to 4 `cache_control` markers (4-breakpoint cap enforced)         |
-| Bedrock Converse        | emits up to 3 `cachePoint` blocks (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) |
 
@@ -104,7 +273,7 @@ const gateway = CloudflareAIGateway.configure({
 }).model("workers-ai/@cf/meta/llama-3.1-8b-instruct")
 ```
 
-Included providers: OpenAI, Anthropic, Google (Gemini), Google Vertex Gemini and Anthropic, Amazon Bedrock, Azure OpenAI, Cloudflare AI Gateway, Cloudflare Workers AI, GitHub Copilot, OpenRouter, xAI, plus generic OpenAI-compatible Chat and Responses entrypoints and an Anthropic Messages-compatible entrypoint.
+Included providers: OpenAI, Anthropic, Google (Gemini), Google Vertex Gemini and Anthropic, Amazon Bedrock, Azure OpenAI, Cloudflare AI Gateway, Cloudflare Workers AI, GitHub Copilot, OpenRouter, xAI, Z.ai, plus generic OpenAI-compatible Chat and Responses entrypoints and an Anthropic Messages-compatible entrypoint.
 
 ### Package-like entrypoints
 
@@ -127,23 +296,37 @@ OpenAI Chat and OpenAI Responses are separate semantic entrypoints:
 - `@opencode-ai/ai/providers/openai/responses`
 - `@opencode-ai/ai/providers/openai-compatible/responses`
 - `@opencode-ai/ai/providers/anthropic-compatible`
-- `@opencode-ai/ai/providers/google-vertex`
+- `@opencode-ai/ai/providers/google-vertex/gemini`
-- `@opencode-ai/ai/providers/google-vertex/anthropic`
+- `@opencode-ai/ai/providers/google-vertex/chat`
+- `@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 and Vertex Anthropic are separate products with separate entrypoints. Both 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 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`.
 
 Tuned Vertex Gemini deployments use model ids shaped like `endpoints/1234567890` and require OAuth or ADC; Vertex express-mode API keys support publisher models only.
 
 ```ts
-import { model } from "@opencode-ai/ai/providers/google-vertex"
+import { model } from "@opencode-ai/ai/providers/google-vertex/gemini"
 
 model("gemini-3.5-flash", { project: "my-project", location: "global" })
 ```
 
 ```ts
-import { model } from "@opencode-ai/ai/providers/google-vertex/anthropic"
+import { model } from "@opencode-ai/ai/providers/google-vertex/chat"
+
+model("deepseek-ai/deepseek-v3.2-maas", { project: "my-project", location: "global" })
+```
+
+```ts
+import { model } from "@opencode-ai/ai/providers/google-vertex/responses"
+
+model("xai/grok-4.20-reasoning", { project: "my-project", location: "global" })
+```
+
+```ts
+import { model } from "@opencode-ai/ai/providers/google-vertex/messages"
 
 model("claude-sonnet-4-6", { project: "my-project", location: "global" })
 ```
@@ -168,7 +351,7 @@ Adding a new model or deployment is usually 5-15 lines using `Route.make({ proto
 
 ## Effect
 
-This package is built on Effect. Public methods return `Effect` or `Stream`; provide `LLMClient.layer` for runtime dispatch and import the provider/protocol modules for the routes you use. The example at `example/tutorial.ts` is a runnable walkthrough.
+This package is built on Effect. Public methods return `Effect` or `Stream`; provide `LLMClient.layer` for LLM dispatch and `ImageClient.layer` for image dispatch, then import the provider/protocol modules for the routes you use. The example at `example/tutorial.ts` is a runnable walkthrough.
 
 ## See also
 

packages/ai/STATUS.md

@@ -1,36 +1,38 @@
 # LLM Provider Parity Status
 
-Last reviewed: 2026-07-15
+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.
 
 ## Existing Status Sources
 
-| File                                 | What it tracks                                                                   | Limitation                                              |
+| File                                | What it tracks                                                          | Limitation                                              |
-| ------------------------------------ | -------------------------------------------------------------------------------- | ------------------------------------------------------- |
+| ----------------------------------- | ----------------------------------------------------------------------- | ------------------------------------------------------- |
-| `packages/ai/DESIGN.md`             | Future clean-break API proposal for `@opencode-ai/ai`.                       | Not a provider parity tracker.                          |
+| `packages/ai/DESIGN.md`             | Future clean-break API proposal for `@opencode-ai/ai`.                  | Not a provider parity tracker.                          |
-| `packages/ai/example/call-sites.md` | Route/value/provider-facade migration checklist and call-site sketches.          | Architecture migration only; not AI SDK package parity. |
+| `packages/ai/example/call-sites.md` | Route/value/provider-facade migration checklist and call-site sketches. | Architecture migration only; not AI SDK package parity. |
 
 ## Current Implementation Snapshot
 
-| Native slice                       | Source                                                                                         | Current state                                                                                                                                                            | Main gaps                                                                                                                                                                                                                                                 |
+| 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 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 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 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 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.                                                                                                                                                                                             |
+| 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.                                                          | No named compatible 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.                                                                                                                                     |
+| 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.                                                                                                                             |
+| 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/google-vertex-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 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 Anthropic Messages          | `src/protocols/google-vertex-anthropic.ts`, `src/providers/google-vertex-anthropic.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.                                                                                                                                  |
+| 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.                                                                                                                                 |
-| 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. |
+| 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.                                                                                                                                 |
-| 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.                                                                                                                          |
+| 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.                                                                                                                                  |
-| 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.                                                                                                                         |
+| 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. |
-| OpenRouter                         | `src/providers/openrouter.ts`                                                                  | Present with OpenRouter-specific usage/reasoning/prompt-cache options over Chat.                                                                                         | Responses-style OpenRouter support is absent.                                                                                                                                                                                                             |
+| 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.                                                                                                                          |
-| xAI                                | `src/providers/xai.ts`                                                                         | Present with Responses and Chat selectors.                                                                                                                               | Needs package-parity review against the AI SDK xAI provider.                                                                                                                                                                                              |
+| 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.                                                                                                                         |
-| 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.                                                                                                                                          |
+| 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
 
@@ -54,8 +56,8 @@ Other `aisdk:` packages, including Google Vertex, Azure, and Bedrock, currently
 | `@ai-sdk/google`                  | Gemini Developer API                                           | Partial / usable | Add typed options for safety, response schema/modalities, cached content, grounding/search/code execution, and non-text output modes where supported.                                                               |
 | `@ai-sdk/google-vertex`           | Vertex Gemini namespace/facade                                 | Partial / usable | Add runner/catalog mapping, recorded coverage, and broader provider-option parity.                                                                                                                                  |
 | `@ai-sdk/google-vertex/anthropic` | Anthropic Messages over Vertex namespace/facade                | Partial / usable | Add runner/catalog mapping, recorded coverage, and Vertex-specific hosted-tool parity.                                                                                                                              |
-| `@ai-sdk/google-vertex/maas`      | Vertex MaaS OpenAI-compatible namespace/facade                 | Missing          | Decide native Chat/Responses selection, endpoint derivation, auth, and catalog mapping.                                                                                                                             |
+| `@ai-sdk/google-vertex/maas`      | Vertex Chat                                                    | Partial / usable | Add runner/catalog mapping, recorded coverage, and MaaS family-specific request parity.                                                                                                                             |
-| `@ai-sdk/google-vertex/xai`       | Vertex xAI OpenAI-compatible namespace/facade                  | Missing          | Decide whether this composes the generic compatible bases or the xAI facade, then add endpoint/auth mapping and tests.                                                                                              |
+| `@ai-sdk/google-vertex/xai`       | Vertex Chat / Responses                                        | Partial / usable | Decide Chat/Responses selection for catalog models, add runner mapping and recorded coverage, and review xAI-specific request options.                                                                              |
 | `@ai-sdk/azure`                   | Azure OpenAI Chat/Responses facade                             | Partial          | Map runner/catalog metadata to native Azure, handle resourceName/baseURL/apiVersion variants, add AAD/token auth story, and verify Chat vs Responses deployment selection.                                          |
 | `@ai-sdk/amazon-bedrock`          | Bedrock Converse                                               | Partial          | Add default AWS credential chain/profile support, region/inference-profile model ID handling, provider option parity via `additionalModelRequestFields`, guardrails/performance config, and runner/catalog mapping. |
 | `@ai-sdk/amazon-bedrock/mantle`   | Bedrock Mantle OpenAI-compatible Chat/Responses namespace      | Missing          | Decide native Mantle shape, likely separate from Converse because it uses OpenAI-compatible Chat/Responses semantics over Bedrock. Add package mapping and tests.                                                   |
@@ -63,34 +65,34 @@ 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 and Vertex Anthropic now have native package entrypoints, but the core runner does not map catalog metadata to them yet and recorded provider coverage is still missing.
+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.
 6. Provider option typing is uneven. OpenAI, Anthropic, Gemini, Bedrock, and OpenRouter each expose a small typed subset plus raw HTTP overlays; this is useful but not equivalent to AI SDK provider option coverage.
 7. Structured output is not provider-native yet. `LLM.generateObject` still uses a synthetic tool strategy, while the future design expects native structured output where reliable and tool fallback where needed.
-8. Package/namespace boundaries for the current native loading set are explicit in docs and exports. Other exported provider facades are not catalog package entrypoints until they implement the contract. Missing native boundaries remain for Vertex MaaS, Vertex xAI, and Bedrock Mantle.
+8. Package/namespace boundaries for the current native loading set are explicit in docs and exports. Other exported provider facades are not catalog package entrypoints until they implement the contract. Vertex xAI still needs catalog API selection; the missing native boundary is Bedrock Mantle.
 9. Recorded coverage is uneven. OpenAI, Anthropic, Gemini, Bedrock Converse, Cloudflare, OpenRouter, and several OpenAI-compatible Chat providers have cassettes. Azure, Vertex, and Mantle need first-class recorded scenarios before switching defaults.
 
 ## Native Namespace Shape
 
 These are implementation/API slices, not separate npm packages.
 
-| API slice                     | Package-like entrypoint                                  | Purpose                                                                      |
+| API slice                     | Package-like entrypoint                                 | Purpose                                                                      |
-| ----------------------------- | -------------------------------------------------------- | ---------------------------------------------------------------------------- |
+| ----------------------------- | ------------------------------------------------------- | ---------------------------------------------------------------------------- |
 | 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`               | Vertex Gemini API.                                                           |
+| Vertex Gemini                 | `@opencode-ai/ai/providers/google-vertex/gemini`        | Vertex Gemini API; `providers/google-vertex` is the default alias.           |
-| Vertex Anthropic Messages     | `@opencode-ai/ai/providers/google-vertex/anthropic`     | Vertex-hosted Anthropic Messages API.                                        |
+| Vertex Chat                   | `@opencode-ai/ai/providers/google-vertex/chat`          | Vertex OpenAI-compatible Chat Completions for MaaS models.                   |
-| Vertex MaaS                   | Missing                                                  | Vertex OpenAI-compatible MaaS APIs.                                          |
+| Vertex Responses              | `@opencode-ai/ai/providers/google-vertex/responses`     | Vertex Open Responses for Grok models.                                       |
-| Vertex xAI                    | Missing                                                  | Vertex-hosted xAI APIs.                                                      |
+| 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.                                   |
+| Bedrock Mantle                | Missing                                                 | AWS Bedrock Mantle OpenAI-compatible APIs.                                   |
 | Azure OpenAI Chat             | `@opencode-ai/ai/providers/azure/chat`                  | Azure specialization of OpenAI Chat.                                         |
 | Azure OpenAI Responses        | `@opencode-ai/ai/providers/azure/responses`             | Azure specialization of OpenAI Responses.                                    |
 
@@ -99,8 +101,8 @@ These are implementation/API slices, not separate npm packages.
 1. Add native runner/catalog mappings for `@ai-sdk/azure`, `@ai-sdk/google`, and `@ai-sdk/amazon-bedrock` where the existing native facades are already close.
 2. Add API-aware runner/catalog selection between OpenAI-compatible Chat and Responses.
 3. Bring Bedrock native auth/config to AI SDK parity: region, profile, default AWS credential chain, bearer token env, endpoint override, and cross-region inference profile handling.
-4. Add runner/catalog mappings and recorded scenarios for the native Vertex Gemini and Vertex Anthropic entrypoints.
+4. Add runner/catalog mappings and recorded scenarios for the native Vertex Gemini, Chat, Responses, and Messages entrypoints.
-5. Add native Vertex MaaS and Vertex xAI entrypoints by composing the compatible bases and shared Vertex auth/endpoint setup.
+5. Decide Chat/Responses selection for `@ai-sdk/google-vertex/xai` catalog models.
 6. Add Bedrock Mantle as a separate OpenAI-compatible Bedrock namespace after deciding whether it uses Chat, Responses, or both by model.
 7. Expand typed provider options from the existing V1 lowerer knowledge in `packages/core/src/v1/config/provider-options.ts` before adding more raw overlay examples.
-8. Add recorded provider tests for Azure, Vertex Gemini, Vertex Anthropic, Bedrock credential-chain behavior, and Mantle before making native runtime the default for those packages.
+8. Add recorded provider tests for Azure, Vertex Gemini, Vertex Chat, Vertex Responses, Vertex Messages, Bedrock credential-chain behavior, and Mantle before making native runtime the default for those packages.

packages/ai/example/call-sites.md

@@ -360,17 +360,29 @@ import { model } from "@opencode-ai/ai/providers/openai/responses"
 model("gpt-4o", { apiKey, transport: "websocket" })
 ```
 
-Vertex keeps Gemini and Anthropic Messages as separate package-like entrypoints,
+Vertex keeps Gemini, Chat, Responses, and Messages as separate package-like entrypoints,
 while sharing project/location resolution and ADC authentication internally:
 
 ```ts
-import { model } from "@opencode-ai/ai/providers/google-vertex"
+import { model } from "@opencode-ai/ai/providers/google-vertex/gemini"
 
 model("gemini-3.5-flash", { project, location: "global" })
 ```
 
 ```ts
-import { model } from "@opencode-ai/ai/providers/google-vertex/anthropic"
+import { model } from "@opencode-ai/ai/providers/google-vertex/chat"
+
+model("deepseek-ai/deepseek-v3.2-maas", { project, location: "global" })
+```
+
+```ts
+import { model } from "@opencode-ai/ai/providers/google-vertex/responses"
+
+model("xai/grok-4.20-reasoning", { project, location: "global" })
+```
+
+```ts
+import { model } from "@opencode-ai/ai/providers/google-vertex/messages"
 
 model("claude-sonnet-4-6", { project, location: "global" })
 ```
@@ -556,7 +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,
+// enabled at a time so the tutorial can demonstrate generate, stream, or
-// or tool-loop behavior without spending tokens on every example.
+// 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

@@ -7,7 +7,7 @@
   "scripts": {
     "setup:recording-env": "bun run script/setup-recording-env.ts",
     "test": "bun test --timeout 30000 --only-failures",
-    "typecheck": "tsgo --noEmit",
+    "typecheck": "tsgo --noEmit && tsgo --noEmit -p tsconfig.types.json",
     "build": "tsc -p tsconfig.build.json"
   },
   "files": [
@@ -15,6 +15,7 @@
   ],
   "exports": {
     ".": "./src/index.ts",
+    "./testing": "./src/testing.ts",
     "./*": "./src/*.ts"
   },
   "devDependencies": {

packages/ai/script/setup-recording-env.ts

@@ -161,6 +161,18 @@ const PROVIDERS: ReadonlyArray<Provider> = [
     vars: [{ name: "TOGETHER_AI_API_KEY" }],
     validate: (env) => validateBearer("https://api.together.xyz/v1/models", Redacted.make(env.TOGETHER_AI_API_KEY)),
   },
+  {
+    id: "minimax",
+    label: "MiniMax",
+    tier: "compatible",
+    note: "Anthropic-compatible Messages text/tool recorded tests",
+    vars: [{ name: "MINIMAX_API_KEY" }],
+    validate: (env) =>
+      HttpClientRequest.get("https://api.minimax.io/anthropic/v1/models").pipe(
+        HttpClientRequest.setHeader("x-api-key", Redacted.value(Redacted.make(env.MINIMAX_API_KEY))),
+        executeRequest,
+      ),
+  },
   {
     id: "mistral",
     label: "Mistral",

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,
+// The default `"auto"` shape places breakpoints at the last tool definition,
-// one at the last system part, and one at the latest user message. This
+// the first and last distinct system parts, and the conversation tail. This
-// matches what production agent harnesses (LangChain's caching middleware,
+// exposes reusable tool, base-agent, project, and session prefixes while
-// kern-ai's 10x cost-reduction playbook) converge on for tool-use loops: the
+// advancing the tail after each tool result keeps the previous cache entry
-// latest user message stays put while a single turn explodes into many
+// within Anthropic's 20-block lookback during long agent turns.
-// assistant/tool round-trips, so caching at that boundary lets every
-// intra-turn API call hit the prefix.
 //
-// Manual `cache: CacheHint` placements on individual parts are preserved —
+// Manual `cache: CacheHint` placements on individual parts are preserved and
-// this function only fills gaps the caller left empty.
+// 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 => {
@@ -39,23 +38,37 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // Protocols whose wire format ignores inline cache markers (OpenAI's implicit
 // prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
 // whole policy pass for these — emitting hints would be harmless but pointless.
-const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "bedrock-converse"])
+const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "bedrock-converse", "openrouter"])
 
 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
+  let changed = false
-  if (system[last]!.cache) return system
+  const next = system.map((part, index) => {
-  return system.map((part, i) => (i === last ? { ...part, cache: hint } : part))
+    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 +105,43 @@ 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-user-message")
-  if (strategy === "latest-assistant") return markMessageAt(messages, lastIndexOfRole(messages, "assistant"), hint)
+    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
+  if (request.model.route.id === "openrouter" && (request.cache === undefined || request.cache === "auto")) 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 budget = { remaining: Math.max(0, BREAKPOINT_CAP - countHints(request)) }
-  const system = policy.system ? markLastSystem(request.system, hint) : request.system
+  const tools = policy.tools ? markLastTool(request.tools, hint, budget) : request.tools
-  const messages = policy.messages ? markMessages(request.messages, policy.messages, hint) : request.messages
+  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/image-client.ts

@@ -0,0 +1,38 @@
+import { Context, Effect, Layer } from "effect"
+import { RequestExecutor } from "./route/executor"
+import type { ImageOptions, ImageRequest, ImageRequestFor, ImageResponse } from "./image"
+import type { LLMError } from "./schema"
+
+export type Execute = RequestExecutor.Interface["execute"]
+
+export interface Interface {
+  readonly generate: <Options extends ImageOptions>(
+    request: ImageRequestFor<Options>,
+  ) => Effect.Effect<ImageResponse, LLMError>
+}
+
+export class Service extends Context.Service<Service, Interface>()("@opencode/ImageClient") {}
+
+export const generate = <Options extends ImageOptions>(
+  request: ImageRequestFor<Options>,
+): Effect.Effect<ImageResponse, LLMError> =>
+  Effect.gen(function* () {
+    const client = yield* Service
+    return yield* client.generate(request)
+  }) as Effect.Effect<ImageResponse, LLMError>
+
+export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer.effect(
+  Service,
+  Effect.gen(function* () {
+    const executor = yield* RequestExecutor.Service
+    return Service.of({
+      generate: (request) => request.model.route.generate(request, executor.execute),
+    })
+  }),
+)
+
+export const ImageClient = {
+  Service,
+  layer,
+  generate,
+} as const

packages/ai/src/image.ts

@@ -0,0 +1,166 @@
+import { Effect, Schema } from "effect"
+import { HttpOptions, InvalidRequestReason, LLMError, ModelID, ProviderID, ProviderMetadata, Usage } from "./schema"
+import { ImageClient, Service, type Execute as ImageExecute } from "./image-client"
+
+export interface ImageRoute<Options extends ImageOptions = ImageOptions> {
+  readonly id: string
+  readonly generate: (
+    request: ImageRequestFor<Options>,
+    execute: ImageExecute,
+  ) => Effect.Effect<ImageResponse, LLMError>
+}
+
+export type ImageOptions = Record<string, unknown>
+
+export class ImageModel<Options extends ImageOptions = ImageOptions> {
+  declare protected readonly _Options: (options: Options) => Options
+  readonly id: ModelID
+  readonly provider: ProviderID
+  readonly route: ImageRoute<Options>
+  readonly http?: HttpOptions
+
+  constructor(input: ImageModel.Input<Options>) {
+    this.id = input.id
+    this.provider = input.provider
+    this.route = input.route
+    this.http = input.http
+  }
+
+  static make<Options extends ImageOptions = ImageOptions>(input: ImageModel.MakeInput<Options>) {
+    return new ImageModel<Options>({
+      id: ModelID.make(input.id),
+      provider: ProviderID.make(input.provider),
+      route: input.route,
+      http: input.http,
+    })
+  }
+}
+
+export namespace ImageModel {
+  export interface Input<Options extends ImageOptions = ImageOptions> {
+    readonly id: ModelID
+    readonly provider: ProviderID
+    readonly route: ImageRoute<Options>
+    readonly http?: HttpOptions
+  }
+
+  export interface MakeInput<Options extends ImageOptions = ImageOptions>
+    extends Omit<Input<Options>, "id" | "provider"> {
+    readonly id: string | ModelID
+    readonly provider: string | ProviderID
+  }
+}
+
+export const ImageModelSchema = Schema.declare((value): value is ImageModel => value instanceof ImageModel, {
+  expected: "Image.Model",
+})
+
+const ImageBytesInput = Schema.Struct({
+  type: Schema.Literal("bytes"),
+  data: Schema.Uint8Array,
+  mediaType: Schema.String,
+})
+const ImageUrlInput = Schema.Struct({
+  type: Schema.Literal("url"),
+  url: Schema.String,
+})
+const ImageFileIDInput = Schema.Struct({
+  type: Schema.Literal("file-id"),
+  id: Schema.String,
+})
+const ImageFileURIInput = Schema.Struct({
+  type: Schema.Literal("file-uri"),
+  uri: Schema.String,
+  mediaType: Schema.String,
+})
+
+export const ImageInputSchema = Schema.Union([
+  ImageBytesInput,
+  ImageUrlInput,
+  ImageFileIDInput,
+  ImageFileURIInput,
+]).pipe(Schema.toTaggedUnion("type"))
+export type ImageInput = Schema.Schema.Type<typeof ImageInputSchema>
+
+export const ImageInput = {
+  bytes: (data: Uint8Array, mediaType: string): ImageInput => ({ type: "bytes", data, mediaType }),
+  url: (url: string): ImageInput => ({ type: "url", url }),
+  file: (id: string): ImageInput => ({ type: "file-id", id }),
+  fileUri: (uri: string, mediaType: string): ImageInput => ({ type: "file-uri", uri, mediaType }),
+} as const
+
+export class ImageRequest extends Schema.Class<ImageRequest>("Image.Request")({
+  model: ImageModelSchema,
+  prompt: Schema.String,
+  images: Schema.optional(Schema.Array(ImageInputSchema)),
+  options: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+  http: Schema.optional(HttpOptions),
+}) {
+  declare protected readonly _ImageRequest: void
+}
+
+export type ImageRequestFor<Options extends ImageOptions = ImageOptions> = Omit<ImageRequest, "model" | "options"> & {
+  readonly model: ImageModel<Options>
+  readonly options?: Options
+}
+
+export type ImageModelOptions<Model> = Model extends ImageModel<infer Options> ? Options : never
+
+export type ImageRequestInput<Model extends object = ImageModel> = Omit<
+  ConstructorParameters<typeof ImageRequest>[0],
+  "model" | "options" | "http"
+> & {
+  readonly model: Model
+  readonly options?: NoInfer<ImageModelOptions<Model>>
+  readonly http?: HttpOptions.Input
+} & (Model extends ImageModel<ImageModelOptions<Model>> ? unknown : never)
+
+export class GeneratedImage extends Schema.Class<GeneratedImage>("Image.Generated")({
+  mediaType: Schema.String,
+  data: Schema.Union([Schema.String, Schema.Uint8Array]),
+  providerMetadata: Schema.optional(ProviderMetadata),
+}) {}
+
+export class ImageResponse extends Schema.Class<ImageResponse>("Image.Response")({
+  images: Schema.Array(GeneratedImage),
+  usage: Schema.optional(Usage),
+  providerMetadata: Schema.optional(ProviderMetadata),
+}) {
+  get image() {
+    return this.images[0]
+  }
+}
+
+export function request<const Model extends object>(
+  input: ImageRequestInput<Model>,
+): ImageRequestFor<ImageModelOptions<Model>>
+export function request(input: ImageRequest): ImageRequest
+export function request(input: ImageRequest | ImageRequestInput) {
+  if (input instanceof ImageRequest) return input
+  return new ImageRequest({
+    ...input,
+    model: input.model as unknown as ImageModel,
+    http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+  })
+}
+
+export function generate<const Model extends object>(
+  input: ImageRequestInput<Model>,
+): Effect.Effect<ImageResponse, LLMError, Service>
+export function generate(input: ImageRequest): Effect.Effect<ImageResponse, LLMError, Service>
+export function generate(input: ImageRequest | ImageRequestInput) {
+  return Effect.try({
+    try: () => (input instanceof ImageRequest ? input : request(input)),
+    catch: (error) =>
+      new LLMError({
+        module: "Image",
+        method: "generate",
+        reason: new InvalidRequestReason({ message: error instanceof Error ? error.message : String(error) }),
+      }),
+  }).pipe(Effect.flatMap((request) => ImageClient.generate(request as unknown as ImageRequestFor<ImageOptions>)))
+}
+
+export const Image = {
+  request,
+  generate,
+} as const

packages/ai/src/index.ts

@@ -1,4 +1,5 @@
 export { LLMClient } from "./route/client"
+export { ImageClient } from "./image-client"
 export { Auth } from "./route/auth"
 export { Provider } from "./provider"
 export { ProviderPackage } from "./provider-package"
@@ -10,6 +11,9 @@ export type {
   Service as LLMClientService,
 } from "./route/client"
 export * from "./schema"
+export { GeneratedImage, ImageInput, ImageInputSchema, ImageModel, ImageRequest, ImageResponse } from "./image"
+export type { ImageModelOptions, ImageOptions, ImageRequestFor, ImageRequestInput, ImageRoute } from "./image"
+export { Image } from "./image"
 export { Tool, ToolFailure, toDefinitions } from "./tool"
 export { ToolRuntime } from "./tool-runtime"
 export type { DispatchResult as ToolDispatchResult, ToolSettlement } from "./tool-runtime"

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 => ({
+export const request = <const SelectedModel extends Model>(input: RequestInput<SelectedModel>) => {
-  ...LLMRequest.input(input),
-})
-
-export const request = (input: RequestInput) => {
   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>>(
+export function generateObject<const SelectedModel extends Model, S extends ToolSchema<any>>(
-  options: GenerateObjectOptions<S>,
+  options: GenerateObjectOptions<S, SelectedModel>,
 ): Effect.Effect<GenerateObjectResponse<Schema.Schema.Type<S>>, LLMError>
-export function generateObject(
+export function generateObject<const SelectedModel extends Model>(
-  options: GenerateObjectDynamicOptions,
+  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"
@@ -27,9 +30,33 @@ import { ToolSchemaProjection } from "./utils/tool-schema"
 import { ToolStream } from "./utils/tool-stream"
 
 const ADAPTER = "anthropic-messages"
+const MEDIA_MIMES = new Set<string>([...ProviderShared.IMAGE_MIMES, ...ProviderShared.PDF_MIMES])
 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
 // =============================================================================
@@ -56,6 +83,17 @@ const AnthropicImageBlock = Schema.Struct({
 })
 type AnthropicImageBlock = Schema.Schema.Type<typeof AnthropicImageBlock>
 
+const AnthropicDocumentBlock = Schema.Struct({
+  type: Schema.tag("document"),
+  source: Schema.Struct({
+    type: Schema.tag("base64"),
+    media_type: Schema.Literal("application/pdf"),
+    data: Schema.String,
+  }),
+  cache_control: Schema.optional(AnthropicCacheControl),
+})
+type AnthropicDocumentBlock = Schema.Schema.Type<typeof AnthropicDocumentBlock>
+
 const AnthropicThinkingBlock = Schema.Struct({
   type: Schema.tag("thinking"),
   thinking: Schema.String,
@@ -63,6 +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,
@@ -101,13 +148,10 @@ const AnthropicServerToolResultBlock = Schema.Struct({
 })
 type AnthropicServerToolResultBlock = Schema.Schema.Type<typeof AnthropicServerToolResultBlock>
 
-// Anthropic accepts either a plain string or an ordered array of text/image
+// Anthropic accepts either a plain string or an ordered array of text, image, and
-// blocks inside `tool_result.content`. The array form is required when a tool
+// document blocks inside `tool_result.content`. The array form keeps media as native
-// returns image bytes (screenshot, image search, etc.) so they can be passed
+// model input instead of JSON-stringifying base64 into prompt text.
-// to the model as proper image inputs instead of being JSON-stringified into
+const AnthropicToolResultContent = Schema.Union([AnthropicTextBlock, AnthropicImageBlock, AnthropicDocumentBlock])
-// the prompt — which silently inflates context by megabytes and can push the
-// conversation over the model's token limit.
-const AnthropicToolResultContent = Schema.Union([AnthropicTextBlock, AnthropicImageBlock])
 
 const AnthropicToolResultBlock = Schema.Struct({
   type: Schema.tag("tool_result"),
@@ -117,11 +161,17 @@ const AnthropicToolResultBlock = Schema.Struct({
   cache_control: Schema.optional(AnthropicCacheControl),
 })
 
-const AnthropicUserBlock = Schema.Union([AnthropicTextBlock, AnthropicImageBlock, AnthropicToolResultBlock])
+const AnthropicUserBlock = Schema.Union([
+  AnthropicTextBlock,
+  AnthropicImageBlock,
+  AnthropicDocumentBlock,
+  AnthropicToolResultBlock,
+])
 type AnthropicUserBlock = Schema.Schema.Type<typeof AnthropicUserBlock>
 const AnthropicAssistantBlock = Schema.Union([
   AnthropicTextBlock,
   AnthropicThinkingBlock,
+  AnthropicRedactedThinkingBlock,
   AnthropicToolUseBlock,
   AnthropicServerToolUseBlock,
   AnthropicServerToolResultBlock,
@@ -145,7 +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 }),
 ])
 
@@ -185,12 +235,27 @@ const AnthropicBodyFields = {
 export const AnthropicMessagesBody = Schema.Struct(AnthropicBodyFields)
 export type AnthropicMessagesBody = Schema.Schema.Type<typeof AnthropicMessagesBody>
 
-const AnthropicUsage = Schema.Struct({
+const AnthropicUsage = Schema.StructWithRest(
-  input_tokens: Schema.optional(Schema.Number),
+  Schema.Struct({
-  output_tokens: Schema.optional(Schema.Number),
+    input_tokens: Schema.optional(Schema.Number),
-  cache_creation_input_tokens: optionalNull(Schema.Number),
+    output_tokens: Schema.optional(Schema.Number),
-  cache_read_input_tokens: optionalNull(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({
@@ -200,6 +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
@@ -237,7 +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
 }
 
@@ -273,6 +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,
@@ -283,7 +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 }),
   })
@@ -316,15 +395,23 @@ 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 lowerImage = Effect.fn("AnthropicMessages.lowerImage")(function* (part: MediaPart) {
+const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (part: MediaPart) {
-  const media = yield* ProviderShared.validateMedia(
+  const media = yield* ProviderShared.validateMedia("Anthropic Messages", part, MEDIA_MIMES)
-    "Anthropic Messages",
+  if (media.mime === "application/pdf")
-    part,
+    return {
-    new Set<string>(ProviderShared.IMAGE_MIMES),
+      type: "document" as const,
-  )
+      source: {
+        type: "base64" as const,
+        media_type: "application/pdf" as const,
+        data: media.base64,
+      },
+    } satisfies AnthropicDocumentBlock
   return {
     type: "image" as const,
     source: {
@@ -335,25 +422,13 @@ const lowerImage = Effect.fn("AnthropicMessages.lowerImage")(function* (part: Me
   } satisfies AnthropicImageBlock
 })
 
-// Tool results may carry structured text/images. Keep media as provider-native
+// 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
-  const media = yield* ProviderShared.validateToolFile(
+  return yield* lowerMedia({ type: "media", mediaType: item.mime, data: item.uri, filename: item.name })
-    "Anthropic Messages",
-    item,
-    new Set<string>(ProviderShared.IMAGE_MIMES),
-  )
-  return {
-    type: "image" as const,
-    source: {
-      type: "base64" as const,
-      media_type: media.mime,
-      data: media.base64,
-    },
-  } satisfies AnthropicImageBlock
 })
 
 const lowerToolResultContent = Effect.fn("AnthropicMessages.lowerToolResultContent")(function* (part: ToolResultPart) {
@@ -361,7 +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)
 })
 
@@ -445,7 +520,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
           continue
         }
         if (part.type === "media") {
-          content.push(yield* lowerImage(part))
+          content.push(yield* lowerMedia(part))
           continue
         }
         return yield* ProviderShared.unsupportedContent("Anthropic Messages", "user", ["text", "media"])
@@ -462,11 +537,16 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
           continue
         }
         if (part.type === "reasoning") {
-          content.push({
+          // Mirrors Vercel's @ai-sdk/anthropic: a signature marks visible
-            type: "thinking",
+          // thinking; only signature-less parts carrying redactedData
-            thinking: part.text,
+          // round-trip as opaque redacted_thinking blocks.
-            signature: part.encrypted ?? signatureFromMetadata(part.providerMetadata),
+          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") {
@@ -503,39 +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 resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function* (input: unknown) {
-  const thinking = anthropicOptions(request)?.thinking
+  if (!ProviderShared.isRecord(input)) return undefined
-  if (!ProviderShared.isRecord(thinking)) return undefined
+  if (input.type === "adaptive") {
-  if (thinking.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 (input.type === "disabled") return { type: "disabled" as const }
-  if (thinking.type !== "enabled") return undefined
+  if (input.type !== "enabled") return undefined
   const budget =
-    typeof thinking.budgetTokens === "number"
+    typeof input.budgetTokens === "number"
-      ? thinking.budgetTokens
+      ? input.budgetTokens
-      : typeof thinking.budget_tokens === "number"
+      : typeof input.budget_tokens === "number"
-        ? thinking.budget_tokens
+        ? 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
@@ -544,7 +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(
@@ -553,6 +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
@@ -567,6 +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,
@@ -579,8 +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),
+    thinking: options.thinking,
-    output_config: outputConfig(request),
+    output_config: options.effort === undefined ? undefined : { effort: options.effort },
   }
 })
 
@@ -589,7 +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"
@@ -599,9 +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
+// thinking tokens are included in `output_tokens`; newer responses also
-// part of `output_tokens`, so `reasoningTokens` stays `undefined` and
+// expose that subset through `output_tokens_details.thinking_tokens`.
-// `outputTokens` carries the combined total.
 const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
   if (!usage) return undefined
   const nonCached = usage.input_tokens
@@ -614,6 +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 },
   })
@@ -632,18 +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: {
+      anthropic:
-        ...left.providerMetadata?.["anthropic"],
+        mergeJsonRecords(left.providerMetadata?.["anthropic"], right.providerMetadata?.["anthropic"]) ?? {},
-        ...right.providerMetadata?.["anthropic"],
-      },
     },
   })
 }
@@ -673,7 +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 }),
   })
 }
 
@@ -700,27 +785,68 @@ 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",
         }),
       },
-      [...events, LLMEvent.toolInputStart({ id: block.id ?? String(event.index), name: block.name ?? "" })],
+      [
+        ...events,
+        LLMEvent.toolInputStart({
+          id: block.id ?? String(event.index),
+          name: block.name ?? "",
+          providerExecuted: block.type === "server_tool_use" ? true : undefined,
+        }),
+      ],
     ]
   }
 
-  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,
     ]
@@ -758,18 +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(
+        reasoningSignatures: { ...state.reasoningSignatures, [index]: delta.signature },
-          state.lifecycle,
-          events,
-          `reasoning-${event.index ?? 0}`,
-          anthropicMetadata({ signature: delta.signature }),
-        ),
       },
-      events,
+      NO_EVENTS,
     ] satisfies StepResult
   }
 
@@ -800,28 +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),
+    reason: state.pendingFinish?.reason ?? {
-    usage,
+      normalized: "unknown",
-    providerMetadata: event.delta?.stop_sequence
+      raw: undefined,
-      ? anthropicMetadata({ stopSequence: event.delta.stop_sequence })
+    },
-      : 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
@@ -846,6 +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])
 }
@@ -866,7 +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,
@@ -52,6 +53,7 @@ const BedrockToolResultContentItem = Schema.Union([
   Schema.Struct({ text: Schema.String }),
   Schema.Struct({ json: Schema.Unknown }),
   BedrockMedia.ImageBlock,
+  BedrockMedia.DocumentBlock,
 ])
 
 const BedrockToolResultBlock = Schema.Struct({
@@ -64,14 +66,15 @@ const BedrockToolResultBlock = Schema.Struct({
 type BedrockToolResultBlock = Schema.Schema.Type<typeof BedrockToolResultBlock>
 
 const BedrockReasoningBlock = Schema.Struct({
-  reasoningContent: Schema.Struct({
+  reasoningContent: Schema.Union([
-    reasoningText: Schema.optional(
+    Schema.Struct({
-      Schema.Struct({
+      reasoningText: Schema.Struct({
         text: Schema.String,
         signature: Schema.optional(Schema.String),
       }),
-    ),
+    }),
-  }),
+    Schema.Struct({ redactedContent: Schema.String }),
+  ]),
 })
 
 const BedrockUserBlock = Schema.Union([
@@ -152,6 +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
@@ -179,6 +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),
             }),
           ),
         }),
@@ -198,11 +212,11 @@ const BedrockEvent = Schema.Struct({
       metrics: Schema.optional(Schema.Unknown),
     }),
   ),
-  internalServerException: Schema.optional(Schema.Struct({ message: Schema.String })),
+  internalServerException: Schema.optional(BedrockStreamException),
-  modelStreamErrorException: Schema.optional(Schema.Struct({ message: Schema.String })),
+  modelStreamErrorException: Schema.optional(BedrockStreamException),
-  validationException: Schema.optional(Schema.Struct({ message: Schema.String })),
+  validationException: Schema.optional(BedrockStreamException),
-  throttlingException: Schema.optional(Schema.Struct({ message: Schema.String })),
+  throttlingException: Schema.optional(BedrockStreamException),
-  serviceUnavailableException: Schema.optional(Schema.Struct({ message: Schema.String })),
+  serviceUnavailableException: Schema.optional(BedrockStreamException),
 })
 type BedrockEvent = Schema.Schema.Type<typeof BedrockEvent>
 
@@ -258,6 +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,
@@ -283,8 +304,6 @@ const lowerToolResultContent = Effect.fn("BedrockConverse.lowerToolResultContent
       data: item.uri,
       filename: item.name,
     })
-    if (!("image" in media))
-      return yield* ProviderShared.invalidRequest("Bedrock Converse only supports image media in tool results")
     content.push(media)
   }
   return content
@@ -349,11 +368,13 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
           continue
         }
         if (part.type === "reasoning") {
-          content.push({
+          const signature = reasoningSignature(part)
-            reasoningContent: {
+          const redactedData = reasoningRedactedData(part)
-              reasoningText: { text: part.text, signature: reasoningSignature(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") {
@@ -393,8 +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"
+    request.tools.length > 0
-      ? { tools: lowerTools(request.model.compatibility?.toolSchema, breakpoints, request.tools), toolChoice }
+      ? {
+          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)
@@ -431,27 +457,29 @@ 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"
 }
 
-// AWS Bedrock Converse reports `inputTokens` (inclusive total) with
+// AWS reports inputTokens separately from cache reads and writes.
-// `cacheReadInputTokens` and `cacheWriteInputTokens` as subsets. Pass
+// Bedrock does not break reasoning out of outputTokens for current models.
-// the total through and derive the non-cached breakdown. Bedrock does
-// not break reasoning out of `outputTokens` for any current model.
 const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
   if (!usage) return undefined
-  const cacheTotal = (usage.cacheReadInputTokens ?? 0) + (usage.cacheWriteInputTokens ?? 0)
+  const inputTokens = ProviderShared.sumTokens(
-  const nonCached = ProviderShared.subtractTokens(usage.inputTokens, cacheTotal)
+    usage.inputTokens,
+    usage.cacheReadInputTokens,
+    usage.cacheWriteInputTokens,
+  )
   return new Usage({
-    inputTokens: usage.inputTokens,
+    inputTokens,
     outputTokens: usage.outputTokens,
-    nonCachedInputTokens: nonCached,
+    nonCachedInputTokens: usage.inputTokens,
     cacheReadInputTokens: usage.cacheReadInputTokens,
     cacheWriteInputTokens: usage.cacheWriteInputTokens,
-    totalTokens: ProviderShared.totalTokens(usage.inputTokens, usage.outputTokens, usage.totalTokens),
+    totalTokens: ProviderShared.totalTokens(inputTokens, usage.outputTokens, usage.totalTokens),
     providerMetadata: { bedrock: usage },
   })
 }
@@ -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-${index}`,
+              reasoning.text ?? "",
+              providerMetadata,
+            )
+          : state.lifecycle
       return [
         {
           ...state,
-          lifecycle: reasoning.text
+          lifecycle,
-            ? Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text)
-            : state.lifecycle,
           reasoningSignatures: reasoning.signature
             ? { ...state.reasoningSignatures, [index]: reasoning.signature }
             : state.reasoningSignatures,
@@ -561,7 +603,9 @@ const step = (state: ParserState, event: BedrockEvent) =>
       return [
         {
           ...state,
-          hasToolCalls: resultEvents.some(LLMEvent.is.toolCall) ? true : state.hasToolCalls,
+          hasToolCalls:
+            resultEvents.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
+            state.hasToolCalls,
           lifecycle,
           tools: result.tools,
           reasoningSignatures: Object.fromEntries(
@@ -576,15 +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)
+      const usage = mapUsage(event.metadata.usage) ?? state.pendingFinish?.usage
-      return [{ ...state, pendingFinish: { reason: state.pendingFinish?.reason ?? "stop", usage } }, []] as const
+      return [
+        {
+          ...state,
+          pendingFinish: {
+            reason: state.pendingFinish?.reason ?? { normalized: "stop" },
+            usage,
+          },
+        },
+        [],
+      ] as const
     }
 
     const exception = (
@@ -601,7 +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],
         }),
       })
@@ -617,8 +676,13 @@ const onHalt = (state: ParserState): ReadonlyArray<LLMEvent> =>
     ? (() => {
         const events: LLMEvent[] = []
         Lifecycle.finish(state.lifecycle, events, {
-          reason:
+          reason: {
-            state.pendingFinish.reason === "stop" && state.hasToolCalls ? "tool-calls" : state.pendingFinish.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 messageType = decoded.headers[":message-type"]?.value
-      const eventType = decoded.headers[":event-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 +27,39 @@ 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 cachedContent?: string
+  readonly safetySettings?: ReadonlyArray<{
+    readonly category:
+      | "HARM_CATEGORY_UNSPECIFIED"
+      | "HARM_CATEGORY_HATE_SPEECH"
+      | "HARM_CATEGORY_DANGEROUS_CONTENT"
+      | "HARM_CATEGORY_HARASSMENT"
+      | "HARM_CATEGORY_SEXUALLY_EXPLICIT"
+      | "HARM_CATEGORY_CIVIC_INTEGRITY"
+      | (string & {})
+    readonly threshold:
+      | "HARM_BLOCK_THRESHOLD_UNSPECIFIED"
+      | "BLOCK_LOW_AND_ABOVE"
+      | "BLOCK_MEDIUM_AND_ABOVE"
+      | "BLOCK_ONLY_HIGH"
+      | "BLOCK_NONE"
+      | "OFF"
+      | (string & {})
+  }>
+  readonly serviceTier?: "standard" | "flex" | "priority" | (string & {})
+  readonly thinkingConfig?: {
+    readonly thinkingBudget?: number
+    readonly includeThoughts?: boolean
+    readonly thinkingLevel?: "minimal" | "low" | "medium" | "high" | (string & {})
+  }
+}
+
+export type ProviderOptionsInput = ProviderOptions & {
+  readonly gemini?: OptionsInput
+}
+
 // =============================================================================
 // Request Body Schema
 // =============================================================================
@@ -41,9 +75,11 @@ const GeminiInlineDataPart = Schema.Struct({
     data: Schema.String,
   }),
 })
+type GeminiInlineDataPart = Schema.Schema.Type<typeof GeminiInlineDataPart>
 
 const GeminiFunctionCallPart = Schema.Struct({
   functionCall: Schema.Struct({
+    id: Schema.optional(Schema.String),
     name: Schema.String,
     args: Schema.Unknown,
   }),
@@ -52,8 +88,10 @@ const GeminiFunctionCallPart = Schema.Struct({
 
 const GeminiFunctionResponsePart = Schema.Struct({
   functionResponse: Schema.Struct({
+    id: Schema.optional(Schema.String),
     name: Schema.String,
     response: Schema.Unknown,
+    parts: Schema.optional(Schema.Array(GeminiInlineDataPart)),
   }),
 })
 
@@ -94,6 +132,12 @@ const GeminiToolConfig = Schema.Struct({
 const GeminiThinkingConfig = Schema.Struct({
   thinkingBudget: Schema.optional(Schema.Number),
   includeThoughts: Schema.optional(Schema.Boolean),
+  thinkingLevel: Schema.optional(Schema.String),
+})
+
+const GeminiSafetySetting = Schema.Struct({
+  category: Schema.String,
+  threshold: Schema.String,
 })
 
 const GeminiGenerationConfig = Schema.Struct({
@@ -106,7 +150,10 @@ const GeminiGenerationConfig = Schema.Struct({
 })
 
 const GeminiBodyFields = {
+  cachedContent: Schema.optional(Schema.String),
   contents: Schema.Array(GeminiContent),
+  safetySettings: optionalArray(GeminiSafetySetting),
+  serviceTier: Schema.optional(Schema.String),
   systemInstruction: Schema.optional(GeminiSystemInstruction),
   tools: optionalArray(GeminiTool),
   toolConfig: Schema.optional(GeminiToolConfig),
@@ -197,8 +244,15 @@ const thoughtSignature = (providerMetadata: ProviderMetadata | undefined) => {
     : undefined
 }
 
+const functionCallId = (providerMetadata: ProviderMetadata | undefined) => {
+  const google = providerMetadata?.google
+  return ProviderShared.isRecord(google) && typeof google.functionCallId === "string"
+    ? google.functionCallId
+    : undefined
+}
+
 const lowerToolCall = (part: ToolCallPart) => ({
-  functionCall: { name: part.name, args: part.input },
+  functionCall: { id: functionCallId(part.providerMetadata), name: part.name, args: part.input },
   thoughtSignature: thoughtSignature(part.providerMetadata),
 })
 
@@ -255,6 +309,7 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
       if (part.result.type !== "content") {
         parts.push({
           functionResponse: {
+            id: functionCallId(part.providerMetadata),
             name: part.name,
             response: {
               name: part.name,
@@ -264,22 +319,25 @@ 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) {
+        if (item.type === "text") continue
+        const value = yield* ProviderShared.validateToolFile("Gemini", item, MEDIA_MIMES)
+        media.push({ inlineData: { mimeType: value.mime, data: value.base64 } })
+      }
       parts.push({
         functionResponse: {
+          id: functionCallId(part.providerMetadata),
           name: part.name,
           response: {
             name: part.name,
             content: text.join("\n"),
           },
+          parts: media.length > 0 ? media : undefined,
         },
       })
-      for (const item of content) {
-        if (item.type === "text") continue
-        const media = yield* ProviderShared.validateToolFile("Gemini", item, MEDIA_MIMES)
-        parts.push({ inlineData: { mimeType: media.mime, data: media.base64 } })
-      }
     }
     contents.push({ role: "user", parts })
   }
@@ -287,21 +345,43 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
   return contents
 })
 
-const geminiOptions = (request: LLMRequest) => request.providerOptions?.gemini
+const resolveOptions = (request: LLMRequest) => {
-
+  const input = request.providerOptions?.gemini
-const thinkingConfig = (request: LLMRequest) => {
+  const value = input?.thinkingConfig
-  const value = geminiOptions(request)?.thinkingConfig
+  const thinkingConfig = {
-  if (!ProviderShared.isRecord(value)) return undefined
+    thinkingBudget:
-  const result = {
+      ProviderShared.isRecord(value) && typeof value.thinkingBudget === "number" ? value.thinkingBudget : undefined,
-    thinkingBudget: typeof value.thinkingBudget === "number" ? value.thinkingBudget : undefined,
+    includeThoughts:
-    includeThoughts: typeof value.includeThoughts === "boolean" ? value.includeThoughts : undefined,
+      ProviderShared.isRecord(value) && typeof value.includeThoughts === "boolean"
+        ? value.includeThoughts
+        : ProviderShared.isRecord(value)
+          ? true
+          : undefined,
+    thinkingLevel:
+      ProviderShared.isRecord(value) && typeof value.thinkingLevel === "string" ? value.thinkingLevel : undefined,
   }
-  return Object.values(result).some((item) => item !== undefined) ? result : undefined
+  return {
+    cachedContent: typeof input?.cachedContent === "string" ? input.cachedContent : undefined,
+    safetySettings: mapSafetySettings(input?.safetySettings),
+    serviceTier: typeof input?.serviceTier === "string" ? input.serviceTier : undefined,
+    thinkingConfig: Object.values(thinkingConfig).some((item) => item !== undefined) ? thinkingConfig : undefined,
+  }
+}
+
+function mapSafetySettings(value: unknown) {
+  if (!Array.isArray(value)) return undefined
+  const settings = value.flatMap((item) =>
+    ProviderShared.isRecord(item) && typeof item.category === "string" && typeof item.threshold === "string"
+      ? [{ category: item.category, threshold: item.threshold }]
+      : [],
+  )
+  return settings
 }
 
 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,
@@ -309,14 +389,17 @@ 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 {
+    cachedContent: options.cachedContent,
     contents: yield* lowerMessages(request),
+    safetySettings: options.safetySettings,
+    serviceTier: options.serviceTier,
     systemInstruction:
       request.system.length === 0 ? undefined : { parts: [{ text: ProviderShared.joinText(request.system) }] },
-    tools: toolsEnabled
+    tools: hasTools
       ? [
           {
             functionDeclarations: request.tools.map((tool) =>
@@ -325,7 +408,7 @@ 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,
@@ -369,10 +452,22 @@ 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"
 }
 
@@ -389,7 +484,10 @@ 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
@@ -441,6 +539,10 @@ const step = (state: ParserState, event: GeminiEvent) => {
     if ("functionCall" in part) {
       const input = part.functionCall.args
       const id = `tool_${nextToolCallId++}`
+      const metadata = {
+        ...(part.functionCall.id === undefined ? {} : { functionCallId: part.functionCall.id }),
+        ...(part.thoughtSignature === undefined ? {} : { thoughtSignature: part.thoughtSignature }),
+      }
       lifecycle = Lifecycle.reasoningEnd(
         lifecycle,
         events,
@@ -453,9 +555,7 @@ const step = (state: ParserState, event: GeminiEvent) => {
           id,
           name: part.functionCall.name,
           input,
-          providerMetadata: part.thoughtSignature
+          providerMetadata: Object.keys(metadata).length > 0 ? googleMetadata(metadata) : undefined,
-            ? googleMetadata({ thoughtSignature: part.thoughtSignature })
-            : undefined,
         }),
       )
       hasToolCalls = true

packages/ai/src/protocols/google-images.ts

@@ -0,0 +1,314 @@
+import { Effect, Encoding, Schema } from "effect"
+import { Headers, HttpClientRequest } from "effect/unstable/http"
+import {
+  GeneratedImage,
+  ImageModel,
+  ImageResponse,
+  type ImageInput,
+  type ImageRequestFor,
+  type ImageRoute,
+} from "../image"
+import { Auth, type Definition as AuthDefinition } from "../route/auth"
+import {
+  InvalidProviderOutputReason,
+  LLMError,
+  Usage,
+  mergeHttpOptions,
+  mergeJsonRecords,
+  type HttpOptions,
+  type ProviderMetadata,
+} from "../schema"
+import { ProviderShared } from "./shared"
+import { ImageInputs } from "./utils/image-input"
+
+const ADAPTER = "google-images"
+export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
+
+export type GoogleImageString<Known extends string> = Known | (string & {})
+
+export type GoogleImageOptions = {
+  readonly aspectRatio?: GoogleImageString<
+    "1:1" | "2:3" | "3:2" | "3:4" | "4:3" | "4:5" | "5:4" | "9:16" | "16:9" | "21:9"
+  >
+  readonly imageSize?: GoogleImageString<"1K" | "2K" | "4K">
+  readonly seed?: number
+  readonly thinkingLevel?: GoogleImageString<"MINIMAL" | "LOW" | "MEDIUM" | "HIGH">
+  readonly includeThoughts?: boolean
+} & Record<string, unknown>
+
+export type GoogleImageBody = Record<string, unknown> & {
+  readonly contents: ReadonlyArray<{
+    readonly role: "user"
+    readonly parts: ReadonlyArray<Record<string, unknown>>
+  }>
+  readonly generationConfig: Record<string, unknown>
+}
+
+const GoogleUsage = Schema.StructWithRest(
+  Schema.Struct({
+    cachedContentTokenCount: Schema.optional(Schema.Number),
+    thoughtsTokenCount: Schema.optional(Schema.Number),
+    promptTokenCount: Schema.optional(Schema.Number),
+    candidatesTokenCount: Schema.optional(Schema.Number),
+    totalTokenCount: Schema.optional(Schema.Number),
+    promptTokensDetails: Schema.optional(Schema.Unknown),
+    candidatesTokensDetails: Schema.optional(Schema.Unknown),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const GoogleImageResponse = Schema.Struct({
+  candidates: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        index: Schema.optional(Schema.Number),
+        content: Schema.optional(
+          Schema.Struct({
+            parts: Schema.Array(
+              Schema.Struct({
+                text: Schema.optional(Schema.String),
+                thought: Schema.optional(Schema.Boolean),
+                thoughtSignature: Schema.optional(Schema.String),
+                inlineData: Schema.optional(
+                  Schema.Struct({
+                    mimeType: Schema.String,
+                    data: Schema.String,
+                  }),
+                ),
+              }),
+            ),
+          }),
+        ),
+        finishReason: Schema.optional(Schema.String),
+        finishMessage: Schema.optional(Schema.String),
+        safetyRatings: Schema.optional(Schema.Unknown),
+        citationMetadata: Schema.optional(Schema.Unknown),
+        groundingMetadata: Schema.optional(Schema.Unknown),
+      }),
+    ),
+  ),
+  usageMetadata: Schema.optional(GoogleUsage),
+  modelVersion: Schema.optional(Schema.String),
+  responseId: Schema.optional(Schema.String),
+  promptFeedback: Schema.optional(Schema.Unknown),
+})
+
+export interface ModelInput {
+  readonly id: string
+  readonly auth: AuthDefinition
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions
+}
+
+const nativeOptions = (options: GoogleImageOptions | undefined) => {
+  const { aspectRatio, imageSize, seed, thinkingLevel, includeThoughts, ...native } = options ?? {}
+  const image = {
+    aspectRatio,
+    imageSize,
+  }
+  const thinkingConfig = {
+    thinkingLevel,
+    includeThoughts,
+  }
+  return (
+    mergeJsonRecords(
+      {
+        responseModalities: ["IMAGE"],
+        imageConfig: Object.values(image).some((value) => value !== undefined) ? image : undefined,
+        seed,
+        thinkingConfig: Object.values(thinkingConfig).some((value) => value !== undefined) ? thinkingConfig : undefined,
+      },
+      native,
+    ) ?? { responseModalities: ["IMAGE"] }
+  )
+}
+
+const invalidOutput = (message: string, providerMetadata?: ProviderMetadata) =>
+  new LLMError({
+    module: ADAPTER,
+    method: "generate",
+    reason: new InvalidProviderOutputReason({ message, route: ADAPTER, providerMetadata }),
+  })
+
+const applyQuery = (url: string, query: Record<string, string> | undefined) => {
+  if (!query) return url
+  const next = new URL(url)
+  Object.entries(query).forEach(([key, value]) => next.searchParams.set(key, value))
+  return next.toString()
+}
+
+export const model = (input: ModelInput) => {
+  const route: ImageRoute<GoogleImageOptions> = {
+    id: ADAPTER,
+    generate: Effect.fn("GoogleImages.generate")(function* (request: ImageRequestFor<GoogleImageOptions>, execute) {
+      const imageParts = yield* Effect.forEach(request.images ?? [], googleImagePart)
+      const http = mergeHttpOptions(request.model.http, request.http)
+      const requestBody = mergeJsonRecords(
+        {
+          contents: [{ role: "user", parts: [{ text: request.prompt }, ...imageParts] }],
+          generationConfig: nativeOptions(request.options),
+        },
+        http?.body,
+      ) as GoogleImageBody
+      const text = ProviderShared.encodeJson(requestBody)
+      const url = applyQuery(
+        `${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}/models/${request.model.id}:generateContent`,
+        http?.query,
+      )
+      const headers = yield* Auth.toEffect(input.auth)({
+        request,
+        method: "POST",
+        url,
+        body: text,
+        headers: Headers.fromInput({ ...input.headers, ...http?.headers }),
+      })
+      const response = yield* execute(
+        HttpClientRequest.post(url).pipe(
+          HttpClientRequest.setHeaders(headers),
+          HttpClientRequest.bodyText(text, "application/json"),
+        ),
+      )
+      const payload = yield* response.json.pipe(
+        Effect.mapError(() => invalidOutput("Failed to read the Google Images response")),
+      )
+      const decoded = yield* Schema.decodeUnknownEffect(GoogleImageResponse)(payload).pipe(
+        Effect.mapError(() => invalidOutput("Google Images returned an invalid response")),
+      )
+      const candidates = decoded.candidates ?? []
+      const candidateMetadata = candidates.map((candidate, candidateIndex) => ({
+        index: candidate.index ?? candidateIndex,
+        finishReason: candidate.finishReason,
+        finishMessage: candidate.finishMessage,
+        safetyRatings: candidate.safetyRatings,
+        citationMetadata: candidate.citationMetadata,
+        groundingMetadata: candidate.groundingMetadata,
+        parts: (candidate.content?.parts ?? []).map((part) =>
+          part.inlineData === undefined
+            ? {
+                type: "text",
+                text: part.text,
+                thought: part.thought,
+                thoughtSignature: part.thoughtSignature,
+              }
+            : {
+                type: "inlineData",
+                mediaType: part.inlineData.mimeType,
+                thought: part.thought,
+                thoughtSignature: part.thoughtSignature,
+              },
+        ),
+      }))
+      const encoded = candidates.flatMap((candidate, candidateIndex) =>
+        (candidate.content?.parts ?? []).flatMap((part, partIndex) =>
+          part.inlineData === undefined || part.thought === true
+            ? []
+            : [{ candidate, candidateIndex, partIndex, inlineData: part.inlineData }],
+        ),
+      )
+      const images = yield* Effect.forEach(encoded, (item) =>
+        Effect.fromResult(Encoding.decodeBase64(item.inlineData.data)).pipe(
+          Effect.mapError(() =>
+            invalidOutput(
+              `Google Images candidate ${item.candidateIndex} part ${item.partIndex} contains invalid base64 data`,
+            ),
+          ),
+          Effect.map(
+            (data) =>
+              new GeneratedImage({
+                mediaType: item.inlineData.mimeType,
+                data,
+                providerMetadata: {
+                  google: {
+                    candidateIndex: item.candidate.index ?? item.candidateIndex,
+                    partIndex: item.partIndex,
+                    finishReason: item.candidate.finishReason,
+                    safetyRatings: item.candidate.safetyRatings,
+                    citationMetadata: item.candidate.citationMetadata,
+                    groundingMetadata: item.candidate.groundingMetadata,
+                    thoughtSignature: item.candidate.content?.parts[item.partIndex]?.thoughtSignature,
+                  },
+                },
+              }),
+          ),
+        ),
+      )
+      if (images.length === 0) {
+        const finishReasons = candidates.flatMap((candidate) =>
+          candidate.finishReason === undefined ? [] : [candidate.finishReason],
+        )
+        return yield* invalidOutput(
+          `Google Images returned no final images${
+            finishReasons.length === 0 ? "" : ` (finish reasons: ${finishReasons.join(", ")})`
+          }; inspect reason.providerMetadata.google for prompt feedback and candidate details`,
+          {
+            google: {
+              promptFeedback: decoded.promptFeedback,
+              candidates: candidateMetadata,
+            },
+          },
+        )
+      }
+      const usage = decoded.usageMetadata
+      const outputTokens =
+        usage?.candidatesTokenCount === undefined
+          ? undefined
+          : usage.candidatesTokenCount + (usage.thoughtsTokenCount ?? 0)
+      return new ImageResponse({
+        images,
+        usage:
+          usage === undefined
+            ? undefined
+            : new Usage({
+                inputTokens: usage.promptTokenCount,
+                outputTokens,
+                nonCachedInputTokens: ProviderShared.subtractTokens(
+                  usage.promptTokenCount,
+                  usage.cachedContentTokenCount,
+                ),
+                cacheReadInputTokens: usage.cachedContentTokenCount,
+                reasoningTokens: usage.thoughtsTokenCount,
+                totalTokens: ProviderShared.totalTokens(usage.promptTokenCount, outputTokens, usage.totalTokenCount),
+                providerMetadata: { google: usage },
+              }),
+        providerMetadata: {
+          google: {
+            modelVersion: decoded.modelVersion,
+            responseId: decoded.responseId,
+            promptFeedback: decoded.promptFeedback,
+            candidates: candidateMetadata,
+          },
+        },
+      })
+    }),
+  }
+  return ImageModel.make<GoogleImageOptions>({ id: input.id, provider: "google", route, http: input.http })
+}
+
+const googleImagePart = (image: ImageInput): Effect.Effect<Record<string, unknown>, LLMError> => {
+  if (image.type === "bytes")
+    return Effect.succeed({ inlineData: { mimeType: image.mediaType, data: Encoding.encodeBase64(image.data) } })
+  if (image.type === "file-uri") return Effect.succeed({ fileData: { mimeType: image.mediaType, fileUri: image.uri } })
+  if (image.type === "url")
+    return ImageInputs.decodeDataUrl(image.url, ADAPTER).pipe(
+      Effect.flatMap((decoded) => {
+        if (decoded === undefined)
+          return Effect.fail(
+            ImageInputs.invalid(
+              ADAPTER,
+              "Google generateContent does not fetch public image URLs; use bytes, a data URL, or a Gemini file URI",
+            ),
+          )
+        return Effect.succeed({
+          inlineData: { mimeType: decoded.mediaType, data: Encoding.encodeBase64(decoded.data) },
+        })
+      }),
+    )
+  return Effect.fail(
+    ImageInputs.invalid(ADAPTER, "Google generateContent requires Gemini file URIs rather than provider file IDs"),
+  )
+}
+
+export const GoogleImages = {
+  model,
+} as const

packages/ai/src/protocols/google-vertex-anthropic.ts

@@ -1,42 +0,0 @@
-import { Effect, Schema, Struct } from "effect"
-import { AnthropicMessages } from "./anthropic-messages"
-import { Auth } from "../route/auth"
-import { Route } from "../route/client"
-import { Endpoint } from "../route/endpoint"
-import { Framing } from "../route/framing"
-import { Protocol } from "../route/protocol"
-
-const VERSION = "vertex-2023-10-16" as const
-
-export const GoogleVertexAnthropicBody = Schema.Struct({
-  ...Struct.omit(AnthropicMessages.AnthropicMessagesBody.fields, ["model"]),
-  anthropic_version: Schema.Literal(VERSION),
-})
-export type GoogleVertexAnthropicBody = Schema.Schema.Type<typeof GoogleVertexAnthropicBody>
-
-export const protocol = Protocol.make({
-  id: "google-vertex-anthropic",
-  body: {
-    schema: GoogleVertexAnthropicBody,
-    from: (request) =>
-      AnthropicMessages.protocol.body.from(request).pipe(
-        Effect.map((body) => ({
-          ...Struct.omit(body, ["model"]),
-          anthropic_version: VERSION,
-        })),
-      ),
-  },
-  stream: AnthropicMessages.protocol.stream,
-})
-
-export const route = Route.make({
-  id: "google-vertex-anthropic",
-  provider: "google-vertex-anthropic",
-  providerMetadataKey: "anthropic",
-  protocol,
-  endpoint: Endpoint.path(({ request }) => `/${request.model.id}:streamRawPredict`),
-  auth: Auth.none,
-  framing: Framing.sse,
-})
-
-export * as GoogleVertexAnthropic from "./google-vertex-anthropic"

packages/ai/src/protocols/google-vertex-gemini.ts

@@ -1,20 +0,0 @@
-import { Gemini } from "./gemini"
-import { Auth } from "../route/auth"
-import { Route } from "../route/client"
-import { Endpoint } from "../route/endpoint"
-import { Framing } from "../route/framing"
-
-export const route = Route.make({
-  id: "google-vertex-gemini",
-  provider: "google-vertex",
-  providerMetadataKey: "google",
-  protocol: Gemini.protocol,
-  endpoint: Endpoint.path(({ request }) => {
-    const model = String(request.model.id)
-    return `/${model.startsWith("endpoints/") ? model : `models/${model}`}:streamGenerateContent?alt=sse`
-  }),
-  auth: Auth.none,
-  framing: Framing.sse,
-})
-
-export * as GoogleVertexGemini from "./google-vertex-gemini"

packages/ai/src/protocols/index.ts

@@ -1,9 +1,9 @@
 export * as AnthropicMessages from "./anthropic-messages"
-export * as GoogleVertexAnthropic from "./google-vertex-anthropic"
-export * as GoogleVertexGemini from "./google-vertex-gemini"
 export * as BedrockConverse from "./bedrock-converse"
 export * as Gemini from "./gemini"
 export * as OpenAIChat from "./openai-chat"
+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 +1,17 @@
 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 CacheHint,
   type JsonSchema,
   type LLMRequest,
   type MediaPart,
@@ -15,8 +19,8 @@ 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"
@@ -25,6 +29,7 @@ import { ToolStream } from "./utils/tool-stream"
 
 const ADAPTER = "openai-chat"
 const IMAGE_MIMES = new Set<string>(ProviderShared.IMAGE_MIMES)
+const RESERVED_REASONING_FIELDS = new Set(["role", "content", "tool_calls"])
 export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
 export const PATH = "/chat/completions"
 
@@ -34,6 +39,11 @@ export const PATH = "/chat/completions"
 // The body schema is the provider-native JSON body. `fromRequest` below builds
 // this shape from the common `LLMRequest`, then `Route.make` validates and
 // JSON-encodes it before transport.
+const OpenAIChatCacheControl = Schema.Struct({
+  type: Schema.Literal("ephemeral"),
+  ttl: Schema.optional(Schema.String),
+})
+
 const OpenAIChatFunction = Schema.Struct({
   name: Schema.String,
   description: Schema.String,
@@ -43,6 +53,7 @@ const OpenAIChatFunction = Schema.Struct({
 const OpenAIChatTool = Schema.Struct({
   type: Schema.tag("function"),
   function: OpenAIChatFunction,
+  cache_control: Schema.optional(OpenAIChatCacheControl),
 })
 type OpenAIChatTool = Schema.Schema.Type<typeof OpenAIChatTool>
 
@@ -57,7 +68,11 @@ const OpenAIChatAssistantToolCall = Schema.Struct({
 type OpenAIChatAssistantToolCall = Schema.Schema.Type<typeof OpenAIChatAssistantToolCall>
 
 const OpenAIChatUserContent = Schema.Union([
-  Schema.Struct({ type: Schema.Literal("text"), text: Schema.String }),
+  Schema.Struct({
+    type: Schema.Literal("text"),
+    text: Schema.String,
+    cache_control: Schema.optional(OpenAIChatCacheControl),
+  }),
   Schema.Struct({
     type: Schema.Literal("image_url"),
     image_url: Schema.Struct({ url: Schema.String }),
@@ -65,18 +80,33 @@ const OpenAIChatUserContent = Schema.Union([
 ])
 
 const OpenAIChatMessage = Schema.Union([
-  Schema.Struct({ role: Schema.Literal("system"), content: Schema.String }),
+  Schema.Struct({
+    role: Schema.Literal("system"),
+    content: Schema.Union([Schema.String, Schema.Array(OpenAIChatUserContent)]),
+  }),
   Schema.Struct({
     role: Schema.Literal("user"),
     content: Schema.Union([Schema.String, Schema.Array(OpenAIChatUserContent)]),
   }),
+  Schema.StructWithRest(
+    Schema.Struct({
+      role: Schema.Literal("assistant"),
+      content: Schema.NullOr(Schema.String),
+      tool_calls: optionalArray(OpenAIChatAssistantToolCall),
+      reasoning_content: Schema.optional(Schema.String),
+      reasoning: Schema.optional(Schema.String),
+      reasoning_text: Schema.optional(Schema.String),
+      reasoning_details: Schema.optional(Schema.Unknown),
+      cache_control: Schema.optional(OpenAIChatCacheControl),
+    }),
+    [Schema.Record(Schema.String, Schema.Unknown)],
+  ),
   Schema.Struct({
-    role: Schema.Literal("assistant"),
+    role: Schema.Literal("tool"),
-    content: Schema.NullOr(Schema.String),
+    tool_call_id: Schema.String,
-    tool_calls: optionalArray(OpenAIChatAssistantToolCall),
+    content: Schema.String,
-    reasoning_content: Schema.optional(Schema.String),
+    cache_control: Schema.optional(OpenAIChatCacheControl),
   }),
-  Schema.Struct({ role: Schema.Literal("tool"), tool_call_id: Schema.String, content: Schema.String }),
 ]).pipe(Schema.toTaggedUnion("role"))
 type OpenAIChatMessage = Schema.Schema.Type<typeof OpenAIChatMessage>
 
@@ -97,6 +127,7 @@ export const bodyFields = {
   stream_options: Schema.optional(Schema.Struct({ include_usage: Schema.Boolean })),
   store: Schema.optional(Schema.Boolean),
   reasoning_effort: Schema.optional(OpenAIOptions.OpenAIReasoningEffort),
+  max_completion_tokens: Schema.optional(Schema.Number),
   max_tokens: Schema.optional(Schema.Number),
   temperature: Schema.optional(Schema.Number),
   top_p: Schema.optional(Schema.Number),
@@ -121,6 +152,7 @@ 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(
@@ -142,30 +174,54 @@ const OpenAIChatToolCallDelta = Schema.Struct({
 })
 type OpenAIChatToolCallDelta = Schema.Schema.Type<typeof OpenAIChatToolCallDelta>
 
-const OpenAIChatDelta = Schema.Struct({
+const OpenAIChatDelta = Schema.StructWithRest(
-  content: optionalNull(Schema.String),
+  Schema.Struct({
-  reasoning_content: optionalNull(Schema.String),
+    content: optionalNull(Schema.String),
-  tool_calls: optionalNull(Schema.Array(OpenAIChatToolCallDelta)),
+    reasoning_content: optionalNull(Schema.String),
-})
+    reasoning: optionalNull(Schema.String),
+    reasoning_text: optionalNull(Schema.String),
+    reasoning_details: optionalNull(Schema.Unknown),
+    tool_calls: optionalNull(Schema.Array(OpenAIChatToolCallDelta)),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
 
 const OpenAIChatChoice = Schema.Struct({
   delta: optionalNull(OpenAIChatDelta),
   finish_reason: optionalNull(Schema.String),
+  native_finish_reason: optionalNull(Schema.String),
+})
+
+const 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]
 
+interface PendingToolDelta {
+  readonly id?: string
+  readonly name?: string
+  readonly input: string
+}
+
 export interface ParserState {
   readonly tools: ToolStream.State<number>
+  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>
+  readonly reasoningDetailsObserved: boolean
+  readonly reasoningEmitted: boolean
 }
 
 // =============================================================================
@@ -174,13 +230,20 @@ export interface ParserState {
 // Lowering is the only place that knows how common LLM messages map onto the
 // OpenAI Chat wire format. Keep provider quirks here instead of leaking native
 // fields into `LLMRequest`.
-const lowerTool = (tool: ToolDefinition, inputSchema: JsonSchema): OpenAIChatTool => ({
+interface LoweringOptions {
+  readonly cacheControl?: (
+    cache: CacheHint | undefined,
+  ) => Schema.Schema.Type<typeof OpenAIChatCacheControl> | undefined
+}
+
+const lowerTool = (tool: ToolDefinition, inputSchema: JsonSchema, options: LoweringOptions): OpenAIChatTool => ({
   type: "function",
   function: {
     name: tool.name,
     description: tool.description,
     parameters: ToolSchemaProjection.openAI(inputSchema),
   },
+  cache_control: options.cacheControl?.(tool.cache),
 })
 
 const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
@@ -208,11 +271,28 @@ const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart
 const openAICompatibleReasoningContent = (native: unknown) =>
   isRecord(native) && typeof native.reasoning_content === "string" ? native.reasoning_content : undefined
 
-const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (message: OpenAIChatRequestMessage) {
+const reasoningField = (part: ReasoningPart) => {
+  const field = part.providerMetadata?.openai?.reasoningField
+  return typeof field === "string" ? field : undefined
+}
+
+const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown) => {
+  const observed = parts.flatMap((part) => {
+    const details = part.providerMetadata?.openai?.reasoningDetails
+    return Array.isArray(details) ? details : []
+  })
+  if (parts.some((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))) return observed
+  if (isRecord(native) && Array.isArray(native.reasoning_details)) return native.reasoning_details
+}
+
+const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (
+  message: OpenAIChatRequestMessage,
+  options: LoweringOptions,
+) {
   const content: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
   for (const part of message.content) {
     if (part.type === "text") {
-      content.push({ type: "text", text: part.text })
+      content.push({ type: "text", text: part.text, cache_control: options.cacheControl?.(part.cache) })
       continue
     }
     if (part.type === "media") {
@@ -221,13 +301,18 @@ const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (mes
     }
     return yield* ProviderShared.unsupportedContent("OpenAI Chat", "user", ["text", "media"])
   }
-  if (content.every((part) => part.type === "text"))
+  if (content.every((part) => part.type === "text" && part.cache_control === undefined))
-    return { role: "user" as const, content: content.map((part) => part.text).join("") }
+    return {
+      role: "user" as const,
+      content: content.map((part) => (part.type === "text" ? part.text : "")).join(""),
+    }
   return { role: "user" as const, content }
 })
 
 const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(function* (
   message: OpenAIChatRequestMessage,
+  configuredField?: string,
+  options: LoweringOptions = {},
 ) {
   const content: TextPart[] = []
   const reasoning: ReasoningPart[] = []
@@ -248,30 +333,61 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
       continue
     }
   }
-  return {
+  const text = reasoning.map((part) => part.text).join("")
+  const details = reasoningDetails(reasoning, message.native?.openaiCompatible)
+  const observedField = reasoning.map(reasoningField).find((value) => value !== undefined)
+  const nativeReasoning = openAICompatibleReasoningContent(message.native?.openaiCompatible)
+  const fullyStructured = reasoning.every((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))
+  const field = (() => {
+    if (configuredField !== undefined) return configuredField
+    if (reasoning.length === 0) return undefined
+    if (observedField !== undefined) return observedField
+    if (nativeReasoning !== undefined) return "reasoning_content"
+    if (!fullyStructured) return "reasoning_content"
+  })()
+  const reasoningText = (() => {
+    if (configuredField !== undefined) return reasoning.length === 0 ? (nativeReasoning ?? "") : text
+    if (reasoning.length === 0) return nativeReasoning
+    return text
+  })()
+  const cached = message.content.findLast((part) => "cache" in part && part.cache !== undefined)
+  const result = {
     role: "assistant" as const,
     content: content.length === 0 ? null : ProviderShared.joinText(content),
     tool_calls: toolCalls.length === 0 ? undefined : toolCalls,
-    reasoning_content:
+    reasoning_details: details,
-      reasoning.length > 0
+    cache_control: options.cacheControl?.(cached && "cache" in cached ? cached.cache : undefined),
-        ? reasoning.map((part) => part.text).join("")
-        : openAICompatibleReasoningContent(message.native?.openaiCompatible),
   }
+  if (field === undefined || reasoningText === undefined) return result
+  return { ...result, [field]: reasoningText }
 })
 
-const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (message: OpenAIChatRequestMessage) {
+const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
+  message: OpenAIChatRequestMessage,
+  options: LoweringOptions,
+) {
   const messages: OpenAIChatMessage[] = []
   const images: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
   for (const part of message.content) {
     if (!ProviderShared.supportsContent(part, ["tool-result"]))
       return yield* ProviderShared.unsupportedContent("OpenAI Chat", "tool", ["tool-result"])
     if (part.result.type !== "content") {
-      messages.push({ role: "tool", tool_call_id: part.id, content: ProviderShared.toolResultText(part) })
+      messages.push({
+        role: "tool",
+        tool_call_id: part.id,
+        content: ProviderShared.toolResultText(part),
+        cache_control: options.cacheControl?.(part.cache),
+      })
       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") })
+    messages.push({
+      role: "tool",
+      tool_call_id: part.id,
+      content: text.join("\n"),
+      cache_control: options.cacheControl?.(part.cache),
+    })
     const files = content.filter((item) => item.type === "file")
     images.push(
       ...(yield* Effect.forEach(files, (item) =>
@@ -282,15 +398,32 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (m
   return { messages, images }
 })
 
-const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (message: OpenAIChatRequestMessage) {
+const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
-  if (message.role === "user") return [yield* lowerUserMessage(message)]
+  message: OpenAIChatRequestMessage,
-  if (message.role === "assistant") return [yield* lowerAssistantMessage(message)]
+  reasoningField?: string,
-  return (yield* lowerToolMessages(message)).messages
+  options: LoweringOptions = {},
+) {
+  if (message.role === "user") return [yield* lowerUserMessage(message, options)]
+  if (message.role === "assistant") return [yield* lowerAssistantMessage(message, reasoningField, options)]
+  return (yield* lowerToolMessages(message, options)).messages
 })
 
-const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request: LLMRequest) {
+const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request: LLMRequest, options: LoweringOptions) {
   const system: OpenAIChatMessage[] =
-    request.system.length === 0 ? [] : [{ role: "system", content: ProviderShared.joinText(request.system) }]
+    request.system.length === 0
+      ? []
+      : request.system.some((part) => part.cache !== undefined) && options.cacheControl !== undefined
+        ? [
+            {
+              role: "system",
+              content: request.system.map((part) => ({
+                type: "text",
+                text: part.text,
+                cache_control: options.cacheControl?.(part.cache),
+              })),
+            },
+          ]
+        : [{ role: "system", content: ProviderShared.joinText(request.system) }]
   const messages = [...system]
   const pendingImages: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
   const flushImages = () => {
@@ -301,67 +434,106 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
     if (message.role === "system") {
       const part = yield* ProviderShared.wrappedSystemUpdate("OpenAI Chat", message)
       if (pendingImages.length > 0) {
-        messages.push({ role: "user", content: [...pendingImages.splice(0), { type: "text", text: part.text }] })
+        messages.push({
+          role: "user",
+          content: [
+            ...pendingImages.splice(0),
+            { type: "text", text: part.text, cache_control: options.cacheControl?.(part.cache) },
+          ],
+        })
         continue
       }
       const previous = messages.at(-1)
       if (previous?.role === "user" && typeof previous.content === "string")
-        messages[messages.length - 1] = { role: "user", content: `${previous.content}\n${part.text}` }
+        messages[messages.length - 1] = options.cacheControl?.(part.cache)
+          ? {
+              role: "user",
+              content: [
+                { type: "text", text: previous.content },
+                { type: "text", text: part.text, cache_control: options.cacheControl(part.cache) },
+              ],
+            }
+          : { role: "user", content: `${previous.content}\n${part.text}` }
       else if (previous?.role === "user" && Array.isArray(previous.content))
         messages[messages.length - 1] = {
           role: "user",
-          content: [...previous.content, { type: "text", text: part.text }],
+          content: [
+            ...previous.content,
+            { type: "text", text: part.text, cache_control: options.cacheControl?.(part.cache) },
+          ],
         }
-      else messages.push({ role: "user", content: part.text })
+      else
+        messages.push(
+          options.cacheControl?.(part.cache)
+            ? {
+                role: "user",
+                content: [{ type: "text", text: part.text, cache_control: options.cacheControl(part.cache) }],
+              }
+            : { role: "user", content: part.text },
+        )
       continue
     }
     if (message.role === "tool") {
-      const lowered = yield* lowerToolMessages(message)
+      const lowered = yield* lowerToolMessages(message, options)
       messages.push(...lowered.messages)
       pendingImages.push(...lowered.images)
       continue
     }
     flushImages()
-    messages.push(...(yield* lowerMessage(message)))
+    messages.push(...(yield* lowerMessage(message, request.model.compatibility?.reasoningField, options)))
   }
   flushImages()
   return messages
 })
 
-const lowerOptions = Effect.fn("OpenAIChat.lowerOptions")(function* (request: LLMRequest) {
+const lowerOptions = (request: LLMRequest) => {
-  const store = OpenAIOptions.store(request)
+  const options = OpenAIOptions.resolve(request)
-  const reasoningEffort = OpenAIOptions.reasoningEffort(request)
   return {
-    ...(store !== undefined ? { store } : {}),
+    ...(options.store !== undefined ? { store: options.store } : {}),
-    ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),
+    ...(options.reasoningEffort ? { reasoning_effort: options.reasoningEffort } : {}),
   }
-})
+}
 
-const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (request: LLMRequest) {
+export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
+  request: LLMRequest,
+  options: LoweringOptions = {},
+) {
   // `fromRequest` returns the provider body only. Endpoint, auth, framing,
   // validation, and HTTP execution are composed by `Route.make`.
+  const reasoningField = request.model.compatibility?.reasoningField
+  if (reasoningField && RESERVED_REASONING_FIELDS.has(reasoningField))
+    return yield* ProviderShared.invalidRequest(
+      `OpenAI Chat reasoning field conflicts with reserved field ${reasoningField}`,
+    )
   const generation = request.generation
   const toolSchemaCompatibility = request.model.compatibility?.toolSchema
+  const maxTokensField = request.model.compatibility?.maxTokensField ?? "max_tokens"
   return {
     model: request.model.id,
-    messages: yield* lowerMessages(request),
+    messages: yield* lowerMessages(request, options),
     tools:
       request.tools.length === 0
         ? undefined
         : request.tools.map((tool) =>
-            lowerTool(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility)),
+            lowerTool(
+              tool,
+              ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility),
+              options,
+            ),
           ),
     tool_choice: request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined,
     stream: true as const,
     stream_options: { include_usage: true },
-    max_tokens: generation?.maxTokens,
+    ...(maxTokensField === "max_completion_tokens"
+      ? { max_completion_tokens: generation?.maxTokens }
+      : { max_tokens: generation?.maxTokens }),
     temperature: generation?.temperature,
     top_p: generation?.topP,
     frequency_penalty: generation?.frequencyPenalty,
     presence_penalty: generation?.presencePenalty,
     seed: generation?.seed,
     stop: generation?.stop,
-    ...(yield* lowerOptions(request)),
+    ...lowerOptions(request),
   }
 })
 
@@ -376,58 +548,171 @@ 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
+// cached-read and cache-write subsets, and `completion_tokens` (inclusive
-// a `reasoning_tokens` subset. We pass the inclusive totals through and
+// total) with a `reasoning_tokens` subset. We pass the inclusive totals
-// derive the non-cached breakdown so the `LLM.Usage` contract is
+// 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 },
   })
 }
 
+const reasoningDelta = (
+  delta: Schema.Schema.Type<typeof OpenAIChatDelta> | null | undefined,
+  configuredField?: string,
+) => {
+  if (!delta) return undefined
+  const fields = new Set([configuredField, "reasoning_content", "reasoning", "reasoning_text"])
+  for (const field of fields) {
+    if (field === undefined) continue
+    const text = delta[field]
+    if (typeof text === "string" && text.length > 0) return { field, text }
+  }
+  return undefined
+}
+
+const detailText = (details: ReadonlyArray<unknown>) => {
+  const text = details.flatMap((detail) => {
+    if (!isRecord(detail)) return []
+    if (detail.type === "reasoning.text" && typeof detail.text === "string" && detail.text) return [detail.text]
+    if (detail.type === "reasoning.summary" && typeof detail.summary === "string" && detail.summary)
+      return [detail.summary]
+    return []
+  })
+  if (text.length > 0) return text.join("")
+}
+
+const appendReasoningDetails = (result: Array<unknown>, details: ReadonlyArray<unknown>) => {
+  for (const detail of details) {
+    const previous = result.at(-1)
+    if (
+      !isRecord(previous) ||
+      previous.type !== "reasoning.text" ||
+      !isRecord(detail) ||
+      detail.type !== "reasoning.text" ||
+      conflictingReasoningTextDetails(previous, detail)
+    ) {
+      result.push(detail)
+      continue
+    }
+    result[result.length - 1] = {
+      ...previous,
+      ...Object.fromEntries(Object.entries(detail).filter((entry) => entry[1] !== undefined)),
+      text: `${typeof previous.text === "string" ? previous.text : ""}${typeof detail.text === "string" ? detail.text : ""}`,
+      signature: mergeDetailValue(previous.signature, detail.signature),
+      format: mergeDetailValue(previous.format, detail.format),
+    }
+  }
+}
+
+const mergeDetailValue = (previous: unknown, current: unknown) =>
+  previous || current || (previous !== undefined ? previous : current)
+
+const conflictingReasoningTextDetails = (previous: Record<string, unknown>, current: Record<string, unknown>) =>
+  conflictingDetailValue(previous.id, current.id) ||
+  conflictingDetailValue(previous.index, current.index) ||
+  conflictingDetailValue(previous.format, current.format) ||
+  (Boolean(previous.signature) && Boolean(current.signature) && previous.signature !== current.signature)
+
+const conflictingDetailValue = (previous: unknown, current: unknown) =>
+  previous !== undefined && previous !== null && current !== undefined && current !== null && previous !== current
+
+const reasoningMetadata = (field: ParserState["reasoningField"], details?: ReadonlyArray<unknown>) => ({
+  openai: {
+    ...(field ? { reasoningField: field } : {}),
+    ...(details ? { reasoningDetails: details } : {}),
+  },
+})
+
 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 choice = event.choices?.[0]
-    const finishReason = choice?.finish_reason ? mapFinishReason(choice.finish_reason) : state.finishReason
+    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
+    let pendingTools = state.pendingTools
 
     let lifecycle = state.lifecycle
 
-    if (delta?.reasoning_content)
+    const reasoning = reasoningDelta(delta, state.reasoningField)
-      lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", delta.reasoning_content)
+    const reasoningField = state.reasoningField ?? (!state.lifecycle.text.has("text-0") ? reasoning?.field : undefined)
+    const detailDelta = Array.isArray(delta?.reasoning_details) ? delta.reasoning_details : undefined
+    if (detailDelta !== undefined) appendReasoningDetails(state.reasoningDetails, detailDelta)
+    const reasoningDetailsObserved = state.reasoningDetailsObserved || detailDelta !== undefined
+    const deltaMetadata = reasoningMetadata(reasoningField)
+    const text = detailDelta?.length ? (detailText(detailDelta) ?? reasoning?.text) : reasoning?.text
+    if (!state.lifecycle.text.has("text-0") && text !== undefined)
+      lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", text, deltaMetadata)
+    else if (
+      reasoningDetailsObserved &&
+      !lifecycle.reasoning.has("reasoning-0") &&
+      (Boolean(delta?.content) || toolDeltas.length > 0)
+    )
+      lifecycle = Lifecycle.reasoningStart(lifecycle, events, "reasoning-0", deltaMetadata)
+    const reasoningEmitted = state.reasoningEmitted || lifecycle.reasoning.has("reasoning-0")
 
     if (delta?.content) {
-      lifecycle = Lifecycle.reasoningEnd(lifecycle, events, "reasoning-0")
+      lifecycle = Lifecycle.reasoningEnd(
+        lifecycle,
+        events,
+        "reasoning-0",
+        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
+      )
       lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)
     }
 
-    if (toolDeltas.length) lifecycle = Lifecycle.reasoningEnd(lifecycle, events, "reasoning-0")
-
     for (const tool of toolDeltas) {
+      const current = tools[tool.index]
+      const pending = pendingTools[tool.index]
+      const id = current?.id ?? pending?.id ?? (tool.id || undefined)
+      const name = current?.name ?? pending?.name ?? (tool.function?.name || undefined)
+      const text = `${pending?.input ?? ""}${tool.function?.arguments ?? ""}`
+      if (!current && (!id || !name)) {
+        pendingTools = { ...pendingTools, [tool.index]: { id: id || undefined, name: name || undefined, input: text } }
+        continue
+      }
+      if (pending) {
+        pendingTools = { ...pendingTools }
+        delete pendingTools[tool.index]
+      }
       const result = ToolStream.appendOrStart(
         ADAPTER,
         tools,
         tool.index,
-        { id: tool.id ?? undefined, name: tool.function?.name ?? undefined, text: tool.function?.arguments ?? "" },
+        { id: id || undefined, name: name || undefined, text },
         "OpenAI Chat tool call delta is missing id or name",
       )
       if (ToolStream.isError(result)) return yield* result
@@ -436,8 +721,11 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
       events.push(...result.events)
     }
 
+    if (finishReason !== undefined && state.finishReason === undefined && Object.keys(pendingTools).length > 0)
+      return yield* ProviderShared.eventError(ADAPTER, "OpenAI Chat tool call delta is missing id or name")
+
     // Finalize accumulated tool inputs eagerly when finish_reason arrives so
-    // JSON parse failures fail the stream at the boundary rather than at halt.
+    // valid calls and malformed local calls settle independently.
     const finished =
       finishReason !== undefined && state.finishReason === undefined && Object.keys(tools).length > 0
         ? yield* ToolStream.finishAll(ADAPTER, tools)
@@ -446,10 +734,15 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
     return [
       {
         tools: finished?.tools ?? tools,
+        pendingTools,
         toolCallEvents: finished?.events ?? state.toolCallEvents,
         usage,
         finishReason,
         lifecycle,
+        reasoningField,
+        reasoningDetails: state.reasoningDetails,
+        reasoningDetailsObserved,
+        reasoningEmitted,
       },
       events,
     ] as const
@@ -458,8 +751,23 @@ 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
-  const lifecycle = state.toolCallEvents.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
+    ? {
+        ...state.finishReason,
+        normalized:
+          state.finishReason.normalized === "stop" && hasToolCalls ? "tool-calls" : state.finishReason.normalized,
+      }
+    : undefined
+  const metadata = reasoningMetadata(
+    state.reasoningField,
+    state.reasoningDetailsObserved ? state.reasoningDetails : undefined,
+  )
+  const started =
+    state.reasoningDetailsObserved && !state.reasoningEmitted
+      ? Lifecycle.reasoningStart(state.lifecycle, events, "reasoning-0", reasoningMetadata(state.reasoningField))
+      : state.lifecycle
+  const ended = Lifecycle.reasoningEnd(started, events, "reasoning-0", metadata)
+  const lifecycle = state.toolCallEvents.length ? Lifecycle.stepStart(ended, events) : ended
   events.push(...state.toolCallEvents)
   if (reason) Lifecycle.finish(lifecycle, events, { reason, usage: state.usage })
   return events
@@ -482,7 +790,16 @@ export const protocol = Protocol.make({
   },
   stream: {
     event: Protocol.jsonEvent(OpenAIChatEvent),
-    initial: () => ({ tools: ToolStream.empty<number>(), toolCallEvents: [], lifecycle: Lifecycle.initial() }),
+    initial: (request) => ({
+      tools: ToolStream.empty<number>(),
+      pendingTools: {},
+      toolCallEvents: [],
+      lifecycle: Lifecycle.initial(),
+      reasoningField: request.model.compatibility?.reasoningField,
+      reasoningDetails: [],
+      reasoningDetailsObserved: false,
+      reasoningEmitted: false,
+    }),
     step,
     onHalt: finishEvents,
   },

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

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

@@ -0,0 +1,270 @@
+import { Effect, Encoding, Schema } from "effect"
+import { Headers, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
+import {
+  ImageModel,
+  GeneratedImage,
+  ImageResponse,
+  type ImageInput,
+  type ImageRequestFor,
+  type ImageRoute,
+} from "../image"
+import { Auth, type Definition as AuthDefinition } from "../route/auth"
+import {
+  InvalidProviderOutputReason,
+  LLMError,
+  Usage,
+  mergeHttpOptions,
+  mergeJsonRecords,
+  type HttpOptions,
+} from "../schema"
+import { ProviderShared } from "./shared"
+import { ImageInputs } from "./utils/image-input"
+import { OpenAIImage } from "./utils/openai-image"
+
+const ADAPTER = "openai-images"
+export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
+export const PATH = "/images/generations"
+export const EDIT_PATH = "/images/edits"
+
+export type OpenAIImageString<Known extends string> = Known | (string & {})
+
+export type OpenAIImageOptions = {
+  readonly mask?: ImageInput
+  readonly n?: number
+  readonly size?: OpenAIImageString<
+    "auto" | "256x256" | "512x512" | "1024x1024" | "1536x1024" | "1024x1536" | "1792x1024" | "1024x1792"
+  >
+  readonly quality?: OpenAIImageString<"auto" | "low" | "medium" | "high" | "standard" | "hd">
+  readonly background?: OpenAIImageString<"auto" | "opaque" | "transparent">
+  readonly moderation?: OpenAIImageString<"auto" | "low">
+  readonly outputFormat?: OpenAIImageString<"png" | "jpeg" | "webp">
+  readonly outputCompression?: number
+} & Record<string, unknown>
+
+export type OpenAIImageBody = Record<string, unknown> & {
+  readonly model: string
+  readonly prompt: string
+}
+
+const OpenAIImageResponse = Schema.Struct({
+  data: Schema.Array(
+    Schema.Struct({
+      b64_json: Schema.optional(Schema.String),
+      url: Schema.optional(Schema.String),
+      revised_prompt: Schema.optional(Schema.String),
+    }),
+  ),
+  output_format: Schema.optional(Schema.String),
+  usage: Schema.optional(
+    Schema.Struct({
+      input_tokens: Schema.optional(Schema.Number),
+      output_tokens: Schema.optional(Schema.Number),
+      total_tokens: Schema.optional(Schema.Number),
+      input_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+      output_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+    }),
+  ),
+})
+
+export interface ModelInput {
+  readonly id: string
+  readonly auth: AuthDefinition
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions
+}
+
+const nativeOptions = (options: OpenAIImageOptions | undefined) => {
+  if (!options) return undefined
+  const { mask: _, outputFormat, outputCompression, ...native } = options
+  return {
+    output_format: outputFormat,
+    output_compression: outputCompression,
+    ...native,
+  }
+}
+
+const invalidOutput = (message: string) =>
+  new LLMError({
+    module: ADAPTER,
+    method: "generate",
+    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
+  })
+
+const applyQuery = (url: string, query: Record<string, string> | undefined) => {
+  if (!query) return url
+  const next = new URL(url)
+  Object.entries(query).forEach(([key, value]) => next.searchParams.set(key, value))
+  return next.toString()
+}
+
+export const model = (input: ModelInput) => {
+  const route: ImageRoute<OpenAIImageOptions> = {
+    id: ADAPTER,
+    generate: Effect.fn("OpenAIImages.generate")(function* (request: ImageRequestFor<OpenAIImageOptions>, execute) {
+      const mask = request.options?.mask
+      if (mask !== undefined && (request.images?.length ?? 0) === 0)
+        return yield* ImageInputs.invalid(ADAPTER, "An OpenAI image mask requires at least one input image")
+      const http = mergeHttpOptions(request.model.http, request.http)
+      const sourceImages = request.images ?? []
+      const multipartImages = yield* Effect.forEach(sourceImages, (image) => {
+        if (image.type === "bytes") return Effect.succeed({ data: image.data, mediaType: image.mediaType })
+        if (image.type === "url") return ImageInputs.decodeDataUrl(image.url, ADAPTER)
+        return Effect.succeed(undefined)
+      })
+      const multipartMask =
+        mask === undefined
+          ? undefined
+          : mask.type === "bytes"
+            ? { data: mask.data, mediaType: mask.mediaType }
+            : mask.type === "url"
+              ? yield* ImageInputs.decodeDataUrl(mask.url, ADAPTER)
+              : undefined
+      const useMultipart =
+        sourceImages.length > 0 &&
+        multipartImages.every((image) => image !== undefined) &&
+        (mask === undefined || multipartMask !== undefined)
+      const path = sourceImages.length === 0 ? PATH : EDIT_PATH
+      const url = applyQuery(`${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}${path}`, http?.query)
+
+      if (useMultipart) {
+        const form = new FormData()
+        form.append("model", request.model.id)
+        form.append("prompt", request.prompt)
+        Object.entries(mergeJsonRecords(nativeOptions(request.options), http?.body) ?? {}).forEach(([key, value]) => {
+          if (["model", "prompt", "image", "image[]", "images", "mask"].includes(key)) return
+          form.append(key, typeof value === "string" ? value : ProviderShared.encodeJson(value))
+        })
+        multipartImages.forEach((image, index) => {
+          if (image === undefined) return
+          form.append("image[]", imageBlob(image.data, image.mediaType), `image-${index}`)
+        })
+        if (multipartMask !== undefined)
+          form.append("mask", imageBlob(multipartMask.data, multipartMask.mediaType), "mask")
+        const headers = yield* Auth.toEffect(input.auth)({
+          request,
+          method: "POST",
+          url,
+          body: "[multipart/form-data]",
+          headers: Headers.remove(Headers.fromInput({ ...input.headers, ...http?.headers }), "content-type"),
+        })
+        const response = yield* execute(
+          HttpClientRequest.post(url).pipe(HttpClientRequest.setHeaders(headers), HttpClientRequest.bodyFormData(form)),
+        )
+        return yield* parseResponse(response, request.options, http?.body)
+      }
+
+      const references = sourceImages.map((image) => {
+        if (image.type === "bytes") return { image_url: ImageInputs.dataUrl(image) }
+        if (image.type === "url") return { image_url: image.url }
+        if (image.type === "file-id") return { file_id: image.id }
+        return undefined
+      })
+      if (references.some((image) => image === undefined))
+        return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts image URLs, data URLs, bytes, and file IDs")
+      const maskReference =
+        mask === undefined
+          ? undefined
+          : mask.type === "bytes"
+            ? { image_url: ImageInputs.dataUrl(mask) }
+            : mask.type === "url"
+              ? { image_url: mask.url }
+              : mask.type === "file-id"
+                ? { file_id: mask.id }
+                : undefined
+      if (mask !== undefined && maskReference === undefined)
+        return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts masks as URLs, data URLs, bytes, or file IDs")
+      const requestBody = mergeJsonRecords(
+        {
+          model: request.model.id,
+          prompt: request.prompt,
+          images: references.length === 0 ? undefined : references,
+          mask: maskReference,
+        },
+        nativeOptions(request.options),
+        http?.body,
+      ) as OpenAIImageBody
+      const text = ProviderShared.encodeJson(requestBody)
+      const headers = yield* Auth.toEffect(input.auth)({
+        request,
+        method: "POST",
+        url,
+        body: text,
+        headers: Headers.fromInput({ ...input.headers, ...http?.headers }),
+      })
+      const response = yield* execute(
+        HttpClientRequest.post(url).pipe(
+          HttpClientRequest.setHeaders(headers),
+          HttpClientRequest.bodyText(text, "application/json"),
+        ),
+      )
+      return yield* parseResponse(response, request.options, http?.body)
+    }),
+  }
+  return ImageModel.make<OpenAIImageOptions>({ id: input.id, provider: "openai", route, http: input.http })
+}
+
+const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  options: OpenAIImageOptions | undefined,
+  overlay: Record<string, unknown> | undefined,
+) {
+  const payload = yield* response.json.pipe(
+    Effect.mapError(() => invalidOutput("Failed to read the OpenAI Images response")),
+  )
+  const decoded = yield* Schema.decodeUnknownEffect(OpenAIImageResponse)(payload).pipe(
+    Effect.mapError(() => invalidOutput("OpenAI Images returned an invalid response")),
+  )
+  const requestBody = mergeJsonRecords(nativeOptions(options), overlay)
+  const format =
+    decoded.output_format ?? (typeof requestBody?.output_format === "string" ? requestBody.output_format : "png")
+  const images = yield* Effect.forEach(decoded.data, (item, index) => {
+    if (item.b64_json)
+      return Effect.fromResult(Encoding.decodeBase64(item.b64_json)).pipe(
+        Effect.mapError(() => invalidOutput(`OpenAI Images result ${index} contains invalid base64 data`)),
+        Effect.map(
+          (data) =>
+            new GeneratedImage({
+              mediaType: `image/${format}`,
+              data,
+              providerMetadata:
+                item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
+            }),
+        ),
+      )
+    if (item.url)
+      return Effect.succeed(
+        new GeneratedImage({
+          mediaType: `image/${format}`,
+          data: item.url,
+          providerMetadata:
+            item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
+        }),
+      )
+    return Effect.fail(invalidOutput(`OpenAI Images result ${index} has neither image data nor a URL`))
+  })
+  if (images.length === 0) return yield* invalidOutput("OpenAI Images returned no images")
+  return new ImageResponse({
+    images,
+    usage:
+      decoded.usage === undefined
+        ? undefined
+        : new Usage({
+            inputTokens: decoded.usage.input_tokens,
+            outputTokens: decoded.usage.output_tokens,
+            totalTokens: decoded.usage.total_tokens,
+            providerMetadata: { openai: decoded.usage },
+          }),
+    providerMetadata: { openai: { outputFormat: format } },
+  })
+})
+
+const imageBlob = (data: Uint8Array, mediaType: string) => {
+  const buffer = new ArrayBuffer(data.byteLength)
+  new Uint8Array(buffer).set(data)
+  return new Blob([buffer], { type: mediaType })
+}
+
+export const OpenAIImages = {
+  model,
+} as const

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"
@@ -158,7 +158,8 @@ export const parseToolInput = (route: string, name: string, raw: string) =>
 export const IMAGE_MIMES = ["image/png", "image/jpeg", "image/gif", "image/webp"] as const
 export const VIDEO_MIMES = ["video/mp4", "video/webm", "video/quicktime"] as const
 export const AUDIO_MIMES = ["audio/wav", "audio/mp3", "audio/aiff", "audio/aac", "audio/ogg", "audio/flac"] as const
-export const MEDIA_MIMES = [...IMAGE_MIMES, ...VIDEO_MIMES, ...AUDIO_MIMES] as const
+export const PDF_MIMES = ["application/pdf"] as const
+export const MEDIA_MIMES = [...IMAGE_MIMES, ...VIDEO_MIMES, ...AUDIO_MIMES, ...PDF_MIMES] as const
 export const MAX_MEDIA_ENCODED_BYTES = 28 * 1024 * 1024
 export const MAX_MEDIA_DECODED_BYTES = 20 * 1024 * 1024
 
@@ -205,7 +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/bedrock-media.ts

@@ -49,10 +49,10 @@ const DOCUMENT_FORMATS = {
   "text/markdown": "md",
 } as const satisfies Record<string, DocumentFormat>
 
-const documentBlock = (part: MediaPart, format: DocumentFormat, bytes: string): DocumentBlock => ({
+const documentBlock = (name: string, format: DocumentFormat, bytes: string): DocumentBlock => ({
   document: {
     format,
-    name: part.filename ?? `document.${format}`,
+    name,
     source: { bytes },
   },
 })
@@ -77,12 +77,14 @@ export const lower = Effect.fn("BedrockMedia.lower")(function* (part: MediaPart)
     return yield* ProviderShared.invalidRequest(`Bedrock Converse does not support image media type ${part.mediaType}`)
   const documentFormat = DOCUMENT_FORMATS[mime as keyof typeof DOCUMENT_FORMATS]
   if (documentFormat) {
+    if (!part.filename)
+      return yield* ProviderShared.invalidRequest("Bedrock Converse document media requires a filename")
     const media = yield* ProviderShared.validateMedia(
       "Bedrock Converse",
       part,
       new Set<string>(Object.keys(DOCUMENT_FORMATS)),
     )
-    return documentBlock(part, documentFormat, media.base64)
+    return documentBlock(part.filename, documentFormat, media.base64)
   }
   return yield* ProviderShared.invalidRequest(`Bedrock Converse does not support media type ${part.mediaType}`)
 })

packages/ai/src/protocols/utils/image-input.ts

@@ -0,0 +1,34 @@
+import { Effect, Encoding } from "effect"
+import type { ImageInput } from "../../image"
+import { InvalidRequestReason, LLMError } from "../../schema"
+
+const invalid = (module: string, message: string) =>
+  new LLMError({
+    module,
+    method: "generate",
+    reason: new InvalidRequestReason({ message }),
+  })
+
+export const dataUrl = (input: Extract<ImageInput, { readonly type: "bytes" }>) =>
+  `data:${input.mediaType};base64,${Encoding.encodeBase64(input.data)}`
+
+export const decodeDataUrl = (
+  url: string,
+  module: string,
+): Effect.Effect<{ readonly mediaType: string; readonly data: Uint8Array } | undefined, LLMError> => {
+  if (!url.startsWith("data:")) return Effect.succeed(undefined)
+  const match = /^data:([^;,]+);base64,(.*)$/s.exec(url)
+  if (!match) return Effect.fail(invalid(module, "Image data URLs must contain a MIME type and base64 data"))
+  return Effect.fromResult(Encoding.decodeBase64(match[2])).pipe(
+    Effect.mapError(() => invalid(module, "Image data URL contains invalid base64 data")),
+    Effect.map((data) => ({ mediaType: match[1], data })),
+  )
+}
+
+export const invalidImageInput = invalid
+
+export const ImageInputs = {
+  dataUrl,
+  decodeDataUrl,
+  invalid: invalidImageInput,
+} as const

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.textStart({ id, providerMetadata }))
-    events.push(LLMEvent.textDelta({ id, text }))
-    return stepped
-  }
-  events.push(LLMEvent.textStart({ id }), LLMEvent.textDelta({ id, text }))
   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[],
@@ -44,7 +47,7 @@ export const reasoningDelta = (
   providerMetadata?: ProviderMetadata,
 ): State => {
   const started = reasoningStart(state, events, id, providerMetadata)
-  events.push(LLMEvent.reasoningDelta({ id, text }))
+  events.push(LLMEvent.reasoningDelta({ id, text, providerMetadata }))
   return started
 }
 
@@ -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-image.ts

@@ -0,0 +1,20 @@
+import { Schema } from "effect"
+
+const dimensions = (value: string) => {
+  const match = /^(\d+)x(\d+)$/.exec(value)
+  if (!match) return undefined
+  return { width: Number(match[1]), height: Number(match[2]) }
+}
+
+export const Size = Schema.String.check(
+  Schema.makeFilter((value) => {
+    if (value === "auto") return undefined
+    const parsed = dimensions(value)
+    if (!parsed) return "image size must be `auto` or `{width}x{height}`"
+    return parsed.width > 0 && parsed.height > 0 ? undefined : "image dimensions must be positive integers"
+  }),
+)
+
+export const OpenAIImage = {
+  Size,
+} as const

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

@@ -1,85 +1,23 @@
-import { Schema } from "effect"
+import { ReasoningEfforts } from "../../schema"
-import type { LLMRequest, TextVerbosity as TextVerbosityValue } from "../../schema"
+import { OpenResponsesOptions } from "./open-responses-options"
-import { ReasoningEfforts, TextVerbosity } from "../../schema"
 
 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 = [
+export const OpenAIResponseIncludables = OpenResponsesOptions.ResponseIncludables
-  "file_search_call.results",
+export type OpenAIResponseIncludable = OpenResponsesOptions.ResponseIncludable
-  "web_search_call.results",
+export const OpenAIServiceTiers = OpenResponsesOptions.ServiceTiers
-  "web_search_call.action.sources",
+export type OpenAIServiceTier = OpenResponsesOptions.ServiceTier
-  "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]
 
-const TEXT_VERBOSITY = new Set<string>(["low", "medium", "high"])
+export const OpenAIReasoningEffort = OpenResponsesOptions.ReasoningEffort
-const INCLUDABLES = new Set<string>(OpenAIResponseIncludables)
+export const OpenAITextVerbosity = OpenResponsesOptions.TextVerbositySchema
-const SERVICE_TIERS = new Set<string>(OpenAIServiceTiers)
+export const OpenAIResponseIncludable = OpenResponsesOptions.ResponseIncludableSchema
-
+export const OpenAIServiceTier = OpenResponsesOptions.ServiceTierSchema
-export const OpenAIReasoningEffort = Schema.String
-export const OpenAITextVerbosity = TextVerbosity
-export const OpenAIResponseIncludable = Schema.Literals(OpenAIResponseIncludables)
-export const OpenAIServiceTier = Schema.Literals(OpenAIServiceTiers)
 
 export const isReasoningEffort = (effort: unknown): effort is OpenAIReasoningEffort => typeof effort === "string"
 
-const isTextVerbosity = (value: unknown): value is TextVerbosityValue =>
+export const resolve = OpenResponsesOptions.resolve
-  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 * 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/protocols/utils/tool-stream.ts

@@ -1,5 +1,5 @@
 import { Effect } from "effect"
-import { LLMError, LLMEvent, type ProviderMetadata, type ToolCall } from "../../schema"
+import { LLMError, LLMEvent, type ProviderMetadata, type ToolCall, type ToolInputError } from "../../schema"
 import { eventError, parseToolInput, type ToolAccumulator } from "../shared"
 
 type StreamKey = string | number
@@ -53,6 +53,7 @@ const inputStart = (tool: PendingTool) =>
   LLMEvent.toolInputStart({
     id: tool.id,
     name: tool.name,
+    providerExecuted: tool.providerExecuted ? true : undefined,
     providerMetadata: tool.providerMetadata,
   })
 
@@ -63,19 +64,36 @@ const inputDelta = (tool: PendingTool, text: string) =>
     text,
   })
 
-const toolCall = (route: string, tool: PendingTool, inputOverride?: string) =>
+const toolCall = (route: string, tool: PendingTool, inputOverride?: string) => {
-  parseToolInput(route, tool.name, inputOverride ?? tool.input).pipe(
+  const raw = inputOverride ?? tool.input
-    Effect.map(
+  return parseToolInput(route, tool.name, raw).pipe(
-      (input): ToolCall =>
+    Effect.map((input): ToolCall | ToolInputError =>
-        LLMEvent.toolCall({
+      LLMEvent.toolCall({
-          id: tool.id,
+        id: tool.id,
-          name: tool.name,
+        name: tool.name,
-          input,
+        input,
-          providerExecuted: tool.providerExecuted ? true : undefined,
+        providerExecuted: tool.providerExecuted ? true : undefined,
-          providerMetadata: tool.providerMetadata,
+        providerMetadata: tool.providerMetadata,
-        }),
+      }),
+    ),
+    Effect.catch((error) =>
+      tool.providerExecuted
+        ? Effect.fail(error)
+        : Effect.succeed(
+            LLMEvent.toolInputError({
+              id: tool.id,
+              name: tool.name,
+              raw,
+            }),
+          ),
     ),
   )
+}
+
+const finishEvents = (tool: PendingTool, event: ToolCall | ToolInputError): ReadonlyArray<LLMEvent> =>
+  event.type === "tool-input-error"
+    ? [event]
+    : [LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }), event]
 
 /** Store the updated tool and produce the optional public delta event. */
 const appendTool = <K extends StreamKey>(
@@ -122,8 +140,8 @@ export const appendOrStart = <K extends StreamKey>(
   missingToolMessage: string,
 ): AppendOutcome<K> | LLMError => {
   const current = tools[key]
-  const id = delta.id ?? current?.id
+  const id = current?.id ?? delta.id
-  const name = delta.name ?? current?.name
+  const name = current?.name ?? delta.name
   if (!id || !name) return eventError(route, missingToolMessage)
 
   const tool = {
@@ -158,8 +176,9 @@ export const appendExisting = <K extends StreamKey>(
 
 /**
  * Finalize one pending tool call: parse the accumulated raw JSON, remove it
- * from state, and return the optional public `tool-call` event. Missing keys are
+ * from state, and return either a call or a non-executable local input error.
- * a no-op because some providers emit stop events for non-tool content blocks.
+ * Missing keys are a no-op because some providers emit stop events for
+ * non-tool content blocks.
  */
 export const finish = <K extends StreamKey>(route: string, tools: State<K>, key: K) =>
   Effect.gen(function* () {
@@ -167,10 +186,7 @@ export const finish = <K extends StreamKey>(route: string, tools: State<K>, key:
     if (!tool) return { tools }
     return {
       tools: withoutTool(tools, key),
-      events: [
+      events: finishEvents(tool, yield* toolCall(route, tool)),
-        LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }),
-        yield* toolCall(route, tool),
-      ],
     }
   })
 
@@ -185,17 +201,14 @@ export const finishWithInput = <K extends StreamKey>(route: string, tools: State
     if (!tool) return { tools }
     return {
       tools: withoutTool(tools, key),
-      events: [
+      events: finishEvents(tool, yield* toolCall(route, tool, input)),
-        LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }),
-        yield* toolCall(route, tool, input),
-      ],
     }
   })
 
 /**
  * Finalize every pending tool call at once. OpenAI Chat has this shape: it does
- * not emit per-tool stop events, so all accumulated calls finish when the choice
+ * not emit per-tool stop events, so all accumulated calls finish independently
- * receives a terminal `finish_reason`.
+ * when the choice receives a terminal `finish_reason`.
  */
 export const finishAll = <K extends StreamKey>(route: string, tools: State<K>) =>
   Effect.gen(function* () {
@@ -205,12 +218,7 @@ export const finishAll = <K extends StreamKey>(route: string, tools: State<K>) =
     return {
       tools: empty<K>(),
       events: yield* Effect.forEach(pending, (tool) =>
-        toolCall(route, tool).pipe(
+        toolCall(route, tool).pipe(Effect.map((event) => finishEvents(tool, event))),
-          Effect.map((call) => [
-            LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }),
-            call,
-          ]),
-        ),
       ).pipe(Effect.map((events) => events.flat())),
     }
   })

packages/ai/src/protocols/xai-images.ts

@@ -0,0 +1,202 @@
+import { Effect, Encoding, Schema } from "effect"
+import { Headers, HttpClientRequest } from "effect/unstable/http"
+import { GeneratedImage, ImageModel, ImageResponse, type ImageRequestFor, type ImageRoute } from "../image"
+import { Auth, type Definition as AuthDefinition } from "../route/auth"
+import {
+  InvalidProviderOutputReason,
+  LLMError,
+  Usage,
+  mergeHttpOptions,
+  mergeJsonRecords,
+  type HttpOptions,
+} from "../schema"
+import { ProviderShared, optionalNull } from "./shared"
+import { ImageInputs } from "./utils/image-input"
+
+const ADAPTER = "xai-images"
+export const DEFAULT_BASE_URL = "https://api.x.ai/v1"
+export const PATH = "/images/generations"
+export const EDIT_PATH = "/images/edits"
+
+export type XAIImageString<Known extends string> = Known | (string & {})
+
+export type XAIImageOptions = {
+  readonly n?: number
+  readonly aspectRatio?: XAIImageString<
+    | "1:1"
+    | "3:4"
+    | "4:3"
+    | "9:16"
+    | "16:9"
+    | "2:3"
+    | "3:2"
+    | "9:19.5"
+    | "19.5:9"
+    | "9:20"
+    | "20:9"
+    | "1:2"
+    | "2:1"
+    | "auto"
+  >
+  readonly aspect_ratio?: XAIImageString<
+    | "1:1"
+    | "3:4"
+    | "4:3"
+    | "9:16"
+    | "16:9"
+    | "2:3"
+    | "3:2"
+    | "9:19.5"
+    | "19.5:9"
+    | "9:20"
+    | "20:9"
+    | "1:2"
+    | "2:1"
+    | "auto"
+  >
+  readonly resolution?: XAIImageString<"1k" | "2k">
+  readonly responseFormat?: XAIImageString<"url" | "b64_json">
+  readonly response_format?: XAIImageString<"url" | "b64_json">
+} & Record<string, unknown>
+
+type XAIImageBody = Record<string, unknown> & {
+  readonly model: string
+  readonly prompt: string
+}
+
+const XAIImageResponse = Schema.Struct({
+  data: Schema.Array(
+    Schema.Struct({
+      b64_json: optionalNull(Schema.String),
+      url: optionalNull(Schema.String),
+      revised_prompt: optionalNull(Schema.String),
+      mime_type: optionalNull(Schema.String),
+    }),
+  ),
+  usage: Schema.optional(Schema.Unknown),
+})
+
+export interface ModelInput {
+  readonly id: string
+  readonly auth: AuthDefinition
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions
+}
+
+const nativeOptions = (options: XAIImageOptions | undefined) => {
+  if (!options) return undefined
+  const { aspectRatio, responseFormat, ...native } = options
+  return {
+    aspect_ratio: aspectRatio,
+    response_format: responseFormat,
+    ...native,
+  }
+}
+
+const invalidOutput = (message: string) =>
+  new LLMError({
+    module: ADAPTER,
+    method: "generate",
+    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
+  })
+
+const applyQuery = (url: string, query: Record<string, string> | undefined) => {
+  if (!query) return url
+  const next = new URL(url)
+  Object.entries(query).forEach(([key, value]) => next.searchParams.set(key, value))
+  return next.toString()
+}
+
+export const model = (input: ModelInput) => {
+  const route: ImageRoute<XAIImageOptions> = {
+    id: ADAPTER,
+    generate: Effect.fn("XAIImages.generate")(function* (request: ImageRequestFor<XAIImageOptions>, execute) {
+      const http = mergeHttpOptions(request.model.http, request.http)
+      const imageReferences = (request.images ?? []).map((image) => {
+        if (image.type === "bytes") return { url: ImageInputs.dataUrl(image), type: "image_url" as const }
+        if (image.type === "url") return { url: image.url, type: "image_url" as const }
+        if (image.type === "file-id") return { file_id: image.id }
+        return undefined
+      })
+      if (imageReferences.some((image) => image === undefined))
+        return yield* ImageInputs.invalid(ADAPTER, "xAI Images accepts image URLs, data URLs, bytes, and file IDs")
+      const requestBody = mergeJsonRecords(
+        {
+          model: request.model.id,
+          prompt: request.prompt,
+          image: imageReferences.length === 1 ? imageReferences[0] : undefined,
+          images: imageReferences.length > 1 ? imageReferences : undefined,
+        },
+        nativeOptions(request.options),
+        http?.body,
+      ) as XAIImageBody
+      const text = ProviderShared.encodeJson(requestBody)
+      const url = applyQuery(
+        `${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}${imageReferences.length === 0 ? PATH : EDIT_PATH}`,
+        http?.query,
+      )
+      const headers = yield* Auth.toEffect(input.auth)({
+        request,
+        method: "POST",
+        url,
+        body: text,
+        headers: Headers.fromInput({ ...input.headers, ...http?.headers }),
+      })
+      const response = yield* execute(
+        HttpClientRequest.post(url).pipe(
+          HttpClientRequest.setHeaders(headers),
+          HttpClientRequest.bodyText(text, "application/json"),
+        ),
+      )
+      const payload = yield* response.json.pipe(
+        Effect.mapError(() => invalidOutput("Failed to read the xAI Images response")),
+      )
+      const decoded = yield* Schema.decodeUnknownEffect(XAIImageResponse)(payload).pipe(
+        Effect.mapError(() => invalidOutput("xAI Images returned an invalid response")),
+      )
+      const images = yield* Effect.forEach(decoded.data, (item, index) => {
+        const mediaType = item.mime_type ?? "application/octet-stream"
+        if (item.b64_json)
+          return Effect.fromResult(Encoding.decodeBase64(item.b64_json)).pipe(
+            Effect.mapError(() => invalidOutput(`xAI Images result ${index} contains invalid base64 data`)),
+            Effect.map(
+              (data) =>
+                new GeneratedImage({
+                  mediaType,
+                  data,
+                  providerMetadata:
+                    item.revised_prompt === undefined || item.revised_prompt === null
+                      ? undefined
+                      : { xai: { revisedPrompt: item.revised_prompt } },
+                }),
+            ),
+          )
+        if (item.url)
+          return Effect.succeed(
+            new GeneratedImage({
+              mediaType,
+              data: item.url,
+              providerMetadata:
+                item.revised_prompt === undefined || item.revised_prompt === null
+                  ? undefined
+                  : { xai: { revisedPrompt: item.revised_prompt } },
+            }),
+          )
+        return Effect.fail(invalidOutput(`xAI Images result ${index} has neither image data nor a URL`))
+      })
+      if (images.length === 0) return yield* invalidOutput("xAI Images returned no images")
+      const usage = ProviderShared.isRecord(decoded.usage) ? decoded.usage : undefined
+      return new ImageResponse({
+        images,
+        usage: usage === undefined ? undefined : new Usage({ providerMetadata: { xai: usage } }),
+        providerMetadata: usage === undefined ? undefined : { xai: { usage } },
+      })
+    }),
+  }
+  return ImageModel.make<XAIImageOptions>({ id: input.id, provider: "xai", route, http: input.http })
+}
+
+export const XAIImages = {
+  model,
+} as const

packages/ai/src/protocols/zai-images.ts

@@ -0,0 +1,132 @@
+import { Effect, Schema } from "effect"
+import { Headers, HttpClientRequest } from "effect/unstable/http"
+import { GeneratedImage, ImageModel, ImageResponse, type ImageRequestFor, type ImageRoute } from "../image"
+import { Auth, type Definition as AuthDefinition } from "../route/auth"
+import { InvalidProviderOutputReason, LLMError, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema"
+import { ProviderShared } from "./shared"
+import { ImageInputs } from "./utils/image-input"
+
+const ADAPTER = "zai-images"
+export const DEFAULT_BASE_URL = "https://api.z.ai/api/paas/v4"
+export const PATH = "/images/generations"
+
+export type ZAIImageString<Known extends string> = Known | (string & {})
+
+export type ZAIImageOptions = {
+  readonly size?: ZAIImageString<
+    "1024x1024" | "768x1344" | "864x1152" | "1344x768" | "1152x864" | "1440x720" | "720x1440"
+  >
+  readonly quality?: ZAIImageString<"hd" | "standard">
+  readonly userID?: string
+} & Record<string, unknown>
+
+type ZAIImageBody = Record<string, unknown> & {
+  readonly model: string
+  readonly prompt: string
+}
+
+const ZAIImageResponse = Schema.Struct({
+  created: Schema.optional(Schema.Int),
+  id: Schema.optional(Schema.String),
+  request_id: Schema.optional(Schema.String),
+  data: Schema.Array(Schema.Struct({ url: Schema.String })),
+  content_filter: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        role: Schema.optional(Schema.String),
+        level: Schema.optional(Schema.Number),
+      }),
+    ),
+  ),
+})
+
+export interface ModelInput {
+  readonly id: string
+  readonly auth: AuthDefinition
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions
+}
+
+const nativeOptions = (options: ZAIImageOptions | undefined) => {
+  if (!options) return undefined
+  const { userID, ...native } = options
+  return {
+    user_id: userID,
+    ...native,
+  }
+}
+
+const invalidOutput = (message: string) =>
+  new LLMError({
+    module: ADAPTER,
+    method: "generate",
+    reason: new InvalidProviderOutputReason({ message, route: ADAPTER }),
+  })
+
+const applyQuery = (url: string, query: Record<string, string> | undefined) => {
+  if (!query) return url
+  const next = new URL(url)
+  Object.entries(query).forEach(([key, value]) => next.searchParams.set(key, value))
+  return next.toString()
+}
+
+export const model = (input: ModelInput) => {
+  const route: ImageRoute<ZAIImageOptions> = {
+    id: ADAPTER,
+    generate: Effect.fn("ZAIImages.generate")(function* (request: ImageRequestFor<ZAIImageOptions>, execute) {
+      if ((request.images?.length ?? 0) > 0)
+        return yield* ImageInputs.invalid(ADAPTER, "Z.ai hosted image generation does not support image inputs")
+      const http = mergeHttpOptions(request.model.http, request.http)
+      const requestBody = mergeJsonRecords(
+        { model: request.model.id, prompt: request.prompt },
+        nativeOptions(request.options),
+        http?.body,
+      ) as ZAIImageBody
+      const text = ProviderShared.encodeJson(requestBody)
+      const url = applyQuery(`${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}${PATH}`, http?.query)
+      const headers = yield* Auth.toEffect(input.auth)({
+        request,
+        method: "POST",
+        url,
+        body: text,
+        headers: Headers.fromInput({ ...input.headers, ...http?.headers }),
+      })
+      const response = yield* execute(
+        HttpClientRequest.post(url).pipe(
+          HttpClientRequest.setHeaders(headers),
+          HttpClientRequest.bodyText(text, "application/json"),
+        ),
+      )
+      const payload = yield* response.json.pipe(
+        Effect.mapError(() => invalidOutput("Failed to read the Z.ai Images response")),
+      )
+      const decoded = yield* Schema.decodeUnknownEffect(ZAIImageResponse)(payload).pipe(
+        Effect.mapError(() => invalidOutput("Z.ai Images returned an invalid response")),
+      )
+      if (decoded.data.length === 0) return yield* invalidOutput("Z.ai Images returned no images")
+      return new ImageResponse({
+        images: decoded.data.map(
+          (item) =>
+            new GeneratedImage({
+              mediaType: "application/octet-stream",
+              data: item.url,
+            }),
+        ),
+        providerMetadata: {
+          zai: {
+            created: decoded.created,
+            id: decoded.id,
+            requestID: decoded.request_id,
+            contentFilter: decoded.content_filter,
+          },
+        },
+      })
+    }),
+  }
+  return ImageModel.make<ZAIImageOptions>({ id: input.id, provider: "zai", route, http: input.http })
+}
+
+export const ZAIImages = {
+  model,
+} as const

packages/ai/src/provider-error.ts

@@ -16,13 +16,17 @@ import {
 
 const patterns = [
   /prompt is too long/i,
+  /request_too_large/i,
   /input is too long for requested model/i,
   /exceeds the context window/i,
+  /exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\d,]+ tokens?|\s*\([\d,]+\))/i,
   /input token count.*exceeds the maximum/i,
   /tokens in request more than max tokens allowed/i,
   /maximum prompt length is \d+/i,
   /reduce the length of the messages/i,
   /maximum context length is \d+ tokens/i,
+  /exceeds (?:the )?maximum allowed input length of [\d,]+ tokens?/i,
+  /input \(\d+ tokens\) is longer than the model'?s context length \(\d+ tokens\)/i,
   /exceeds the limit of \d+/i,
   /exceeds the available context size/i,
   /greater than the context length/i,
@@ -34,11 +38,17 @@ const patterns = [
   /input length.*exceeds.*context length/i,
   /prompt too long; exceeded (?:max )?context length/i,
   /too large for model with \d+ maximum context length/i,
+  /prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i,
   /model_context_window_exceeded/i,
+  /too many tokens/i,
+  /token limit exceeded/i,
 ]
 
+const exclusions = [/^(throttling error|service unavailable):/i, /rate limit/i, /too many requests/i]
+
 export const isContextOverflow = (message: string) =>
-  patterns.some((pattern) => pattern.test(message)) || /^4(00|13)\s*(status code)?\s*\(no body\)/i.test(message)
+  !exclusions.some((pattern) => pattern.test(message)) &&
+  (patterns.some((pattern) => pattern.test(message)) || /^4(00|13)\s*(status code)?\s*\(no body\)/i.test(message))
 
 export const isContextOverflowFailure = (failure: unknown) =>
   failure instanceof LLMError
@@ -125,7 +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,
@@ -135,7 +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

@@ -1,16 +1,21 @@
-import type { Model } from "./schema"
+import type { Model, ProviderOptions } from "./schema"
 
 export interface Settings extends Readonly<Record<string, unknown>> {
+  readonly baseURL?: string
   readonly headers?: Readonly<Record<string, string>>
   readonly body?: Readonly<Record<string, unknown>>
   readonly limits?: {
     readonly context: number
+    readonly input?: number
     readonly output: number
   }
 }
 
-export interface Definition<ProviderSettings extends Settings = Settings> {
+export interface Definition<
-  readonly model: (modelID: string, settings: ProviderSettings) => Model
+  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) =>
+export const responsesModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
-  configure(config(settings)).responses(modelID)
+  modelID,
-export const chatModel: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+  settings,
-  configure(config(settings)).chat(modelID)
+) => 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

@@ -1,16 +1,18 @@
 import type { ProviderPackage } from "../provider-package"
-import { GoogleVertexAnthropic } from "../protocols/google-vertex-anthropic"
+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-anthropic")
+export const id = ProviderID.make("google-vertex")
 
 export type Config = RouteDefaultsInput &
   GoogleVertexShared.OAuthOptions & {
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -19,14 +21,18 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly location?: string
   readonly project?: string
-  readonly providerOptions?: ProviderOptions
+  readonly providerOptions?: OpenAIProviderOptionsInput
 }
 
-export const routes = [GoogleVertexAnthropic.route]
+const route = OpenAICompatibleChat.route.with({
+  id: "google-vertex-chat",
+  provider: id,
+})
+
+export const routes = [route]
 
 const configuredRoute = (input: Config) => {
-  if ("apiKey" in input && input.apiKey !== undefined)
+  if ("apiKey" in input && input.apiKey !== undefined) throw new Error("Google Vertex Chat does not support API keys")
-    throw new Error("Google Vertex Anthropic does not support API keys")
   const {
     accessToken: _accessToken,
     auth: _auth,
@@ -37,12 +43,12 @@ const configuredRoute = (input: Config) => {
   } = input
   const location = GoogleVertexShared.location(inputLocation, "global")
   const project = GoogleVertexShared.project(inputProject)
-  return GoogleVertexAnthropic.route.with({
+  return route.with({
     ...rest,
     endpoint: {
       baseURL:
         baseURL ??
-        `https://${GoogleVertexShared.host(location)}/v1/projects/${GoogleVertexShared.requireProject(project)}/locations/${location}/publishers/anthropic/models`,
+        `https://aiplatform.googleapis.com/v1/projects/${GoogleVertexShared.requireProject(project)}/locations/${location}/endpoints/openapi`,
     },
     auth: GoogleVertexShared.oauth(input, project),
   })
@@ -52,7 +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,
   }
 }
@@ -62,8 +68,8 @@ 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 Anthropic does not support API keys")
+  if (settings.apiKey !== undefined) throw new Error("Google Vertex Chat does not support API keys")
   return configure({
     accessToken: settings.accessToken,
     baseURL: settings.baseURL,

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

@@ -0,0 +1,119 @@
+import { Effect, Schema, Struct } from "effect"
+import type { ProviderPackage } from "../provider-package"
+import { AnthropicMessages } from "../protocols/anthropic-messages"
+import { Auth } from "../route/auth"
+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 } 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.
+export const id = ProviderID.make("google-vertex-anthropic")
+
+export type Config = RouteDefaultsInput &
+  GoogleVertexShared.OAuthOptions & {
+    readonly baseURL?: string
+    readonly location?: string
+    readonly project?: string
+    readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly accessToken?: string
+  readonly apiKey?: never
+  readonly baseURL?: string
+  readonly location?: string
+  readonly project?: string
+  readonly providerOptions?: AnthropicMessages.ProviderOptionsInput
+}
+
+const route = Route.make({
+  id: "google-vertex-messages",
+  provider: id,
+  providerMetadataKey: "anthropic",
+  protocol: Protocol.make({
+    id: AnthropicMessages.protocol.id,
+    body: {
+      schema: Schema.Struct({
+        ...Struct.omit(AnthropicMessages.AnthropicMessagesBody.fields, ["model"]),
+        anthropic_version: Schema.Literal(VERSION),
+      }),
+      from: (request) =>
+        AnthropicMessages.protocol.body.from(request).pipe(
+          Effect.map((body) => ({
+            ...Struct.omit(body, ["model"]),
+            anthropic_version: VERSION,
+          })),
+        ),
+    },
+    stream: AnthropicMessages.protocol.stream,
+  }),
+  endpoint: Endpoint.path(({ request }) => `/${request.model.id}:streamRawPredict`),
+  auth: Auth.none,
+  framing: Framing.sse,
+})
+
+export const routes = [route]
+
+const configuredRoute = (input: Config) => {
+  if ("apiKey" in input && input.apiKey !== undefined)
+    throw new Error("Google Vertex Messages does not support API keys")
+  const {
+    accessToken: _accessToken,
+    auth: _auth,
+    baseURL,
+    location: inputLocation,
+    project: inputProject,
+    ...rest
+  } = input
+  const location = GoogleVertexShared.location(inputLocation, "global")
+  const project = GoogleVertexShared.project(inputProject)
+  return route.with({
+    ...rest,
+    endpoint: {
+      baseURL:
+        baseURL ??
+        `https://${GoogleVertexShared.host(location)}/v1/projects/${GoogleVertexShared.requireProject(project)}/locations/${location}/publishers/anthropic/models`,
+    },
+    auth: GoogleVertexShared.oauth(input, project),
+  })
+}
+
+export const configure = (input: Config = {}) => {
+  const route = configuredRoute(input)
+  return {
+    id,
+    model: (modelID: string | ModelID) => route.model<AnthropicMessages.ProviderOptionsInput>({ id: modelID }),
+    configure,
+  }
+}
+
+export const provider = {
+  id,
+  configure,
+}
+
+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,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    limits: settings.limits,
+    location: settings.location,
+    project: settings.project,
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
+}

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

@@ -0,0 +1,88 @@
+import type { ProviderPackage } from "../provider-package"
+import { OpenAICompatibleResponses } from "../protocols/openai-compatible-responses"
+import type { RouteDefaultsInput } from "../route/client"
+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")
+
+export type Config = RouteDefaultsInput &
+  GoogleVertexShared.OAuthOptions & {
+    readonly baseURL?: string
+    readonly location?: string
+    readonly project?: string
+    readonly providerOptions?: OpenResponsesProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly accessToken?: string
+  readonly apiKey?: never
+  readonly baseURL?: string
+  readonly location?: string
+  readonly project?: string
+  readonly providerOptions?: OpenResponsesProviderOptionsInput
+}
+
+const route = OpenAICompatibleResponses.route.with({
+  id: "google-vertex-responses",
+  provider: id,
+  providerOptions: { openresponses: { store: false } },
+})
+
+export const routes = [route]
+
+const configuredRoute = (input: Config) => {
+  if ("apiKey" in input && input.apiKey !== undefined)
+    throw new Error("Google Vertex Responses does not support API keys")
+  const {
+    accessToken: _accessToken,
+    auth: _auth,
+    baseURL,
+    location: inputLocation,
+    project: inputProject,
+    ...rest
+  } = input
+  const location = GoogleVertexShared.location(inputLocation, "global")
+  const project = GoogleVertexShared.project(inputProject)
+  return route.with({
+    ...rest,
+    endpoint: {
+      baseURL:
+        baseURL ??
+        `https://aiplatform.googleapis.com/v1/projects/${GoogleVertexShared.requireProject(project)}/locations/${location}/endpoints/openapi`,
+    },
+    auth: GoogleVertexShared.oauth(input, project),
+  })
+}
+
+export const configure = (input: Config = {}) => {
+  const route = configuredRoute(input)
+  return {
+    id,
+    model: (modelID: string | ModelID) => route.model<OpenResponsesProviderOptionsInput>({ id: modelID }),
+    configure,
+  }
+}
+
+export const provider = {
+  id,
+  configure,
+}
+
+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,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    limits: settings.limits,
+    location: settings.location,
+    project: settings.project,
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
+}

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

@@ -1,10 +1,15 @@
 import type { ProviderPackage } from "../provider-package"
-import { GoogleVertexGemini } from "../protocols/google-vertex-gemini"
+import { Gemini } from "../protocols/gemini"
 import { Auth } from "../route/auth"
-import type { RouteDefaultsInput } from "../route/client"
+import { Route, type RouteDefaultsInput } from "../route/client"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { Endpoint } from "../route/endpoint"
+import { Framing } from "../route/framing"
+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 &
@@ -12,6 +17,7 @@ export type Config = RouteDefaultsInput &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: Gemini.ProviderOptionsInput
   }
 
 export type Settings = ProviderPackage.Settings &
@@ -22,10 +28,23 @@ export type Settings = ProviderPackage.Settings &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
-    readonly providerOptions?: ProviderOptions
+    readonly providerOptions?: Gemini.ProviderOptionsInput
   }
 
-export const routes = [GoogleVertexGemini.route]
+const route = Route.make({
+  id: "google-vertex-gemini",
+  provider: id,
+  providerMetadataKey: "google",
+  protocol: Gemini.protocol,
+  endpoint: Endpoint.path(({ request }) => {
+    const model = String(request.model.id)
+    return `/${model.startsWith("endpoints/") ? model : `models/${model}`}:streamGenerateContent?alt=sse`
+  }),
+  auth: Auth.none,
+  framing: Framing.sse,
+})
+
+export const routes = [route]
 
 const configuredRoute = (input: Config, modelID: string | ModelID) => {
   const {
@@ -48,7 +67,7 @@ const configuredRoute = (input: Config, modelID: string | ModelID) => {
     (apiKey
       ? "https://aiplatform.googleapis.com/v1/publishers/google"
       : `https://${GoogleVertexShared.host(location)}/v1beta1/projects/${GoogleVertexShared.requireProject(project)}/locations/${location}${endpointModel ? "" : "/publishers/google"}`)
-  return GoogleVertexGemini.route.with({
+  return route.with({
     ...rest,
     endpoint: { baseURL: endpoint },
     auth: apiKey === undefined ? GoogleVertexShared.oauth(input, project) : Auth.header("x-goog-api-key", apiKey),
@@ -58,7 +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,
   }
 }
@@ -67,7 +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-vertex/anthropic.ts

@@ -1,2 +0,0 @@
-export { model } from "../google-vertex-anthropic"
-export type { Settings } from "../google-vertex-anthropic"

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

@@ -0,0 +1,2 @@
+export { model } from "../google-vertex-chat"
+export type { Settings } from "../google-vertex-chat"

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

@@ -0,0 +1,2 @@
+export { model } from "../google-vertex"
+export type { Settings } from "../google-vertex"

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

@@ -0,0 +1,2 @@
+export { model } from "../google-vertex-messages"
+export type { Settings } from "../google-vertex-messages"

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

@@ -0,0 +1,2 @@
+export { model } from "../google-vertex-responses"
+export type { Settings } from "../google-vertex-responses"

packages/ai/src/providers/google.ts

@@ -2,19 +2,28 @@ 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 { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { HttpOptions, ProviderID, mergeHttpOptions, type ModelID } from "../schema"
-import * as Gemini from "../protocols/gemini"
+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")
 
 export const routes = [Gemini.route]
 
-export type Config = RouteDefaultsInput & ProviderAuthOption<"optional"> & { readonly baseURL?: string }
+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">) => {
@@ -31,15 +40,24 @@ const configuredRoute = (input: Config) => {
 
 export const configure = (input: Config = {}) => {
   const route = configuredRoute(input)
+  const image = (modelID: string | ModelID) =>
+    GoogleImages.model({
+      id: modelID,
+      auth: auth(input),
+      baseURL: input.baseURL,
+      headers: input.headers,
+      http: mergeHttpOptions(input.http === undefined ? undefined : HttpOptions.make(input.http)),
+    })
   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,
@@ -48,3 +66,5 @@ export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, se
     limits: settings.limits,
     providerOptions: settings.providerOptions,
   }).model(modelID)
+
+export const image = provider.image

packages/ai/src/providers/index.ts

@@ -7,9 +7,12 @@ export { CloudflareAIGateway, CloudflareWorkersAI } from "./cloudflare"
 export * as GitHubCopilot from "./github-copilot"
 export * as Google from "./google"
 export * as GoogleVertex from "./google-vertex"
-export * as GoogleVertexAnthropic from "./google-vertex-anthropic"
+export * as GoogleVertexChat from "./google-vertex-chat"
+export * as GoogleVertexMessages from "./google-vertex-messages"
+export * as GoogleVertexResponses from "./google-vertex-responses"
 export * as OpenAI from "./openai"
 export * as OpenAICompatible from "./openai-compatible"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses"
 export * as OpenRouter from "./openrouter"
 export * as XAI from "./xai"
+export * as ZAI from "./zai"

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 {
+export type OpenAIOptionsInput = OpenResponsesOptionsInput
-  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 OpenAIProviderOptionsInput = ProviderOptions & {
   readonly openai?: OpenAIOptionsInput

packages/ai/src/providers/openai.ts

@@ -1,12 +1,14 @@
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options"
 import type { Route, RouteDefaultsInput } from "../route/client"
 import type { ProviderPackage } from "../provider-package"
-import { ProviderID, type ModelID } from "../schema"
+import { HttpOptions, ProviderID, ToolDefinition, mergeHttpOptions, type ModelID } from "../schema"
 import * as OpenAIChat from "../protocols/openai-chat"
 import * as OpenAIResponses from "../protocols/openai-responses"
 import { withOpenAIOptions, type OpenAIProviderOptionsInput } from "./openai-options"
+import { OpenAIImages, type OpenAIImageString } from "../protocols/openai-images"
 
 export type { OpenAIOptionsInput, OpenAIResponseIncludable } from "./openai-options"
+export type { OpenAIImageOptions } from "../protocols/openai-images"
 
 export const id = ProviderID.make("openai")
 
@@ -22,6 +24,39 @@ export type Config = RouteDefaultsInput &
     readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
+export interface ImageGenerationOptions {
+  readonly action?: OpenAIImageString<"auto" | "generate" | "edit">
+  readonly background?: OpenAIImageString<"auto" | "opaque" | "transparent">
+  readonly inputFidelity?: OpenAIImageString<"low" | "high">
+  readonly outputCompression?: number
+  readonly outputFormat?: OpenAIImageString<"png" | "jpeg" | "webp">
+  readonly partialImages?: number
+  readonly quality?: OpenAIImageString<"auto" | "low" | "medium" | "high" | "standard" | "hd">
+  readonly size?: OpenAIImageString<
+    "auto" | "256x256" | "512x512" | "1024x1024" | "1536x1024" | "1024x1536" | "1792x1024" | "1024x1792"
+  >
+}
+
+export const imageGeneration = (options: ImageGenerationOptions = {}) =>
+  ToolDefinition.make({
+    name: "image_generation",
+    description: "Generate or edit an image using OpenAI's hosted image generation tool.",
+    inputSchema: { type: "object", properties: {}, additionalProperties: false },
+    native: {
+      openai: {
+        type: "image_generation",
+        action: options.action,
+        background: options.background,
+        input_fidelity: options.inputFidelity,
+        output_compression: options.outputCompression,
+        output_format: options.outputFormat,
+        partial_images: options.partialImages,
+        quality: options.quality,
+        size: options.size,
+      },
+    },
+  })
+
 export interface Settings extends ProviderPackage.Settings {
   readonly apiKey?: string
   readonly baseURL?: string
@@ -51,10 +86,26 @@ 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 })
+    responsesWebSocketRoute
-  const chat = (id: string | ModelID) => chatRoute.with(withOpenAIOptions(id, modelDefaults)).model({ id })
+      .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,
+      auth: auth(input),
+      baseURL: input.baseURL,
+      headers: input.headers,
+      http: mergeHttpOptions(
+        input.http === undefined ? undefined : HttpOptions.make(input.http),
+        input.queryParams === undefined ? undefined : new HttpOptions({ query: input.queryParams }),
+      ),
+    })
 
   return {
     id,
@@ -62,6 +113,7 @@ export const configure = (input: Config = {}) => {
     responses,
     responsesWebSocket,
     chat,
+    image,
     configure,
   }
 }
@@ -85,15 +137,18 @@ 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) =>
+export const chatModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
-  configure(config(settings)).chat(modelID)
+  modelID,
+  settings,
+) => configure(config(settings)).chat(modelID)
 export const responses = provider.responses
 export const responsesWebSocket = provider.responsesWebSocket
 export const chat = provider.chat
+export const image = provider.image

packages/ai/src/providers/openrouter.ts

@@ -4,20 +4,72 @@ import { Endpoint } from "../route/endpoint"
 import { Framing } from "../route/framing"
 import { Protocol } from "../route/protocol"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { ProviderID, type CacheHint, type ModelID, type ProviderOptions } from "../schema"
+import type { ProviderPackage } from "../provider-package"
 import * as OpenAICompatibleProfiles from "./openai-compatible-profile"
 import * as OpenAIChat from "../protocols/openai-chat"
+import { newBreakpoints, ttlBucket } from "../protocols/utils/cache"
 import { isRecord } from "../protocols/shared"
 
 export const profile = OpenAICompatibleProfiles.profiles.openrouter
 export const id = ProviderID.make(profile.provider)
 const ADAPTER = "openrouter"
 
+type OpenRouterString<Known extends string> = Known | (string & {})
+
+export interface OpenRouterProviderRouting {
+  readonly [key: string]: unknown
+  readonly order?: ReadonlyArray<string>
+  readonly allow_fallbacks?: boolean
+  readonly require_parameters?: boolean
+  readonly data_collection?: OpenRouterString<"allow" | "deny">
+  readonly only?: ReadonlyArray<string>
+  readonly ignore?: ReadonlyArray<string>
+  readonly quantizations?: ReadonlyArray<string>
+  readonly sort?: OpenRouterString<"price" | "throughput" | "latency">
+  readonly max_price?: Readonly<{
+    prompt?: number | string
+    completion?: number | string
+    image?: number | string
+    audio?: number | string
+    request?: number | string
+  }>
+  readonly zdr?: boolean
+}
+
+export type OpenRouterPlugin =
+  | Readonly<{
+      id: "web"
+      max_results?: number
+      search_prompt?: string
+      engine?: OpenRouterString<"native" | "exa">
+    }>
+  | Readonly<{ id: "file-parser"; max_files?: number; pdf?: { engine?: string } }>
+  | Readonly<{ id: "moderation" }>
+  | Readonly<{ id: "response-healing" }>
+  | Readonly<{ id: "auto-router"; allowed_models?: ReadonlyArray<string> }>
+  | Readonly<{ id: string & {}; [key: string]: unknown }>
+
 export interface OpenRouterOptions {
   readonly [key: string]: unknown
-  readonly usage?: boolean | Record<string, unknown>
+  readonly debug?: Readonly<{ echo_upstream_body?: boolean }>
-  readonly reasoning?: Record<string, unknown>
+  readonly models?: ReadonlyArray<string>
+  readonly plugins?: ReadonlyArray<OpenRouterPlugin>
   readonly promptCacheKey?: string
+  readonly provider?: OpenRouterProviderRouting
+  readonly reasoning?: Readonly<{
+    enabled?: boolean
+    exclude?: boolean
+    effort?: OpenRouterString<"none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max">
+    max_tokens?: number
+  }>
+  readonly usage?: boolean | Readonly<{ include: boolean }>
+  readonly user?: string
+  readonly web_search_options?: Readonly<{
+    max_results?: number
+    search_prompt?: string
+    engine?: OpenRouterString<"native" | "exa">
+  }>
 }
 
 export type OpenRouterProviderOptionsInput = ProviderOptions & {
@@ -30,6 +82,12 @@ export type ModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
     readonly providerOptions?: OpenRouterProviderOptionsInput
   }
 
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenRouterProviderOptionsInput
+}
+
 const OpenRouterBody = Schema.StructWithRest(Schema.Struct(OpenAIChat.bodyFields), [
   Schema.Record(Schema.String, Schema.Any),
 ])
@@ -40,29 +98,70 @@ export const protocol = Protocol.make({
   body: {
     schema: OpenRouterBody,
     from: (request) =>
-      OpenAIChat.protocol.body.from(request).pipe(
+      OpenAIChat.fromRequest(request, { cacheControl: cacheControl() }).pipe(
-        Effect.map(
+        Effect.map((body) => {
-          (body) =>
+          const sourceAssistants = request.messages.filter((message) => message.role === "assistant")
-            ({
+          let assistantIndex = 0
-              ...body,
+          const messages = body.messages.map((message) => {
-              ...bodyOptions(request.providerOptions?.openrouter),
+            if (message.role !== "assistant") return message
-            }) as OpenRouterBody,
+            const source = sourceAssistants[assistantIndex++]
-        ),
+            const reasoning = source?.content
+              .filter((part) => part.type === "reasoning")
+              .map((part) => part.text)
+              .join("")
+            const reasoningDetails = Array.isArray(message.reasoning_details) ? message.reasoning_details : undefined
+            return {
+              ...message,
+              reasoning_content: undefined,
+              reasoning_text: undefined,
+              reasoning: reasoning && reasoningDetails && reasoningDetails.length > 0 ? reasoning : undefined,
+              reasoning_details: reasoningDetails,
+            }
+          })
+          return {
+            ...body,
+            messages,
+            ...bodyOptions(request.providerOptions?.openrouter),
+          } as OpenRouterBody
+        }),
       ),
   },
   stream: OpenAIChat.protocol.stream,
 })
 
+const cacheControl = () => {
+  const breakpoints = newBreakpoints(4)
+  return (cache: CacheHint | undefined) => {
+    if (cache === undefined || breakpoints.remaining === 0) return undefined
+    breakpoints.remaining -= 1
+    return {
+      type: "ephemeral" as const,
+      ...(ttlBucket(cache.ttlSeconds) === "1h" ? { ttl: "1h" } : {}),
+    }
+  }
+}
+
 const bodyOptions = (input: unknown) => {
   const openrouter = isRecord(input) ? input : {}
+  const { usage, models, provider, plugins, web_search_options, debug, user, reasoning, promptCacheKey, ...options } =
+    openrouter
   return {
-    ...(openrouter.usage === true
+    ...options,
+    ...(usage === undefined || usage === true
       ? { usage: { include: true } }
-      : isRecord(openrouter.usage)
+      : usage === false
-        ? { usage: openrouter.usage }
+        ? { usage: { include: false } }
-        : {}),
+        : isRecord(usage)
-    ...(isRecord(openrouter.reasoning) ? { reasoning: openrouter.reasoning } : {}),
+          ? { usage }
-    ...(typeof openrouter.promptCacheKey === "string" ? { prompt_cache_key: openrouter.promptCacheKey } : {}),
+          : {}),
+    ...(Array.isArray(models) ? { models } : {}),
+    ...(isRecord(provider) ? { provider } : {}),
+    ...(Array.isArray(plugins) ? { plugins } : {}),
+    ...(isRecord(web_search_options) ? { web_search_options } : {}),
+    ...(isRecord(debug) ? { debug } : {}),
+    ...(typeof user === "string" ? { user } : {}),
+    ...(isRecord(reasoning) ? { reasoning } : {}),
+    ...(typeof promptCacheKey === "string" ? { prompt_cache_key: promptCacheKey } : {}),
   }
 }
 
@@ -89,10 +188,21 @@ 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,
   }
 }
 
 export const provider = configure()
-export const model = provider.model
+export const model: ProviderPackage.Definition<Settings, OpenRouterProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers,
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    limits: settings.limits,
+    providerOptions: settings.providerOptions,
+  }).model(modelID)

packages/ai/src/providers/xai.ts

@@ -1,26 +1,62 @@
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options"
-import type { RouteDefaultsInput } from "../route/client"
+import { Route, type RouteDefaultsInput } from "../route/client"
-import { ProviderID, type ModelID } from "../schema"
+import { Endpoint } from "../route/endpoint"
+import { HttpOptions, ProviderID, type ModelID, type ProviderOptions } from "../schema"
 import * as OpenAICompatibleProfiles from "./openai-compatible-profile"
 import * as OpenAICompatibleChat from "../protocols/openai-compatible-chat"
+import * as OpenAIChat from "../protocols/openai-chat"
 import * as OpenAIResponses from "../protocols/openai-responses"
+import { XAIImages } from "../protocols/xai-images"
+import type { OpenAIOptionsInput } from "./openai-options"
+import type { ProviderPackage } from "../provider-package"
 
 export const id = ProviderID.make("xai")
 
-export type ModelOptions = RouteDefaultsInput &
+export type XAIProviderOptionsInput = ProviderOptions & {
+  readonly xai?: OpenAIOptionsInput
+}
+
+export type ModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     readonly baseURL?: string
+    readonly providerOptions?: XAIProviderOptionsInput
   }
 
-export const routes = [OpenAIResponses.route, OpenAICompatibleChat.route]
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: XAIProviderOptionsInput
+}
+
+export type { XAIImageOptions } from "../protocols/xai-images"
+
+const responsesRoute = Route.make({
+  id: "openai-responses",
+  provider: id,
+  providerMetadataKey: "xai",
+  protocol: OpenAIResponses.protocol,
+  endpoint: Endpoint.path("/responses", { baseURL: OpenAICompatibleProfiles.profiles.xai.baseURL }),
+  transport: OpenAIResponses.httpTransport,
+  defaults: { providerOptions: { xai: { store: false } } },
+})
+
+const chatRoute = Route.make({
+  id: "openai-compatible-chat",
+  provider: id,
+  providerMetadataKey: "xai",
+  protocol: OpenAIChat.protocol,
+  endpoint: Endpoint.path("/chat/completions", { baseURL: OpenAICompatibleProfiles.profiles.xai.baseURL }),
+  transport: OpenAICompatibleChat.route.transport,
+})
+
+export const routes = [responsesRoute, chatRoute]
 
 const auth = (options: ProviderAuthOption<"optional">) => AuthOptions.bearer(options, "XAI_API_KEY")
 
 const configuredResponsesRoute = (input: ModelOptions) => {
   const { apiKey: _, auth: _auth, baseURL, ...rest } = input
-  return OpenAIResponses.route.with({
+  return responsesRoute.with({
     ...rest,
-    provider: id,
     endpoint: { baseURL: baseURL ?? OpenAICompatibleProfiles.profiles.xai.baseURL },
     auth: auth(input),
   })
@@ -28,9 +64,8 @@ const configuredResponsesRoute = (input: ModelOptions) => {
 
 const configuredChatRoute = (input: ModelOptions) => {
   const { apiKey: _, auth: _auth, baseURL, ...rest } = input
-  return OpenAICompatibleChat.route.with({
+  return chatRoute.with({
     ...rest,
-    provider: id,
     endpoint: { baseURL: baseURL ?? OpenAICompatibleProfiles.profiles.xai.baseURL },
     auth: auth(input),
   })
@@ -39,18 +74,36 @@ 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 responses = (modelID: string | ModelID) => responsesRoute.model<XAIProviderOptionsInput>({ id: modelID })
-  const chat = (modelID: string | ModelID) => chatRoute.model({ id: modelID })
+  const chat = (modelID: string | ModelID) => chatRoute.model<XAIProviderOptionsInput>({ id: modelID })
+  const image = (modelID: string | ModelID) =>
+    XAIImages.model({
+      id: modelID,
+      auth: auth(input),
+      baseURL: input.baseURL ?? OpenAICompatibleProfiles.profiles.xai.baseURL,
+      headers: input.headers,
+      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+    })
   return {
     id,
     model: responses,
     responses,
     chat,
+    image,
     configure,
   }
 }
 
 export const provider = configure()
-export const model = provider.model
+export const model: ProviderPackage.Definition<Settings, XAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers,
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    limits: settings.limits,
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
 export const responses = provider.responses
 export const chat = provider.chat
+export const image = provider.image

packages/ai/src/providers/zai.ts

@@ -0,0 +1,35 @@
+import { ZAIImages } from "../protocols/zai-images"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options"
+import { HttpOptions, ProviderID, type ModelID } from "../schema"
+
+export const id = ProviderID.make("zai")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+export type { ZAIImageOptions } from "../protocols/zai-images"
+
+const auth = (options: ProviderAuthOption<"optional">) => AuthOptions.bearer(options, "ZAI_API_KEY")
+
+export const configure = (input: Config = {}) => {
+  const image = (modelID: string | ModelID) =>
+    ZAIImages.model({
+      id: modelID,
+      auth: auth(input),
+      baseURL: input.baseURL,
+      headers: input.headers,
+      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+    })
+
+  return {
+    id,
+    image,
+    configure,
+  }
+}
+
+export const provider = configure()
+export const image = provider.image

packages/ai/src/route/auth.ts

@@ -1,6 +1,6 @@
 import { Config, Effect, Redacted } from "effect"
 import { Headers } from "effect/unstable/http"
-import { AuthenticationReason, InvalidRequestReason, LLMError, type LLMRequest } from "../schema"
+import { AuthenticationReason, InvalidRequestReason, LLMError, type HttpOptions } from "../schema"
 
 export class MissingCredentialError extends Error {
   readonly _tag = "MissingCredentialError"
@@ -15,7 +15,7 @@ export type AuthError = CredentialError | LLMError
 type Secret = string | Redacted.Redacted | Config.Config<string | Redacted.Redacted>
 
 export interface AuthInput {
-  readonly request: LLMRequest
+  readonly request: { readonly http?: HttpOptions }
   readonly method: "POST" | "GET"
   readonly url: string
   readonly body: string

packages/ai/src/route/client.ts

@@ -5,12 +5,12 @@ import { Endpoint, type EndpointPatch } from "./endpoint"
 import { RequestExecutor } from "./executor"
 import { Framing } from "./framing"
 import { HttpTransport } from "./transport"
-import type { Transport, TransportRuntime } from "./transport"
+import type { HttpRequestTransform, Transport, TransportRuntime } from "./transport"
 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,
@@ -46,8 +45,12 @@ 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 prepareTransport: (
+    body: Body,
+    request: LLMRequest,
+    options?: StreamOptions,
+  ) => Effect.Effect<Prepared, LLMError>
   readonly streamPrepared: (
     prepared: Prepared,
     request: LLMRequest,
@@ -93,12 +96,12 @@ 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,
@@ -142,27 +145,20 @@ 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
 }
 
+export interface StreamOptions {
+  readonly transform?: HttpRequestTransform
+}
+
 export interface StreamMethod {
-  (request: LLMRequest): Stream.Stream<LLMEvent, LLMError>
+  (request: LLMRequest, options?: StreamOptions): Stream.Stream<LLMEvent, LLMError>
 }
 
 export interface GenerateMethod {
-  (request: LLMRequest): Effect.Effect<LLMResponse, LLMError>
+  (request: LLMRequest, options?: StreamOptions): Effect.Effect<LLMResponse, LLMError>
 }
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/LLMClient") {}
@@ -296,8 +292,9 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
           defaults: mergeRouteDefaults(route.defaults, defaults),
         })
       },
-      model: (input) => makeRouteModel(route, input),
+      model: <Options extends ProviderOptions = ProviderOptions>(input: RouteMappedModelInput) =>
-      prepareTransport: (body, request) =>
+        makeRouteModel<Options>(route, input),
+      prepareTransport: (body, request, options) =>
         routeInput.transport.prepare({
           body,
           request,
@@ -305,6 +302,7 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
           auth: routeInput.auth ?? Auth.none,
           encodeBody,
           headers: routeInput.headers,
+          transform: options?.transform,
         }),
       streamPrepared: (prepared: Prepared, request: LLMRequest, runtime: TransportRuntime) => {
         const route = `${request.model.provider}/${request.model.route.id}`
@@ -370,17 +368,14 @@ export function make<Body, Prepared, Frame, Event, State>(
   })
 }
 
-// `compile` is the important boundary: it turns a common `LLMRequest` into a
+const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest, options?: StreamOptions) {
-// 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
 
   const body = yield* route.body
     .from(resolved)
     .pipe(Effect.flatMap(ProviderShared.validateWith(Schema.decodeUnknownEffect(route.body.schema))))
-  const prepared = yield* route.prepareTransport(body, resolved)
+  const prepared = yield* route.prepareTransport(body, resolved, options)
 
   return {
     request: resolved,
@@ -390,30 +385,30 @@ 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 {
-  return new PreparedRequest({
     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) =>
+const streamRequestWith = (runtime: TransportRuntime) => (request: LLMRequest, options?: StreamOptions) =>
   Stream.unwrap(
     Effect.gen(function* () {
-      const compiled = yield* compile(request)
+      const compiled = yield* compile(request, options)
       return compiled.route.streamPrepared(compiled.prepared, compiled.request, runtime)
     }),
   )
 
 const generateWith = (stream: Interface["stream"]) =>
-  Effect.fn("LLM.generate")(function* (request: LLMRequest) {
+  Effect.fn("LLM.generate")(function* (request: LLMRequest, options?: StreamOptions) {
-    const state = yield* stream(request).pipe(Stream.runFold(LLMResponse.empty, LLMResponse.reduce))
+    const state = yield* stream(request, options).pipe(Stream.runFold(LLMResponse.empty, LLMResponse.reduce))
     const response = LLMResponse.complete(state)
     if (response) return response
     return yield* ProviderShared.eventError(
@@ -422,27 +417,24 @@ const generateWith = (stream: Interface["stream"]) =>
     )
   })
 
-export const prepare = <Body = unknown>(request: LLMRequest) =>
+export function stream(request: LLMRequest, options?: StreamOptions): Stream.Stream<LLMEvent, LLMError> {
-  prepareWith(request) as Effect.Effect<PreparedRequestOf<Body>, LLMError>
-
-export function stream(request: LLMRequest): Stream.Stream<LLMEvent, LLMError> {
   return Stream.unwrap(
     Effect.gen(function* () {
-      return (yield* Service).stream(request)
+      return (yield* Service).stream(request, options)
     }),
   ) as Stream.Stream<LLMEvent, LLMError>
 }
 
-export function generate(request: LLMRequest): Effect.Effect<LLMResponse, LLMError> {
+export function generate(request: LLMRequest, options?: StreamOptions): Effect.Effect<LLMResponse, LLMError> {
   return Effect.gen(function* () {
-    return yield* (yield* Service).generate(request)
+    return yield* (yield* Service).generate(request, options)
   }) as Effect.Effect<LLMResponse, LLMError>
 }
 
-export const streamRequest = (request: LLMRequest) =>
+export const streamRequest = (request: LLMRequest, options?: StreamOptions) =>
   Stream.unwrap(
     Effect.gen(function* () {
-      return (yield* Service).stream(request)
+      return (yield* Service).stream(request, options)
     }),
   )
 
@@ -453,7 +445,7 @@ 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 +454,6 @@ export const Route = { make } as const
 export const LLMClient = {
   Service,
   layer,
-  prepare,
   stream,
   generate,
 } as const

packages/ai/src/route/index.ts

@@ -8,6 +8,7 @@ export type {
   AnyRoute,
   Interface as LLMClientShape,
   Service as LLMClientService,
+  StreamOptions,
 } from "./client"
 export * from "./executor"
 export { Auth } from "./auth"
@@ -22,4 +23,4 @@ export type { ApiKeyMode, AuthOverride, ProviderAuthOption } from "./auth-option
 export type { Definition as EndpointFn, EndpointInput } from "./endpoint"
 export type { Definition as FramingDef } from "./framing"
 export type { Protocol as ProtocolDef } from "./protocol"
-export type { Transport as TransportDef, TransportRuntime } from "./transport"
+export type { HttpRequest, HttpRequestTransform, Transport as TransportDef, TransportRuntime } from "./transport"

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/route/transport/http.ts

@@ -28,57 +28,9 @@ const applyQuery = (url: string, query: Record<string, string> | undefined) => {
   return next.toString()
 }
 
-const PROTOCOL_BODY_OVERLAY_DENYLIST = new Set([
-  "anthropic_version",
-  "content",
-  "contents",
-  "frequencyPenalty",
-  "frequency_penalty",
-  "generationConfig",
-  "inferenceConfig",
-  "input",
-  "maxTokens",
-  "max_tokens",
-  "messages",
-  "model",
-  "presencePenalty",
-  "presence_penalty",
-  "responseFormat",
-  "response_format",
-  "seed",
-  "stop",
-  "stopSequences",
-  "stop_sequences",
-  "stream",
-  "streamOptions",
-  "stream_options",
-  "system",
-  "systemInstruction",
-  "system_instruction",
-  "temperature",
-  "thinking",
-  "toolChoice",
-  "toolConfig",
-  "tool_choice",
-  "tool_config",
-  "tools",
-  "topK",
-  "topP",
-  "top_k",
-  "top_p",
-])
-
-const forbiddenBodyOverlayKeys = (body: Record<string, unknown>) =>
-  Object.keys(body).filter((key) => PROTOCOL_BODY_OVERLAY_DENYLIST.has(key))
-
 const bodyWithOverlay = <Body>(body: Body, request: LLMRequest, encodeBody: (body: Body) => string) =>
   Effect.gen(function* () {
     if (request.http?.body === undefined) return { jsonBody: body, bodyText: encodeBody(body) }
-    const forbiddenKeys = forbiddenBodyOverlayKeys(request.http.body)
-    if (forbiddenKeys.length > 0)
-      return yield* ProviderShared.invalidRequest(
-        `http.body cannot overlay protocol-owned field(s): ${forbiddenKeys.join(", ")}`,
-      )
     if (ProviderShared.isRecord(body)) {
       const overlaid = mergeJsonRecords(body, request.http.body) ?? {}
       return { jsonBody: overlaid, bodyText: ProviderShared.encodeJson(overlaid) }
@@ -120,14 +72,19 @@ export const httpJson = <Body, Frame>(input: HttpJsonInput<Body, Frame>): HttpJs
   id: "http-json",
   with: (patch) => httpJson({ ...input, ...patch }),
   prepare: (prepareInput) =>
-    jsonRequestParts({
+    Effect.gen(function* () {
-      ...prepareInput,
+      const parts = yield* jsonRequestParts({ ...prepareInput })
-    }).pipe(
+      const request = { url: parts.url, method: "POST", headers: { ...parts.headers }, body: parts.bodyText }
-      Effect.map((parts) => ({
+      yield* (prepareInput.transform?.(request) ?? Effect.void)
-        request: ProviderShared.jsonPost({ url: parts.url, body: parts.bodyText, headers: parts.headers }),
+      return {
+        request: ProviderShared.jsonPost({
+          url: request.url,
+          body: request.body ?? "",
+          headers: Headers.fromInput(request.headers),
+        }),
         framing: input.framing,
-      })),
+      }
-    ),
+    }),
   frames: (prepared, request, runtime) =>
     Stream.unwrap(
       runtime.http

packages/ai/src/route/transport/index.ts

@@ -10,6 +10,15 @@ export interface TransportRuntime {
   readonly webSocket?: WebSocketExecutorInterface
 }
 
+export interface HttpRequest {
+  url: string
+  readonly method: string
+  headers: Record<string, string>
+  body: string | undefined
+}
+
+export type HttpRequestTransform = (request: HttpRequest) => Effect.Effect<void>
+
 export interface Transport<Body, Prepared, Frame> {
   readonly id: string
   readonly prepare: (input: TransportPrepareInput<Body>) => Effect.Effect<Prepared, LLMError>
@@ -27,6 +36,7 @@ export interface TransportPrepareInput<Body> {
   readonly auth: Auth.Definition
   readonly encodeBody: (body: Body) => string
   readonly headers?: (input: { readonly request: LLMRequest }) => Record<string, string>
+  readonly transform?: HttpRequestTransform
 }
 
 export * as HttpTransport from "./http"

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", {
+export class ToolFailure extends Tool.Error {}
-  message: Schema.String,
-  error: Schema.optional(Schema.Defect()),
-  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
-}) {}

packages/ai/src/schema/events.ts

@@ -1,6 +1,5 @@
 import { Schema } from "effect"
-import { ContentBlockID, FinishReason, ProtocolID, ProviderMetadata, RouteID, ToolCallID } from "./ids"
+import { ContentBlockID, FinishReason, ProviderMetadata, ToolCallID } from "./ids"
-import { ModelSchema } from "./options"
 import { Message, ToolCallPart, ToolOutput, ToolResultPart, ToolResultValue, type ContentPart } from "./messages"
 import { ProviderFailureClassification } from "./errors"
 
@@ -34,14 +33,15 @@ import { ProviderFailureClassification } from "./errors"
  *
  * **Semantics by provider**:
  *
- * - OpenAI Chat / Responses / Gemini / Bedrock: provider reports inclusive
+ * - OpenAI Chat / Responses / Gemini: provider reports inclusive
  *   `inputTokens` and an inclusive `outputTokens`; mapper subtracts to
  *   derive the breakdown.
- * - Anthropic: provider reports the breakdown natively (`input_tokens` is
+ * - Anthropic and Bedrock report the input breakdown natively: Anthropic's
- *   non-cached only); mapper sums to derive the inclusive `inputTokens`.
+ *   `input_tokens` and Bedrock's `inputTokens` are non-cached only. Their
- *   Anthropic does *not* break extended-thinking out of `output_tokens`, so
+ *   mappers sum the breakdown to derive the inclusive `inputTokens`.
- *   `reasoningTokens` is `undefined` and `outputTokens` carries the
+ *   Anthropic's `outputTokens` includes extended thinking. Newer responses
- *   combined total — a documented limitation of the Anthropic API.
+ *   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.)
@@ -129,6 +129,7 @@ export const ToolInputStart = Schema.Struct({
   type: Schema.tag("tool-input-start"),
   id: ToolCallID,
   name: Schema.String,
+  providerExecuted: Schema.optional(Schema.Boolean),
   providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.ToolInputStart" })
 export type ToolInputStart = Schema.Schema.Type<typeof ToolInputStart>
@@ -149,6 +150,15 @@ export const ToolInputEnd = Schema.Struct({
 }).annotate({ identifier: "LLM.Event.ToolInputEnd" })
 export type ToolInputEnd = Schema.Schema.Type<typeof ToolInputEnd>
 
+/** A local tool call whose final input could not be decoded. */
+export const ToolInputError = Schema.Struct({
+  type: Schema.tag("tool-input-error"),
+  id: ToolCallID,
+  name: Schema.String,
+  raw: Schema.String,
+}).annotate({ identifier: "LLM.Event.ToolInputError" })
+export type ToolInputError = Schema.Schema.Type<typeof ToolInputError>
+
 export const ToolCall = Schema.Struct({
   type: Schema.tag("tool-call"),
   id: ToolCallID,
@@ -180,10 +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" })
@@ -191,7 +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" })
@@ -216,6 +232,7 @@ const llmEventTagged = Schema.Union([
   ToolInputStart,
   ToolInputDelta,
   ToolInputEnd,
+  ToolInputError,
   ToolCall,
   ToolResult,
   ToolError,
@@ -253,6 +270,8 @@ export const LLMEvent = Object.assign(llmEventTagged, {
   toolInputDelta: (input: WithID<ToolInputDelta, ToolCallID>) =>
     ToolInputDelta.make({ ...input, id: toolCallID(input.id) }),
   toolInputEnd: (input: WithID<ToolInputEnd, ToolCallID>) => ToolInputEnd.make({ ...input, id: toolCallID(input.id) }),
+  toolInputError: (input: WithID<ToolInputError, ToolCallID>) =>
+    ToolInputError.make({ ...input, id: toolCallID(input.id) }),
   toolCall: (input: WithID<ToolCall, ToolCallID>) => ToolCall.make({ ...input, id: toolCallID(input.id) }),
   toolResult: (input: WithID<ToolResult, ToolCallID>) =>
     ToolResult.make({
@@ -283,6 +302,7 @@ export const LLMEvent = Object.assign(llmEventTagged, {
     toolInputStart: llmEventTagged.guards["tool-input-start"],
     toolInputDelta: llmEventTagged.guards["tool-input-delta"],
     toolInputEnd: llmEventTagged.guards["tool-input-end"],
+    toolInputError: llmEventTagged.guards["tool-input-error"],
     toolCall: llmEventTagged.guards["tool-call"],
     toolResult: llmEventTagged.guards["tool-result"],
     toolError: llmEventTagged.guards["tool-error"],
@@ -293,29 +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)
@@ -350,7 +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>>
@@ -378,7 +375,7 @@ const appendEvent = (state: ResponseState, event: LLMEvent): ResponseState => {
     return {
       ...state,
       events,
-      finishReason: state.finishReason ?? "error",
+      finishReason: state.finishReason ?? { normalized: "error" },
     }
   }
   return {
@@ -548,6 +545,10 @@ const reduceResponseState = (state: ResponseState, event: LLMEvent): ResponseSta
       return reduceToolInputDelta(next, event)
     case "tool-input-end":
       return reduceToolInputEnd(next, event)
+    case "tool-input-error": {
+      const { [event.id]: _finished, ...toolInputs } = next.toolInputs
+      return { ...next, toolInputs }
+    }
     case "tool-call":
       return reduceToolCall(next, event)
     case "tool-result":
@@ -561,7 +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":
@@ -126,6 +124,7 @@ export const ToolCallPart = Object.assign(
     name: Schema.String,
     input: Schema.Unknown,
     providerExecuted: Schema.optional(Schema.Boolean),
+    cache: Schema.optional(CacheHint),
     metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
     providerMetadata: Schema.optional(ProviderMetadata),
   }).annotate({ identifier: "LLM.Content.ToolCall" }),
@@ -170,6 +169,7 @@ export const ReasoningPart = Schema.Struct({
   type: Schema.Literal("reasoning"),
   text: Schema.String,
   encrypted: Schema.optional(Schema.String),
+  cache: Schema.optional(CacheHint),
   metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
   providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Content.Reasoning" })
@@ -261,13 +261,6 @@ 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 +271,6 @@ 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 +288,6 @@ 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

@@ -123,6 +123,7 @@ export const mergeGenerationOptions = (...items: ReadonlyArray<GenerationOptions
 
 export class ModelLimits extends Schema.Class<ModelLimits>("LLM.ModelLimits")({
   context: Schema.optional(Schema.Number),
+  input: Schema.optional(Schema.Number),
   output: Schema.optional(Schema.Number),
 }) {}
 
@@ -166,8 +167,13 @@ export namespace ModelDefaults {
 export const ModelToolSchemaCompatibility = Schema.Literals(["gemini", "moonshot"])
 export type ModelToolSchemaCompatibility = Schema.Schema.Type<typeof ModelToolSchemaCompatibility>
 
+export const ModelMaxTokensFieldCompatibility = Schema.Literals(["max_completion_tokens", "max_tokens"])
+export type ModelMaxTokensFieldCompatibility = Schema.Schema.Type<typeof ModelMaxTokensFieldCompatibility>
+
 export class ModelCompatibility extends Schema.Class<ModelCompatibility>("LLM.ModelCompatibility")({
   toolSchema: Schema.optional(ModelToolSchemaCompatibility),
+  reasoningField: Schema.optional(Schema.String),
+  maxTokensField: Schema.optional(ModelMaxTokensFieldCompatibility),
 }) {}
 
 export namespace ModelCompatibility {
@@ -177,7 +183,8 @@ 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
@@ -192,8 +199,8 @@ export class Model {
     this.compatibility = input.compatibility
   }
 
-  static make(input: Model.Input) {
+  static make<Options extends ProviderOptions = ProviderOptions>(input: Model.Input) {
-    return new Model({
+    return new Model<Options>({
       id: ModelID.make(input.id),
       provider: ProviderID.make(input.provider),
       route: input.route,
@@ -202,7 +209,7 @@ 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,
@@ -212,9 +219,9 @@ 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,
     })
@@ -240,6 +247,8 @@ 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")({
@@ -250,11 +259,11 @@ 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
+// usual. `"auto"` is the recommended default for agent loops — it places
-// breakpoint at the last tool definition, one at the last system part, and one
+// breakpoints at the last tool definition, the first and last distinct system
-// at the latest user message. The combination of provider invalidation
+// parts, and the conversation tail. The rolling message breakpoint keeps a
-// hierarchy (tools → system → messages) and Anthropic/Bedrock's 20-block
+// prior cache entry within Anthropic/Bedrock's 20-block lookback during long
-// lookback means three trailing breakpoints reliably cover the static prefix.
+// 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)),
     ),
   )
@@ -68,10 +68,29 @@ const result = (call: ToolCallPart, value: ToolResultValueType | ToolSettlement,
     events:
       settlement.result.type === "error"
         ? [
-            LLMEvent.toolError({ id: call.id, name: call.name, message: String(settlement.result.value), error }),
+            LLMEvent.toolError({
-            LLMEvent.toolResult({ id: call.id, name: call.name, result: settlement.result }),
+              id: call.id,
+              name: call.name,
+              message: String(settlement.result.value),
+              error,
+              providerMetadata: call.providerMetadata,
+            }),
+            LLMEvent.toolResult({
+              id: call.id,
+              name: call.name,
+              result: settlement.result,
+              providerMetadata: call.providerMetadata,
+            }),
           ]
-        : [LLMEvent.toolResult({ id: call.id, name: call.name, result: settlement.result, output: settlement.output })],
+        : [
+            LLMEvent.toolResult({
+              id: call.id,
+              name: call.name,
+              result: settlement.result,
+              output: settlement.output,
+              providerMetadata: call.providerMetadata,
+            }),
+          ],
   }
 }
 

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* compileRequest(
-      const prepared = yield* llm.prepare(
+        LLMRequest.update(request, { model: updateModel(request.model, { route: configuredGemini }) }),
-        LLM.updateRequest(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(
+      const prepared = yield* compileRequest(
-        LLM.updateRequest(request, { model: updateModel(request.model, { route: duplicate }) }),
+        LLMRequest.update(request, { model: updateModel(request.model, { route: duplicate }) }),
       )
 
       expect(prepared.body).toEqual({ body: "late-default" })

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

@@ -1,7 +1,6 @@
 import { Config } from "effect"
-import type { Auth } from "../src/route/auth"
+import { Auth } from "../src/route"
 import type { ModelFactory } from "../src/route/auth-options"
-import { Auth as RuntimeAuth } from "../src/route/auth"
 import * as OpenAIChat from "../src/protocols/openai-chat"
 import * as AmazonBedrock from "../src/providers/amazon-bedrock"
 import * as Anthropic from "../src/providers/anthropic"
@@ -11,7 +10,9 @@ import * as Cloudflare from "../src/providers/cloudflare"
 import * as GitHubCopilot from "../src/providers/github-copilot"
 import * as Google from "../src/providers/google"
 import * as GoogleVertex from "../src/providers/google-vertex"
-import * as GoogleVertexAnthropic from "../src/providers/google-vertex-anthropic"
+import * as GoogleVertexChat from "../src/providers/google-vertex-chat"
+import * as GoogleVertexMessages from "../src/providers/google-vertex-messages"
+import * as GoogleVertexResponses from "../src/providers/google-vertex-responses"
 import * as OpenAI from "../src/providers/openai"
 import * as OpenAICompatible from "../src/providers/openai-compatible"
 import * as OpenRouter from "../src/providers/openrouter"
@@ -26,7 +27,7 @@ type Model = {
   readonly id: string
 }
 
-declare const auth: Auth
+declare const auth: Auth.Definition
 declare const optionalAuthModel: ModelFactory<BaseOptions, "optional", Model>
 declare const requiredAuthModel: ModelFactory<BaseOptions, "required", Model>
 const configApiKey = Config.redacted("OPENAI_API_KEY")
@@ -74,9 +75,9 @@ OpenAI.responses("gpt-4.1-mini")
 OpenAI.configure({}).responses("gpt-4.1-mini")
 OpenAI.configure({ apiKey: "sk-test" }).responses("gpt-4.1-mini")
 OpenAI.configure({ apiKey: configApiKey }).responses("gpt-4.1-mini")
-OpenAI.configure({ auth: RuntimeAuth.bearer("oauth-token") }).responses("gpt-4.1-mini")
+OpenAI.configure({ auth: Auth.bearer("oauth-token") }).responses("gpt-4.1-mini")
 OpenAI.configure({
-  auth: RuntimeAuth.headers({ authorization: "Bearer gateway" }),
+  auth: Auth.headers({ authorization: "Bearer gateway" }),
   baseURL: "https://gateway.example.com/v1",
 }).responses("gpt-4.1-mini")
 OpenAI.configure({
@@ -100,51 +101,62 @@ OpenAI.configure({ generation: { maxTokens: "many" } })
 OpenAI.configure({ providerOptions: { openai: { store: "false" } } })
 
 // @ts-expect-error auth is an override, so OpenAI rejects apiKey with auth.
-OpenAI.configure({ apiKey: "sk-test", auth: RuntimeAuth.bearer("oauth-token") })
+OpenAI.configure({ apiKey: "sk-test", auth: Auth.bearer("oauth-token") })
 
 OpenAI.chat("gpt-4.1-mini")
 OpenAI.configure({ apiKey: "sk-test" }).chat("gpt-4.1-mini")
 OpenAI.configure({ apiKey: configApiKey }).chat("gpt-4.1-mini")
-OpenAI.configure({ auth: RuntimeAuth.bearer("oauth-token") }).chat("gpt-4.1-mini")
+OpenAI.configure({ auth: Auth.bearer("oauth-token") }).chat("gpt-4.1-mini")
 
 // @ts-expect-error OpenAI chat selectors only accept model ids.
 OpenAI.configure({ apiKey: "sk-test" }).chat("gpt-4.1-mini", {})
 
 // @ts-expect-error auth is an override, so OpenAI Chat rejects apiKey with auth.
-OpenAI.configure({ apiKey: "sk-test", auth: RuntimeAuth.bearer("oauth-token") })
+OpenAI.configure({ apiKey: "sk-test", auth: Auth.bearer("oauth-token") })
 
 // @ts-expect-error Azure requires at least one of `resourceName` or `baseURL`.
 Azure.configure()
 Azure.configure({ apiKey: "azure-key", resourceName: "resource" }).responses("deployment")
 Azure.configure({ apiKey: configApiKey, resourceName: "resource" }).responses("deployment")
-Azure.configure({ auth: RuntimeAuth.header("api-key", "azure-key"), resourceName: "resource" }).responses("deployment")
+Azure.configure({ auth: Auth.header("api-key", "azure-key"), resourceName: "resource" }).responses("deployment")
 
 // @ts-expect-error Azure model selectors only accept deployment ids.
 Azure.configure({ apiKey: "azure-key", resourceName: "resource" }).responses("deployment", {})
 
 // @ts-expect-error auth is an override, so Azure rejects apiKey with auth.
-Azure.configure({ resourceName: "resource", apiKey: "azure-key", auth: RuntimeAuth.header("api-key", "override") })
+Azure.configure({ resourceName: "resource", apiKey: "azure-key", auth: Auth.header("api-key", "override") })
 
 Azure.configure({ apiKey: "azure-key", resourceName: "resource" }).chat("deployment")
 Azure.configure({ apiKey: configApiKey, resourceName: "resource" }).chat("deployment")
-Azure.configure({ auth: RuntimeAuth.header("api-key", "azure-key"), resourceName: "resource" }).chat("deployment")
+Azure.configure({ auth: Auth.header("api-key", "azure-key"), resourceName: "resource" }).chat("deployment")
 
 // @ts-expect-error Azure chat model selectors only accept deployment ids.
 Azure.configure({ apiKey: "azure-key", resourceName: "resource" }).chat("deployment", {})
 
 // @ts-expect-error auth is an override, so Azure Chat rejects apiKey with auth.
-Azure.configure({ resourceName: "resource", apiKey: "azure-key", auth: RuntimeAuth.header("api-key", "override") })
+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" })
@@ -158,12 +170,21 @@ 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: RuntimeAuth.bearer("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.
 GoogleVertex.configure({ apiKey: "vertex-key" }).model("gemini-3.5-flash", {})
 // @ts-expect-error Vertex Gemini config accepts only one auth source.
@@ -171,21 +192,59 @@ GoogleVertex.configure({ accessToken: "vertex-token", apiKey: "vertex-key", proj
 // @ts-expect-error Vertex Gemini package settings accept only one auth source.
 GoogleVertex.model("gemini-3.5-flash", { accessToken: "vertex-token", apiKey: "vertex-key", project: "project" })
 
-GoogleVertexAnthropic.configure({ accessToken: "vertex-token", project: "project" }).model("claude-sonnet-4-6")
+GoogleVertexChat.configure({ accessToken: "vertex-token", project: "project" }).model("deepseek-ai/deepseek-v3.2-maas")
-// @ts-expect-error Vertex Anthropic package settings do not accept API keys.
+GoogleVertexChat.configure({ auth: Auth.bearer("vertex-token"), project: "project" }).model(
-GoogleVertexAnthropic.model("claude-sonnet-4-6", { apiKey: "vertex-key", project: "project" })
+  "deepseek-ai/deepseek-v3.2-maas",
-GoogleVertexAnthropic.configure({ auth: RuntimeAuth.bearer("vertex-token"), project: "project" }).model(
-  "claude-sonnet-4-6",
 )
-GoogleVertexAnthropic.configure({ accessToken: "vertex-token", project: "project" }).model(
+// @ts-expect-error Vertex Chat package settings do not accept API keys.
-  "claude-sonnet-4-6",
+GoogleVertexChat.model("deepseek-ai/deepseek-v3.2-maas", { apiKey: "vertex-key", project: "project" })
-  // @ts-expect-error Vertex Anthropic model selectors only accept model ids.
+GoogleVertexChat.configure({ accessToken: "vertex-token", project: "project" }).model(
+  "deepseek-ai/deepseek-v3.2-maas",
+  // @ts-expect-error Vertex Chat model selectors only accept model ids.
   {},
 )
-GoogleVertexAnthropic.configure({
+GoogleVertexChat.configure({
   accessToken: "vertex-token",
-  // @ts-expect-error Vertex Anthropic config accepts only one auth source.
+  // @ts-expect-error Vertex Chat config accepts only one auth source.
-  auth: RuntimeAuth.bearer("vertex-token"),
+  auth: Auth.bearer("vertex-token"),
+  project: "project",
+})
+
+GoogleVertexResponses.configure({ accessToken: "vertex-token", project: "project" }).model("xai/grok-4.20-reasoning")
+GoogleVertexResponses.configure({ auth: Auth.bearer("vertex-token"), project: "project" }).model(
+  "xai/grok-4.20-reasoning",
+)
+// @ts-expect-error Vertex Responses package settings do not accept API keys.
+GoogleVertexResponses.model("xai/grok-4.20-reasoning", { apiKey: "vertex-key", project: "project" })
+GoogleVertexResponses.configure({ accessToken: "vertex-token", project: "project" }).model(
+  "xai/grok-4.20-reasoning",
+  // @ts-expect-error Vertex Responses model selectors only accept model ids.
+  {},
+)
+GoogleVertexResponses.configure({
+  accessToken: "vertex-token",
+  // @ts-expect-error Vertex Responses config accepts only one auth source.
+  auth: Auth.bearer("vertex-token"),
+  project: "project",
+})
+
+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")
+GoogleVertexMessages.configure({ accessToken: "vertex-token", project: "project" }).model(
+  "claude-sonnet-4-6",
+  // @ts-expect-error Vertex Messages model selectors only accept model ids.
+  {},
+)
+GoogleVertexMessages.configure({
+  accessToken: "vertex-token",
+  // @ts-expect-error Vertex Messages config accepts only one auth source.
+  auth: Auth.bearer("vertex-token"),
   project: "project",
 })
 

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
+      // A single system block is both the first and last boundary, so the auto
-      // user message both get cache_control markers.
+      // 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.",
@@ -136,24 +137,61 @@ describe("request option precedence", () => {
     ),
   )
 
-  it.effect("rejects raw body overlays for protocol-owned roots", () =>
+  it.effect("transforms the final HTTP request after serialization and authentication", () =>
-    Effect.gen(function* () {
+    LLMClient.generate(
-      const model = OpenAIChat.route
+      LLM.request({
-        .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
+        model: OpenAIChat.route
-        .model({ id: "gpt-4o-mini" })
+          .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("fresh-key") })
-      const error = yield* LLMClient.prepare(
+          .model({ id: "gpt-4o-mini" }),
-        LLM.request({
+        prompt: "Say hello.",
-          model,
+      }),
-          prompt: "Say hello.",
+      {
-          http: { body: { model: "gpt-5", messages: [], tools: [] } },
+        transform: (request) =>
-        }),
+          Effect.sync(() => {
-      ).pipe(Effect.flip)
+            expect(request.headers.authorization).toBe("Bearer fresh-key")
+            request.url = "https://proxy.test/v1/chat/completions"
+            request.headers["x-plugin"] = "transformed"
+            request.body = JSON.stringify({ transformed: true })
+          }),
+      },
+    ).pipe(
+      Effect.provide(
+        dynamicResponse((input) =>
+          Effect.gen(function* () {
+            const web = yield* HttpClientRequest.toWeb(input.request).pipe(Effect.orDie)
+            expect(web.url).toBe("https://proxy.test/v1/chat/completions")
+            expect(web.headers.get("x-plugin")).toBe("transformed")
+            expect(decodeJson(input.text)).toEqual({ transformed: true })
+            return input.respond(sseEvents(deltaChunk({}, "stop")), {
+              headers: { "content-type": "text/event-stream" },
+            })
+          }),
+        ),
+      ),
+    ),
+  )
 
-      expect(error.reason).toMatchObject({
+  it.effect("applies raw body overlays after protocol lowering", () =>
-        _tag: "InvalidRequest",
+    LLMClient.generate(
-        message: "http.body cannot overlay protocol-owned field(s): model, messages, tools",
+      LLM.request({
-      })
+        model: OpenAIChat.route
-    }),
+          .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
+          .model({ id: "gpt-4o-mini" }),
+        prompt: "Say hello.",
+        http: { body: { model: "gpt-5", messages: [], tools: [] } },
+      }),
+    ).pipe(
+      Effect.provide(
+        dynamicResponse((input) =>
+          Effect.gen(function* () {
+            expect(decodeJson(input.text)).toMatchObject({ model: "gpt-5", messages: [], tools: [] })
+            return input.respond(sseEvents(deltaChunk({}, "stop")), {
+              headers: { "content-type": "text/event-stream" },
+            })
+          }),
+        ),
+      ),
+    ),
   )
 
   it.effect("uses model output limits after route limits and before call maxTokens", () =>
@@ -164,10 +202,8 @@ 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>(
+      const withoutMaxTokens = yield* compileRequest(LLM.request({ model, prompt: "Say hello.", cache: "none" }))
-        LLM.request({ model, prompt: "Say hello.", cache: "none" }),
+      const withMaxTokens = yield* compileRequest(
-      )
-      const withMaxTokens = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
         LLM.request({ model, prompt: "Say hello.", cache: "none", generation: { maxTokens: 32 } }),
       )
 

packages/ai/test/exports.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, test } from "bun:test"
-import { LLM, LLMClient, Provider } from "@opencode-ai/ai"
+import { ImageInput, LLM, LLMClient, Provider } from "@opencode-ai/ai"
 import { Route, Protocol } from "@opencode-ai/ai/route"
 import { Provider as ProviderSubpath } from "@opencode-ai/ai/provider"
 import {
@@ -16,16 +16,20 @@ import {
   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", () => {
     expect(LLM.request).toBeFunction()
     expect(LLMClient.Service).toBeFunction()
     expect(LLMClient.layer).toBeDefined()
+    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", () => {
@@ -49,9 +53,7 @@ describe("public exports", () => {
     expect(CloudflareWorkersAI.configure).toBeFunction()
     expect(CloudflareWorkersAI.configure({ accountId: "fixture", apiKey: "fixture" }).model).toBeFunction()
     expect(OpenRouter.model).toBeFunction()
-    expect(OpenRouter.provider.model).toBe(OpenRouter.model)
     expect(XAI.model).toBeFunction()
-    expect(XAI.provider.model).toBe(XAI.model)
     expect(XAI.provider.responses).toBe(XAI.responses)
     expect(XAI.provider.chat).toBe(XAI.chat)
     expect(XAI.configure({ apiKey: "fixture" }).responses("grok-4.3").route.id).toBe("openai-responses")
@@ -78,7 +80,9 @@ 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")