feat(cli): continue run after rejecting a permission ask

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

.changeset/directory-page-heading.md

@@ -1,5 +0,0 @@
----
-"@opencode/core": patch
----
-
-Correct directory page headings when the read offset is zero.

.github/workflows/publish.yml

@@ -24,6 +24,10 @@ on:
         description: "Override version (optional)"
         required: false
         type: string
+      release_notes:
+        description: "Reviewed V2 release notes for the Discord announcement (optional)"
+        required: false
+        type: string
 
 concurrency: ${{ github.workflow }}-${{ github.ref }}-${{ (github.ref_name == 'v2' && (inputs.version || inputs.bump) && 'release') || inputs.version || inputs.bump }}
 
@@ -653,3 +657,19 @@ jobs:
           OPENCODE_DESKTOP_DIST: /tmp/desktop
           CLOUDFLARE_ACCOUNT_ID: 15d29c8639fd3733b1b5486a2acfd968
           CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}
+
+  notify-discord-v2:
+    needs:
+      - version
+      - publish
+    if: github.repository == 'anomalyco/opencode' && github.ref_name == 'v2' && needs.version.outputs.release && needs.publish.result == 'success'
+    runs-on: blacksmith-4vcpu-ubuntu-2404
+    steps:
+      # Unlike dev, V2 publishes a tag rather than a GitHub Release event.
+      - name: Announce V2 release in Discord
+        uses: SethCohen/github-releases-to-discord@24d166886aee4646d448c8a389ff9e1ebcab3682 # v1.20.0
+        with:
+          webhook_url: ${{ secrets.DISCORD_WEBHOOK }}
+          release_name: OpenCode V2 ${{ needs.version.outputs.tag }}
+          release_body: ${{ inputs.release_notes }}
+          release_html_url: https://github.com/${{ github.repository }}/tree/${{ needs.version.outputs.tag }}

flake.lock

@@ -2,11 +2,11 @@
   "nodes": {
     "nixpkgs": {
       "locked": {
-        "lastModified": 1776683584,
-        "narHash": "sha256-NuTLMrr10Tng72hurYG8jYQ4XKK8wnpJmOGcPiis96g=",
+        "lastModified": 1790510107,
+        "narHash": "sha256-EVMNYv7hYDDD9TGVT/hIyTYgpiXA8y3m5xIEIxuGNU0=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "9dd5558b06dbdacbf635a3dd36dce1b1a7ee3a89",
+        "rev": "3181085bfd08663b6b9e60bc7a8395c2aaa741bd",
         "type": "github"
       },
       "original": {

packages/ai/AGENTS.md

@@ -98,11 +98,11 @@ When a provider supports multiple physical transports, selection remains executi
 
 Media does not fit the SSE-frames-to-event-state-machine LLM route. `MediaRoute.inline(...)` / `queued(...)` / `stream(...)` (`src/route/media.ts`) compose a `MediaProtocol` kind with `Endpoint` and `Auth` and own the transport plumbing: `http` option merging, URL/query rendering, auth headers, JSON vs multipart encoding, and handing the response back to the protocol. `MediaProtocol.inline` (`src/route/media-protocol.ts`) is `body.from(request)` plus `response.decode(response, context)`; each protocol declares `const route = MediaProtocol.identity({ id, name, provider })` once and decodes through `route.decodeJson` / `route.text` / `route.decodeStarted` so decode failures retain the raw body and HTTP context, raising `route.unsupported(operation, message)` for requests it cannot lower, and passes `route` as the first argument to `MediaProtocol.inline` / `queued` / `stream`. `Generation` (`src/generation.ts`) is the provider-neutral handle for a queued generation over a `GenerationRoute` (`status`, `result`, `cancel`). Image protocol files follow the same section order as LLM protocols and declare unsupported common fields once through the protocol's `unsupported` list.
 
-`MediaProtocol.queued` is the submit-then-poll kind every video route uses: `start` (body + decode into `{ token, snapshot }`), `status`, `result`, and optional `cancel` (with `activeOnly` when the provider's cancel endpoint deletes finished work, as Runway's does: the route refreshes status first and skips terminal generations), each addressed by a route-owned `token` whose `Schema.Codec` makes it serializable. `MediaRoute.inline` and `MediaRoute.queued` compose the two kinds with `Endpoint` and `Auth`; the queued route decodes the token once at the boundary (`start` output or `resume` input) and closes over it in a token-free `GenerationRoute` (`status`/`result`/`cancel` are plain Effects), so `Generation` never sees the token's shape and only carries the encoded JSON for persistence. Polls reuse the route's auth and deployment headers plus the request's `http` overlay after `start`, and resolve relative paths against the route base URL (provider-issued absolute URLs such as fal's `status_url` pass through). `result` is always its own GET even when the provider returns output inside the status document, so `Generation.await` behaves the same after `start` and after `resume`. `PollContext.auth` carries only what `Auth` added or changed so protocols can hand download credentials to output assets as transient `Media.Asset.headers` (Veo) — never part of `source` or JSON. Status strings map through a per-protocol `STATUS` table via `MediaProtocol.status`; terminal generations without output fail through `output.ended` / `output.contentPolicy` with the provider document on `reason.body`. `GenerationAwaitOptions` (`AwaitOptions` in `src/generation.ts`, `{ poll?: Poll }`) is the one options type for `await`, `events`, `Video.generate`, and `Video.stream`.
+`MediaProtocol.queued` is the submit-then-poll kind every video route uses: `start` (body + decode into `{ token, snapshot }`), `status`, `result`, and optional `cancel` (with `activeOnly` when the provider's cancel endpoint deletes finished work, as Runway's does: the route refreshes status first and skips terminal generations), each addressed by a route-owned `token` whose `Schema.Codec` makes it serializable. `MediaRoute.inline` and `MediaRoute.queued` compose the two kinds with `Endpoint` and `Auth`; the queued route decodes the token once at the boundary (`start` output or `resume` input) and closes over it in a token-free `GenerationRoute` (`status`/`result`/`cancel` are plain Effects), so `Generation` never sees the token's shape and only carries the encoded JSON for persistence. Polls reuse the route's auth and deployment headers plus the request's `http` overlay after `start`, and resolve relative paths against the route base URL (provider-issued absolute URLs such as fal's `status_url` pass through). `result` is always its own GET even when the provider returns output inside the status document, so `Generation.await` behaves the same after `start` and after `resume`. `PollContext.auth` carries only what `Auth` added or changed so protocols can hand download credentials to output assets as transient `Media.Asset.headers` (Veo) — never part of `source` or JSON. Status strings map through a per-protocol `STATUS` table via `MediaProtocol.status`; terminal generations without output fail through `output.ended` / `output.contentPolicy` with the provider document on `reason.body`; a `failed` generation maps the provider's error code through a per-protocol `FAILURE` table via `MediaProtocol.failure` so rejected inputs are not reported as retryable `ProviderInternal`. `GenerationAwaitOptions` (`AwaitOptions` in `src/generation.ts`, `{ poll?: Poll }`) is the one options type for `await`, `events`, `Video.generate`, and `Video.stream`.
 
 `MediaProtocol.stream` is the incremental kind every speech route uses, with the same discipline as LLM protocols. `MediaRoute.stream` submits the caller's request as `MediaProtocol.Addressed<Request>` (`{ ...request, mode }`, `mode: "generate" | "stream"`), so one provider stays one protocol: `body.from`, the endpoint path, and `frames` read `request.mode` to pick the body, path, and framing. `frames(bytes, context)` returns frames — `Framing.sse`, `Framing.lines`, `Framing.document` (a single-document response shaped like a streamed record), or the raw `bytes` for chunked audio. `initial()` is fresh per-response parser state; `step` folds each frame into it and emits modality events; `finish(state, context)` runs once after the last frame with the request, body, and observed `http` (header-only usage lives there) and emits exactly one terminal event or fails with `route.incomplete()`. Keep parser state to real accumulators and derive anything the request or body determines in `finish`. `generate` runs the same stream and folds it with the modality's `collect`. Request-derived URL parameters go on the body's `query` (array values repeat the parameter), applied before route and caller `http.query`. Decode frames with `route.decodeFrame` and raise stream-time failures with `route.frameError` (the frame stays on `reason.body`); protocols never thread HTTP context, because the route fills `reason.http` on stream errors that lack it. Speech protocols share `protocols/utils/speech-stream.ts` for deltas, timestamps, voice ids, PCM and container descriptions, and the terminal asset.
 
-Every modality route is the inline | stream | queued union (transcription uses all three: OpenAI and Gemini stream, Deepgram is inline, AssemblyAI is queued), every client is `MediaClient.make(Service, { modality, responseEvents })` (`src/media-client.ts`), which dispatches on the route's `kind`, and every model composes through `composeRoute`. fal queue protocols come from `protocols/utils/fal-queue.ts`, bodies are `json`, `multipart`, or `binary` (a raw upload), and a queued protocol that must upload media before submitting implements `start.prepare` (`MediaProtocol.Prepare`; AssemblyAI `/v2/upload`).
+Every modality route is the inline | stream | queued union (transcription uses all three: OpenAI and Gemini stream, Deepgram and ElevenLabs are inline, AssemblyAI is queued), every client is `MediaClient.make(Service, { modality, responseEvents })` (`src/media-client.ts`), which dispatches on the route's `kind`, and every model composes through `composeRoute`. fal queue protocols come from `protocols/utils/fal-queue.ts`, bodies are `json`, `multipart`, or `binary` (a raw upload), and a queued protocol that must upload media before submitting implements `start.prepare` (`MediaProtocol.Prepare`; AssemblyAI `/v2/upload`).
 
 ### URL Construction
 

packages/ai/README.md

@@ -129,8 +129,9 @@ VercelAIGateway.configure().experimental.evaluation("typesafe-ai/jev")
 
 OpenRouter reads `OPENROUTER_API_KEY`. Vercel reads `AI_GATEWAY_API_KEY`, then `VERCEL_OIDC_TOKEN`.
 The common API uses `boolean`; System One routes lower it to native `noul`.
-Choice and score confidence plus score legends remain available in provider metadata, and the
-provider's rounded probabilities are returned unchanged.
+Choice and score answers include `confidence` when the provider returns it, such as
+`response.answers.department.confidence`. Score legends remain available in provider metadata, and
+the provider's rounded probabilities are returned unchanged.
 
 ## Alibaba Cloud Model Studio
 
@@ -752,7 +753,10 @@ const events = Video.stream({ model: Runway.configure({ apiKey }).video("gen4.5"
 
 Status polls, result fetches, cancels, and asset downloads all run through the same request executor with the route's
 auth. `Generation.await` and `Generation.events` fail with a
-`Timeout` reason when `poll.timeout` (default 10 minutes) elapses. Failed,
+`Timeout` reason when `poll.timeout` (default 10 minutes) elapses. Status polls and result fetches retry transient
+failures (rate limits, provider 5xx, network errors) with backoff that honors `retry-after`, always within
+`poll.timeout`; submits and cancels never retry. Interrupting a wait (or aborting its `signal`) does not cancel the
+provider job, which keeps running and billing: call `cancel()` to stop it. Failed,
 cancelled, and expired generations fail typed with the provider's terminal document on `reason.body`; moderation
 outcomes (Veo `raiMediaFilteredReasons`, xAI `respect_moderation`, Runway `SAFETY.*` codes) surface as `notices` when
 a video is still returned and as a `ContentPolicy` reason when nothing is.
@@ -773,7 +777,9 @@ Provider notes:
 The promise client exposes the same surface: `ai.video.start(...)` resolves to a handle with `await`, `events`,
 `result`, `refresh`, `cancel`, and `token`; `ai.video.generate`, `ai.video.resume(model, token)`, and
 `ai.video.stream` mirror the Effect API. The handle's `status` and `progress` are a snapshot from when it was
-created; `refresh()` resolves to a new handle.
+created; `refresh()` resolves to a new handle. Every promise method and stream accepts `{ signal }`: like `fetch`,
+aborting rejects the Promise or throws from the `for await` loop with `signal.reason` (an `AbortError` `DOMException`
+unless `abort(reason)` passed one), while `break` stops a stream without throwing.
 
 ```ts
 import { ai } from "@opencode/ai/promise"
@@ -871,11 +877,12 @@ for await (const event of ai.speech.stream({ model, text: "Hello from OpenCode."
 ## Transcription
 
 Transcription (speech-to-text) is the one modality whose providers use every route kind: OpenAI and Gemini stream,
-Deepgram answers inline, and AssemblyAI is queued. `Transcription.generate` and `Transcription.stream` work on all of
-them; `Transcription.start` / `resume` return a `Generation` on queued routes and fail with `UnsupportedOperation`
-elsewhere. Models come from `.transcription(...)` selectors on the `OpenAI`, `Google`, `Deepgram`, and `AssemblyAI`
-facades. Common fields (`language`, `prompt`, `timestamps: "none" | "segment" | "word"`, `diarize`, `speakers`) lower
-natively or fail with a typed `AIError` before any network call; a route may return more than asked.
+Deepgram and ElevenLabs answer inline, and AssemblyAI is queued. `Transcription.generate` and `Transcription.stream`
+work on all of them; `Transcription.start` / `resume` return a `Generation` on queued routes and fail with
+`UnsupportedOperation` elsewhere. Models come from `.transcription(...)` selectors on the `OpenAI`, `Google`,
+`Deepgram`, `ElevenLabs`, and `AssemblyAI` facades. Common fields (`language`, `prompt`,
+`timestamps: "none" | "segment" | "word"`, `diarize`, `speakers`) lower natively or fail with a typed `AIError` before
+any network call; a route may return more than asked.
 
 ```ts
 import { Console, Effect, Stream } from "effect"
@@ -887,7 +894,7 @@ const openai = OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY })
 const program = Effect.gen(function* () {
   const audio = yield* Media.file("./call.mp3")
 
-  // Speaker-labelled segments; labels are provider-native strings ("A", "0", "spk:0").
+  // Speaker-labelled segments; labels are provider-native strings ("A", "0", "spk:0", "speaker_0").
   const response = yield* Transcription.generate({
     model: Deepgram.configure({ apiKey }).transcription("nova-3"),
     audio,
@@ -897,7 +904,7 @@ const program = Effect.gen(function* () {
   response.text // "Hello from OpenCode."
   response.segments // [{ text, startSeconds, endSeconds, speaker: "0" }]
   response.words // [{ text, startSeconds, endSeconds, speaker, confidence }]
-  response.language // the provider's own value, lowercased ("en", "english", "en_us")
+  response.language // the provider's own value, lowercased ("en", "eng", "english", "en_us")
 
   // Text deltas as the model transcribes, then one finish carrying the whole transcript.
   yield* Transcription.stream({ model: openai.transcription("gpt-4o-mini-transcribe"), audio }).pipe(
@@ -921,7 +928,12 @@ Provider notes:
 - **OpenAI** takes inline audio only; `diarize` needs `gpt-4o-transcribe-diarize`, timestamps need `whisper-1`, and `whisper-1` does not stream.
 - **Gemini** needs a transcribe model (`gemini-3.5-transcribe`); `prompt` and `speakers` fail typed.
 - **Deepgram** detects the language unless `language` is set; vocabulary goes in `providerOptions.keyterm`.
-- **AssemblyAI** uploads inline audio before submitting and is the only route that accepts `speakers`.
+- **ElevenLabs** (`scribe_v2`) uploads inline audio as the multipart `file` and sends a URL as `source_url`. Words
+  always carry timestamps, and segments are speaker turns, so `diarize`, `timestamps: "segment"`, or `speakers` turns
+  on diarization. `speakers` is an upper bound (`num_speakers`); `prompt` fails typed (vocabulary goes in
+  `providerOptions.keyterms`), as do webhook delivery and per-channel output (`use_multi_channel` without
+  `multichannel_output_style: "combined"`).
+- **AssemblyAI** uploads inline audio before submitting and treats `speakers` as the exact speaker count.
 
 The promise client mirrors the Effect API:
 

packages/ai/docs/media-design.md

@@ -1,7 +1,6 @@
 # Media generation in `@opencode/ai` — public API direction
 
-Status: phases 1–4 implemented (through Image queued routes and partial images; ElevenLabs Scribe transcription
-pending); phase 5 proposal.
+Status: phases 1–4 implemented (through Image queued routes and partial images); phase 5 proposal.
 
 ## Goal
 
@@ -271,8 +270,8 @@ Deferred: `Speech.session(...)` — input-streaming TTS where text arrives incre
 #### Transcription (STT)
 
 Shipped as the second half of phase 3 (`src/transcription.ts`, `src/transcription-client.ts`, protocols
-`openai-transcription`, `google-transcription`, `deepgram-transcription`, `assemblyai-transcription`; new `AssemblyAI`
-facade).
+`openai-transcription`, `google-transcription`, `deepgram-transcription`, `elevenlabs-transcription`,
+`assemblyai-transcription`; new `AssemblyAI` facade).
 
 ```ts
 const request = Transcription.request({
@@ -281,7 +280,7 @@ const request = Transcription.request({
   language: "en",                                  // provider-native passthrough
   timestamps: "segment",                           // none | segment | word
   diarize: true,
-  speakers: 2,                                     // exact speaker count (AssemblyAI only)
+  speakers: 2,                                     // speaker count (AssemblyAI exact, ElevenLabs maximum)
   providerOptions: { known_speaker_names: ["agent"] },
 })
 
@@ -309,17 +308,23 @@ upload); `packages/ai/AGENTS.md` (Media Routes) describes both.
 Settled rules:
 
 - **Timestamps.** A granularity the selected route or model cannot produce fails as `UnsupportedOperation`
-  (`media.timestamps`), following Speech; a route that returns more than asked (Deepgram and AssemblyAI always return
-  words) is not stripped. Segments always carry start and end times: Gemini times each transcription part from its
+  (`media.timestamps`), following Speech; a route that returns more than asked (Deepgram, ElevenLabs, and AssemblyAI
+  always return words) is not stripped. Segments always carry start and end times: Gemini times each transcription part from its
   word offsets, so segment timestamps and diarization also request word offsets there.
 - **Diarization.** `diarize` means segments (and words, where the provider labels them) carry `speaker`. Labels are
-  provider-native strings — OpenAI `A` or a known speaker name, Deepgram `0`, Gemini `spk:0`, AssemblyAI `A` — with no
-  cross-provider speaker model. `speakers` is the exact number of speakers to label, which AssemblyAI (`speakers_expected`, the only route that
-  accepts it) treats as a constraint rather than a hint.
+  provider-native strings — OpenAI `A` or a known speaker name, Deepgram `0`, Gemini `spk:0`, AssemblyAI `A`,
+  ElevenLabs `speaker_0` — with no cross-provider speaker model. `speakers` is the number of speakers to label:
+  AssemblyAI (`speakers_expected`) treats it as an exact constraint rather than a hint, and ElevenLabs
+  (`num_speakers`) as the maximum. Both turn on diarization for it; the other routes reject it.
+- **Segments from words.** ElevenLabs returns only a token list (`word`, `spacing`, `audio_event`), so its segments
+  are speaker turns: consecutive words and spacing with one `speaker_id`, text joined from the provider's own spacing
+  tokens. `words` drops spacing and audio events. Segments therefore need diarization, which `timestamps: "segment"`
+  turns on, as AssemblyAI's utterances need speaker labels.
 - **Language** is passed through (`language`, OpenAI `gpt-transcribe` `languages[]`, Gemini `languageCodes`,
-  AssemblyAI `language_code`). `response.language` is the provider's own value, lowercased but not normalized: an
-  ISO code on most routes (AssemblyAI's detection returns `en`), `english` from whisper-1. Deepgram and AssemblyAI
-  assume English unless asked to detect, so a missing `language` enables their detection.
+  AssemblyAI and ElevenLabs `language_code`). `response.language` is the provider's own value, lowercased but not
+  normalized: an ISO code on most routes (AssemblyAI's detection returns `en`, ElevenLabs ISO 639-3 `eng`), `english`
+  from whisper-1. Deepgram and AssemblyAI assume English unless asked to detect, so a missing `language` enables their
+  detection.
 - **Gemini** requires a transcribe model; other model ids fail with `UnsupportedOperation` before the call, because
   general models ignore `audioTranscriptionConfig` and answer conversationally. Streamed chunks carry whole speaker
   turns (one part per turn), which join with a space.
@@ -332,11 +337,12 @@ Settled rules:
 | OpenAI | stream (`stream: true` in `stream` mode; `whisper-1` ignores `stream`, so it emits only `finish`) | multipart `file` (inline only) | `whisper-1` (`verbose_json`); diarize model: `segment` | `gpt-4o-transcribe-diarize` (`diarized_json`) | `speakers`; `prompt` on the diarize model | `tokens` or `seconds` |
 | Gemini | stream (`generateContent` / `streamGenerateContent`) | `inlineData` or Gemini Files `fileData` | `audioTranscriptionConfig.wordTimestamp` | `audioTranscriptionConfig.diarization` | `prompt`, `speakers` | `tokens` |
 | Deepgram | inline | raw body, or JSON `{ url }` | words always; `segment` → `utterances` | `diarize_model=latest` + `utterances` | `prompt`, `speakers` | `seconds` (`metadata.duration`) |
+| ElevenLabs | inline | multipart `file`, or `source_url` | words always; `segment` → `diarize` (speaker turns) | `diarize` | `prompt`; `webhook`, per-channel `use_multi_channel` | `seconds` (`audio_duration_secs`) |
 | AssemblyAI | queued (upload → submit → poll) | `/v2/upload` then `audio_url`, or a URL | words always; `segment` → `speaker_labels` | `speaker_labels` | — | `seconds` (`audio_duration`) |
 
 Deferred: `Transcription.session(...)` — realtime STT over WebSocket (Deepgram live, AssemblyAI streaming, ElevenLabs
 realtime, OpenAI realtime transcription) — is the same future scoped `session` shape as input-streaming TTS and ships
-with the realtime work in phase 5. ElevenLabs Scribe is not implemented yet.
+with the realtime work in phase 5.
 
 ### `Generation` — shared async execution
 
@@ -361,6 +367,10 @@ Poll = { interval?: Duration; timeout?: Duration }
 
 `Generation` is not video-specific. Image routes on BFL, fal, Replicate, and Stability `upscale()` are queued; `Image.start` exists for them. A route declares itself `inline` or `queued`; `generate` on a queued route is `start` then `await`.
 
+Status polls and result reads retry transient failures (rate limits, provider 5xx, and transport errors, classified by the same `isRetryable` the Session runner uses) inside `MediaRoute.queued`. Only the HTTP exchange retries, never the decoded document: a terminal `failed` generation also surfaces as `ProviderInternal` and must not be re-read. Gaps grow exponentially from 1s with jitter, up to 30s each, honoring a provider `retry-after` up to that cap, for at most 8 retries. `await`, `events`, and `Video.stream` cut retries off at `poll.timeout` and fail with `Timeout`, so retries never extend the caller's deadline; a direct `result()` or `resume` read is bounded by the retry cap alone. `start` and `cancel` never retry: a repeated submit can start and bill a second job. The policy is internal; there is no option for it.
+
+Interrupting `await`, `events`, or `Video.stream` (or aborting the promise API's `signal`) stops waiting only. The provider job keeps running and billing; call `cancel()` explicitly to stop it.
+
 ### Usage
 
 ```ts
@@ -402,7 +412,7 @@ for await (const event of ai.llm.stream(request)) { … }
 await ai.dispose()
 ```
 
-Streams become `AsyncIterable` via `Stream.toAsyncIterable`. `AIError` is thrown as-is. `AbortSignal` maps to interruption. Nothing in `src/*` except this entrypoint knows about promises.
+Streams become `AsyncIterable` via `Stream.toAsyncIterable`. `AIError` is thrown as-is. Aborting an `AbortSignal` interrupts the work and, like `fetch`, rejects the Promise or throws from the stream with `signal.reason` instead of ending the stream as if complete. Nothing in `src/*` except this entrypoint knows about promises.
 
 ### Providers
 
@@ -414,7 +424,7 @@ implemented):
 | `OpenAI` | responses (default), chat | Images API (stream) | *Sora skipped (decision 8)* | ✓ | ✓ | |
 | `Google` | Gemini | Gemini-native | Veo | Gemini TTS | `gemini-3.5-transcribe` | |
 | `XAI` | ✓ | ✓ | ✓ | | | |
-| `ElevenLabs` | | | | ✓ | *Scribe (pending)* | *soundEffect, music (phase 5)* |
+| `ElevenLabs` | | | | ✓ | Scribe | *soundEffect, music (phase 5)* |
 | `Cartesia` | | | | ✓ | | |
 | `Deepgram` | | | | Aura | ✓ | |
 | `Fal` | | ✓ (queued) | ✓ | | | |
@@ -467,7 +477,7 @@ Foundation + Image ship together as the reference implementation, serially. Vide
 
 1. **Foundation** — per-modality selectors, `Media`, `Generation`, `Poll`, `Usage` union, `MediaProtocol` kinds, `@opencode/ai/promise` with `llm` + `image`. Port the five existing image protocols onto it. Unify `MediaPart` and add the `media` LLM event (fixes Gemini image output being dropped).
 2. **Video** — ✅ Veo, xAI, fal, Runway shipped (`MediaProtocol.queued`, `Video.start/generate/resume/stream`, promise `ai.video`). Deferred: `Video.complete` (webhooks), Luma, Kling, MiniMax, Replicate.
-3. **Speech + Transcription** — ✅ Speech: OpenAI, Gemini TTS, ElevenLabs, Cartesia, Deepgram shipped (`MediaProtocol.stream`, `Speech.generate/stream`, promise `ai.speech`). ✅ Transcription: OpenAI, Gemini, Deepgram, AssemblyAI shipped across all three route kinds (`Transcription.generate/stream/start/resume`, promise `ai.transcription`). Pending: ElevenLabs Scribe. Deferred: `Speech.session` and `Transcription.session` (WebSocket streaming).
+3. **Speech + Transcription** — ✅ Speech: OpenAI, Gemini TTS, ElevenLabs, Cartesia, Deepgram shipped (`MediaProtocol.stream`, `Speech.generate/stream`, promise `ai.speech`). ✅ Transcription: OpenAI, Gemini, Deepgram, ElevenLabs Scribe, AssemblyAI shipped across all three route kinds (`Transcription.generate/stream/start/resume`, promise `ai.transcription`). Deferred: `Speech.session` and `Transcription.session` (WebSocket streaming).
 4. **Image queued routes and partials** — ✅ BFL, fal, Replicate, and Stability creative upscale queued; Stability generate inline; OpenAI `partial_images` streaming (`image-partial` restored). Imagen dropped: shut down on the Gemini API and discontinued on Vertex (2026-06-30). Deferred: Stability's synchronous edit and fast/conservative upscale endpoints.
 5. **Later** — ElevenLabs music/SFX, Lyria, `Speech.session` / `Transcription.session`, realtime.
 

packages/ai/src/experimental/evaluation.ts

@@ -63,6 +63,7 @@ export const ChoiceAnswer = Schema.Struct({
   type: Schema.Literal("choice"),
   choice: Schema.String,
   probabilities: Schema.optional(Schema.Record(Schema.String, Probability)),
+  confidence: Schema.optional(Probability),
 })
 export type ChoiceAnswer = Schema.Schema.Type<typeof ChoiceAnswer>
 
@@ -70,6 +71,7 @@ export const ScoreAnswer = Schema.Struct({
   type: Schema.Literal("score"),
   score: Schema.Number,
   probabilities: Schema.optional(Schema.Record(Schema.String, Probability)),
+  confidence: Schema.optional(Probability),
 })
 export type ScoreAnswer = Schema.Schema.Type<typeof ScoreAnswer>
 
@@ -92,6 +94,7 @@ export type AnswerFor<Question extends EvaluationQuestion> = Question extends {
       readonly type: "choice"
       readonly choice: Extract<keyof Criteria, string>
       readonly probabilities?: Readonly<Record<Extract<keyof Criteria, string>, number>>
+      readonly confidence?: number
     }
   : Question extends { readonly type: "score" }
     ? ScoreAnswer

packages/ai/src/experimental/system-one.ts

@@ -142,32 +142,37 @@ export const model = <Options extends EvaluationOptions = EvaluationOptions>(cfg
             Effect.mapError((cause) => fail("System One returned an invalid response", cause, text)),
           )
 
-          const confidence: Record<string, number> = {}
           const legend: Record<string, Record<string, Schema.Json>> = {}
           const answers = Object.fromEntries(
             Object.entries(data.answers).map(([id, answer]): [string, EvaluationAnswer] => {
               if (answer.type === "noul") return [id, { type: "boolean", probability: answer.noul }]
               if (answer.type === "choice") {
-                if (answer.confidence !== undefined) confidence[id] = answer.confidence
                 return [
                   id,
                   {
                     type: "choice",
                     choice: answer.choice,
                     probabilities: answer.probabilities,
+                    ...(answer.confidence === undefined ? {} : { confidence: answer.confidence }),
                   },
                 ]
               }
-              if (answer.confidence !== undefined) confidence[id] = answer.confidence
               if (answer.legend !== undefined) legend[id] = answer.legend
-              return [id, { type: "score", score: answer.score, probabilities: answer.probabilities }]
+              return [
+                id,
+                {
+                  type: "score",
+                  score: answer.score,
+                  probabilities: answer.probabilities,
+                  ...(answer.confidence === undefined ? {} : { confidence: answer.confidence }),
+                },
+              ]
             }),
           )
           const meta = {
             ...(data.id === undefined ? {} : { responseId: data.id }),
             ...(data.provider === undefined ? {} : { provider: data.provider }),
             ...data.provider_metadata?.[cfg.providerMetadataKey],
-            ...(Object.keys(confidence).length === 0 ? {} : { confidence }),
             ...(Object.keys(legend).length === 0 ? {} : { legend }),
           }
           return new EvaluationResponse({

packages/ai/src/generation.ts

@@ -102,7 +102,7 @@ export class Generation<Response> {
     return settled.pipe(
       // Non-completed terminal states also go through `result` so the route can surface its provider failure body.
       Effect.flatMap((generation) => generation.result()),
-      Effect.timeoutOrElse({ duration: timeout, orElse: () => this.timeoutError(timeout) }),
+      Effect.timeoutOrElse({ duration: timeout, orElse: () => timeoutError(this.id, timeout) }),
     )
   }
 
@@ -123,20 +123,7 @@ export class Generation<Response> {
       Clock.currentTimeMillis.pipe(
         Effect.map((start) => {
           const deadline = start + Duration.toMillis(timeout)
-          // Fail before polling once the deadline has passed: a fast status request could otherwise win the zero-budget
-          // race and schedule another zero-delay poll.
-          const refresh = Clock.currentTimeMillis.pipe(
-            Effect.flatMap((now) =>
-              now >= deadline
-                ? this.timeoutError(timeout)
-                : this.refresh().pipe(
-                    Effect.timeoutOrElse({
-                      duration: Duration.millis(deadline - now),
-                      orElse: () => this.timeoutError(timeout),
-                    }),
-                  ),
-            ),
-          )
+          const refresh = within(this.refresh(), this.id, timeout, deadline)
           const schedule = this.schedule(options?.poll).pipe(
             Schedule.modifyDelay((meta) =>
               Effect.succeed(Duration.min(meta.duration, Duration.millis(Math.max(0, deadline - meta.now)))),
@@ -157,15 +144,6 @@ export class Generation<Response> {
     return { type: "generation-progress", id: this.id, progress: this.progress }
   }
 
-  private timeoutError(timeout: Duration.Duration) {
-    return new AIError({
-      reason: new TimeoutError({
-        message: `Generation ${this.id} did not finish within ${Duration.format(timeout)}`,
-        timeoutMs: Duration.toMillis(timeout),
-      }),
-    })
-  }
-
   private poll(poll: Poll | undefined) {
     return this.refresh().pipe(
       Effect.repeat({ schedule: this.schedule(poll), until: (generation) => generation.terminal }),
@@ -177,12 +155,53 @@ export class Generation<Response> {
   }
 }
 
+/** `events` followed by the expanded result, with the result fetch bounded by the same `poll.timeout` deadline. */
 export const resultEvents = <Response, A>(
   generation: Generation<Response>,
   expand: (response: Response) => ReadonlyArray<A>,
   options?: AwaitOptions,
-): Stream.Stream<Observation | A, AIError> =>
-  generation.events(options).pipe(
-    Stream.filter((event): event is Observation => event.type !== "generation-finished"),
-    Stream.concat(Stream.fromIterableEffect(Effect.map(generation.result(), expand))),
+): Stream.Stream<Observation | A, AIError> => {
+  const timeout = Duration.fromInputUnsafe(options?.poll?.timeout ?? DEFAULT_POLL_TIMEOUT)
+  return Stream.unwrap(
+    Clock.currentTimeMillis.pipe(
+      Effect.map((start) =>
+        generation.events(options).pipe(
+          Stream.filter((event): event is Observation => event.type !== "generation-finished"),
+          Stream.concat(
+            Stream.fromIterableEffect(
+              within(generation.result(), generation.id, timeout, start + Duration.toMillis(timeout)).pipe(
+                Effect.map(expand),
+              ),
+            ),
+          ),
+        ),
+      ),
+    ),
   )
+}
+
+/**
+ * Run `effect` within the time left until `deadline`. Fails before starting once the deadline has passed: a fast
+ * request could otherwise win the zero-budget race and schedule another zero-delay poll.
+ */
+const within = <A>(effect: Effect.Effect<A, AIError>, id: string, timeout: Duration.Duration, deadline: number) =>
+  Clock.currentTimeMillis.pipe(
+    Effect.flatMap((now) =>
+      now >= deadline
+        ? Effect.fail(timeoutError(id, timeout))
+        : effect.pipe(
+            Effect.timeoutOrElse({
+              duration: Duration.millis(deadline - now),
+              orElse: () => Effect.fail(timeoutError(id, timeout)),
+            }),
+          ),
+    ),
+  )
+
+const timeoutError = (id: string, timeout: Duration.Duration) =>
+  new AIError({
+    reason: new TimeoutError({
+      message: `Generation ${id} did not finish within ${Duration.format(timeout)}`,
+      timeoutMs: Duration.toMillis(timeout),
+    }),
+  })

packages/ai/src/index.ts

@@ -4,7 +4,7 @@ export { ImageClient } from "./image-client.js"
 export { Auth } from "./route/auth.js"
 export { Provider } from "./provider.js"
 export { ProviderPackage } from "./provider-package.js"
-export { isContextOverflow, isContextOverflowFailure } from "./provider-error.js"
+export { isContextOverflow, isContextOverflowFailure, isRetryable } from "./provider-error.js"
 export type {
   RouteLanguageModelInput,
   RouteRoutedLanguageModelInput,

packages/ai/src/promise.ts

@@ -42,7 +42,7 @@ export type GenerationHandle<Response> = Snapshot & {
   /** Serializable JSON; pass it back to `resume` from another process. */
   readonly token: unknown
   readonly await: (options?: AwaitOptions & RunOptions) => Promise<Response>
-  /** Status observations until the first terminal one, polling like `await`; abort ends iteration without throwing. */
+  /** Status observations until the first terminal one, polling like `await`; abort throws `signal.reason`. */
   readonly events: (options?: AwaitOptions & RunOptions) => AsyncIterable<Event>
   /** The result without polling; fails when the generation has not completed. */
   readonly result: (options?: RunOptions) => Promise<Response>
@@ -50,15 +50,16 @@ export type GenerationHandle<Response> = Snapshot & {
   readonly cancel: (options?: RunOptions) => Promise<void>
 }
 
+// Fails with `signal.reason` so aborted calls reject and aborted streams throw like `fetch`: an `AbortError` by default.
 const abortEffect = (signal: AbortSignal | undefined) =>
   signal === undefined
     ? Effect.never
-    : Effect.callback<void>((resume) => {
+    : Effect.callback<never, unknown>((resume) => {
         if (signal.aborted) {
-          resume(Effect.void)
+          resume(Effect.fail(signal.reason))
           return
         }
-        const onAbort = () => resume(Effect.void)
+        const onAbort = () => resume(Effect.fail(signal.reason))
         signal.addEventListener("abort", onAbort, { once: true })
         return Effect.sync(() => signal.removeEventListener("abort", onAbort))
       })
@@ -68,14 +69,14 @@ export const make = (options: Options = {}) => {
 
   /** Run any package Effect (for example `LLMClient.compact(...)`) inside this runtime. */
   const run = <A, E>(effect: Effect.Effect<A, E, Services>, options?: RunOptions) =>
-    runtime.runPromise(effect, { signal: options?.signal })
+    runtime.runPromise(Effect.raceFirst(effect, abortEffect(options?.signal)))
 
   const iterate = <A, E>(stream: Stream.Stream<A, E, Services>, options?: RunOptions): AsyncIterable<A> =>
     Stream.toAsyncIterable(
       Stream.unwrap(
         runtime.contextEffect.pipe(
           Effect.map(
-            (context): Stream.Stream<A, E> =>
+            (context): Stream.Stream<A, unknown> =>
               stream.pipe(Stream.interruptWhen(abortEffect(options?.signal)), Stream.provideContext(context)),
           ),
         ),

packages/ai/src/protocols/deepgram-transcription.ts

@@ -6,6 +6,7 @@ import { mergeJsonRecords, type OpenString } from "../schema/index.js"
 import { TranscriptionModel, TranscriptionResponse, type TranscriptionRequestFor } from "../transcription.js"
 import { ProviderShared } from "./shared.js"
 import { MediaInput } from "./utils/media-input.js"
+import { SpeakerTurns } from "./utils/speaker-turns.js"
 
 const route = MediaProtocol.identity({ id: "deepgram-transcription", name: "Deepgram", provider: "deepgram" })
 export const DEFAULT_BASE_URL = "https://api.deepgram.com"
@@ -115,16 +116,6 @@ const speaker = (value: number | undefined) => (value === undefined ? undefined
 
 const wordText = (word: typeof Word.Type) => word.punctuated_word ?? word.word
 
-// Utterances split on pauses, not speakers: the v2 diarizer labels a whole utterance with one speaker even when its
-// words change speaker, so segments split each utterance at speaker changes.
-const speakerTurns = (words: ReadonlyArray<typeof Word.Type>) =>
-  words.reduce<Array<Array<typeof Word.Type>>>((turns, word) => {
-    const last = turns.at(-1)
-    if (last === undefined || last[0].speaker !== word.speaker) return [...turns, [word]]
-    last.push(word)
-    return turns
-  }, [])
-
 const decodeResponse = Effect.fn("DeepgramTranscription.decodeResponse")(function* (
   response: HttpClientResponse.HttpClientResponse,
 ) {
@@ -136,6 +127,8 @@ const decodeResponse = Effect.fn("DeepgramTranscription.decodeResponse")(functio
   const requestID = output.value.metadata?.request_id
   return new TranscriptionResponse({
     text: alternative.transcript,
+    // Utterances split on pauses, not speakers: the v2 diarizer labels a whole utterance with one speaker even when
+    // its words change speaker, so segments split each utterance at speaker changes.
     segments: output.value.results.utterances?.flatMap((utterance) =>
       utterance.words === undefined || utterance.words.length === 0
         ? [
@@ -146,7 +139,7 @@ const decodeResponse = Effect.fn("DeepgramTranscription.decodeResponse")(functio
               speaker: speaker(utterance.speaker),
             },
           ]
-        : speakerTurns(utterance.words).map((turn) => ({
+        : SpeakerTurns.group(utterance.words, (word) => word.speaker).map((turn) => ({
             text: turn.map(wordText).join(" "),
             startSeconds: turn[0].start,
             endSeconds: turn[turn.length - 1].end,

packages/ai/src/protocols/elevenlabs-transcription.ts

@@ -0,0 +1,211 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { TranscriptionModel, TranscriptionResponse, type TranscriptionRequestFor } from "../transcription.js"
+import { mediaTypeExtension } from "../utils/media-type.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+import { SpeakerTurns } from "./utils/speaker-turns.js"
+
+const route = MediaProtocol.identity({
+  id: "elevenlabs-transcription",
+  name: "ElevenLabs Transcription",
+  provider: "elevenlabs",
+})
+export const DEFAULT_BASE_URL = "https://api.elevenlabs.io"
+export const PATH = "/v1/speech-to-text"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type ElevenLabsTranscriptionOptions = {
+  readonly tag_audio_events?: boolean
+  readonly timestamps_granularity?: OpenString<"none" | "word" | "character">
+  readonly diarization_threshold?: number
+  readonly file_format?: OpenString<"pcm_s16le_16" | "other">
+  readonly temperature?: number
+  readonly seed?: number
+  readonly keyterms?: ReadonlyArray<string>
+  readonly no_verbatim?: boolean
+  readonly detect_speaker_roles?: boolean
+  readonly use_speaker_library?: boolean
+  readonly entity_detection?: string | ReadonlyArray<string>
+  readonly entity_redaction?: string | ReadonlyArray<string>
+  readonly entity_redaction_mode?: OpenString<"redacted" | "entity_type" | "enumerated_entity_type">
+} & Record<string, unknown>
+
+export type Request = TranscriptionRequestFor<ElevenLabsTranscriptionOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Response schema
+// ---------------------------------------------------------------------------
+
+/** `type` is `word`, `spacing` (the whitespace between words), or `audio_event` (`(laughter)`). */
+const Token = Schema.Struct({
+  text: Schema.String,
+  type: Schema.String,
+  start: optionalNull(Schema.Number),
+  end: optionalNull(Schema.Number),
+  speaker_id: optionalNull(Schema.String),
+  logprob: optionalNull(Schema.Number),
+})
+type Token = Schema.Schema.Type<typeof Token>
+
+const Transcript = Schema.Struct({
+  language_code: optionalNull(Schema.String),
+  text: Schema.String,
+  words: optionalNull(Schema.Array(Token)),
+  transcription_id: optionalNull(Schema.String),
+  audio_duration_secs: optionalNull(Schema.Number),
+})
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+/** Speaker turns are the only segments ElevenLabs can produce, and `num_speakers` only applies to diarization. */
+const diarizes = (request: Request) =>
+  request.diarize === true || request.timestamps === "segment" || request.speakers !== undefined
+
+const RESERVED_FORM_FIELDS = new Set([
+  "file",
+  "cloud_storage_url",
+  "source_url",
+  "model_id",
+  "language_code",
+  "diarize",
+  "num_speakers",
+])
+
+const validate = (request: Request, overlay: Record<string, unknown>) => {
+  // Webhook requests return 202 with no transcript; the result arrives at a configured webhook instead.
+  if (overlay.webhook === true)
+    return Effect.fail(route.unsupported("transcription.webhook", `${route.name} does not deliver to webhooks`))
+  // Separate multichannel output replaces the transcript with one transcript per channel.
+  if (overlay.use_multi_channel === true && overlay.multichannel_output_style !== "combined")
+    return Effect.fail(
+      route.unsupported(
+        "transcription.multichannel",
+        `${route.name} returns a single transcript; set multichannel_output_style: "combined" to merge channels`,
+      ),
+    )
+  if (overlay.timestamps_granularity === "none" && (request.timestamps === "word" || diarizes(request)))
+    return Effect.fail(
+      route.unsupported(
+        "media.timestamps",
+        `${route.name} cannot return word timestamps or speaker turns with timestamps_granularity: "none"`,
+      ),
+    )
+  return Effect.void
+}
+
+const fromRequest = Effect.fn("ElevenLabsTranscription.fromRequest")(function* (request: Request) {
+  const overlay = mergeJsonRecords(request.providerOptions, request.http?.body) ?? {}
+  yield* validate(request, overlay)
+  const form = new FormData()
+  const url = ProviderShared.mediaUrl(request.audio)
+  if (url === undefined) {
+    const extension = mediaTypeExtension(request.audio.mediaType)
+    const audio = yield* MediaInput.inlineBytes(route.id, request.audio)
+    form.append(
+      "file",
+      MediaInput.blob(audio, request.audio.mediaType),
+      extension === undefined ? "audio" : `audio.${extension}`,
+    )
+  }
+  MediaInput.appendFields(
+    form,
+    {
+      model_id: request.model.id,
+      // `cloud_storage_url` is deprecated in favor of `source_url`, which accepts any hosted audio or video URL.
+      source_url: url,
+      language_code: request.language,
+      diarize: diarizes(request) ? true : undefined,
+      num_speakers: request.speakers,
+    },
+    { overlay, reserved: RESERVED_FORM_FIELDS, repeatArrays: "key" },
+  )
+  return MediaProtocol.multipart(form)
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeTranscript = route.decodeJson(Transcript)
+
+type TimedWord = Token & { readonly start: number; readonly end: number }
+
+const isTimedWord = (token: Token): token is TimedWord =>
+  token.type === "word" && typeof token.start === "number" && typeof token.end === "number"
+
+/** Turn text keeps the provider's own spacing tokens, so languages written without spaces are not re-spaced. */
+const speakerTurns = (tokens: ReadonlyArray<Token>) =>
+  SpeakerTurns.group(
+    tokens.filter((token) => token.type === "word" || token.type === "spacing"),
+    (token) => token.speaker_id,
+  ).flatMap((turn) => {
+    const words = turn.filter(isTimedWord)
+    if (words.length === 0) return []
+    return [
+      {
+        text: turn
+          .map((token) => token.text)
+          .join("")
+          .trim(),
+        startSeconds: words[0].start,
+        endSeconds: words[words.length - 1].end,
+        speaker: turn[0].speaker_id ?? undefined,
+      },
+    ]
+  })
+
+const decodeResponse = Effect.fn("ElevenLabsTranscription.decodeResponse")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.DecodeContext<Request>,
+) {
+  const output = yield* decodeTranscript(response)
+  const transcript = output.value
+  const tokens = transcript.words ?? []
+  const duration = transcript.audio_duration_secs ?? undefined
+  const transcriptionID = transcript.transcription_id ?? undefined
+  return new TranscriptionResponse({
+    text: transcript.text,
+    segments: diarizes(context.request) ? speakerTurns(tokens) : undefined,
+    words: tokens.filter(isTimedWord).map((word) => ({
+      text: word.text,
+      startSeconds: word.start,
+      endSeconds: word.end,
+      speaker: word.speaker_id ?? undefined,
+      confidence: typeof word.logprob === "number" ? Math.exp(word.logprob) : undefined,
+    })),
+    language: transcript.language_code?.toLowerCase(),
+    durationSeconds: duration,
+    usage: duration === undefined ? undefined : { type: "seconds", seconds: duration },
+    providerMetadata: transcriptionID === undefined ? undefined : { elevenlabs: { transcriptionId: transcriptionID } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.inline<Request, TranscriptionResponse>(route, {
+  unsupported: ["prompt"],
+  body: { from: fromRequest },
+  response: { decode: decodeResponse },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  TranscriptionModel.fromRoute<ElevenLabsTranscriptionOptions>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: PATH },
+    input,
+  )
+
+export const ElevenLabsTranscription = {
+  protocol,
+  model,
+} as const

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

@@ -36,7 +36,9 @@ const StartResponse = Schema.Struct({ name: Schema.String })
 
 const Operation = Schema.Struct({
   done: Schema.optional(Schema.Boolean),
-  error: Schema.optional(Schema.Struct({ message: Schema.optional(Schema.String) })),
+  error: Schema.optional(
+    Schema.Struct({ code: Schema.optional(Schema.Number), message: Schema.optional(Schema.String) }),
+  ),
   response: Schema.optional(
     Schema.Struct({
       generateVideoResponse: Schema.optional(
@@ -60,6 +62,16 @@ const Operation = Schema.Struct({
   metadata: Schema.optional(Schema.Unknown),
 })
 
+// Operation errors are `google.rpc.Status`; unlisted codes (INTERNAL, UNAVAILABLE, ...) are provider-side.
+const FAILURE = {
+  3: "InvalidRequest", // INVALID_ARGUMENT
+  7: "Authentication", // PERMISSION_DENIED
+  8: "RateLimit", // RESOURCE_EXHAUSTED
+  9: "InvalidRequest", // FAILED_PRECONDITION
+  11: "InvalidRequest", // OUT_OF_RANGE
+  16: "Authentication", // UNAUTHENTICATED
+} as const satisfies Record<number, MediaProtocol.Failure>
+
 // ---------------------------------------------------------------------------
 // 5. Request body construction
 // ---------------------------------------------------------------------------
@@ -154,6 +166,7 @@ const decodeResult = Effect.fn("GoogleVideo.decodeResult")(function* (
     return yield* output.ended(
       "failed",
       `${route.name} operation failed${operation.error?.message === undefined ? "" : `: ${operation.error.message}`}`,
+      MediaProtocol.failure(FAILURE, operation.error?.code),
     )
   const generated = operation.response?.generateVideoResponse
   // Downloads require the same API key as the poll; the asset carries it transiently and follows the redirect.

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

@@ -64,6 +64,14 @@ const OpenAIChatTool = Schema.Struct({
 })
 type OpenAIChatTool = Schema.Schema.Type<typeof OpenAIChatTool>
 
+// Gemini's OpenAI-compatible surface carries thought signatures in tool call
+// `extra_content` and rejects replayed parallel calls without them:
+// https://ai.google.dev/gemini-api/docs/thinking#signatures
+const ExtraContent = Schema.Struct({
+  google: Schema.Struct({ thought_signature: Schema.String }),
+})
+const decodeExtraContent = (value: unknown) => Option.getOrUndefined(Schema.decodeUnknownOption(ExtraContent)(value))
+
 const OpenAIChatAssistantToolCall = Schema.Struct({
   id: Schema.String,
   type: Schema.tag("function"),
@@ -71,6 +79,7 @@ const OpenAIChatAssistantToolCall = Schema.Struct({
     name: Schema.String,
     arguments: Schema.String,
   }),
+  extra_content: Schema.optional(ExtraContent),
 })
 type OpenAIChatAssistantToolCall = Schema.Schema.Type<typeof OpenAIChatAssistantToolCall>
 
@@ -112,12 +121,6 @@ const decodeReasoningDetail = Schema.decodeUnknownOption(ReasoningDetail)
 const knownReasoningDetails = (details: ReadonlyArray<unknown>) =>
   details.flatMap((detail) => Option.toArray(decodeReasoningDetail(detail)))
 
-// Intentionally omit Gemini's provider-specific `extra_content.google.thought_signature`
-// extension until direct Google OpenAI-compatible routing is supported here:
-// https://github.com/vercel/ai/issues/11590
-// https://github.com/vercel/ai/pull/11745
-// https://ai.google.dev/gemini-api/docs/thought-signatures#openai
-
 const OpenAIChatUserContent = Schema.Union([
   Schema.Struct({
     type: Schema.Literal("text"),
@@ -242,6 +245,7 @@ const OpenAIChatToolCallDelta = Schema.Struct({
   index: optionalNull(Schema.Number),
   id: optionalNull(Schema.String),
   function: optionalNull(OpenAIChatToolCallDeltaFunction),
+  extra_content: optionalNull(Schema.Unknown),
 })
 type OpenAIChatToolCallDelta = Schema.Schema.Type<typeof OpenAIChatToolCallDelta>
 
@@ -294,6 +298,7 @@ interface PendingToolDelta {
   readonly id?: string
   readonly name?: string
   readonly input: string
+  readonly extraContent?: Schema.Schema.Type<typeof ExtraContent>
 }
 
 export interface ParserState {
@@ -347,13 +352,17 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
     tool: (name) => ({ type: "function" as const, function: { name } }),
   })
 
-const lowerToolCall = (part: ToolCallPart, options: LoweringOptions): OpenAIChatAssistantToolCall => ({
+const lowerToolCall = (
+  part: ToolCallPart,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
+): OpenAIChatAssistantToolCall => ({
   id: options.toolCallID?.(part.id) ?? part.id,
   type: "function",
   function: {
     name: part.name,
     arguments: ProviderShared.encodeJson(part.input === undefined ? {} : part.input),
   },
+  extra_content: decodeExtraContent(part.providerMetadata?.[options.providerMetadataKey]?.extraContent),
 })
 
 const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart) {
@@ -721,7 +730,9 @@ const detectSupportsStore = (provider: string, baseURL: string | undefined): boo
     p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
   const isAntLing = p === "ant-ling" || url.includes("api.ant-ling.com")
   const isOpencode = p === "opencode" || url.includes("opencode.ai")
+  const isGemini = url.includes("generativelanguage.googleapis.com")
   const isNonStandard =
+    isGemini ||
     isNvidia ||
     isCerebras ||
     isXai ||
@@ -1114,12 +1125,13 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
       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 ?? ""}`
+      const extraContent = pending?.extraContent ?? decodeExtraContent(tool.extra_content)
       latestToolIndex = index
       nextToolIndex = Math.max(nextToolIndex, index + 1)
       if (!current && (!id || !name)) {
         pendingTools = {
           ...pendingTools,
-          [index]: { id: id || undefined, name: name || undefined, input: text },
+          [index]: { id: id || undefined, name: name || undefined, input: text, extraContent },
         }
         continue
       }
@@ -1131,7 +1143,12 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
         ADAPTER,
         tools,
         index,
-        { id: id || undefined, name: name || undefined, text },
+        {
+          id: id || undefined,
+          name: name || undefined,
+          text,
+          providerMetadata: extraContent && { [state.providerMetadataKey]: { extraContent } },
+        },
         "OpenAI Chat tool call delta is missing id or name",
       )
       if (ToolStream.isError(result))

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

@@ -193,7 +193,7 @@ const fromRequest = Effect.fn("OpenAITranscription.fromRequest")(function* (requ
     {
       overlay: mergeJsonRecords(request.providerOptions, request.http?.body),
       reserved: RESERVED_FORM_FIELDS,
-      repeatArrays: true,
+      repeatArrays: "key[]",
     },
   )
   return MediaProtocol.multipart(form)

packages/ai/src/protocols/runway-video.ts

@@ -137,7 +137,12 @@ const decodeResult = Effect.fn("RunwayVideo.decodeResult")(function* (
     const message = `${route.name} task failed${code === undefined ? "" : ` (${code})`}${task.failure ? `: ${task.failure}` : ""}`
     // Runway failure codes are dotted paths; every moderation outcome carries a SAFETY segment.
     if (code !== undefined && /(^|\.)SAFETY(\.|$)/.test(code)) return yield* output.contentPolicy(message)
-    return yield* output.ended("failed", message)
+    // ASSET.INVALID rejects the caller's input media; Runway documents it as not retryable.
+    return yield* output.ended(
+      "failed",
+      message,
+      code !== undefined && /^ASSET\.INVALID(\.|$)/.test(code) ? "InvalidRequest" : "ProviderInternal",
+    )
   }
   if (status === "cancelled")
     return yield* output.ended("cancelled", `${route.name} task ${context.token.taskID} was cancelled`)

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

@@ -71,8 +71,9 @@ export const imageOutput = (
 }
 
 /**
- * Append multipart text fields: strings as-is, other values as JSON, or arrays as repeated `key[]` parts with
- * `repeatArrays`. `overlay` keys in `reserved` are dropped so `http.body` cannot replace route-owned fields.
+ * Append multipart text fields: strings as-is, other values as JSON, or scalar arrays as one part per item with
+ * `repeatArrays`, named `key[]` or `key`. `overlay` keys in `reserved` are dropped so `http.body` cannot replace
+ * route-owned fields.
  */
 export const appendFields = (
   form: FormData,
@@ -80,13 +81,13 @@ export const appendFields = (
   options: {
     readonly overlay?: Record<string, unknown>
     readonly reserved: ReadonlySet<string>
-    readonly repeatArrays?: true
+    readonly repeatArrays?: "key[]" | "key"
   },
 ) => {
   const overlay = Object.entries(options.overlay ?? {}).filter(([key]) => !options.reserved.has(key))
   Object.entries(mergeJsonRecords(fields, Object.fromEntries(overlay)) ?? {}).forEach(([key, value]) => {
-    if (Array.isArray(value) && options.repeatArrays)
-      return value.forEach((item) => form.append(`${key}[]`, String(item)))
+    if (Array.isArray(value) && value.every(isScalar) && options.repeatArrays !== undefined)
+      return value.forEach((item) => form.append(options.repeatArrays === "key[]" ? `${key}[]` : key, String(item)))
     form.append(key, typeof value === "string" ? value : encodeJson(value))
   })
 }

packages/ai/src/protocols/utils/speaker-turns.ts

@@ -0,0 +1,10 @@
+/** Split an ordered token list into runs of consecutive tokens with the same speaker. */
+export const group = <Item>(items: ReadonlyArray<Item>, speaker: (item: Item) => unknown) =>
+  items.reduce<Array<Array<Item>>>((turns, item) => {
+    const last = turns.at(-1)
+    if (last === undefined || speaker(last[0]) !== speaker(item)) return [...turns, [item]]
+    last.push(item)
+    return turns
+  }, [])
+
+export * as SpeakerTurns from "./speaker-turns.js"

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

@@ -147,7 +147,12 @@ export const appendOrStart = <K extends StreamKey>(
   route: string,
   tools: State<K>,
   key: K,
-  delta: { readonly id?: string; readonly name?: string; readonly text: string },
+  delta: {
+    readonly id?: string
+    readonly name?: string
+    readonly text: string
+    readonly providerMetadata?: ProviderMetadata
+  },
   missingToolMessage: string,
 ): AppendOutcome<K> | AIError => {
   const current = tools[key]
@@ -161,7 +166,7 @@ export const appendOrStart = <K extends StreamKey>(
     namespace: current?.namespace,
     input: `${current?.input ?? ""}${delta.text}`,
     providerExecuted: current?.providerExecuted,
-    providerMetadata: current?.providerMetadata,
+    providerMetadata: current?.providerMetadata ?? delta.providerMetadata,
   }
   if (current && delta.text.length === 0 && current.id === id && current.name === name)
     return { tools, tool: current, events: [] }

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

@@ -65,6 +65,13 @@ const STATUS = {
   expired: "expired",
 } as const satisfies Record<string, Status>
 
+// Documented video error codes; `service_unavailable`, `internal_error`, and unknown codes are provider-side.
+const FAILURE = {
+  invalid_argument: "InvalidRequest",
+  failed_precondition: "InvalidRequest",
+  permission_denied: "Authentication",
+} as const satisfies Record<string, MediaProtocol.Failure>
+
 // ---------------------------------------------------------------------------
 // 5. Request body construction
 // ---------------------------------------------------------------------------
@@ -143,6 +150,7 @@ const decodeResult = Effect.fn("XAIVideo.decodeResult")(function* (
     return yield* output.ended(
       "failed",
       `${route.name} generation failed${code === undefined ? "" : ` (${code})`}${message === undefined ? "" : `: ${message}`}`,
+      MediaProtocol.failure(FAILURE, code),
     )
   }
   if (status !== "completed")

packages/ai/src/provider-error.ts

@@ -58,6 +58,47 @@ export const isContextOverflowFailure = (failure: unknown) =>
     ? failure.reason._tag === "InvalidRequest" && failure.reason.classification === "context-overflow"
     : Schema.is(ProviderErrorEvent)(failure) && failure.classification === "context-overflow"
 
+/**
+ * Whether a failed call may succeed when sent again: rate limits, provider-side failures, transport failures that did
+ * not deliver an accepted write, and unrecognized failures. Callers decide which calls are safe to repeat.
+ */
+export const isRetryable = (error: AIError) => {
+  const override = error.reason.http?.headers["x-should-retry"]
+  if (override === "true") return true
+  if (override === "false") return false
+  switch (error.reason._tag) {
+    case "RateLimit":
+    case "ProviderInternal":
+      return true
+    // A WebSocket acknowledgment marks delivery accepted before model output may exist.
+    // Read failures can still recover; the caller chooses retry versus continuation from durable output.
+    case "Transport":
+      return (
+        error.reason.delivery !== "rejected" &&
+        (error.reason.delivery !== "accepted" || error.reason.operation === "read")
+      )
+    case "InvalidProviderOutput":
+      return error.reason.classification === "incomplete-stream"
+    // Unrecognized failures retry: classification records affirmative
+    // deterministic evidence, and transient failures are exactly the ones
+    // that arrive in shapes no classifier anticipates.
+    case "UnknownProvider":
+      return true
+    case "Authentication":
+    case "QuotaExceeded":
+    case "ContentPolicy":
+    case "InvalidRequest":
+    case "UnsupportedOperation":
+    case "NoRoute":
+    case "Timeout":
+      return false
+    default: {
+      const exhaustive: never = error.reason
+      return exhaustive
+    }
+  }
+}
+
 const decodeJson = Schema.decodeUnknownOption(Schema.fromJsonString(Schema.Unknown))
 // OpenCode Zen reports account caps as typed 429/402 errors that are not throttles.
 const QUOTA_CODES = new Set([
@@ -68,7 +109,8 @@ const QUOTA_CODES = new Set([
   "freeusagelimiterror",
   "creditlimitexceeded",
 ])
-const AUTH_CODES = new Set(["authentication_error", "permission_error"])
+// Google reports an invalid API key as HTTP 400 INVALID_ARGUMENT with this `details[].reason`.
+const AUTH_CODES = new Set(["authentication_error", "permission_error", "api_key_invalid"])
 const SERVER_CODES = new Set([
   "api_error",
   "internal_error",
@@ -218,6 +260,10 @@ function providerCodes(value: unknown) {
     error?.type,
     error?.status,
     error?.error_type,
+    // Google `google.rpc.ErrorInfo` details carry the specific reason.
+    ...(Array.isArray(error?.details)
+      ? error.details.map((detail) => (isRecord(detail) ? detail.reason : undefined))
+      : []),
     inner?.code,
     metadata?.error_type,
     responseError?.code,

packages/ai/src/providers/elevenlabs.ts

@@ -3,8 +3,10 @@ import type { ProviderAuthOption } from "../route/auth-options.js"
 import { MediaRoute } from "../route/media.js"
 import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
 import { ElevenLabsSpeech } from "../protocols/elevenlabs-speech.js"
+import { ElevenLabsTranscription } from "../protocols/elevenlabs-transcription.js"
 
 export type { ElevenLabsOutputFormat, ElevenLabsSpeechOptions } from "../protocols/elevenlabs-speech.js"
+export type { ElevenLabsTranscriptionOptions } from "../protocols/elevenlabs-transcription.js"
 
 export const id = ProviderID.make("elevenlabs")
 
@@ -24,12 +26,15 @@ const auth = (options: ProviderAuthOption<"optional">) => {
 export const configure = (input: Config = {}) => {
   const media = MediaRoute.deployment(input, auth(input))
   const speech = (modelID: string | ModelID) => ElevenLabsSpeech.model({ ...media, id: modelID })
+  const transcription = (modelID: string | ModelID) => ElevenLabsTranscription.model({ ...media, id: modelID })
   return {
     id,
     speech,
+    transcription,
     configure,
   }
 }
 
 export const provider = configure()
 export const speech = provider.speech
+export const transcription = provider.transcription

packages/ai/src/route/media-protocol.ts

@@ -5,12 +5,14 @@ import { Media } from "../media.js"
 import type { AuthInput } from "./auth.js"
 import {
   AIError,
+  AuthenticationError,
   ContentPolicyError,
   HttpContext,
   InvalidProviderOutputError,
   InvalidRequestError,
   ProviderID,
   ProviderInternalError,
+  RateLimitError,
   UnsupportedOperationError,
 } from "../schema/index.js"
 
@@ -188,6 +190,16 @@ export const stream = <Request, Event, Frame, State>(
 // Response helpers
 // ---------------------------------------------------------------------------
 
+/** Reasons a provider can report for a `failed` generation; anything it does not classify is `ProviderInternal`. */
+const FAILURES = {
+  InvalidRequest: InvalidRequestError,
+  Authentication: AuthenticationError,
+  RateLimit: RateLimitError,
+  ProviderInternal: ProviderInternalError,
+}
+
+export type Failure = keyof typeof FAILURES
+
 const context = (response: HttpClientResponse.HttpClientResponse) =>
   new HttpContext({ url: response.request.url, status: response.status, headers: response.headers })
 
@@ -199,9 +211,10 @@ export const identity = (input: { readonly id: string; readonly name: string; re
 
   /**
    * Read a text body while retaining the original payload and HTTP context on every downstream error. `invalid` is a
-   * malformed provider document; `ended` is a generation that reached a terminal status without output (`failed` is
-   * provider-side, `cancelled`/`expired` mean the result will never exist); `pending` is a `result()` read before the
-   * generation finished, which is caller misuse; `contentPolicy` is a moderated result.
+   * malformed provider document; `ended` is a generation that reached a terminal status without output (`failed`
+   * carries the provider's classification, defaulting to `ProviderInternal`; `cancelled`/`expired` mean the result
+   * will never exist); `pending` is a `result()` read before the generation finished, which is caller misuse;
+   * `contentPolicy` is a moderated result.
    */
   const text = Effect.fn("MediaProtocol.text")(function* (response: HttpClientResponse.HttpClientResponse) {
     const http = context(response)
@@ -223,11 +236,15 @@ export const identity = (input: { readonly id: string; readonly name: string; re
       http,
       invalid: (message: string, cause?: unknown) =>
         new AIError({ reason: new InvalidProviderOutputError({ route: input.id, message, body, http, cause }) }),
-      ended: (status: Exclude<Status, "queued" | "running" | "completed">, message: string) =>
+      ended: (
+        status: Exclude<Status, "queued" | "running" | "completed">,
+        message: string,
+        failure: Failure = "ProviderInternal",
+      ) =>
         new AIError({
           reason:
             status === "failed"
-              ? new ProviderInternalError({ message, body, http })
+              ? new FAILURES[failure]({ message, body, http })
               : new InvalidRequestError({ message, body, http }),
         }),
       pending: (id: string) =>
@@ -303,6 +320,10 @@ export const status = <Table extends Record<string, Status>>(
   return Effect.succeed(table[raw])
 }
 
+/** Map a provider error code through the protocol's table; missing or unmapped codes are `ProviderInternal`. */
+export const failure = (table: Readonly<Record<string, Failure>>, code: string | number | undefined): Failure =>
+  code !== undefined && Object.hasOwn(table, code) ? table[code] : "ProviderInternal"
+
 /** A `url` asset whose provider-declared retention window starts now. */
 export const expiringUrl = (url: string, retention: Duration.Duration, options?: Parameters<typeof Media.url>[1]) =>
   Clock.currentTimeMillis.pipe(

packages/ai/src/route/media.ts

@@ -1,4 +1,4 @@
-import { Effect, Schema, Stream } from "effect"
+import { Duration, Effect, Schedule, Schema, Stream } from "effect"
 import { Headers, HttpClientRequest, type HttpClientResponse } from "effect/unstable/http"
 import { Auth, type AuthInput } from "./auth.js"
 import { Endpoint } from "./endpoint.js"
@@ -7,6 +7,7 @@ import { RequestExecutor } from "./executor.js"
 import { MediaProtocol } from "./media-protocol.js"
 import { Generation, isTerminal } from "../generation.js"
 import type { Media } from "../media.js"
+import { isRetryable } from "../provider-error.js"
 import {
   AIError,
   AIErrorReason,
@@ -137,6 +138,32 @@ export const inline = <Request extends MediaRequest, Response>(
   }
 }
 
+const READ_RETRY_MAX_DELAY = Duration.seconds(30)
+
+/**
+ * Status and result reads retry transient failures; `start` and `cancel` never do. Gaps grow exponentially from 1s,
+ * jittered, up to 30s each, for at most 8 retries (about two minutes when every attempt fails), so a direct
+ * `Generation.result()` stays bounded; `await` and `events` also cut retries off at `poll.timeout`. A provider
+ * `retryAfterMs` raises the gap, still capped at 30s.
+ */
+const READ_RETRY = Schedule.max([
+  Schedule.min([Schedule.exponential("1 second"), Schedule.spaced(READ_RETRY_MAX_DELAY)]),
+  Schedule.recurs(8),
+]).pipe(
+  Schedule.jittered,
+  Schedule.setInputType<AIError>(),
+  Schedule.modifyDelay(({ input, duration }) =>
+    Effect.succeed(
+      Duration.min(
+        input.reason._tag === "RateLimit" || input.reason._tag === "ProviderInternal"
+          ? Duration.max(duration, Duration.millis(input.reason.retryAfterMs ?? 0))
+          : duration,
+        READ_RETRY_MAX_DELAY,
+      ),
+    ),
+  ),
+)
+
 /**
  * Compose a queued media protocol the same way, adding `start`/`resume` handles whose polls reuse the route's auth,
  * deployment headers, and (for `start`) the request's `http` overlay. The token is decoded once at the boundary and
@@ -154,6 +181,8 @@ export const queued = <Request extends MediaRequest, Response, Token>(
   const generationRoute = (token: Token, http: HttpOptions | undefined, execute: Execute) => {
     const materialize = (asset: Media.Asset) =>
       asset.materialize().pipe(Effect.provideService(RequestExecutorService, { execute }))
+    // Only the GET exchange retries: a decoded terminal failure (`output.ended`) can be a `ProviderInternal` too, and
+    // re-reading it would spin until the caller's deadline.
     const poll = <A>(operation: {
       readonly path: (token: Token) => string
       readonly decode: (
@@ -161,9 +190,10 @@ export const queued = <Request extends MediaRequest, Response, Token>(
         context: MediaProtocol.PollContext<Token>,
       ) => Effect.Effect<A, AIError>
     }) =>
-      transport
-        .call("GET", operation.path(token), http, execute)
-        .pipe(Effect.flatMap((sent) => operation.decode(sent.response, { token, auth: sent.auth, materialize })))
+      transport.call("GET", operation.path(token), http, execute).pipe(
+        Effect.retry({ schedule: READ_RETRY, while: isRetryable }),
+        Effect.flatMap((sent) => operation.decode(sent.response, { token, auth: sent.auth, materialize })),
+      )
     const status = poll(protocol.status)
     const cancel = protocol.cancel
     const send =

packages/ai/src/transcription.ts

@@ -103,7 +103,7 @@ export type TranscriptionRequestInput<Model extends TranscriptionModel = Transcr
 // Response and events
 // ---------------------------------------------------------------------------
 
-/** Speaker labels are provider-native (`A`, `0`, `spk:0`, or a known speaker name). */
+/** Speaker labels are provider-native (`A`, `0`, `spk:0`, `speaker_0`, or a known speaker name). */
 export const TranscriptionSegment = Schema.Struct({
   text: Schema.String,
   startSeconds: Schema.Number,

packages/ai/test/evaluation.test.ts

@@ -39,17 +39,18 @@ describe("experimental Evaluation", () => {
         type: "choice",
         choice: "billing",
         probabilities: { billing: 0.9, technical: 0.1 },
+        confidence: 0.8,
       })
       expect(response.answers.urgency).toEqual({
         type: "score",
         score: 1.2,
         probabilities: { "0": 0, "1": 0.8, "2": 0.2 },
+        confidence: 0.6,
       })
       expect(response.answers.refund).toEqual({ type: "boolean", probability: 0.97 })
       expect(response.usage?.totalTokens).toBe(36)
       expect(response.providerMetadata).toEqual({
         typesafe: {
-          confidence: { department: 0.8, urgency: 0.6 },
           legend: { urgency: { "0": "Can wait", "1": "Needs attention", "2": "Blocking" } },
         },
       })

packages/ai/test/evaluation.types.ts

@@ -26,8 +26,10 @@ const request = Evaluation.request({
 const result = EvaluationClient.evaluate(request)
 type Result = Success<typeof result>
 type Choice = Assert<Equal<Result["answers"]["topic"]["choice"], "billing" | "support">>
+type Confidence = Assert<Equal<Result["answers"]["topic"]["confidence"], number | undefined>>
 type ClientRequirements = Assert<Equal<Requirements<typeof result>, Service>>
 void (true satisfies Choice)
+void (true satisfies Confidence)
 void (true satisfies ClientRequirements)
 
 Effect.gen(function* () {

packages/ai/test/exports.test.ts

@@ -197,6 +197,11 @@ describe("public exports", () => {
     expect(Google.configure({ apiKey: "fixture" }).transcription("gemini-3.5-transcribe").route.kind).toBe("stream")
     expect(Deepgram.configure({ apiKey: "fixture" }).transcription("nova-3").route.kind).toBe("inline")
     expect(AssemblyAI.configure({ apiKey: "fixture" }).transcription("universal-3-5-pro").route.kind).toBe("queued")
+    expect(ElevenLabs.configure({ apiKey: "fixture" }).transcription("scribe_v2").route.id).toBe(
+      "elevenlabs-transcription",
+    )
+    expect(ElevenLabs.configure({ apiKey: "fixture" }).transcription("scribe_v2").route.kind).toBe("inline")
+    expect(ElevenLabs.provider.transcription).toBe(ElevenLabs.transcription)
   })
 
   test("protocol barrels expose supported low-level routes", () => {

packages/ai/test/fixtures/recordings/elevenlabs-transcription/groups-diarized-words-into-speaker-turns.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "tags": [
+      "prefix:elevenlabs-transcription",
+      "provider:elevenlabs",
+      "protocol:elevenlabs-transcription"
+    ],
+    "name": "elevenlabs-transcription/groups-diarized-words-into-speaker-turns",
+    "recordedAt": "2026-09-27T09:35:28.265Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.elevenlabs.io/v1/speech-to-text",
+        "headers": {
+          "content-type": "multipart/form-data; boundary=----WebKitFormBoundary356bdc14864a477dbacbfcf60d1ecceb"
+        },
+        "body": "--BOUNDARY\r\nContent-Disposition: form-data; name=\"file\"; filename=\"audio.mp3\"\r\nContent-Type: audio/mpeg\r\n\r\n[audio]\r\n--BOUNDARY\r\nContent-Disposition: form-data; name=\"model_id\"\r\n\r\nscribe_v2\r\n--BOUNDARY\r\nContent-Disposition: form-data; name=\"diarize\"\r\n\r\ntrue\r\n--BOUNDARY--\r\n"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"language_code\":\"eng\",\"language_probability\":0.9495430588722229,\"text\":\"Did the release ship? Yes, it shipped this morning\",\"words\":[{\"text\":\"Did\",\"start\":0.34,\"end\":0.44,\"type\":\"word\",\"speaker_id\":\"speaker_0\",\"logprob\":-1.7881377516459906e-6},{\"text\":\" \",\"start\":0.44,\"end\":0.48,\"type\":\"spacing\",\"speaker_id\":\"speaker_0\",\"logprob\":-1.1920928244535389e-7},{\"text\":\"the\",\"start\":0.48,\"end\":0.56,\"type\":\"word\",\"speaker_id\":\"speaker_0\",\"logprob\":-1.1920928244535389e-7},{\"text\":\" \",\"start\":0.56,\"end\":0.6,\"type\":\"spacing\",\"speaker_id\":\"speaker_0\",\"logprob\":-7.152531907195225e-6},{\"text\":\"release\",\"start\":0.6,\"end\":0.92,\"type\":\"word\",\"speaker_id\":\"speaker_0\",\"logprob\":-7.152531907195225e-6},{\"text\":\" \",\"start\":0.92,\"end\":0.94,\"type\":\"spacing\",\"speaker_id\":\"speaker_0\",\"logprob\":-8.344646857949556e-7},{\"text\":\"ship?\",\"start\":0.94,\"end\":1.26,\"type\":\"word\",\"speaker_id\":\"speaker_0\",\"logprob\":-7.414704032271402e-6},{\"text\":\" \",\"start\":1.26,\"end\":1.26,\"type\":\"spacing\",\"speaker_id\":\"speaker_0\",\"logprob\":-0.0009363081189803779},{\"text\":\"Yes,\",\"start\":1.68,\"end\":2.02,\"type\":\"word\",\"speaker_id\":\"speaker_1\",\"logprob\":-0.003542040009030245},{\"text\":\" \",\"start\":2.02,\"end\":2.48,\"type\":\"spacing\",\"speaker_id\":\"speaker_1\",\"logprob\":-3.933898824470816e-6},{\"text\":\"it\",\"start\":2.48,\"end\":2.62,\"type\":\"word\",\"speaker_id\":\"speaker_1\",\"logprob\":-3.933898824470816e-6},{\"text\":\" \",\"start\":2.62,\"end\":2.64,\"type\":\"spacing\",\"speaker_id\":\"speaker_1\",\"logprob\":-0.000013589766240329482},{\"text\":\"shipped\",\"start\":2.66,\"end\":2.9,\"type\":\"word\",\"speaker_id\":\"speaker_1\",\"logprob\":-0.000013589766240329482},{\"text\":\" \",\"start\":2.9,\"end\":2.94,\"type\":\"spacing\",\"speaker_id\":\"speaker_1\",\"logprob\":0.0},{\"text\":\"this\",\"start\":2.94,\"end\":3.12,\"type\":\"word\",\"speaker_id\":\"speaker_1\",\"logprob\":0.0},{\"text\":\" \",\"start\":3.12,\"end\":3.18,\"type\":\"spacing\",\"speaker_id\":\"speaker_1\",\"logprob\":-3.576278118089249e-7},{\"text\":\"morning\",\"start\":3.18,\"end\":3.5,\"type\":\"word\",\"speaker_id\":\"speaker_1\",\"logprob\":-3.576278118089249e-7}],\"transcription_id\":\"cs3I2282TH8hjw12brNg\",\"audio_duration_secs\":3.5526875}"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/elevenlabs-transcription/transcribes-audio-with-word-timestamps.json

@@ -0,0 +1,50 @@
+{
+  "version": 1,
+  "metadata": {
+    "tags": [
+      "prefix:elevenlabs-transcription",
+      "provider:elevenlabs",
+      "protocol:elevenlabs-transcription"
+    ],
+    "name": "elevenlabs-transcription/transcribes-audio-with-word-timestamps",
+    "recordedAt": "2026-09-27T09:35:27.686Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.elevenlabs.io/v1/speech-to-text",
+        "headers": {
+          "content-type": "multipart/form-data; boundary=----WebKitFormBoundarye2be7b31e94441bbbeb35a9c890a9d74"
+        },
+        "body": "--BOUNDARY\r\nContent-Disposition: form-data; name=\"file\"; filename=\"audio.mp3\"\r\nContent-Type: audio/mpeg\r\n\r\n[audio]\r\n--BOUNDARY\r\nContent-Disposition: form-data; name=\"model_id\"\r\n\r\nscribe_v2\r\n--BOUNDARY--\r\n"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"language_code\":\"eng\",\"language_probability\":0.6618340611457825,\"text\":\"Hello from OpenCode\",\"words\":[{\"text\":\"Hello\",\"start\":0.4,\"end\":0.66,\"type\":\"word\",\"logprob\":-0.000014781842764932662},{\"text\":\" \",\"start\":0.66,\"end\":0.74,\"type\":\"spacing\",\"logprob\":-3.814689989667386e-6},{\"text\":\"from\",\"start\":0.74,\"end\":0.84,\"type\":\"word\",\"logprob\":-3.814689989667386e-6},{\"text\":\" \",\"start\":0.84,\"end\":0.9,\"type\":\"spacing\",\"logprob\":-0.018268775194883347},{\"text\":\"OpenCode\",\"start\":0.9,\"end\":1.44,\"type\":\"word\",\"logprob\":-0.1251817401498556}],\"transcription_id\":\"D4VfnANM2ArCHTujIb9q\",\"audio_duration_secs\":1.54125}"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.elevenlabs.io/v1/speech-to-text",
+        "headers": {
+          "content-type": "multipart/form-data; boundary=----WebKitFormBoundaryfb80d0e44d9e44d299416ed546a04056"
+        },
+        "body": "--BOUNDARY\r\nContent-Disposition: form-data; name=\"file\"; filename=\"audio.mp3\"\r\nContent-Type: audio/mpeg\r\n\r\n[audio]\r\n--BOUNDARY\r\nContent-Disposition: form-data; name=\"model_id\"\r\n\r\nscribe_v2\r\n--BOUNDARY--\r\n"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"language_code\":\"eng\",\"language_probability\":0.6618340611457825,\"text\":\"Hello from OpenCode\",\"words\":[{\"text\":\"Hello\",\"start\":0.4,\"end\":0.66,\"type\":\"word\",\"logprob\":-0.000023007127310847864},{\"text\":\" \",\"start\":0.66,\"end\":0.74,\"type\":\"spacing\",\"logprob\":-2.3841830625315197e-6},{\"text\":\"from\",\"start\":0.74,\"end\":0.84,\"type\":\"word\",\"logprob\":-2.3841830625315197e-6},{\"text\":\" \",\"start\":0.84,\"end\":0.9,\"type\":\"spacing\",\"logprob\":-0.008306833915412426},{\"text\":\"OpenCode\",\"start\":0.9,\"end\":1.44,\"type\":\"word\",\"logprob\":-0.1075385226868093}],\"transcription_id\":\"SkYplzfq1DW8Ae3bWnoy\",\"audio_duration_secs\":1.54125}"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/openai-compatible-chat/gemini-parallel-tool-signatures.json

@@ -0,0 +1,54 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "gemini-3.8-flash",
+    "tags": [
+      "prefix:openai-compatible-chat",
+      "provider:google",
+      "protocol:openai-chat",
+      "tool",
+      "tool-loop",
+      "continuation"
+    ],
+    "name": "gemini-parallel-tool-signatures",
+    "recordedAt": "2026-09-28T03:12:05.083Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gemini-3.8-flash\",\"messages\":[{\"role\":\"system\",\"content\":\"Call get_weather for every requested city in parallel, then answer in one short sentence.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris and in Tokyo?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true}}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"role\":\"assistant\",\"tool_calls\":[{\"extra_content\":{\"google\":{\"thought_signature\":\"ErYCCrMCAWkUfRNGh+/Zbc8YUSzk1yfWAfROcjA4HyF1x69jz6167w8zd4n6kZQQ5FDeBZ5HZMEbEkQ4ENOpzsQL8roCR6wONkhXpiduWrTD6XwbP8KGNkf6D1tX/JlBh7G5Cl+0rdjiSOl/mdY1lcjbkfyCRFs5T8odNWMG7WD3rCrXJDFQ/5QfOl+tqVTceKGz2yyXBhOhvsQDU33ulR9tHQJo/Fmx3HNyDdwvmyKUXm+kgqHsYZnkdv6y6xZwy9zBXGUO4QwxelMw3Rrc24Mp2rNbEDZS3YEeP72Jn/hIP5a8XWOx6+zop/4/CjYxxSfN/tJRY5d48NAKRNFzUN1Az8SvAs/N5X2I3/5ZMaDnQWW/BVdS9fm2KFYJ2baqtBiRRnJxMq4ELArkKjGZzHpd97XG8YjV6g==\"}},\"function\":{\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\",\"name\":\"get_weather\"},\"id\":\"call_723181\",\"type\":\"function\"}]},\"index\":0}],\"created\":1790565123,\"id\":\"A9u5aoufBp3rz7IPke_3oAo\",\"model\":\"gemini-3.8-flash\",\"object\":\"chat.completion.chunk\",\"usage\":{\"completion_tokens\":16,\"prompt_tokens\":74,\"total_tokens\":132}}\n\ndata: {\"choices\":[{\"delta\":{\"role\":\"assistant\",\"tool_calls\":[{\"function\":{\"arguments\":\"{\\\"city\\\":\\\"Tokyo\\\"}\",\"name\":\"get_weather\"},\"id\":\"call_723184\",\"type\":\"function\"}]},\"index\":0}],\"created\":1790565123,\"id\":\"A9u5aoufBp3rz7IPke_3oAo\",\"model\":\"gemini-3.8-flash\",\"object\":\"chat.completion.chunk\",\"usage\":{\"completion_tokens\":32,\"prompt_tokens\":74,\"total_tokens\":148}}\n\ndata: {\"choices\":[{\"delta\":{\"role\":\"assistant\"},\"finish_reason\":\"stop\",\"index\":0}],\"created\":1790565124,\"id\":\"A9u5aoufBp3rz7IPke_3oAo\",\"model\":\"gemini-3.8-flash\",\"object\":\"chat.completion.chunk\",\"usage\":{\"completion_tokens\":32,\"prompt_tokens\":74,\"total_tokens\":148}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gemini-3.8-flash\",\"messages\":[{\"role\":\"system\",\"content\":\"Call get_weather for every requested city in parallel, then answer in one short sentence.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris and in Tokyo?\"},{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"call_723181\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},\"extra_content\":{\"google\":{\"thought_signature\":\"ErYCCrMCAWkUfRNGh+/Zbc8YUSzk1yfWAfROcjA4HyF1x69jz6167w8zd4n6kZQQ5FDeBZ5HZMEbEkQ4ENOpzsQL8roCR6wONkhXpiduWrTD6XwbP8KGNkf6D1tX/JlBh7G5Cl+0rdjiSOl/mdY1lcjbkfyCRFs5T8odNWMG7WD3rCrXJDFQ/5QfOl+tqVTceKGz2yyXBhOhvsQDU33ulR9tHQJo/Fmx3HNyDdwvmyKUXm+kgqHsYZnkdv6y6xZwy9zBXGUO4QwxelMw3Rrc24Mp2rNbEDZS3YEeP72Jn/hIP5a8XWOx6+zop/4/CjYxxSfN/tJRY5d48NAKRNFzUN1Az8SvAs/N5X2I3/5ZMaDnQWW/BVdS9fm2KFYJ2baqtBiRRnJxMq4ELArkKjGZzHpd97XG8YjV6g==\"}}},{\"id\":\"call_723184\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Tokyo\\\"}\"}}]},{\"role\":\"tool\",\"tool_call_id\":\"call_723181\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"},{\"role\":\"tool\",\"tool_call_id\":\"call_723184\",\"content\":\"{\\\"temperature\\\":0,\\\"condition\\\":\\\"unknown\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true}}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"It is currently sunny and 22°C in Paris,\",\"role\":\"assistant\"},\"index\":0}],\"created\":1790565125,\"id\":\"BNu5auWeB-2fz7IPxY6l4QY\",\"model\":\"gemini-3.8-flash\",\"object\":\"chat.completion.chunk\",\"usage\":{\"completion_tokens\":13,\"prompt_tokens\":150,\"total_tokens\":226}}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\" while Tokyo is 0°C.\",\"role\":\"assistant\"},\"index\":0}],\"created\":1790565125,\"id\":\"BNu5auWeB-2fz7IPxY6l4QY\",\"model\":\"gemini-3.8-flash\",\"object\":\"chat.completion.chunk\",\"usage\":{\"completion_tokens\":21,\"prompt_tokens\":150,\"total_tokens\":234}}\n\ndata: {\"choices\":[{\"delta\":{\"extra_content\":{\"google\":{\"thought_signature\":\"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\"}},\"role\":\"assistant\"},\"finish_reason\":\"stop\",\"index\":0}],\"created\":1790565125,\"id\":\"BNu5auWeB-2fz7IPxY6l4QY\",\"model\":\"gemini-3.8-flash\",\"object\":\"chat.completion.chunk\",\"usage\":{\"completion_tokens\":21,\"prompt_tokens\":192,\"total_tokens\":276}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/promise.test.ts

@@ -22,7 +22,8 @@ const chatBody = sseEvents(
 /**
  * Executor layer that answers chat completions with SSE text, image generations with one base64 PNG, Runway video
  * tasks with a queued submission that succeeds on the second poll, speech with raw audio or SSE audio deltas, OpenAI
- * transcription with JSON or SSE text deltas, and AssemblyAI transcripts that complete on the first poll.
+ * transcription with JSON or SSE text deltas, AssemblyAI transcripts that complete on the first poll, and `slow.test`
+ * chat completions that send one text delta and never finish.
  */
 const executor = (seen: Array<string>) =>
   RequestExecutor.layer.pipe(
@@ -55,6 +56,18 @@ const executor = (seen: Array<string>) =>
               output: "https://replicate.test/a.webp",
               urls: { get: "https://replicate.test/p_1", cancel: "https://replicate.test/p_1/cancel" },
             })
+          if (web.url.startsWith("https://slow.test"))
+            return input.respond(
+              new ReadableStream({
+                start: (controller) =>
+                  controller.enqueue(
+                    new TextEncoder().encode(
+                      `data: ${JSON.stringify({ choices: [{ delta: { content: "Hello" } }] })}\n\n`,
+                    ),
+                  ),
+              }),
+              { headers: { "content-type": "text/event-stream" } },
+            )
           if (web.url.endsWith("/chat/completions"))
             return input.respond(chatBody, { headers: { "content-type": "text/event-stream" } })
           if (web.url.endsWith("/audio/speech"))
@@ -304,6 +317,77 @@ describe("AI promise client", () => {
     await ai.dispose()
   })
 
+  test("aborted calls reject and aborted streams throw with the signal's reason", async () => {
+    const ai = AI.make({ layer: executor([]) })
+    const slow = OpenAI.configure({ apiKey: "test", baseURL: "https://slow.test/v1" }).chat("gpt-4o-mini")
+    const aborted = new AbortController()
+    aborted.abort()
+    const reason = new Error("mine")
+
+    const rejected = await ai.run(Effect.never, { signal: aborted.signal }).catch((error: unknown) => error)
+    expect(rejected).toBe(aborted.signal.reason)
+    expect(rejected).toMatchObject({ name: "AbortError" })
+
+    const inFlight = new AbortController()
+    setTimeout(() => inFlight.abort(reason), 10)
+    expect(
+      await ai.llm
+        .generate({ model: slow, prompt: "Hello" }, { signal: inFlight.signal })
+        .catch((error: unknown) => error),
+    ).toBe(reason)
+
+    const preAborted = await Array.fromAsync(
+      ai.speech.stream({ model: openai.speech("gpt-4o-mini-tts"), text: "Hello" }, { signal: aborted.signal }),
+    ).catch((error: unknown) => error)
+    expect(preAborted).toBe(aborted.signal.reason)
+    expect(preAborted).toMatchObject({ name: "AbortError" })
+
+    const midStream = new AbortController()
+    const deltas: Array<string> = []
+    const midStreamFailure = await Array.fromAsync(
+      ai.llm.stream({ model: slow, prompt: "Hello" }, { signal: midStream.signal }),
+      (event) => {
+        if (!LLMEvent.is.textDelta(event)) return
+        deltas.push(event.text)
+        midStream.abort()
+      },
+    ).catch((error: unknown) => error)
+    expect(deltas).toEqual(["Hello"])
+    expect(midStreamFailure).toBe(midStream.signal.reason)
+    expect(midStreamFailure).toMatchObject({ name: "AbortError" })
+
+    const model = Runway.configure({ apiKey: "test", baseURL: "https://runway.test/v1" }).video("gen4.5")
+    const generation = await ai.video.start({ model, prompt: "A kite" })
+    const polling = new AbortController()
+    const events: Array<string> = []
+    const eventsFailure = await Array.fromAsync(
+      generation.events({ poll: { interval: 60_000 }, signal: polling.signal }),
+      (event) => {
+        events.push(event.type)
+        polling.abort(reason)
+      },
+    ).catch((error: unknown) => error)
+    expect(events).toEqual(["generation-progress"])
+    expect(eventsFailure).toBe(reason)
+
+    await ai.dispose()
+  })
+
+  test("breaking out of an abortable stream cleans up without throwing", async () => {
+    const ai = AI.make({ layer: executor([]) })
+    const slow = OpenAI.configure({ apiKey: "test", baseURL: "https://slow.test/v1" }).chat("gpt-4o-mini")
+    const controller = new AbortController()
+    const deltas: Array<string> = []
+    for await (const event of ai.llm.stream({ model: slow, prompt: "Hello" }, { signal: controller.signal })) {
+      if (!LLMEvent.is.textDelta(event)) continue
+      deltas.push(event.text)
+      break
+    }
+    controller.abort()
+    expect(deltas).toEqual(["Hello"])
+    await ai.dispose()
+  })
+
   test("the default client is created lazily and can be disposed", async () => {
     expect(typeof AI.ai.llm.generate).toBe("function")
     expect(typeof AI.ai.image.generate).toBe("function")

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

@@ -355,6 +355,32 @@ describe("provider error rawBody classification", () => {
     }
   })
 
+  test("classifies Google invalid API keys as authentication failures", () => {
+    const rawBody = JSON.stringify({
+      error: {
+        code: 400,
+        message: "API key not valid. Please pass a valid API key.",
+        status: "INVALID_ARGUMENT",
+        details: [
+          {
+            "@type": "type.googleapis.com/google.rpc.ErrorInfo",
+            reason: "API_KEY_INVALID",
+            domain: "googleapis.com",
+          },
+          {
+            "@type": "type.googleapis.com/google.rpc.LocalizedMessage",
+            locale: "en-US",
+            message: "API key not valid. Please pass a valid API key.",
+          },
+        ],
+      },
+    })
+    expect(
+      classifyProviderFailure({ message: "API key not valid. Please pass a valid API key.", status: 400, rawBody })
+        ._tag,
+    ).toBe("Authentication")
+  })
+
   test("classifies overflow signals buried in the raw payload when the summary is vague", () => {
     const reason = classifyProviderFailure({
       message: "Request failed",

packages/ai/test/provider/elevenlabs-transcription.recorded.test.ts

@@ -0,0 +1,50 @@
+import { describe, expect } from "bun:test"
+import { Effect, Stream } from "effect"
+import { Transcription } from "../../src/index.js"
+import { ElevenLabs } from "../../src/providers.js"
+import { recordedTests } from "../recorded-test.js"
+import { TRANSCRIPT, audio, audioRecording, dialog } from "./transcription-recording.js"
+
+const model = ElevenLabs.configure({ apiKey: process.env.ELEVENLABS_API_KEY ?? "fixture" }).transcription("scribe_v2")
+
+const recorded = recordedTests({
+  prefix: "elevenlabs-transcription",
+  provider: "elevenlabs",
+  protocol: "elevenlabs-transcription",
+  requires: ["ELEVENLABS_API_KEY"],
+  options: audioRecording,
+})
+
+describe("ElevenLabs Transcription recorded", () => {
+  recorded.effect("transcribes audio with word timestamps", () =>
+    Effect.gen(function* () {
+      const request = Transcription.request({ model, audio: yield* audio, timestamps: "word" })
+      const response = yield* Transcription.generate(request)
+
+      expect(response.text).toMatch(TRANSCRIPT)
+      expect(response.words?.map((word) => word.text)).toEqual(["Hello", "from", "OpenCode"])
+      expect(response.words?.every((word) => word.speaker === undefined && (word.confidence ?? 0) > 0)).toBe(true)
+      expect(response.segments).toBeUndefined()
+      expect(response.language).toBe("eng")
+      expect(response.durationSeconds).toBeGreaterThan(0)
+      expect(response.usage).toEqual({ type: "seconds", seconds: response.durationSeconds })
+      expect(response.providerMetadata?.elevenlabs?.transcriptionId).toEqual(expect.any(String))
+
+      const events = Array.from(yield* Stream.runCollect(Transcription.stream(request)))
+      expect(events.map((event) => event.type)).toEqual(["finish"])
+    }),
+  )
+
+  recorded.effect("groups diarized words into speaker turns", () =>
+    Effect.gen(function* () {
+      const response = yield* Transcription.generate({ model, audio: yield* dialog, diarize: true })
+
+      expect(response.segments?.map((segment) => segment.speaker)).toEqual(["speaker_0", "speaker_1"])
+      expect(response.segments?.[0].text).toMatch(/^Did the release ship\?$/)
+      expect(response.segments?.[1].text).toMatch(/^Yes, it shipped this morning\.?$/)
+      expect(response.segments?.map((segment) => segment.text).join(" ")).toBe(response.text)
+      expect(response.words?.some((word) => word.text.trim() === "")).toBe(false)
+      expect(new Set(response.words?.map((word) => word.speaker))).toEqual(new Set(["speaker_0", "speaker_1"]))
+    }),
+  )
+})

packages/ai/test/provider/evaluation.recorded.test.ts

@@ -92,5 +92,6 @@ const assertEvaluation = <Options extends EvaluationOptions>(
     expect(response.answers.refund.probability).toBeGreaterThan(0.5)
     expect(response.usage?.inputTokens).toBeGreaterThan(0)
     expect(response.usage?.outputTokens).toBeGreaterThan(0)
-    expect(response.providerMetadata?.[metadataKey]?.confidence).toBeDefined()
+    expect(response.answers.department.confidence).toBeGreaterThan(0)
+    expect(response.answers.urgency.confidence).toBeGreaterThan(0)
   })

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

@@ -0,0 +1,68 @@
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMEvent, LLMRequest, Message, ToolRuntime, toDefinitions } from "../../src/index.js"
+import * as OpenAICompatible from "../../src/providers/openai-compatible.js"
+import { LLMClient } from "../../src/route.js"
+import { compileRequest } from "../../src/route/client.js"
+import { recordedTests } from "../recorded-test.js"
+import { weatherRuntimeTool, weatherToolName } from "../recorded-scenarios.js"
+
+const model = OpenAICompatible.configure({
+  provider: "google",
+  baseURL: "https://generativelanguage.googleapis.com/v1beta/openai",
+  apiKey: process.env.GOOGLE_GENERATIVE_AI_API_KEY ?? "fixture",
+}).model("gemini-3.8-flash")
+
+const recorded = recordedTests({
+  prefix: "openai-compatible-chat",
+  provider: "google",
+  protocol: "openai-chat",
+  requires: ["GOOGLE_GENERATIVE_AI_API_KEY"],
+  tags: ["tool", "tool-loop", "continuation"],
+  metadata: { model: model.id },
+})
+
+describe("Gemini OpenAI-compatible Chat recorded", () => {
+  recorded.effect.with(
+    "replays thought signatures through a parallel tool loop",
+    { cassette: "openai-compatible-chat/gemini-parallel-tool-signatures" },
+    () =>
+      Effect.gen(function* () {
+        const tools = { [weatherToolName]: weatherRuntimeTool }
+        const request = LLM.request({
+          model,
+          system: "Call get_weather for every requested city in parallel, then answer in one short sentence.",
+          prompt: "What is the weather in Paris and in Tokyo?",
+          tools: toDefinitions(tools),
+          cache: "none",
+        })
+        const first = yield* LLMClient.generate(request)
+        const calls = first.events.filter(LLMEvent.is.toolCall)
+        expect(calls.map((call) => call.input)).toEqual([{ city: "Paris" }, { city: "Tokyo" }])
+        const extraContent = calls[0]?.providerMetadata?.google?.extraContent
+        expect(extraContent).toEqual({ google: { thought_signature: expect.any(String) } })
+
+        const results = yield* Effect.forEach(calls, (call) => ToolRuntime.dispatch(tools, call))
+        const continuation = LLMRequest.update(request, {
+          messages: [
+            ...request.messages,
+            first.message,
+            ...calls.map((call, index) =>
+              Message.tool({ id: call.id, name: call.name, result: results[index]!.result }),
+            ),
+          ],
+        })
+        const prepared = yield* compileRequest(continuation)
+        const assistant = prepared.body.messages.find((message) => message.role === "assistant")
+        expect(assistant?.role === "assistant" ? assistant.tool_calls?.[0]?.extra_content : undefined).toEqual(
+          extraContent,
+        )
+
+        const second = yield* LLMClient.generate(continuation)
+        expect(second.events.filter(LLMEvent.is.toolCall)).toHaveLength(0)
+        expect(second.text).toMatch(/Paris/)
+        expect(second.text).toMatch(/Tokyo/)
+      }),
+    60_000,
+  )
+})

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

@@ -472,6 +472,45 @@ describe("OpenAI Chat route", () => {
     }),
   )
 
+  it.effect("replays Gemini thought signatures as tool call extra content", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("Weather in Paris and Tokyo?"),
+            Message.assistant([
+              ToolCallPart.make({
+                id: "call_1",
+                name: "lookup",
+                input: { city: "Paris" },
+                providerMetadata: { openai: { extraContent: { google: { thought_signature: "sig_1" } } } },
+              }),
+              ToolCallPart.make({ id: "call_2", name: "lookup", input: { city: "Tokyo" } }),
+            ]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Sunny" }),
+            Message.tool({ id: "call_2", name: "lookup", result: "Rainy" }),
+          ],
+        }),
+      )
+
+      const assistant = prepared.body.messages[1]
+      expect(assistant?.role === "assistant" ? assistant.tool_calls : undefined).toEqual([
+        {
+          id: "call_1",
+          type: "function",
+          function: { name: "lookup", arguments: encodeJson({ city: "Paris" }) },
+          extra_content: { google: { thought_signature: "sig_1" } },
+        },
+        {
+          id: "call_2",
+          type: "function",
+          function: { name: "lookup", arguments: encodeJson({ city: "Tokyo" }) },
+        },
+      ])
+    }),
+  )
+
   it.effect("limits OpenAI and Azure Chat tool call IDs to 40 characters", () =>
     Effect.gen(function* () {
       const id = `call_${"a".repeat(48)}`
@@ -1805,6 +1844,78 @@ describe("OpenAI Chat route", () => {
     }),
   )
 
+  it.effect("preserves Gemini thought signatures on streamed parallel tool calls", () =>
+    Effect.gen(function* () {
+      // Gemini's OpenAI-compatible endpoint omits `index`, streams each call whole,
+      // and signs only the first call of a parallel batch.
+      const body = sseEvents(
+        deltaChunk({
+          role: "assistant",
+          tool_calls: [
+            {
+              extra_content: { google: { thought_signature: "sig_1" } },
+              id: "call_1",
+              type: "function",
+              function: { name: "lookup", arguments: '{"city":"Paris"}' },
+            },
+          ],
+        }),
+        deltaChunk({
+          role: "assistant",
+          tool_calls: [{ id: "call_2", type: "function", function: { name: "lookup", arguments: '{"city":"Tokyo"}' } }],
+        }),
+        deltaChunk({}, "stop"),
+      )
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        {
+          type: "tool-call",
+          id: "call_1",
+          name: "lookup",
+          input: { city: "Paris" },
+          providerExecuted: undefined,
+          providerMetadata: { openai: { extraContent: { google: { thought_signature: "sig_1" } } } },
+        },
+        {
+          type: "tool-call",
+          id: "call_2",
+          name: "lookup",
+          input: { city: "Tokyo" },
+          providerExecuted: undefined,
+          providerMetadata: undefined,
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps extra content that arrives before the tool identity", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        deltaChunk({
+          tool_calls: [
+            { index: 0, extra_content: { google: { thought_signature: "sig_1" } }, function: { arguments: "{" } },
+          ],
+        }),
+        deltaChunk({ tool_calls: [{ index: 0, id: "call_1", function: { name: "lookup", arguments: "}" } }] }),
+        deltaChunk({}, "tool_calls"),
+      )
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events.filter(LLMEvent.is.toolCall).map((event) => event.providerMetadata)).toEqual([
+        { openai: { extraContent: { google: { thought_signature: "sig_1" } } } },
+      ])
+    }),
+  )
+
   it.effect("does not finalize streamed tool calls when content is filtered", () =>
     Effect.gen(function* () {
       const body = sseEvents(

packages/ai/test/transcription.test.ts

@@ -3,7 +3,7 @@ import { Effect, Fiber, Layer, Stream } from "effect"
 import * as TestClock from "effect/testing/TestClock"
 import { HttpClientRequest } from "effect/unstable/http"
 import { Media, Transcription, TranscriptionClient, type TranscriptionEvent } from "../src/index.js"
-import { AssemblyAI, Deepgram, Google, OpenAI } from "../src/providers.js"
+import { AssemblyAI, Deepgram, ElevenLabs, Google, OpenAI } from "../src/providers.js"
 import { it } from "./lib/effect.js"
 import { dynamicResponse, json, observe, type Call } from "./lib/http.js"
 import { sseEvents } from "./lib/sse.js"
@@ -35,6 +35,9 @@ const formFields = (call: Call) =>
 const assemblyai = AssemblyAI.configure({ apiKey: "aai-key", baseURL: "https://assemblyai.test" }).transcription(
   "universal-3-5-pro",
 )
+const elevenlabs = ElevenLabs.configure({ apiKey: "test", baseURL: "https://elevenlabs.test" }).transcription(
+  "scribe_v2",
+)
 
 describe("Transcription", () => {
   it.effect("rejects what a route cannot honor before sending anything", () =>
@@ -459,6 +462,115 @@ describe("Transcription", () => {
     }),
   )
 
+  it.effect("rejects ElevenLabs prompts, webhooks, per-channel transcripts, and untimed diarization", () =>
+    Effect.gen(function* () {
+      const errors = yield* Effect.all(
+        [
+          Transcription.generate({ model: elevenlabs, audio, prompt: "OpenCode" }),
+          Transcription.generate({ model: elevenlabs, audio, providerOptions: { webhook: true } }),
+          Transcription.generate({ model: elevenlabs, audio, http: { body: { use_multi_channel: true } } }),
+          Transcription.generate({
+            model: elevenlabs,
+            audio,
+            diarize: true,
+            providerOptions: { timestamps_granularity: "none" },
+          }),
+          Transcription.generate({
+            model: elevenlabs,
+            audio: Media.ref("file_1", { provider: "elevenlabs", mediaType: "audio/mpeg" }),
+          }),
+        ].map((effect) => Effect.flip(effect)),
+      )
+      expect(errors.map((error) => [error.reason._tag, "operation" in error.reason && error.reason.operation])).toEqual(
+        [
+          ["UnsupportedOperation", "media.prompt"],
+          ["UnsupportedOperation", "transcription.webhook"],
+          ["UnsupportedOperation", "transcription.multichannel"],
+          ["UnsupportedOperation", "media.timestamps"],
+          ["InvalidRequest", false],
+        ],
+      )
+    }).pipe(Effect.provide(layer(() => Effect.die("an unsupported request reached the network")))),
+  )
+
+  it.effect("sends ElevenLabs URL audio as source_url and groups diarized words into speaker turns", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const token = (text: string, type: string, start: number, end: number, speaker_id?: string) => ({
+        text,
+        type,
+        start,
+        end,
+        speaker_id,
+        logprob: 0,
+      })
+      const response = yield* Transcription.generate({
+        model: elevenlabs,
+        audio: Media.url("https://a.test/call.mp3"),
+        language: "en",
+        speakers: 2,
+        providerOptions: { keyterms: ["OpenCode", "Scribe"], tag_audio_events: true, diarize: false },
+      }).pipe(
+        Effect.provide(
+          layer((input) =>
+            observe(calls, input).pipe(
+              Effect.as(
+                json(input, {
+                  language_code: "ENG",
+                  text: "Ready? (laughs) Yes. Go",
+                  words: [
+                    token("Ready?", "word", 0, 0.5, "speaker_0"),
+                    token(" ", "spacing", 0.5, 0.6, "speaker_0"),
+                    token("(laughs)", "audio_event", 0.6, 1, "speaker_0"),
+                    token(" ", "spacing", 1, 1.1, "speaker_0"),
+                    token("Yes.", "word", 1.2, 1.5, "speaker_1"),
+                    token(" ", "spacing", 1.5, 1.6, "speaker_1"),
+                    token("Go", "word", 1.6, 1.9, "speaker_0"),
+                  ],
+                  transcription_id: "tr_1",
+                  audio_duration_secs: 2,
+                }),
+              ),
+            ),
+          ),
+        ),
+      )
+
+      // `observe` re-encodes the FormData with a new boundary, so read the boundary from the sent body.
+      const boundary = /^--(\S+)/.exec(calls[0].body)?.[1]
+      const form = yield* Effect.promise(() =>
+        new Response(calls[0].body, {
+          headers: { "content-type": `multipart/form-data; boundary=${boundary}` },
+        }).formData(),
+      )
+      expect(calls[0].url).toBe("https://elevenlabs.test/v1/speech-to-text")
+      expect(calls[0].headers.get("xi-api-key")).toBe("test")
+      expect(Array.from(form.entries())).toEqual([
+        ["model_id", "scribe_v2"],
+        ["source_url", "https://a.test/call.mp3"],
+        ["language_code", "en"],
+        ["diarize", "true"],
+        ["num_speakers", "2"],
+        ["keyterms", "OpenCode"],
+        ["keyterms", "Scribe"],
+        ["tag_audio_events", "true"],
+      ])
+      expect(response.segments).toEqual([
+        { text: "Ready?", startSeconds: 0, endSeconds: 0.5, speaker: "speaker_0" },
+        { text: "Yes.", startSeconds: 1.2, endSeconds: 1.5, speaker: "speaker_1" },
+        { text: "Go", startSeconds: 1.6, endSeconds: 1.9, speaker: "speaker_0" },
+      ])
+      expect(response.words?.map((word) => [word.text, word.speaker, word.confidence])).toEqual([
+        ["Ready?", "speaker_0", 1],
+        ["Yes.", "speaker_1", 1],
+        ["Go", "speaker_0", 1],
+      ])
+      expect(response.language).toBe("eng")
+      expect(response.usage).toEqual({ type: "seconds", seconds: 2 })
+      expect(response.providerMetadata).toEqual({ elevenlabs: { transcriptionId: "tr_1" } })
+    }),
+  )
+
   it.effect("rejects reading an AssemblyAI result before the transcript finishes", () =>
     Effect.gen(function* () {
       const generation = yield* Transcription.resume(assemblyai, { transcriptID: "tr_1" })

packages/ai/test/video.test.ts

@@ -1,9 +1,10 @@
 import { describe, expect } from "bun:test"
-import { Effect, Layer, Stream } from "effect"
+import { Effect, Fiber, Layer, Stream } from "effect"
+import * as TestClock from "effect/testing/TestClock"
 import { Media, Video, VideoClient, type GenerationEvent, type VideoEvent } from "../src/index.js"
 import { Fal, Google, Runway, XAI } from "../src/providers.js"
 import { it } from "./lib/effect.js"
-import { dynamicResponse, json, observe, settle, type Call } from "./lib/http.js"
+import { dynamicResponse, json, observe, settle, type Call, type HandlerInput } from "./lib/http.js"
 
 const layer = (handler: Parameters<typeof dynamicResponse>[0]) =>
   VideoClient.layer.pipe(Layer.provideMerge(dynamicResponse(handler)))
@@ -162,27 +163,39 @@ describe("Video / Google Veo", () => {
     ),
   )
 
-  it.effect("surfaces an operation error as a failed generation with the provider body", () =>
-    Effect.gen(function* () {
-      const failure = {
-        name: operation,
-        done: true,
-        error: { code: 3, message: "Prompt violates policy", status: "INVALID_ARGUMENT" },
-      }
-      const error = yield* Video.generate({ model, prompt: "nope" }).pipe(
-        Effect.flip,
-        Effect.provide(
-          layer((input) =>
-            Effect.succeed(input.request.method === "POST" ? json(input, { name: operation }) : json(input, failure)),
-          ),
-        ),
-      )
-      expect(error.reason._tag).toBe("ProviderInternal")
-      expect(error.message).toBe("Google Veo operation failed: Prompt violates policy")
-      expect(error.reason.body).toBe(JSON.stringify(failure))
-      expect(error.reason.http?.status).toBe(200)
-    }),
-  )
+  for (const terminal of [
+    { error: { code: 3, message: "Prompt violates policy", status: "INVALID_ARGUMENT" }, tag: "InvalidRequest" },
+    { error: { code: 9, message: "Unsupported resolution", status: "FAILED_PRECONDITION" }, tag: "InvalidRequest" },
+    { error: { code: 11, message: "Duration out of range", status: "OUT_OF_RANGE" }, tag: "InvalidRequest" },
+    { error: { code: 7, message: "Permission denied", status: "PERMISSION_DENIED" }, tag: "Authentication" },
+    { error: { code: 16, message: "Invalid credentials", status: "UNAUTHENTICATED" }, tag: "Authentication" },
+    { error: { code: 8, message: "Quota exceeded", status: "RESOURCE_EXHAUSTED" }, tag: "RateLimit" },
+    { error: { code: 13, message: "Internal error", status: "INTERNAL" }, tag: "ProviderInternal" },
+    { error: { code: 14, message: "Service unavailable", status: "UNAVAILABLE" }, tag: "ProviderInternal" },
+    { error: { message: "Something broke" }, tag: "ProviderInternal" },
+  ]) {
+    it.effect(
+      `surfaces ${terminal.error.status ?? "an uncoded"} operation error as ${terminal.tag} with the provider body`,
+      () =>
+        Effect.gen(function* () {
+          const failure = { name: operation, done: true, error: terminal.error }
+          const error = yield* Video.generate({ model, prompt: "nope" }).pipe(
+            Effect.flip,
+            Effect.provide(
+              layer((input) =>
+                Effect.succeed(
+                  input.request.method === "POST" ? json(input, { name: operation }) : json(input, failure),
+                ),
+              ),
+            ),
+          )
+          expect(error.reason._tag).toBe(terminal.tag)
+          expect(error.message).toBe(`Google Veo operation failed: ${terminal.error.message}`)
+          expect(error.reason.body).toBe(JSON.stringify(failure))
+          expect(error.reason.http?.status).toBe(200)
+        }),
+    )
+  }
 
   it.effect("reports fully filtered output as a content policy failure", () =>
     Effect.gen(function* () {
@@ -332,12 +345,37 @@ describe("Video / xAI", () => {
   for (const terminal of [
     {
       body: { status: "failed", error: { code: "invalid_argument", message: "Prompt cannot be empty." } },
-      tag: "ProviderInternal",
+      tag: "InvalidRequest",
       message: "xAI Video generation failed (invalid_argument): Prompt cannot be empty.",
     },
+    {
+      body: { status: "failed", error: { code: "failed_precondition", message: "Extension is not supported." } },
+      tag: "InvalidRequest",
+      message: "xAI Video generation failed (failed_precondition): Extension is not supported.",
+    },
+    {
+      body: { status: "failed", error: { code: "permission_denied", message: "Team lacks access." } },
+      tag: "Authentication",
+      message: "xAI Video generation failed (permission_denied): Team lacks access.",
+    },
+    {
+      body: { status: "failed", error: { code: "service_unavailable", message: "Overloaded." } },
+      tag: "ProviderInternal",
+      message: "xAI Video generation failed (service_unavailable): Overloaded.",
+    },
+    {
+      body: { status: "failed", error: { code: "internal_error", message: "Generation failed." } },
+      tag: "ProviderInternal",
+      message: "xAI Video generation failed (internal_error): Generation failed.",
+    },
+    {
+      body: { status: "failed", error: { code: "constructor", message: "Future code." } },
+      tag: "ProviderInternal",
+      message: "xAI Video generation failed (constructor): Future code.",
+    },
     { body: { status: "expired" }, tag: "InvalidRequest", message: "xAI Video request req_1 expired" },
   ]) {
-    it.effect(`surfaces ${terminal.body.status} generations with the provider body`, () =>
+    it.effect(`surfaces ${terminal.body.error?.code ?? terminal.body.status} generations with the provider body`, () =>
       Effect.gen(function* () {
         const error = yield* Video.generate({ model, prompt: "x" }).pipe(Effect.flip)
         expect(error.reason._tag).toBe(terminal.tag)
@@ -558,7 +596,10 @@ describe("Video / fal", () => {
   ]) {
     it.effect(`fails await for ${failure.name} with the response_url body and HTTP context`, () =>
       Effect.gen(function* () {
-        const error = yield* Video.generate({ model, prompt: "x" }).pipe(Effect.flip)
+        // A transient 500 on the result fetch is retried first; the body and HTTP context survive the final failure.
+        const fiber = yield* Effect.forkChild(Video.generate({ model, prompt: "x" }).pipe(Effect.flip))
+        yield* TestClock.adjust("5 minutes")
+        const error = yield* Fiber.join(fiber)
         expect(error.reason._tag).toBe(failure.tag)
         expect(error.reason.body).toBe(JSON.stringify(failure.result.body))
         expect(error.reason.http).toMatchObject({ url: urls.response, status: failure.result.status })
@@ -765,6 +806,11 @@ describe("Video / Runway", () => {
       tag: "ProviderInternal",
       message: "Runway task failed (INTERNAL.BAD_OUTPUT.CODE01): Something broke",
     },
+    {
+      body: { status: "FAILED", failure: "Unsupported dimensions", failureCode: "ASSET.INVALID" },
+      tag: "InvalidRequest",
+      message: "Runway task failed (ASSET.INVALID): Unsupported dimensions",
+    },
     { body: { status: "CANCELLED" }, tag: "InvalidRequest", message: "Runway task task_1 was cancelled" },
   ]) {
     it.effect(`surfaces ${terminal.body.failureCode ?? terminal.body.status} with the task body`, () =>
@@ -908,6 +954,169 @@ describe("Video / Runway", () => {
   )
 })
 
+// ---------------------------------------------------------------------------
+// Transient read failures
+// ---------------------------------------------------------------------------
+
+describe("Video / transient read failures", () => {
+  const model = Runway.configure({ apiKey: "test", baseURL: "https://runway.test/v1" }).video("gen4.5")
+  const succeeded = { id: "task_1", status: "SUCCEEDED", output: ["https://runway.test/out.mp4"] }
+  const failure = (input: HandlerInput, status: number, headers?: Record<string, string>) =>
+    json(input, { error: `HTTP ${status}` }, { status, headers })
+  const methods = (calls: ReadonlyArray<Call>) => calls.map((call) => call.method)
+
+  it.effect("retries a 503 status poll and a 503 result read, then returns the result", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const response = yield* settle(
+        Video.generate({ model, prompt: "x" }, { poll: { interval: "1 second" } }),
+        5,
+      ).pipe(
+        Effect.provide(
+          layer((input) =>
+            Effect.gen(function* () {
+              const { call, nth } = yield* observe(calls, input)
+              if (call.method === "POST") return json(input, { id: "task_1" })
+              // 1: status fails, 2: status succeeds, 3: result fails, 4: result succeeds.
+              if (nth === 1 || nth === 3) return failure(input, 503)
+              return json(input, succeeded)
+            }),
+          ),
+        ),
+      )
+      expect(response.video.source).toMatchObject({ type: "url", url: "https://runway.test/out.mp4" })
+      expect(methods(calls)).toEqual(["POST", "GET", "GET", "GET", "GET"])
+    }),
+  )
+
+  it.effect("waits for a 429 retry-after before polling again", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const fiber = yield* Effect.forkChild(
+        Video.generate({ model, prompt: "x" }, { poll: { interval: "1 second" } }).pipe(
+          Effect.provide(
+            layer((input) =>
+              Effect.gen(function* () {
+                const { call, nth } = yield* observe(calls, input)
+                if (call.method === "POST") return json(input, { id: "task_1" })
+                if (nth === 1) return failure(input, 429, { "retry-after": "10" })
+                return json(input, succeeded)
+              }),
+            ),
+          ),
+        ),
+      )
+      yield* TestClock.adjust("9 seconds")
+      expect(methods(calls)).toEqual(["POST", "GET"])
+      yield* TestClock.adjust("1 second")
+      yield* Fiber.join(fiber)
+      expect(methods(calls)).toEqual(["POST", "GET", "GET", "GET"])
+    }),
+  )
+
+  it.effect("fails a 400 status poll without retrying", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const error = yield* Video.generate({ model, prompt: "x" }).pipe(
+        Effect.flip,
+        Effect.provide(
+          layer((input) =>
+            Effect.gen(function* () {
+              const { call } = yield* observe(calls, input)
+              return call.method === "POST" ? json(input, { id: "task_1" }) : failure(input, 400)
+            }),
+          ),
+        ),
+      )
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(methods(calls)).toEqual(["POST", "GET"])
+    }),
+  )
+
+  it.effect("stops retrying at poll.timeout with a Timeout reason", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const error = yield* settle(
+        Video.generate({ model, prompt: "x" }, { poll: { interval: "1 second", timeout: "5 seconds" } }).pipe(
+          Effect.flip,
+        ),
+        6,
+      ).pipe(
+        Effect.provide(
+          layer((input) =>
+            Effect.gen(function* () {
+              const { call } = yield* observe(calls, input)
+              return call.method === "POST" ? json(input, { id: "task_1" }) : failure(input, 503)
+            }),
+          ),
+        ),
+      )
+      expect(error.reason._tag).toBe("Timeout")
+      expect(calls.filter((call) => call.method === "GET").length).toBeGreaterThan(1)
+    }),
+  )
+
+  it.effect("bounds a streamed result read's retries by poll.timeout", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const error = yield* settle(
+        Video.stream({ model, prompt: "x" }, { poll: { interval: "1 second", timeout: "5 seconds" } }).pipe(
+          Stream.runCollect,
+          Effect.flip,
+        ),
+        6,
+      ).pipe(
+        Effect.provide(
+          layer((input) =>
+            Effect.gen(function* () {
+              const { call, nth } = yield* observe(calls, input)
+              if (call.method === "POST") return json(input, { id: "task_1" })
+              return nth === 1 ? json(input, succeeded) : failure(input, 503)
+            }),
+          ),
+        ),
+      )
+      expect(error.reason._tag).toBe("Timeout")
+      expect(calls.filter((call) => call.method === "GET").length).toBeGreaterThan(2)
+    }),
+  )
+
+  it.effect("never retries a failed submit", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const error = yield* Video.generate({ model, prompt: "x" }).pipe(
+        Effect.flip,
+        Effect.provide(layer((input) => observe(calls, input).pipe(Effect.map(() => failure(input, 503))))),
+      )
+      expect(error.reason._tag).toBe("ProviderInternal")
+      expect(methods(calls)).toEqual(["POST"])
+    }),
+  )
+
+  it.effect("never retries a failed cancel", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      const error = yield* Effect.gen(function* () {
+        const generation = yield* Video.start({ model, prompt: "x" })
+        return yield* generation.cancel().pipe(Effect.flip)
+      }).pipe(
+        Effect.provide(
+          layer((input) =>
+            Effect.gen(function* () {
+              const { call } = yield* observe(calls, input)
+              if (call.method === "POST") return json(input, { id: "task_1" })
+              if (call.method === "DELETE") return failure(input, 503)
+              return json(input, { id: "task_1", status: "RUNNING" })
+            }),
+          ),
+        ),
+      )
+      expect(error.reason._tag).toBe("ProviderInternal")
+      expect(methods(calls)).toEqual(["POST", "GET", "DELETE"])
+    }),
+  )
+})
+
 // ---------------------------------------------------------------------------
 // Shared queued behavior
 // ---------------------------------------------------------------------------

packages/cli/src/commands/commands.ts

@@ -51,7 +51,7 @@ const Root = Spec.make(typeof OPENCODE_CLI_NAME === "string" ? OPENCODE_CLI_NAME
     ),
     session: Flag.string("session").pipe(
       Flag.withAlias("s"),
-      Flag.withDescription("Session ID to continue"),
+      Flag.withDescription("Session ID to continue, or to create if it does not exist"),
       Flag.optional,
     ),
     prompt: Flag.string("prompt").pipe(Flag.withDescription("Prompt to use"), Flag.optional),
@@ -328,7 +328,7 @@ const Root = Spec.make(typeof OPENCODE_CLI_NAME === "string" ? OPENCODE_CLI_NAME
         ),
         session: Flag.string("session").pipe(
           Flag.withAlias("s"),
-          Flag.withDescription("Session ID to continue"),
+          Flag.withDescription("Session ID to continue, or to create if it does not exist"),
           Flag.optional,
         ),
         fork: Flag.boolean("fork").pipe(
@@ -368,7 +368,7 @@ const Root = Spec.make(typeof OPENCODE_CLI_NAME === "string" ? OPENCODE_CLI_NAME
         ),
         session: Flag.string("session").pipe(
           Flag.withAlias("s"),
-          Flag.withDescription("Session ID to continue"),
+          Flag.withDescription("Session ID to continue, or to create if it does not exist"),
           Flag.optional,
         ),
         fork: Flag.boolean("fork").pipe(

packages/cli/src/commands/handlers/default.ts

@@ -11,6 +11,10 @@ import { UpdatePreflight } from "../../services/update-preflight"
 import { Npm } from "@opencode/util/npm"
 import { OPENCODE_ARTIFACT, OPENCODE_CHANNEL, OPENCODE_VERSION } from "../../version"
 import { Env } from "../../env"
+import { Service } from "@opencode/client/effect/service"
+import { OpenCode } from "@opencode/client/promise"
+import { findSession } from "../../session-target"
+import { errorMessage } from "../../util/error"
 
 export default Runtime.handler(Commands, (input) =>
   Effect.gen(function* () {
@@ -46,6 +50,15 @@ export default Runtime.handler(Commands, (input) =>
         Effect.promise(() => preflight.fail("OpenCode update could not start the new background service")),
       ),
     )
+    const session = Option.getOrUndefined(input.session)
+    // A missing --session ID becomes the ID of the session the first prompt creates.
+    const sessionExists =
+      session !== undefined &&
+      (yield* Effect.tryPromise({
+        try: () =>
+          findSession(OpenCode.make({ baseUrl: server.endpoint.url, headers: Service.headers(server.endpoint) }), session),
+        catch: (cause) => new Error(errorMessage(cause)),
+      })) !== undefined
     const updater = yield* Updater.Service
     let installing: string | undefined
     const updateListeners = new Set<(version: string) => void>()
@@ -81,7 +94,8 @@ export default Runtime.handler(Commands, (input) =>
       },
       args: {
         continue: input.continue,
-        sessionID: Option.getOrUndefined(input.session),
+        sessionID: sessionExists ? session : undefined,
+        newSessionID: sessionExists ? undefined : session,
         prompt: Option.getOrUndefined(input.prompt),
         auto: input.auto || input.yolo || input.dangerouslySkipPermissions,
       },

packages/cli/src/run/noninteractive.ts

@@ -67,6 +67,9 @@ type FormRequest = Extract<V2Event, { type: "form.created" }>["data"]["form"]
 // attached client must not cancel input that may belong to another session.
 const GLOBAL_FORM_SESSION_ID = "global"
 
+const PERMISSION_REJECTED_FEEDBACK =
+  "This non-interactive run cannot ask the user for permission, so the request was rejected. Continue without this action."
+
 export async function runNonInteractivePrompt(input: Input) {
   const controller = new AbortController()
   const stream = input.client.event.subscribe({ signal: controller.signal })[Symbol.asyncIterator]()
@@ -133,8 +136,11 @@ export async function runNonInteractivePrompt(input: Input) {
   }
 
   const replyPermission = async (request: { id: string; action: string; resources: ReadonlyArray<string> }) => {
+    // Nobody can approve here. Outside V1 compatibility, reject with feedback so the tool fails
+    // as ordinary model-visible output and the model continues without the action.
+    const continuing = !input.auto && input.compatibility !== "v1"
     if (!input.auto) {
-      permissionRejected = true
+      if (!continuing) permissionRejected = true
       UI.println(
         UI.Style.TEXT_WARNING_BOLD + "!",
         UI.Style.TEXT_NORMAL +
@@ -146,9 +152,10 @@ export async function runNonInteractivePrompt(input: Input) {
         sessionID: input.sessionID,
         requestID: request.id,
         decision: input.auto ? "once" : "reject",
+        ...(continuing ? { message: PERMISSION_REJECTED_FEEDBACK } : {}),
       })
       .catch(() => {})
-    if (!input.auto) {
+    if (!input.auto && !continuing) {
       await input.client.session.interrupt({ sessionID: input.sessionID }).catch(() => {})
     }
   }
@@ -163,6 +170,7 @@ export async function runNonInteractivePrompt(input: Input) {
       if (!formAlreadySettled(error)) throw error
     }
     formCancelled = true
+    if (input.compatibility !== "v1") process.exitCode = 1
   }
 
   const consume = async () => {
@@ -493,7 +501,8 @@ export async function runNonInteractivePrompt(input: Input) {
       if (event.type === "session.execution.interrupted") {
         if (input.compatibility === "v1" && (permissionRejected || formCancelled)) return
         if (event.data.reason === "user" && interrupted) process.exitCode = 130
-        if (event.data.reason !== "user" && !emittedError) {
+        // A declined tool call ends the step with an interruption; it was already reported above.
+        if (event.data.reason !== "user" && !emittedError && !permissionRejected && !formCancelled) {
           emittedError = true
           process.exitCode = 1
           const error = { type: "aborted" as const, message: `Session interrupted: ${event.data.reason}` }
@@ -620,7 +629,9 @@ export async function runNonInteractivePrompt(input: Input) {
         UI.error(item.state.error.message)
       }
 
-      if (message.error && !emittedError) {
+      // A declined tool call ends its step with an interrupted-step error that is
+      // only a consequence of our own rejection; it was already reported above.
+      if (message.error && !emittedError && !permissionRejected && !formCancelled) {
         emittedError = true
         process.exitCode = 1
         if (!emit("error", timestamp, { error: message.error })) UI.error(message.error.message)

packages/cli/src/session-target.ts

@@ -63,6 +63,7 @@ export async function resolveSessionTarget(input: {
     (await input.client.session
       .create(
         {
+          id: input.session,
           agent: prepared.agent,
           model: prepared.model,
           location: { directory: location.directory },
@@ -101,14 +102,11 @@ async function selectSession(input: {
   fork?: boolean
   signal?: AbortSignal
 }) {
-  const explicit = input.session
-    ? await input.client.session.get({ sessionID: input.session }, ...requestOptions(input.signal)).catch((error) => {
-        if (error && typeof error === "object" && "_tag" in error && error._tag === "SessionNotFoundError")
-          return undefined
-        throw error
-      })
-    : undefined
-  if (input.session && !explicit) throw new Error("Session not found")
+  const explicit = input.session ? await findSession(input.client, input.session, input.signal) : undefined
+  if (input.session && !explicit) {
+    if (input.fork) throw new Error("Session not found")
+    return { session: undefined }
+  }
   if (explicit)
     return {
       session: input.fork
@@ -133,6 +131,13 @@ async function selectSession(input: {
   }
 }
 
+export function findSession(client: OpenCodeClient, sessionID: string, signal?: AbortSignal) {
+  return client.session.get({ sessionID }, ...requestOptions(signal)).catch((error) => {
+    if (error && typeof error === "object" && "_tag" in error && error._tag === "SessionNotFoundError") return undefined
+    throw error
+  })
+}
+
 async function latestSession(
   client: OpenCodeClient,
   location: LocationGetOutput,

packages/cli/test/run/noninteractive.test.ts

@@ -309,6 +309,7 @@ async function capture(input: Parameters<typeof run>[0]) {
 
 afterEach(() => {
   mock.restore()
+  process.exitCode = 0
 })
 
 describe("runNonInteractivePrompt", () => {
@@ -433,6 +434,7 @@ describe("runNonInteractivePrompt", () => {
     expect(sdk.form.list).toHaveBeenCalledWith({
       location: { directory: "/work tree" },
     })
+    expect(process.exitCode).toBe(1)
   })
 
   test("attach mode cancels only session-owned forms", async () => {
@@ -448,6 +450,7 @@ describe("runNonInteractivePrompt", () => {
       { sessionID: "global", formID: "frm_pending_global" },
       expect.anything(),
     )
+    expect(process.exitCode).toBe(1)
   })
 
   test("V1 JSON output flushes step_start before an unrelated step failure", async () => {

packages/cli/test/session-target.test.ts

@@ -42,6 +42,28 @@ describe("session target resolver", () => {
     })
   })
 
+  test("creates a missing explicit Session with its ID", async () => {
+    const client = OpenCode.make({ baseUrl: "https://opencode.test" })
+    spyOn(client.session, "get").mockRejectedValue({ _tag: "SessionNotFoundError" })
+    spyOn(client.location, "get").mockResolvedValue(location("/project"))
+    const create = spyOn(client.session, "create").mockResolvedValue(session("ses_chosen", "/project"))
+
+    const target = await resolveSessionTarget({ client, session: "ses_chosen", prepare })
+    expect(create).toHaveBeenCalledWith(expect.objectContaining({ id: "ses_chosen" }))
+    expect(target).toMatchObject({ session: { id: "ses_chosen" }, resume: false })
+  })
+
+  test("does not create a missing explicit Session to fork", async () => {
+    const client = OpenCode.make({ baseUrl: "https://opencode.test" })
+    spyOn(client.session, "get").mockRejectedValue({ _tag: "SessionNotFoundError" })
+    const create = spyOn(client.session, "create")
+
+    await expect(resolveSessionTarget({ client, session: "ses_chosen", fork: true, prepare })).rejects.toThrow(
+      "Session not found",
+    )
+    expect(create).not.toHaveBeenCalled()
+  })
+
   test("paginates to continue the exact directory", async () => {
     const client = OpenCode.make({ baseUrl: "https://opencode.test" })
     spyOn(client.location, "get").mockResolvedValue(location("/project"))

packages/core/package.json

@@ -66,12 +66,6 @@
       "node": "./src/shell/parser-wasm.node.ts",
       "default": "./src/shell/parser-wasm.bun.ts"
     },
-    "#process-lock-ffi": {
-      "workerd": "./src/util/process-lock-ffi.workerd.ts",
-      "bun": "./src/util/process-lock-ffi.bun.ts",
-      "node": "./src/util/process-lock-ffi.node.ts",
-      "default": "./src/util/process-lock-ffi.bun.ts"
-    },
     "#v1-migration": {
       "types": "./src/database/v1-migration.bun.ts",
       "bun": "./src/database/v1-migration.bun.ts",

packages/core/script/build.ts

@@ -26,7 +26,6 @@ const result = await Bun.build({
     "#fff",
     "#photon-wasm",
     "#shell-parser-wasm",
-    "#process-lock-ffi",
     "#v1-migration",
   ],
   splitting: true,

packages/core/src/file-access.ts

@@ -30,12 +30,6 @@ export interface ExternalDirectoryAuthorization {
   readonly save: string
 }
 
-export const externalDirectoryPermission = (input: ExternalDirectoryAuthorization) => ({
-  action: input.action,
-  resources: [input.resource],
-  save: [input.save],
-})
-
 export interface Target {
   readonly absolute: AbsolutePath
   /** Location-relative for internal paths, absolute for external paths. */

packages/core/src/file.ts

@@ -1,6 +0,0 @@
-export * as File from "./file.js"
-
-import { FileDiff } from "@opencode/schema/file-diff"
-
-export const Diff = FileDiff.Info
-export type Diff = typeof Diff.Type

packages/core/src/generate.ts

@@ -1,6 +1,7 @@
 export * as Generate from "./generate.js"
 
 import { LLM, LLMClient, AIError } from "@opencode/ai"
+import { SessionID } from "@opencode/schema/session-id"
 import { Context, Effect, Layer, Schema } from "effect"
 import { makeLocationNode } from "@opencode/util/effect/app-node"
 import { llmClient } from "./effect/app-node-platform.js"
@@ -60,15 +61,24 @@ export const layer = Layer.effect(
             ? `Model unavailable: ${input.model.providerID}/${input.model.id}`
             : "No model specified and no supported model is available",
         })
-      const response = yield* llm.generate(LLM.request({ model: resolved.model, prompt: input.prompt })).pipe(
-        Effect.mapError(
-          (error: AIError) =>
-            new UnavailableError({
-              message: error.message,
-              service: resolved.ref.providerID,
-            }),
-        ),
-      )
+      const response = yield* llm
+        .generate(
+          LLM.request({
+            model: resolved.model,
+            prompt: input.prompt,
+            // Gateways require session attribution even for a stateless call; no Session is stored.
+            http: { headers: { "x-opencode-session": SessionID.create() } },
+          }),
+        )
+        .pipe(
+          Effect.mapError(
+            (error: AIError) =>
+              new UnavailableError({
+                message: error.message,
+                service: resolved.ref.providerID,
+              }),
+          ),
+        )
       return response.text
     })
 

packages/core/src/git.ts

@@ -7,7 +7,7 @@ import { AbsolutePath, RelativePath } from "./schema.js"
 import { FSUtil } from "@opencode/util/fs-util"
 import { AppProcess } from "@opencode/util/process"
 import { makeGlobalNode } from "@opencode/util/effect/app-node"
-import { File } from "./file.js"
+import { FileDiff } from "@opencode/schema/file-diff"
 import { KeyedMutex } from "./effect/keyed-mutex.js"
 import { VcsPatch } from "./vcs/patch.js"
 import { gitExecutable } from "./util/git-executable.js"
@@ -152,7 +152,7 @@ export interface Interface {
       to: TreeID
       context?: number
       paths?: readonly RelativePath[]
-    }) => Effect.Effect<readonly File.Diff[], OperationError>
+    }) => Effect.Effect<readonly FileDiff.Info[], OperationError>
     readonly restore: (input: {
       repository: Repository
       files: ReadonlyMap<RelativePath, TreeID>
@@ -571,7 +571,7 @@ const layer = Layer.effect(
           additions: stat?.additions ?? 0,
           deletions: stat?.deletions ?? 0,
           patch: stat?.binary ? "" : (patches.get(entry.file) ?? VcsPatch.emptyPatch(entry.file)),
-        } satisfies File.Diff
+        } satisfies FileDiff.Info
       })
     })
 

packages/core/src/instance.ts

@@ -152,7 +152,7 @@ export function layer(ref: Location.Ref, options: Options = {}): Layer.Layer<Ser
   const replacements: LayerNode.Replacements = [
     ...(options.discovery === false ? vanillaReplacements : []),
     ...(options.replacements ?? []),
-    Location.node.replace(Location.boundNode(ref, { discovery: options.discovery })),
+    Location.node.replace(Location.boundNode(ref)),
     InstancePlugins.node.replace(InstancePlugins.bound(options.plugins ?? [])),
   ]
 

packages/core/src/location-mutation.ts

@@ -1,3 +0,0 @@
-/** @deprecated Use FileAccess for path resolution and authorization. */
-export { FileAccess as LocationMutation } from "./file-access.js"
-export * from "./file-access.js"

packages/core/src/location-services.ts

@@ -8,7 +8,6 @@ import { LocationServiceMap } from "./location-service-map.js"
 export { LocationServiceMap } from "./location-service-map.js"
 
 export type LocationServices = Instance.Services
-export type LocationError = Instance.Error
 
 export function buildLocationServiceMap(
   replacements: LayerNode.Replacements = [],

packages/core/src/location.ts

@@ -16,12 +16,12 @@ export class Service extends Context.Service<Service, Interface>()("@opencode/Lo
 
 export const node = LayerNode.unbound(Service, tags.values.location)
 
-const layer = (ref: Ref, options?: { readonly discovery?: boolean }) =>
+const layer = (ref: Ref) =>
   Layer.effect(
     Service,
     Effect.gen(function* () {
       const project = yield* Project.Service
-      const resolved = yield* project.resolve(ref.directory, options)
+      const resolved = yield* project.resolve(ref.directory)
       return Service.of({
         directory: ref.directory,
         workspaceID: ref.workspaceID,
@@ -31,9 +31,9 @@ const layer = (ref: Ref, options?: { readonly discovery?: boolean }) =>
     }),
   )
 
-export const boundNode = (ref: Ref, options?: { readonly discovery?: boolean }) =>
+export const boundNode = (ref: Ref) =>
   makeLocationNode({
     service: Service,
-    layer: layer(ref, options),
+    layer: layer(ref),
     deps: [Project.node],
   })

packages/core/src/mcp/index.ts

@@ -45,8 +45,6 @@ export const ResourceTemplate = Mcp.ResourceTemplate
 export type ResourceTemplate = Mcp.ResourceTemplate
 export const ResourceCatalog = Mcp.ResourceCatalog
 export type ResourceCatalog = Mcp.ResourceCatalog
-export const ResourceContentPart = Mcp.ResourceContentPart
-export type ResourceContentPart = Mcp.ResourceContentPart
 export const ResourceContent = Mcp.ResourceContent
 export type ResourceContent = Mcp.ResourceContent
 

packages/core/src/project.ts

@@ -65,7 +65,7 @@ export interface Interface {
   /** Records Project activity for recency ordering, at most once per minute per Project. */
   readonly activate: (projectID: ID) => Effect.Effect<void>
   /** Resolves and persists the owning Project. */
-  readonly resolve: (input: AbsolutePath, options?: { readonly discovery?: boolean }) => Effect.Effect<Resolved>
+  readonly resolve: (input: AbsolutePath) => Effect.Effect<Resolved>
 }
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/Project") {}
@@ -334,10 +334,7 @@ const layer = Layer.effect(
       }
     })
 
-    const resolve = Effect.fn("Project.resolve")(function* (
-      input: AbsolutePath,
-      _options?: { readonly discovery?: boolean },
-    ) {
+    const resolve = Effect.fn("Project.resolve")(function* (input: AbsolutePath) {
       const directory = AbsolutePath.make(yield* fs.resolve(input))
       const native = yield* fs.up({ targets: [".git", ".hg"], start: directory, mode: "first" }).pipe(
         Effect.map((matches) => matches[0]),

packages/core/src/pty.ts

@@ -6,7 +6,6 @@ import { Context, Effect, Layer, Schema, Types } from "effect"
 import { Pty } from "@opencode/schema/pty"
 import { Bus } from "./bus.js"
 import { Location } from "./location.js"
-import { PtyID } from "./pty/schema.js"
 import { ShellSelect } from "./shell/select.js"
 import { lazy } from "./util/lazy.js"
 
@@ -35,6 +34,9 @@ type Active = {
   listeners: Disp[]
 }
 
+export const ID = Pty.ID
+export type ID = Pty.ID
+
 export const Info = Pty.Info
 export type Info = Types.DeepMutable<typeof Info.Type>
 
@@ -69,21 +71,21 @@ export type Attachment = {
 }
 
 export class NotFoundError extends Schema.TaggedError<NotFoundError>()("Pty.NotFoundError", {
-  ptyID: PtyID,
+  ptyID: ID,
 }) {}
 
 export class ExitedError extends Schema.TaggedError<ExitedError>()("Pty.ExitedError", {
-  ptyID: PtyID,
+  ptyID: ID,
 }) {}
 
 export interface Interface {
   readonly list: () => Effect.Effect<Info[]>
-  readonly get: (id: PtyID) => Effect.Effect<Info, NotFoundError>
+  readonly get: (id: ID) => Effect.Effect<Info, NotFoundError>
   readonly create: (input: CreateInput) => Effect.Effect<Info>
-  readonly update: (id: PtyID, input: UpdateInput) => Effect.Effect<Info, NotFoundError>
-  readonly remove: (id: PtyID) => Effect.Effect<void, NotFoundError>
-  readonly write: (id: PtyID, data: string) => Effect.Effect<void, NotFoundError>
-  readonly attach: (id: PtyID, input: AttachInput) => Effect.Effect<Attachment, NotFoundError | ExitedError>
+  readonly update: (id: ID, input: UpdateInput) => Effect.Effect<Info, NotFoundError>
+  readonly remove: (id: ID) => Effect.Effect<void, NotFoundError>
+  readonly write: (id: ID, data: string) => Effect.Effect<void, NotFoundError>
+  readonly attach: (id: ID, input: AttachInput) => Effect.Effect<Attachment, NotFoundError | ExitedError>
 }
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/Pty") {}
@@ -96,8 +98,8 @@ const layer = Layer.effect(
     const shell = yield* ShellSelect.Service
     const context = yield* Effect.context()
     const runFork = Effect.runForkWith(context)
-    const sessions = new Map<PtyID, Active>()
-    const exitOrder: PtyID[] = []
+    const sessions = new Map<ID, Active>()
+    const exitOrder: ID[] = []
 
     function notifyEnd(session: Active, event: { exitCode?: number }) {
       for (const subscriber of session.subscribers.values()) {
@@ -131,13 +133,13 @@ const layer = Layer.effect(
       }),
     )
 
-    const requireSession = Effect.fn("Pty.requireSession")(function* (id: PtyID) {
+    const requireSession = Effect.fn("Pty.requireSession")(function* (id: ID) {
       const session = sessions.get(id)
       if (!session) return yield* new NotFoundError({ ptyID: id })
       return session
     })
 
-    const removeSession = Effect.fnUntraced(function* (id: PtyID) {
+    const removeSession = Effect.fnUntraced(function* (id: ID) {
       const session = sessions.get(id)
       if (!session) return
       sessions.delete(id)
@@ -148,7 +150,7 @@ const layer = Layer.effect(
       yield* bus.publish(Pty.Event.Deleted, { id: session.info.id })
     })
 
-    const remove = Effect.fn("Pty.remove")(function* (id: PtyID) {
+    const remove = Effect.fn("Pty.remove")(function* (id: ID) {
       yield* requireSession(id)
       yield* removeSession(id)
     })
@@ -157,12 +159,12 @@ const layer = Layer.effect(
       return Array.from(sessions.values()).map((session) => session.info)
     })
 
-    const get = Effect.fn("Pty.get")(function* (id: PtyID) {
+    const get = Effect.fn("Pty.get")(function* (id: ID) {
       return (yield* requireSession(id)).info
     })
 
     const create = Effect.fn("Pty.create")(function* (input: CreateInput) {
-      const id = PtyID.ascending()
+      const id = ID.ascending()
       const command = input.command || (yield* shell.resolve({ priority: "config" }))
       const args = ShellSelect.login(command) ? [...(input.args ?? []), "-l"] : [...(input.args ?? [])]
       const cwd = input.cwd || location.directory
@@ -242,7 +244,7 @@ const layer = Layer.effect(
       return info
     })
 
-    const update = Effect.fn("Pty.update")(function* (id: PtyID, input: UpdateInput) {
+    const update = Effect.fn("Pty.update")(function* (id: ID, input: UpdateInput) {
       const session = yield* requireSession(id)
       if (input.title) session.info.title = input.title
       if (input.size && session.info.status === "running") session.process.resize(input.size.cols, input.size.rows)
@@ -250,12 +252,12 @@ const layer = Layer.effect(
       return session.info
     })
 
-    const write = Effect.fn("Pty.write")(function* (id: PtyID, data: string) {
+    const write = Effect.fn("Pty.write")(function* (id: ID, data: string) {
       const session = yield* requireSession(id)
       if (session.info.status === "running") session.process.write(data)
     })
 
-    const attach = Effect.fn("Pty.attach")(function* (id: PtyID, input: AttachInput) {
+    const attach = Effect.fn("Pty.attach")(function* (id: ID, input: AttachInput) {
       const session = yield* requireSession(id)
       if (session.info.status !== "running") return yield* new ExitedError({ ptyID: id })
       yield* Effect.logInfo("client attached to session", { id, directory: location.directory })

packages/core/src/pty/schema.ts

@@ -1 +0,0 @@
-export { ID as PtyID } from "@opencode/schema/pty"

packages/core/src/pty/ticket.ts

@@ -2,7 +2,7 @@ export * as PtyTicket from "./ticket.js"
 
 import type { Workspace } from "@opencode/schema/workspace"
 import { PtyTicket } from "@opencode/schema/pty-ticket"
-import { PtyID } from "./schema.js"
+import type { Pty } from "@opencode/schema/pty"
 import { Cache, Context, Duration, Effect, Layer } from "effect"
 import { makeGlobalNode } from "@opencode/util/effect/app-node"
 
@@ -12,7 +12,7 @@ const CAPACITY = 10_000
 export const ConnectToken = PtyTicket.ConnectToken
 
 export type Scope = {
-  readonly ptyID: PtyID
+  readonly ptyID: Pty.ID
   readonly directory?: string
   readonly workspaceID?: Workspace.ID
 }

packages/core/src/repository.ts

@@ -145,8 +145,9 @@ function parts(input: string) {
     .filter(Boolean)
 }
 
+// cachePath makes each `:`-separated host part a directory.
 function safeHost(input: string) {
-  return Boolean(input) && !input.startsWith("-") && !/[\s/\\]/.test(input)
+  return Boolean(input) && !input.startsWith("-") && input.split(":").every(safeSegment)
 }
 
 function safeSegment(input: string) {

packages/core/src/session/error.ts

@@ -50,15 +50,6 @@ export class StepFailedError extends Schema.TaggedError<StepFailedError>()("Sess
   }
 }
 
-export class UserInterruptedError extends Schema.TaggedError<UserInterruptedError>()(
-  "Session.UserInterruptedError",
-  {},
-) {
-  override get message() {
-    return "Session interrupted by user"
-  }
-}
-
 export class PromptConflictError extends Schema.TaggedError<PromptConflictError>()("Session.PromptConflictError", {
   sessionID: SessionSchema.ID,
   messageID: SessionMessage.ID,

packages/core/src/session/execution.ts

@@ -12,7 +12,6 @@ import { SessionRunner } from "./runner/index.js"
 import { SessionSchema } from "./schema.js"
 import { SessionStore } from "./store.js"
 import { toSessionError } from "./to-session-error.js"
-import { UserInterruptedError } from "./error.js"
 import { SessionInbox } from "./inbox.js"
 
 export interface Interface {
@@ -51,9 +50,7 @@ type InterruptReason = "user" | "shutdown" | "inactivity"
 export function terminal(exit: Exit.Exit<void, SessionRunner.RunError>, reason?: InterruptReason) {
   if (Exit.isSuccess(exit)) return { type: "succeeded" as const }
   if (Cause.hasInterrupts(exit.cause)) return { type: "interrupted" as const, reason: reason ?? "shutdown" }
-  const failure = Cause.squash(exit.cause)
-  if (failure instanceof UserInterruptedError) return { type: "interrupted" as const, reason: "user" as const }
-  return { type: "failed" as const, error: toSessionError(failure) }
+  return { type: "failed" as const, error: toSessionError(Cause.squash(exit.cause)) }
 }
 
 /** Process-local execution: drains run in this process using the selected instance. */

packages/core/src/session/model-request.ts

@@ -46,6 +46,8 @@ const IMAGE_REMOVED =
 const GENERATION_KEYS = new Set(Object.keys(GenerationOptions.fields))
 // Used when the catalog has no output limit for the model.
 const OUTPUT_TOKEN_FALLBACK = 32_000
+// No reply needs more, however much the model allows.
+const OUTPUT_TOKEN_MAX = 256_000
 // A summary never needs more, and a request asking for more cannot be shrunk to fit a window the catalog overstates.
 const SUMMARY_OUTPUT_MAX = 32_000
 // Prompt text is estimated at about 4 characters per token, which can run low on dense text such as code.
@@ -87,7 +89,7 @@ const outputLimit = (
   kind: "primary" | "compaction",
   inputTokens?: Input["inputTokens"],
 ) => {
-  const model = limit.output > 0 ? limit.output : OUTPUT_TOKEN_FALLBACK
+  const model = Math.min(limit.output > 0 ? limit.output : OUTPUT_TOKEN_FALLBACK, OUTPUT_TOKEN_MAX)
   const requested = kind === "compaction" ? Math.min(model, SUMMARY_OUTPUT_MAX) : model
   if (inputTokens === undefined || limit.context <= 0) return requested
   const room = limit.context - inputTokens.measured - Math.ceil(inputTokens.estimated * (1 + ESTIMATE_ERROR))
@@ -273,12 +275,12 @@ export const layer = Layer.effect(
         model: model.model,
         http: {
           headers: {
-            "x-session-affinity": session.id,
-            "X-Session-Id": session.id,
+            "x-session-affinity": affinity,
+            "X-Session-Id": affinity,
             ...(session.parentID ? { "x-parent-session-id": session.parentID } : {}),
             "User-Agent": App.useragent(app),
             "x-opencode-project": session.projectID,
-            "x-opencode-session": session.id,
+            "x-opencode-session": affinity,
             "x-opencode-client": app.name,
           },
         },

packages/core/src/session/runner/index.ts

@@ -4,7 +4,7 @@ import type { AIError } from "@opencode/ai"
 import { Context, Data, Effect } from "effect"
 import { SessionSchema } from "../schema.js"
 import type { Promotable } from "../inbox.js"
-import type { AgentNotFoundError, MessageDecodeError, StepFailedError, UserInterruptedError } from "../error.js"
+import type { AgentNotFoundError, MessageDecodeError, StepFailedError } from "../error.js"
 import { SessionRunnerModel } from "./model.js"
 import type { Instructions } from "../../instructions/index.js"
 
@@ -14,7 +14,6 @@ export type RunError =
   | MessageDecodeError
   | AgentNotFoundError
   | StepFailedError
-  | UserInterruptedError
   | Instructions.InitializationBlocked
 
 export type Continuation = { readonly step: number }

packages/core/src/session/runner/model.ts

@@ -29,19 +29,6 @@ export class ModelUnavailableError extends Schema.TaggedError<ModelUnavailableEr
     return `Model unavailable: ${this.providerID}/${this.modelID}`
   }
 }
-export const VariantUnavailableError = ModelResolver.VariantUnavailableError
-export type VariantUnavailableError = ModelResolver.VariantUnavailableError
-export const UnsupportedPackageError = ModelResolver.UnsupportedPackageError
-export type UnsupportedPackageError = ModelResolver.UnsupportedPackageError
-export const ModelConfigurationError = ModelResolver.ModelConfigurationError
-export type ModelConfigurationError = ModelResolver.ModelConfigurationError
-export const ModelInitializationError = ModelResolver.ModelInitializationError
-export type ModelInitializationError = ModelResolver.ModelInitializationError
-export const UnresolvedProviderVariablesError = ModelResolver.UnresolvedProviderVariablesError
-export type UnresolvedProviderVariablesError = ModelResolver.UnresolvedProviderVariablesError
-export const UnsupportedCompactionError = ModelResolver.UnsupportedCompactionError
-export type UnsupportedCompactionError = ModelResolver.UnsupportedCompactionError
-
 export type Error = ModelNotSelectedError | ModelUnavailableError | ModelResolver.Error
 export type Resolved = ModelResolver.Resolved
 

packages/core/src/session/runner/retry.ts

@@ -1,6 +1,6 @@
 export * as SessionRunnerRetry from "./retry.js"
 
-import { AIError, isContextOverflowFailure } from "@opencode/ai"
+import { AIError, isRetryable } from "@opencode/ai"
 import { Agent } from "@opencode/schema/agent"
 import { Model } from "@opencode/schema/model"
 import { SessionError } from "@opencode/schema/session-error"
@@ -10,7 +10,8 @@ import type { PluginHooks } from "../../plugin/hooks.js"
 import { SessionEvent } from "../event.js"
 import { SessionMessage } from "../message.js"
 import { SessionSchema } from "../schema.js"
-import { toSessionError } from "../to-session-error.js"
+
+export { isRetryable }
 
 interface Input {
   readonly cause: AIError
@@ -27,43 +28,6 @@ export interface Decision {
   readonly delay: number
 }
 
-export function isRetryable(error: AIError) {
-  const override = error.reason.http?.headers["x-should-retry"]
-  if (override === "true") return true
-  if (override === "false") return false
-  switch (error.reason._tag) {
-    case "RateLimit":
-    case "ProviderInternal":
-      return true
-    // A WebSocket acknowledgment marks delivery accepted before model output may exist.
-    // Read failures can still recover; the Step chooses retry versus continuation from durable output.
-    case "Transport":
-      return (
-        error.reason.delivery !== "rejected" &&
-        (error.reason.delivery !== "accepted" || error.reason.operation === "read")
-      )
-    case "InvalidProviderOutput":
-      return error.reason.classification === "incomplete-stream"
-    // Unrecognized failures retry: classification records affirmative
-    // deterministic evidence, and transient failures are exactly the ones
-    // that arrive in shapes no classifier anticipates.
-    case "UnknownProvider":
-      return true
-    case "Authentication":
-    case "QuotaExceeded":
-    case "ContentPolicy":
-    case "InvalidRequest":
-    case "UnsupportedOperation":
-    case "NoRoute":
-    case "Timeout":
-      return false
-    default: {
-      const exhaustive: never = error.reason
-      return exhaustive
-    }
-  }
-}
-
 /** Bound provider-requested delays so a hostile or buggy retry-after cannot stall a session for hours. */
 const RETRY_AFTER_MAX = Duration.toMillis("15 minutes")
 
@@ -119,24 +83,6 @@ export const policy = (sessionID: SessionSchema.ID) =>
       })
   })
 
-/**
- * Retries one auxiliary request's transient failures under a shared `policy` allowance, letting the
- * session retry hook adjust each decision. Context overflow is never transient: callers recover it.
- */
-export const transient =
-  (decide: Effect.Success<ReturnType<typeof policy>>, input: Pick<Input, "agent" | "model" | "hook">) =>
-  <A, R>(effect: Effect.Effect<A, AIError, R>) =>
-    Effect.retry(effect, {
-      while: (cause) =>
-        Effect.gen(function* () {
-          if (isContextOverflowFailure(cause)) return false
-          const decision = yield* decide({ ...input, cause, error: toSessionError(cause), retry: isRetryable(cause) })
-          if (!decision.retry) return false
-          yield* Effect.sleep(decision.delay)
-          return true
-        }),
-    })
-
 export const make = (bus: Bus.Interface, sessionID: SessionSchema.ID) =>
   Effect.gen(function* () {
     const decide = yield* policy(sessionID)

packages/core/src/session/to-session-error.ts

@@ -3,7 +3,8 @@ import { Tool } from "@opencode/schema/tool"
 import { SessionError } from "@opencode/schema/session-error"
 import { Permission } from "../permission.js"
 import { Integration } from "../integration.js"
-import { AgentNotFoundError, StepFailedError, UserInterruptedError } from "./error.js"
+import { AgentNotFoundError, StepFailedError } from "./error.js"
+import { ModelResolver } from "../model-resolver.js"
 import { SessionRunnerModel } from "./runner/model.js"
 
 export function toSessionError(cause: unknown): SessionError.Error {
@@ -40,6 +41,7 @@ export function toSessionError(cause: unknown): SessionError.Error {
     }
   }
   if (cause instanceof Permission.BlockedError) return { type: "permission.rejected", message: cause.message }
+  if (cause instanceof Permission.CorrectedError) return { type: "permission.rejected", message: cause.feedback }
   if (cause instanceof ToolFailure || cause instanceof Tool.Error) {
     if (cause.error === undefined) return { type: "tool.execution", message: cause.message }
     // The canonical error is the sole model-visible representation, so a cause
@@ -48,18 +50,17 @@ export function toSessionError(cause: unknown): SessionError.Error {
     return unwrapped.message === "" ? { ...unwrapped, type: "tool.execution", message: cause.message } : unwrapped
   }
   if (cause instanceof StepFailedError) return cause.error
-  if (cause instanceof SessionRunnerModel.UnsupportedCompactionError)
+  if (cause instanceof ModelResolver.UnsupportedCompactionError)
     return { type: "provider.unsupported-operation", message: cause.message }
   if (cause instanceof AgentNotFoundError) return { type: "unknown", message: cause.message }
-  if (cause instanceof UserInterruptedError) return { type: "aborted", message: cause.message }
   if (
     cause instanceof SessionRunnerModel.ModelNotSelectedError ||
     cause instanceof SessionRunnerModel.ModelUnavailableError ||
-    cause instanceof SessionRunnerModel.VariantUnavailableError ||
-    cause instanceof SessionRunnerModel.UnsupportedPackageError ||
-    cause instanceof SessionRunnerModel.ModelConfigurationError ||
-    cause instanceof SessionRunnerModel.ModelInitializationError ||
-    cause instanceof SessionRunnerModel.UnresolvedProviderVariablesError
+    cause instanceof ModelResolver.VariantUnavailableError ||
+    cause instanceof ModelResolver.UnsupportedPackageError ||
+    cause instanceof ModelResolver.ModelConfigurationError ||
+    cause instanceof ModelResolver.ModelInitializationError ||
+    cause instanceof ModelResolver.UnresolvedProviderVariablesError
   )
     return { type: "provider.no-route", message: cause.message }
   if (cause instanceof Integration.AuthorizationError) return { type: "provider.auth", message: cause.message }

packages/core/src/snapshot.ts

@@ -3,7 +3,7 @@ export * as Snapshot from "./snapshot.js"
 import { makeLocationNode } from "@opencode/util/effect/app-node"
 import path from "path"
 import { Context, Effect, Fiber, Layer, Schema, Scope } from "effect"
-import { File } from "./file.js"
+import { FileDiff } from "@opencode/schema/file-diff"
 import { FSUtil } from "@opencode/util/fs-util"
 import { Git } from "./git.js"
 import { Global } from "@opencode/util/global"
@@ -58,7 +58,7 @@ export interface Interface extends State.Transformable<Editor> {
    * Generate structured per-file diffs between two captured trees. `context`
    * controls unchanged lines around each unified diff hunk.
    */
-  readonly diff: (input: DiffInput) => Effect.Effect<readonly File.Diff[], Error>
+  readonly diff: (input: DiffInput) => Effect.Effect<readonly FileDiff.Info[], Error>
 
   /**
    * Restore selected project-relative paths from their associated trees. A path

packages/core/src/util/array.ts

@@ -1,10 +0,0 @@
-export function findLast<T>(
-  items: readonly T[],
-  predicate: (item: T, index: number, items: readonly T[]) => boolean,
-): T | undefined {
-  for (let i = items.length - 1; i >= 0; i -= 1) {
-    const item = items[i]
-    if (predicate(item, i, items)) return item
-  }
-  return undefined
-}

packages/core/src/util/process-lock-ffi.bun.ts

@@ -1,50 +0,0 @@
-import { dlopen, read, type Pointer } from "bun:ffi"
-import { existsSync } from "node:fs"
-
-export type LockResult =
-  | { readonly acquired: true }
-  | { readonly acquired: false; readonly held: true }
-  | { readonly acquired: false; readonly held: false; readonly code: number }
-
-const LOCK_EX = 2
-const LOCK_NB = 4
-const DARWIN_EWOULDBLOCK = 35
-const LINUX_EWOULDBLOCK = 11
-
-export function lockDarwin(fd: number): LockResult {
-  const library = dlopen("/usr/lib/libSystem.B.dylib", {
-    flock: { args: ["i32", "i32"], returns: "i32" },
-    __error: { args: [], returns: "ptr" },
-  })
-  try {
-    const result = library.symbols.flock(fd, LOCK_EX | LOCK_NB)
-    const code = result === 0 ? 0 : errorCode(library.symbols.__error())
-    if (result === 0) return { acquired: true }
-    if (code === DARWIN_EWOULDBLOCK) return { acquired: false, held: true }
-    return { acquired: false, held: false, code }
-  } finally {
-    library.close()
-  }
-}
-
-export function lockLinux(fd: number): LockResult {
-  const musl = `/lib/libc.musl-${process.arch === "arm64" ? "aarch64" : "x86_64"}.so.1`
-  const library = dlopen(existsSync(musl) ? musl : "libc.so.6", {
-    flock: { args: ["i32", "i32"], returns: "i32" },
-    __errno_location: { args: [], returns: "ptr" },
-  })
-  try {
-    const result = library.symbols.flock(fd, LOCK_EX | LOCK_NB)
-    const code = result === 0 ? 0 : errorCode(library.symbols.__errno_location())
-    if (result === 0) return { acquired: true }
-    if (code === LINUX_EWOULDBLOCK) return { acquired: false, held: true }
-    return { acquired: false, held: false, code }
-  } finally {
-    library.close()
-  }
-}
-
-function errorCode(pointer: Pointer | bigint | null) {
-  if (pointer === null) throw new Error("Failed to read process lock error code")
-  return read.i32(pointer, 0)
-}

packages/core/src/util/process-lock-ffi.node.ts

@@ -1,43 +0,0 @@
-import { dlopen, getInt32 } from "node:ffi"
-
-export type LockResult =
-  | { readonly acquired: true }
-  | { readonly acquired: false; readonly held: true }
-  | { readonly acquired: false; readonly held: false; readonly code: number }
-
-const LOCK_EX = 2
-const LOCK_NB = 4
-const DARWIN_EWOULDBLOCK = 35
-const LINUX_EWOULDBLOCK = 11
-
-export function lockDarwin(fd: number): LockResult {
-  const library = dlopen("/usr/lib/libSystem.B.dylib", {
-    flock: { arguments: ["int32", "int32"], return: "int32" },
-    __error: { arguments: [], return: "pointer" },
-  })
-  try {
-    const result = library.functions.flock(fd, LOCK_EX | LOCK_NB)
-    const code = result === 0 ? 0 : getInt32(library.functions.__error(), 0)
-    if (result === 0) return { acquired: true }
-    if (code === DARWIN_EWOULDBLOCK) return { acquired: false, held: true }
-    return { acquired: false, held: false, code }
-  } finally {
-    library.lib.close()
-  }
-}
-
-export function lockLinux(fd: number): LockResult {
-  const library = dlopen("libc.so.6", {
-    flock: { arguments: ["int32", "int32"], return: "int32" },
-    __errno_location: { arguments: [], return: "pointer" },
-  })
-  try {
-    const result = library.functions.flock(fd, LOCK_EX | LOCK_NB)
-    const code = result === 0 ? 0 : getInt32(library.functions.__errno_location(), 0)
-    if (result === 0) return { acquired: true }
-    if (code === LINUX_EWOULDBLOCK) return { acquired: false, held: true }
-    return { acquired: false, held: false, code }
-  } finally {
-    library.lib.close()
-  }
-}

packages/core/src/util/process-lock-ffi.workerd.ts

@@ -1,14 +0,0 @@
-export type LockResult =
-  | { readonly acquired: true }
-  | { readonly acquired: false; readonly held: true }
-  | { readonly acquired: false; readonly held: false; readonly code: number }
-
-// workerd has no FFI and no cross-process file locking; a Durable Object is
-// already single-threaded per instance, so nothing on this runtime should
-// reach these.
-const unavailable = (_fd: number): LockResult => {
-  throw new Error("Process locks are unavailable on the workerd runtime")
-}
-
-export const lockDarwin = unavailable
-export const lockLinux = unavailable

packages/core/src/util/process-lock.ts

@@ -1,131 +0,0 @@
-import { lockDarwin, lockLinux, type LockResult } from "#process-lock-ffi"
-import { closeSync, mkdirSync, openSync } from "node:fs"
-import { connect, createServer, type Server, type Socket } from "node:net"
-import path from "node:path"
-import { Effect, Schema } from "effect"
-import { Hash } from "@opencode/util/hash"
-
-export namespace ProcessLock {
-  export class HeldError extends Schema.TaggedError<HeldError>()("ProcessLockHeldError", {
-    file: Schema.String,
-  }) {
-    override get message() {
-      return `Process lock is already held: ${this.file}`
-    }
-  }
-
-  export class SystemError extends Schema.TaggedError<SystemError>()("ProcessLockSystemError", {
-    file: Schema.String,
-    operation: Schema.Literals(["open", "acquire"]),
-    code: Schema.String,
-  }) {
-    override get message() {
-      return `Process lock ${this.operation} failed for ${this.file}: ${this.code}`
-    }
-  }
-
-  export type LockError = HeldError | SystemError
-
-  const acquirePosix = Effect.fnUntraced(function* (file: string) {
-    const fd = yield* Effect.try({
-      try: () => {
-        mkdirSync(path.dirname(file), { recursive: true })
-        return openSync(file, "a+", 0o600)
-      },
-      catch: (cause) =>
-        new SystemError({
-          file,
-          operation: "open",
-          code: cause instanceof Error ? cause.message : String(cause),
-        }),
-    })
-    const result = yield* Effect.try({
-      try: () => lock(fd),
-      catch: (cause) =>
-        new SystemError({
-          file,
-          operation: "acquire",
-          code: cause instanceof Error ? cause.message : String(cause),
-        }),
-    }).pipe(
-      Effect.tapError(() =>
-        Effect.sync(() => {
-          closeSync(fd)
-        }),
-      ),
-    )
-    if (result.acquired) return fd
-    closeSync(fd)
-    return yield* result.held
-      ? new HeldError({ file })
-      : new SystemError({ file, operation: "acquire", code: String(result.code) })
-  })
-
-  export const acquire = Effect.fn("ProcessLock.acquire")(function* (file: string) {
-    if (process.platform === "win32") {
-      yield* Effect.acquireRelease(acquireWindows(file), closeWindows)
-      return
-    }
-    yield* Effect.acquireRelease(acquirePosix(file), (fd) =>
-      Effect.sync(() => {
-        closeSync(fd)
-      }),
-    )
-  })
-}
-
-function lock(fd: number): LockResult {
-  if (process.platform === "darwin") return lockDarwin(fd)
-  if (process.platform === "linux") return lockLinux(fd)
-  throw new Error(`Unsupported process lock platform: ${process.platform}`)
-}
-
-function acquireWindows(file: string) {
-  return Effect.callback<Server, ProcessLock.LockError>((resume) => {
-    const server = createServer()
-    let probe: Socket | undefined
-    const pipe = `\\\\.\\pipe\\opencode-process-lock-${Hash.sha256(path.resolve(file).toLowerCase())}`
-    const onError = (cause: NodeJS.ErrnoException) => {
-      server.off("listening", onListening)
-      probe = connect(pipe)
-      const onProbeError = () => {
-        probe?.off("connect", onConnect)
-        resume(
-          Effect.fail(
-            new ProcessLock.SystemError({
-              file,
-              operation: "acquire",
-              code: cause.code ?? cause.message,
-            }),
-          ),
-        )
-      }
-      const onConnect = () => {
-        probe?.off("error", onProbeError)
-        probe?.destroy()
-        resume(Effect.fail(new ProcessLock.HeldError({ file })))
-      }
-      probe.once("connect", onConnect)
-      probe.once("error", onProbeError)
-    }
-    const onListening = () => {
-      server.off("error", onError)
-      resume(Effect.succeed(server))
-    }
-    server.once("error", onError)
-    server.once("listening", onListening)
-    server.on("connection", (socket) => socket.destroy())
-    server.listen(pipe)
-    return Effect.sync(() => {
-      probe?.destroy()
-      server.close()
-    })
-  })
-}
-
-function closeWindows(server: Server) {
-  return Effect.callback<void>((resume) => {
-    if (!server.listening) return resume(Effect.void)
-    server.close((error) => resume(error ? Effect.die(error) : Effect.void))
-  })
-}

packages/core/test/fixture/process-lock-worker.ts

@@ -1,17 +0,0 @@
-import { ProcessLock } from "@opencode/core/util/process-lock"
-import { Effect, Schema } from "effect"
-import fs from "node:fs/promises"
-
-const input = Schema.decodeUnknownSync(
-  Schema.fromJsonString(Schema.Struct({ file: Schema.String, ready: Schema.String })),
-)(process.argv[2])
-
-await Effect.runPromise(
-  Effect.scoped(
-    Effect.gen(function* () {
-      yield* ProcessLock.acquire(input.file)
-      yield* Effect.promise(() => fs.writeFile(input.ready, String(process.pid)))
-      return yield* Effect.never
-    }),
-  ),
-)

packages/core/test/generate.test.ts

@@ -1,5 +1,6 @@
 import { expect } from "bun:test"
-import { LanguageModel } from "@opencode/ai"
+import { LanguageModel, LLMClient } from "@opencode/ai"
+import { RequestExecutor } from "@opencode/ai/route"
 import { OpenAIChat } from "@opencode/ai/protocols"
 import { TestLLM } from "@opencode/ai/testing"
 import { AISDK } from "@opencode/core/aisdk"
@@ -10,6 +11,7 @@ import { ID, Info, Model, Ref } from "@opencode/core/model"
 import { Provider } from "@opencode/core/provider"
 import { Npm } from "@opencode/util/npm"
 import { Effect, Layer } from "effect"
+import { HttpClient, HttpClientResponse } from "effect/unstable/http"
 import { testEffect } from "./lib/effect"
 
 const selected = Info.make({
@@ -89,3 +91,56 @@ resolverIt.effect("resolves dynamic models with their catalog metadata", () =>
     })
   }),
 )
+
+testEffect(Layer.empty).effect("attributes each stateless completion without creating a stored session", () =>
+  Effect.gen(function* () {
+    const sessions: string[] = []
+    const http = Layer.succeed(
+      HttpClient.HttpClient,
+      HttpClient.make((request) =>
+        Effect.sync(() => {
+          const session = request.headers["x-opencode-session"]
+          if (!session)
+            return HttpClientResponse.fromWeb(
+              request,
+              Response.json(
+                {
+                  error: { type: "MissingSessionID", message: "Session ID is required" },
+                },
+                { status: 400 },
+              ),
+            )
+          sessions.push(session)
+          return HttpClientResponse.fromWeb(
+            request,
+            new Response(
+              `data: ${JSON.stringify({
+                id: "completion",
+                object: "chat.completion.chunk",
+                created: 1,
+                model: "gemini",
+                choices: [{ index: 0, delta: { content: "OK" }, finish_reason: "stop" }],
+              })}\n\ndata: [DONE]\n\n`,
+              { headers: { "content-type": "text/event-stream" } },
+            ),
+          )
+        }),
+      ),
+    )
+    const native = LLMClient.layer.pipe(Layer.provide(RequestExecutor.layer.pipe(Layer.provide(http))))
+    yield* Effect.gen(function* () {
+      const generate = yield* Generate.Service
+      for (let index = 0; index < 2; index++) {
+        expect(
+          yield* generate.text({
+            prompt: "Return exactly OK",
+            model: Ref.make({ providerID: selected.providerID, id: selected.id }),
+          }),
+        ).toBe("OK")
+      }
+    }).pipe(Effect.provide(Generate.layer.pipe(Layer.provide(Layer.merge(resolver, native)))))
+    expect(sessions).toHaveLength(2)
+    expect(sessions[0]).toStartWith("ses_")
+    expect(sessions[1]).not.toBe(sessions[0])
+  }),
+)

packages/core/test/pty/pty-session.test.ts

@@ -5,13 +5,12 @@ import { LayerNode } from "@opencode/util/effect/layer-node"
 import { Bus } from "@opencode/core/bus"
 import { Location } from "@opencode/core/location"
 import { Pty } from "@opencode/core/pty"
-import { PtyID } from "@opencode/core/pty/schema"
 import { AbsolutePath } from "@opencode/core/schema"
 import { ShellSelect } from "@opencode/core/shell/select"
 import { location } from "../fixture/location"
 import { testEffect } from "../lib/effect"
 
-type PtyEvent = { type: "created" | "exited" | "deleted"; id: PtyID }
+type PtyEvent = { type: "created" | "exited" | "deleted"; id: Pty.ID }
 
 const locationLayer = Layer.succeed(
   Location.Service,
@@ -46,7 +45,7 @@ const createPty = Effect.fn("PtySessionTest.createPty")(function* (command: stri
   )
 })
 
-const waitForEvents = (events: Queue.Queue<PtyEvent>, id: PtyID, count: number) =>
+const waitForEvents = (events: Queue.Queue<PtyEvent>, id: Pty.ID, count: number) =>
   Effect.gen(function* () {
     const picked: Array<PtyEvent["type"]> = []
     while (picked.length < count) {
@@ -61,7 +60,7 @@ const waitForEvents = (events: Queue.Queue<PtyEvent>, id: PtyID, count: number)
     }),
   )
 
-const attachCollecting = Effect.fn("PtySessionTest.attachCollecting")(function* (id: PtyID, cursor?: number) {
+const attachCollecting = Effect.fn("PtySessionTest.attachCollecting")(function* (id: Pty.ID, cursor?: number) {
   const pty = yield* Pty.Service
   const output = yield* Queue.unbounded<string>()
   const ended = yield* Deferred.make<{ exitCode?: number }>()
@@ -90,7 +89,7 @@ describe("pty", () => {
   it.live("returns typed not found errors for missing sessions", () =>
     Effect.gen(function* () {
       const pty = yield* Pty.Service
-      const id = PtyID.make("pty_missing")
+      const id = Pty.ID.make("pty_missing")
 
       for (const result of [
         yield* pty.get(id).pipe(Effect.asVoid, Effect.exit),

packages/core/test/pty/ticket.test.ts

@@ -1,7 +1,7 @@
 import { describe, expect } from "bun:test"
 import { Effect, Layer } from "effect"
 import { LayerNode } from "@opencode/util/effect/layer-node"
-import { PtyID } from "@opencode/core/pty/schema"
+import { Pty } from "@opencode/core/pty"
 import { PtyTicket } from "@opencode/core/pty/ticket"
 import { Workspace } from "@opencode/core/workspace"
 import { testEffect } from "../lib/effect"
@@ -17,7 +17,7 @@ describe("PTY websocket tickets", () => {
   it.live("consumes tickets once", () =>
     Effect.gen(function* () {
       const tickets = yield* PtyTicket.Service
-      const scope = { ptyID: PtyID.ascending(), directory: "/tmp/a" }
+      const scope = { ptyID: Pty.ID.ascending(), directory: "/tmp/a" }
       const issued = yield* tickets.issue(scope)
 
       expect(yield* tickets.consume({ ...scope, ticket: issued.ticket })).toBe(true)
@@ -28,7 +28,7 @@ describe("PTY websocket tickets", () => {
   it.live("rejects tickets scoped to a different request", () =>
     Effect.gen(function* () {
       const tickets = yield* PtyTicket.Service
-      const ptyID = PtyID.ascending()
+      const ptyID = Pty.ID.ascending()
       const issued = yield* tickets.issue({ ptyID, directory: "/tmp/a" })
 
       expect(yield* tickets.consume({ ptyID, directory: "/tmp/b", ticket: issued.ticket })).toBe(false)
@@ -39,7 +39,7 @@ describe("PTY websocket tickets", () => {
   itExpiring.live("rejects tickets after the TTL elapses", () =>
     Effect.gen(function* () {
       const tickets = yield* PtyTicket.Service
-      const ptyID = PtyID.ascending()
+      const ptyID = Pty.ID.ascending()
       const issued = yield* tickets.issue({ ptyID })
 
       yield* Effect.promise(() => new Promise((resolve) => setTimeout(resolve, 25)))
@@ -51,7 +51,7 @@ describe("PTY websocket tickets", () => {
   it.live("rejects tickets scoped to a different workspace", () =>
     Effect.gen(function* () {
       const tickets = yield* PtyTicket.Service
-      const ptyID = PtyID.ascending()
+      const ptyID = Pty.ID.ascending()
       const workspaceID = Workspace.ID.ascending()
       const issued = yield* tickets.issue({ ptyID, workspaceID })
 

packages/core/test/repository.test.ts

@@ -41,6 +41,12 @@ describe("Repository", () => {
     })
   })
 
+  test("caches a host with a port under a directory per host part", () => {
+    expect(Repository.cachePath("/cache", Repository.parseRemote("ssh://git@example.com:2222/owner/repo"))).toBe(
+      path.join("/cache", "example.com", "2222", "owner", "repo"),
+    )
+  })
+
   test("keeps local file repositories distinct from remote repositories", () => {
     const localPath = path.resolve("repo.git")
     const reference = Repository.parse(pathToFileURL(localPath).href)
@@ -62,6 +68,18 @@ describe("Repository", () => {
     expect(() => Repository.validateBranch("bad branch")).toThrow(Repository.InvalidBranchError)
   })
 
+  test.each([
+    "..:repo",
+    "git@..:repo",
+    "../owner/repo",
+    "ssh://../repo",
+    "ssh://..:22/repo",
+    "https://%2e%2e/owner/repo",
+    ".:repo",
+  ])("rejects %s because its host contains a relative path segment", (input) => {
+    expect(() => Repository.parseRemote(input)).toThrow(Repository.InvalidReferenceError)
+  })
+
   test("compares cache identity independent of input spelling", () => {
     const shorthand = Repository.parseRemote("owner/repo")
 

packages/core/test/session-compaction.test.ts

@@ -445,12 +445,12 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
     expect(requests).toHaveLength(1)
     expect(requests[0]?.promptCacheKey).toBe(parentID)
     expect(requests[0]?.http?.headers).toEqual({
-      "x-session-affinity": sessionID,
-      "X-Session-Id": sessionID,
+      "x-session-affinity": parentID,
+      "X-Session-Id": parentID,
       "x-parent-session-id": parentID,
       "User-Agent": App.useragent(App.make()),
       "x-opencode-project": Project.ID.global,
-      "x-opencode-session": sessionID,
+      "x-opencode-session": parentID,
       "x-opencode-client": "opencode",
     })
     expect(requests[0]?.generation).toEqual(GenerationOptions.make({ maxTokens: 20_000 }))

packages/core/test/session-execution.test.ts

@@ -15,7 +15,6 @@ import { AbsolutePath } from "@opencode/core/schema"
 import { Session } from "@opencode/core/session"
 import { SessionExecution } from "@opencode/core/session/execution"
 import { SessionRestart } from "@opencode/core/session/execution/restart"
-import { UserInterruptedError } from "@opencode/core/session/error"
 import { SessionEvent } from "@opencode/core/session/event"
 import { SessionInbox } from "@opencode/core/session/inbox"
 import { SessionMessage } from "@opencode/core/session/message"
@@ -50,10 +49,6 @@ describe("SessionExecution lifecycle", () => {
     const interrupted = Effect.runSyncExit(Effect.interrupt)
     expect(SessionExecution.terminal(interrupted)).toEqual({ type: "interrupted", reason: "shutdown" })
     expect(SessionExecution.terminal(interrupted, "user")).toEqual({ type: "interrupted", reason: "user" })
-    expect(SessionExecution.terminal(Exit.fail(new UserInterruptedError()))).toEqual({
-      type: "interrupted",
-      reason: "user",
-    })
   })
 
   it.effect("the sweep only lists claimed top-level Sessions", () =>

packages/core/test/session-runner.test.ts

@@ -128,6 +128,7 @@ const undersizedContextModel = testModel("undersized-context", { context: 1, out
 const recoveryModel = testModel("recovery", { context: 200_000, output: 1_000 })
 const fittedOutputModel = testModel("fitted-output", { context: 100_000, output: 64_000 })
 const smallWindowModel = testModel("small-window", { context: 64_000, output: 16_000 })
+const largeOutputModel = testModel("large-output", { context: 1_000_000, output: 1_000_000 })
 
 test("calculates step cost using the matching context tier", () => {
   expect(
@@ -3424,6 +3425,14 @@ describe("SessionRunnerLLM", () => {
     expect(s.requests[1]?.generation?.maxTokens).toBeGreaterThan(100_000 - 50_000 - 200)
   })
 
+  scenario("caps the output limit a large model advertises", function* (s) {
+    s.currentModel = largeOutputModel
+    yield* s.llm.push(TestLLM.text("Answer", "text-large-output"))
+    yield* s.runPrompt("Question")
+
+    expect(s.requests[0]?.generation?.maxTokens).toBe(256_000)
+  })
+
   scenario("gives the summary its full output limit when the conversation overshot the threshold", function* (s) {
     // The conversation overshot the threshold, so the prepared summary request exceeds the budget and `deliver`
     // shrinks it before sending. The output limit must follow the budget, not the oversized prepared request.
@@ -4693,7 +4702,12 @@ describe("SessionRunnerLLM", () => {
       .pipe(Effect.orDie)
     yield* s.runPrompt("Run child request")
 
-    expect(s.requests[0]?.http?.headers?.["x-parent-session-id"]).toBe(parentID)
+    expect(s.requests[0]?.http?.headers).toMatchObject({
+      "x-session-affinity": parentID,
+      "X-Session-Id": parentID,
+      "x-parent-session-id": parentID,
+      "x-opencode-session": parentID,
+    })
     expect(s.requests[0]?.promptCacheKey).toBe(parentID)
   })
 

packages/core/test/util/process-lock.test.ts

@@ -1,71 +0,0 @@
-import { expect } from "bun:test"
-import { ProcessLock } from "@opencode/core/util/process-lock"
-import { Effect } from "effect"
-import fs from "node:fs/promises"
-import os from "node:os"
-import path from "node:path"
-import { it } from "../lib/effect"
-
-const worker = path.join(import.meta.dir, "../fixture/process-lock-worker.ts")
-
-it.live(
-  "releases ownership when the scope closes",
-  Effect.gen(function* () {
-    const root = yield* temp("opencode-process-lock-")
-    const file = path.join(root, "service.lock")
-    yield* Effect.scoped(ProcessLock.acquire(file))
-    yield* Effect.scoped(ProcessLock.acquire(file))
-  }),
-)
-
-it.live(
-  "releases ownership when the process dies",
-  Effect.gen(function* () {
-    const root = yield* temp("opencode-process-lock-death-")
-    const file = path.join(root, "service.lock")
-    const ready = path.join(root, "ready")
-    const child = yield* Effect.acquireRelease(
-      Effect.sync(() =>
-        Bun.spawn([process.execPath, worker, JSON.stringify({ file, ready })], {
-          stdout: "ignore",
-          stderr: "pipe",
-        }),
-      ),
-      (child) =>
-        Effect.promise(async () => {
-          kill(child)
-          await child.exited
-        }),
-    )
-    yield* Effect.promise(async () => {
-      for (let attempt = 0; attempt < 100 && !(await Bun.file(ready).exists()); attempt++) await Bun.sleep(20)
-    })
-    expect(yield* Effect.promise(() => Bun.file(ready).exists())).toBe(true)
-
-    const error = yield* Effect.scoped(ProcessLock.acquire(file)).pipe(Effect.flip)
-    expect(error._tag).toBe("ProcessLockHeldError")
-
-    if (process.platform !== "win32") {
-      process.kill(child.pid, "SIGSTOP")
-      const paused = yield* Effect.scoped(ProcessLock.acquire(file)).pipe(Effect.flip)
-      expect(paused._tag).toBe("ProcessLockHeldError")
-      process.kill(child.pid, "SIGCONT")
-    }
-
-    kill(child)
-    yield* Effect.promise(() => child.exited)
-    yield* Effect.scoped(ProcessLock.acquire(file))
-  }),
-)
-
-function temp(prefix: string) {
-  return Effect.acquireRelease(
-    Effect.promise(() => fs.mkdtemp(path.join(os.tmpdir(), prefix))),
-    (root) => Effect.promise(() => fs.rm(root, { recursive: true, force: true })),
-  )
-}
-
-function kill(child: Bun.Subprocess) {
-  if (process.platform === "win32") return child.kill()
-  return child.kill("SIGKILL")
-}

packages/server/src/workerd.ts

@@ -23,8 +23,8 @@ import type { ServerOptions } from "./options"
  *
  * - Database runs on the injected `DurableObjectStorage` SQLite.
  * - Watcher and fff are disabled through their existing option flags; pty, fff,
- *   shell-parser, photon, and process-lock native modules resolve to inert
- *   stubs under the `workerd` bundle condition.
+ *   shell-parser, and photon native modules resolve to inert stubs under the
+ *   `workerd` bundle condition.
  * - Bare locations use a typed no-execution-plane process spawner; FileSystem,
  *   FileSystemSearch, and Pty fail with a clear defect until a remote sandbox
  *   backs them; Snapshot and Vcs degrade to no-op results.

packages/tui/src/component/dialog-open.tsx

@@ -208,6 +208,8 @@ export function DialogOpen(props: { sessions: SessionInfo[]; onLoad: (sessions:
     const projectOptions = data.project
       .list()
       .filter((project) => project.canonical !== "/")
+      // Historical project identities can share a checkout. The list is newest-active first.
+      .filter((project, index, projects) => projects.findIndex((item) => item.canonical === project.canonical) === index)
       .map((project) => ({ directory: project.canonical, project }))
       .map((item) => {
         const title =

packages/tui/src/component/prompt/index.tsx

@@ -53,6 +53,7 @@ import { usePromptMove } from "./move"
 import { resolvePastedAttachments } from "./local-attachment"
 import { locationKey, useData } from "../../context/data"
 import { useLocation } from "../../context/location"
+import { useArgs } from "../../context/args"
 import { Keymap, type KeymapCommand } from "../../context/keymap"
 import { useInteractivity } from "../../context/interactivity"
 import { abbreviateHome } from "../../runtime"
@@ -206,6 +207,7 @@ export function Prompt(props: PromptProps) {
   const directoryRecents = useDirectoryRecents()
   const keymapCommands = Keymap.useCommands()
   const currentLocation = useLocation()
+  const args = useArgs()
   const config = useConfig().data
   const dialog = useDialog()
   const toast = useToast()
@@ -1229,7 +1231,9 @@ export function Prompt(props: PromptProps) {
       // a local session record synchronously, so the navigation below happens
       // immediately — enter feels sent even while the create round-trip is in
       // flight. Sends against the new session gate on the request.
+      const newSessionID = args.takeNewSessionID()
       const created = data.session.create({
+        id: newSessionID,
         location: directory ? { directory } : location,
         agent: agent.id,
         model: {
@@ -1238,6 +1242,7 @@ export function Prompt(props: PromptProps) {
           variant,
         },
       })
+      if (newSessionID !== undefined) created.request.catch(() => args.restoreNewSessionID(newSessionID))
       sessionID = created.id
       session = data.session.get(created.id)
       newSession = {

packages/tui/src/context/args.tsx

@@ -1,3 +1,4 @@
+import { mergeProps } from "solid-js"
 import { createSimpleContext } from "./helper"
 
 export interface Args {
@@ -6,11 +7,25 @@ export interface Args {
   prompt?: string
   continue?: boolean
   sessionID?: string
+  newSessionID?: string
   fork?: boolean
   auto?: boolean
 }
 
 export const { use: useArgs, provider: ArgsProvider } = createSimpleContext({
   name: "Args",
-  init: (props: Args) => props,
+  init: (props: Args) => {
+    // The first new session created from home takes this ID; later ones mint their own.
+    let pending = props.newSessionID
+    return mergeProps(props, {
+      takeNewSessionID() {
+        const id = pending
+        pending = undefined
+        return id
+      },
+      restoreNewSessionID(id: string) {
+        pending ??= id
+      },
+    })
+  },
 })

packages/tui/src/feature-plugins/prompt/btw.tsx

@@ -130,8 +130,8 @@ export function Answer(props: {
         <text attributes={TextAttributes.BOLD} fg={theme.text.base} flexShrink={0}>
           /btw
         </text>
-        <text fg={theme.text.muted} wrapMode="word" flexGrow={1}>
-          {props.question}
+        <text fg={theme.text.muted} wrapMode="none" flexGrow={1} truncate>
+          {props.question.replace(/\s+/g, " ")}
         </text>
         <text fg={theme.text.muted} flexShrink={0} onMouseUp={() => dialog.clear()}>
           esc

packages/tui/src/routes/home.tsx

@@ -65,12 +65,13 @@ export function Home() {
     untrack(() => composer.set(prompt))
   })
 
-  // Wait for the model store to be ready before auto-submitting --prompt.
+  // Wait for everything submit needs before auto-submitting --prompt; it runs once.
   createEffect(() => {
     const r = ref()
     if (sent) return
     if (!r) return
-    if (!local.model.ready) return
+    if (!local.model.ready || !local.model.catalogReady) return
+    if (!local.agent.current() || !local.model.current()) return
     if (!args.prompt) return
     if (r.current.text !== args.prompt) return
     sent = true

packages/tui/src/routes/session/index.tsx

@@ -2808,7 +2808,7 @@ function Shell(props: ToolProps) {
       command={stringValue(props.input.command)}
       workdir={stringValue(props.input.workdir)}
       status={props.part.state.status}
-      background={Boolean(stringValue(props.metadata.shellID)) && props.part.state.status !== "running"}
+      background={props.part.state.status === "completed" && props.metadata.status === "running"}
       output={stringValue(props.metadata.shellID) ? undefined : props.output}
     />
   )

packages/tui/test/app-lifecycle.test.tsx

@@ -876,6 +876,61 @@ test("session startup prompt is submitted exactly once", async () => {
   }
 })
 
+test("home startup prompt is submitted exactly once", async () => {
+  await using state = await tmpdir()
+  const cwd = process.cwd()
+  const location = { directory: cwd, project: { id: "project", directory: cwd, canonical: cwd } }
+  const bodies: unknown[] = []
+  const submitted = Promise.withResolvers<void>()
+  let session: unknown
+  await using setup = await createAppFixture({
+    state: state.path,
+    args: { prompt: "HOME_READY" },
+    config: { animations: false, tabs: { mode: "off" } },
+    fetch: async (url, request) => {
+      if (url.pathname === "/api/location") return json(location)
+      if (url.pathname === "/api/fs/list") return json({ location, data: [] })
+      if (url.pathname === "/api/agent")
+        return json({ location, data: [{ id: "build", mode: "primary", hidden: false, permissions: [] }] })
+      if (url.pathname === "/api/model")
+        return json({ location, data: [{ id: "model", providerID: "provider", name: "Model", variants: [] }] })
+      if (url.pathname === "/api/provider") return json({ location, data: [{ id: "provider", name: "Provider" }] })
+      if (url.pathname === "/api/session" && request.method === "POST") {
+        const input: unknown = await request.json()
+        if (typeof input !== "object" || input === null) throw new Error("Expected a session input")
+        session = {
+          ...input,
+          projectID: "project",
+          cost: 0,
+          tokens: { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } },
+          time: { created: 0, updated: 0 },
+        }
+        return json({ data: session })
+      }
+      if (/^\/api\/session\/[^/]+\/prompt$/.test(url.pathname)) {
+        bodies.push(await request.json())
+        submitted.resolve()
+        return json({ data: {} })
+      }
+      if (/^\/api\/session\/[^/]+\/(message|inbox|permission)$/.test(url.pathname))
+        return json({ data: [], cursor: {} })
+      if (session && /^\/api\/session\/[^/]+$/.test(url.pathname)) return json({ data: session })
+      return undefined
+    },
+  })
+
+  await setup.ready
+  await Promise.race([
+    submitted.promise,
+    Bun.sleep(2000).then(() => {
+      throw new Error("startup prompt was not submitted")
+    }),
+  ])
+  await Bun.sleep(20)
+  expect(bodies).toHaveLength(1)
+  expect(bodies[0]).toMatchObject({ text: "HOME_READY" })
+})
+
 test.each([false, true])("uses the resolved launch directory for new prompts (fallback: %s)", async (fallback) => {
   await using state = await tmpdir()
   const target = fallback ? directory : process.cwd()

packages/tui/test/new-session-id.test.tsx

@@ -0,0 +1,100 @@
+import { expect, test } from "bun:test"
+import { createAppFixture } from "./fixture/app"
+import { json } from "./fixture/tui-client"
+
+type SessionInput = { id?: string; location?: { directory: string } }
+
+const location = {
+  directory: "/tmp/opencode/packages/tui",
+  project: { id: "project", directory: "/tmp/opencode", canonical: "/tmp/opencode" },
+}
+
+function sessionInfo(record: SessionInput) {
+  return {
+    ...record,
+    location: record.location ?? { directory: location.directory },
+    projectID: "project",
+    cost: 0,
+    tokens: { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } },
+    time: { created: 0, updated: 0 },
+  }
+}
+
+async function launch(options: { failCreates?: number } = {}) {
+  const attempts: (string | undefined)[] = []
+  const created: SessionInput[] = []
+  const prompts: string[] = []
+  let failures = options.failCreates ?? 0
+  const setup = await createAppFixture({
+    args: { newSessionID: "ses_chosen" },
+    config: { animations: false, keybinds: { "session.new": "f6" } },
+    fetch: async (url, request) => {
+      if (url.pathname === "/api/agent")
+        return json({ location, data: [{ id: "build", mode: "primary", hidden: false, permissions: [] }] })
+      if (url.pathname === "/api/provider") return json({ location, data: [{ id: "demo", name: "Demo" }] })
+      if (url.pathname === "/api/model")
+        return json({ location, data: [{ id: "model", providerID: "demo", name: "Demo Model", variants: [] }] })
+      if (url.pathname === "/api/session" && request.method === "POST") {
+        const record: SessionInput = await request.json()
+        attempts.push(record.id)
+        if (failures > 0) {
+          failures--
+          return json({ message: "create failed" }, { status: 500 })
+        }
+        created.push(record)
+        return json({ data: sessionInfo(record) })
+      }
+      if (/^\/api\/session\/[^/]+\/prompt$/.test(url.pathname)) {
+        prompts.push(url.pathname.split("/")[3] ?? "")
+        return json({ data: {} })
+      }
+      if (/^\/api\/session\/[^/]+\/(message|inbox|permission)$/.test(url.pathname))
+        return json({ data: [], cursor: {} })
+      if (/^\/api\/session\/[^/]+\/(agent|model)$/.test(url.pathname)) return new Response(null, { status: 204 })
+      if (/^\/api\/session\/[^/]+$/.test(url.pathname)) {
+        const record = created.find((item) => item.id === url.pathname.split("/")[3])
+        if (!record) return json({ message: "not found" }, { status: 404 })
+        return json({ data: sessionInfo(record) })
+      }
+      return undefined
+    },
+  })
+  return { setup, attempts, created, prompts }
+}
+
+test("the first new session uses the launch session ID and later ones mint their own", async () => {
+  const run = await launch()
+  await using setup = run.setup
+
+  await setup.ready
+  await setup.waitForFrame((frame) => frame.includes("Demo Model"))
+  await setup.mockInput.typeText("hello")
+  setup.mockInput.pressEnter()
+  await setup.waitForFrame(() => run.prompts.length === 1)
+  expect(run.prompts[0]).toBe("ses_chosen")
+  expect(run.created.map((item) => item.id)).toEqual(["ses_chosen"])
+
+  setup.mockInput.pressKey("F6")
+  await setup.renderOnce()
+  await setup.mockInput.typeText("again")
+  setup.mockInput.pressEnter()
+  await setup.waitForFrame(() => run.created.length === 2)
+  expect(run.created[1]?.id).toMatch(/^ses/)
+  expect(run.created[1]?.id).not.toBe("ses_chosen")
+})
+
+test("a failed first create keeps the launch session ID for the retry", async () => {
+  const run = await launch({ failCreates: 1 })
+  await using setup = run.setup
+
+  await setup.ready
+  await setup.waitForFrame((frame) => frame.includes("Demo Model"))
+  await setup.mockInput.typeText("hello")
+  setup.mockInput.pressEnter()
+  await setup.waitForFrame((frame) => frame.includes("Creating a session failed") && frame.includes("hello"))
+  expect(run.attempts).toEqual(["ses_chosen"])
+
+  setup.mockInput.pressEnter()
+  await setup.waitForFrame(() => run.created.length === 1)
+  expect(run.attempts).toEqual(["ses_chosen", "ses_chosen"])
+})

packages/web/src/content/docs/ar/zen.mdx

@@ -283,6 +283,7 @@ https://opencode.ai/zen/v1/models
 - Nemotron 3 Ultra Free (نقاط نهاية NVIDIA المجانية): للاستخدام التجريبي فقط — لا ترسل بيانات شخصية أو سرية. يُسجَّل استخدامك لأغراض أمنية ولتحسين منتجات وخدمات NVIDIA. بيانات الجلسة المُسجَّلة لأغراض التحسين غير مرتبطة بهويتك أو بأي مُعرِّف دائم. لمزيد من المعلومات حول ممارسات معالجة البيانات لدينا، راجع [سياسة الخصوصية](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf). بتفاعلك مع نقطة النهاية هذه، فإنك توافق على جمعنا لهذه المعلومات وتسجيلها واستخدامها وعلى [شروط خدمة النسخة التجريبية من واجهة NVIDIA API](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf).
 - OpenAI APIs: يتم الاحتفاظ بالطلبات لمدة 30 يوما وفقا لـ [OpenAI's Data Policies](https://platform.openai.com/docs/guides/your-data).
 - Anthropic APIs: يتم الاحتفاظ بالطلبات لمدة 30 يوما وفقا لـ [Anthropic's Data Policies](https://docs.anthropic.com/en/docs/claude-code/data-usage).
+- Jev APIs: لا تُستخدم المطالبات والمدخلات الأخرى في التدريب. يُحتفظ بالبيانات وفقًا لـ [سياسة الخصوصية الخاصة بـ TypeSafe AI](https://typesafe.ai/legal/privacy-policy).
 
 ---
 

packages/web/src/content/docs/bs/zen.mdx

@@ -295,6 +295,7 @@ i ne koriste vaše podatke za treniranje modela, uz sljedeće izuzetke:
 - Nemotron 3 Ultra Free (besplatni NVIDIA endpointi): Samo za probnu upotrebu — nemojte slati lične ili povjerljive podatke. Vaše korištenje se bilježi radi sigurnosti i poboljšanja NVIDIA proizvoda i usluga. Zabilježeni podaci sesije koji se koriste u svrhu poboljšanja nisu povezani s vašim identitetom niti bilo kojim trajnim identifikatorom. Za više informacija o našim praksama obrade podataka pogledajte našu [Politiku privatnosti](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf). Interakcijom s ovim endpointom pristajete na naše prikupljanje, bilježenje i korištenje takvih informacija te na [NVIDIA API Trial Terms of Service](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf).
 - OpenAI APIs: Requests are retained for 30 days in accordance with [OpenAI's Data Policies](https://platform.openai.com/docs/guides/your-data).
 - Anthropic APIs: Requests are retained for 30 days in accordance with [Anthropic's Data Policies](https://docs.anthropic.com/en/docs/claude-code/data-usage).
+- Jev APIs: Promptovi i drugi ulazni podaci ne koriste se za treniranje modela. Podaci se čuvaju u skladu s [Politikom privatnosti TypeSafe AI](https://typesafe.ai/legal/privacy-policy).
 
 ---
 

packages/web/src/content/docs/da/zen.mdx

@@ -293,6 +293,7 @@ Alle vores modeller hostes i US. Vores udbydere følger en nul-opbevaringspoliti
 - Nemotron 3 Ultra Free (gratis NVIDIA-endpoints): Kun til prøvebrug — indsend ikke personlige eller fortrolige data. Din brug logges af sikkerhedshensyn og for at forbedre NVIDIAs produkter og tjenester. De loggede sessionsdata, der bruges til forbedringsformål, er ikke knyttet til din identitet eller nogen vedvarende identifikator. For mere information om vores databehandlingspraksis, se vores [privatlivspolitik](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf). Ved at interagere med dette endpoint giver du samtykke til vores indsamling, registrering og brug af sådanne oplysninger samt [NVIDIA API Trial Terms of Service](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf).
 - OpenAI APIs: Anmodninger opbevares i 30 dage i overensstemmelse med [OpenAI's Data Policies](https://platform.openai.com/docs/guides/your-data).
 - Anthropic APIs: Anmodninger opbevares i 30 dage i overensstemmelse med [Anthropic's Data Policies](https://docs.anthropic.com/en/docs/claude-code/data-usage).
+- Jev APIs: Prompts og andre input bruges ikke til træning. Data opbevares i overensstemmelse med [TypeSafe AI's privatlivspolitik](https://typesafe.ai/legal/privacy-policy).
 
 ---
 

packages/web/src/content/docs/de/zen.mdx

@@ -279,6 +279,7 @@ Alle unsere Modelle werden in den USA gehostet. Unsere Provider folgen einer Zer
 - Nemotron 3 Ultra Free (kostenlose NVIDIA-Endpunkte): Nur für Testzwecke — übermitteln Sie keine personenbezogenen oder vertraulichen Daten. Ihre Nutzung wird zu Sicherheitszwecken und zur Verbesserung der Produkte und Dienste von NVIDIA protokolliert. Die zu Verbesserungszwecken protokollierten Sitzungsdaten sind nicht mit Ihrer Identität oder einem dauerhaften Identifikator verknüpft. Weitere Informationen zu unseren Datenverarbeitungspraktiken finden Sie in unserer [Datenschutzrichtlinie](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf). Durch die Interaktion mit diesem Endpunkt stimmen Sie unserer Erhebung, Aufzeichnung und Nutzung solcher Informationen sowie den [NVIDIA API Trial Terms of Service](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf) zu.
 - OpenAI APIs: Anfragen werden in Übereinstimmung mit [OpenAI's Data Policies](https://platform.openai.com/docs/guides/your-data) 30 Tage lang gespeichert.
 - Anthropic APIs: Anfragen werden in Übereinstimmung mit [Anthropic's Data Policies](https://docs.anthropic.com/en/docs/claude-code/data-usage) 30 Tage lang gespeichert.
+- Jev APIs: Prompts und andere Eingaben werden nicht für das Training verwendet. Daten werden gemäß der [Datenschutzrichtlinie von TypeSafe AI](https://typesafe.ai/legal/privacy-policy) gespeichert.
 
 ---
 

packages/web/src/content/docs/es/zen.mdx

@@ -293,6 +293,7 @@ Todos nuestros modelos están alojados en US. Nuestros proveedores siguen una po
 - Nemotron 3 Ultra Free (endpoints gratuitos de NVIDIA): Solo para uso de prueba — no envíes datos personales ni confidenciales. Tu uso se registra con fines de seguridad y para mejorar los productos y servicios de NVIDIA. Los datos de sesión registrados con fines de mejora no están vinculados a tu identidad ni a ningún identificador persistente. Para obtener más información sobre nuestras prácticas de procesamiento de datos, consulta nuestra [Política de privacidad](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf). Al interactuar con este endpoint, aceptas que recopilemos, registremos y usemos dicha información, así como los [NVIDIA API Trial Terms of Service](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf).
 - OpenAI APIs: Las solicitudes se conservan durante 30 días de acuerdo con [OpenAI's Data Policies](https://platform.openai.com/docs/guides/your-data).
 - Anthropic APIs: Las solicitudes se conservan durante 30 días de acuerdo con [Anthropic's Data Policies](https://docs.anthropic.com/en/docs/claude-code/data-usage).
+- Jev APIs: Los prompts y otras entradas no se usan para el entrenamiento. Los datos se conservan de acuerdo con la [política de privacidad de TypeSafe AI](https://typesafe.ai/legal/privacy-policy).
 
 ---
 

packages/web/src/content/docs/fr/zen.mdx

@@ -279,6 +279,7 @@ Tous nos modèles sont hébergés aux US. Nos fournisseurs suivent une politique
 - Nemotron 3 Ultra Free (endpoints NVIDIA gratuits) : Réservé à un usage d'essai — n'envoyez pas de données personnelles ou confidentielles. Votre utilisation est journalisée à des fins de sécurité et pour améliorer les produits et services de NVIDIA. Les données de session journalisées à des fins d'amélioration ne sont pas liées à votre identité ni à un quelconque identifiant persistant. Pour plus d'informations sur nos pratiques de traitement des données, consultez notre [Politique de confidentialité](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf). En interagissant avec cet endpoint, vous consentez à notre collecte, à notre enregistrement et à notre utilisation de ces informations ainsi qu'aux [NVIDIA API Trial Terms of Service](https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf).
 - OpenAI APIs : Les requêtes sont conservées pendant 30 jours conformément à [OpenAI's Data Policies](https://platform.openai.com/docs/guides/your-data).
 - Anthropic APIs : Les requêtes sont conservées pendant 30 jours conformément à [Anthropic's Data Policies](https://docs.anthropic.com/en/docs/claude-code/data-usage).
+- Jev APIs : Les prompts et autres données saisies ne sont pas utilisés pour l'entraînement. Les données sont conservées conformément à la [politique de confidentialité de TypeSafe AI](https://typesafe.ai/legal/privacy-policy).
 
 ---