fix(ai): keep OpenAI image output settings and Z.ai URL expiry (#51380 )

fix(ai): harden OpenAI transcription stream parsing (#51378 )
fix(ai): tighten media error consistency (#51374 )
2026-09-25 18:17:35 +00:00 · 2026-09-25 22:51:46 +05:30 · 2026-09-25 22:51:22 +05:30 · 2026-09-25 22:46:04 +05:30 · 2026-09-25 22:37:20 +05:30 · 2026-09-25 22:36:02 +05:30
53 changed files with 832 additions and 563 deletions
@@ -12,9 +12,9 @@

 Per-type constructors live on the type, not as top-level re-exports. Use `Message.system(...)`, `Message.user(...)`, `Message.assistant(...)`, `Message.tool(...)`, `Message.media(...)`, `LanguageModel.make(...)`, `ToolDefinition.make(...)`, `ToolCallPart.make(...)`, `ToolResultPart.make(...)`, `ToolChoice.make(...)`, `ToolChoice.named(...)`, `SystemPart.make(...)`, and `GenerationOptions.make(...)` directly. The top-level `LLM` namespace is reserved for request-shaped call APIs: `LLM.request`, `LLM.generate`, `LLM.stream`, and `LLM.generateObject`. `LLM.generate`/`LLM.stream` and Promise `ai.llm.generate`/`ai.llm.stream` accept ergonomic input or an `LLMRequest`; both paths use the same canonical request. Core still builds, logs, replays, and updates that durable `LLMRequest` boundary. Use `LLMRequest.update(...)` when deriving canonical request data; do not add a duplicate `LLM.updateRequest(...)` path.

-Modality namespaces mirror `LLM` exactly: `Image.request`, `Image.generate`, `Image.stream` (later `Video`, `Speech`, `Transcription`). Common request fields (`images`, `mask`, `n`, `size`, `aspectRatio`, `seed`, `format`) lower natively or fail with a typed `AIError`; provider-native controls always live under `providerOptions`, never under a modality-specific `options` key.
+Modality namespaces mirror `LLM` exactly: `Image.request`, `Image.generate`, `Image.stream`, and the same for `Video`, `Speech`, and `Transcription`. Common request fields (`images`, `mask`, `n`, `size`, `aspectRatio`, `seed`, `format`) lower natively or fail with a typed `AIError`; provider-native controls always live under `providerOptions`, never under a modality-specific `options` key.

-Media payloads are always `Media.Asset` (`src/media.ts`). Construct them with `Media.bytes`, `Media.base64`, `Media.url`, `Media.ref`, `Media.fromDataUrl`, or `Media.file`; never introduce a parallel `data: string | Uint8Array` shape. `MediaPart.media`, `ImageRequest.images`/`mask`, `ImageResponse.images`, and the `media` `LLMEvent` all share it. Protocols branch on `asset.source.type` and `asset.kind` and use `ProviderShared.inlineMedia` / `requireInlineMedia` / `mediaUrl` / `MediaInput.refID` rather than re-deriving base64 or URL handling.
+Media payloads are always `Media.Asset` (`src/media.ts`). Construct them with `Media.bytes`, `Media.base64`, `Media.url`, `Media.ref`, `Media.fromDataUrl`, or `Media.file`; never introduce a parallel `data: string | Uint8Array` shape. `MediaPart.media`, `ImageRequest.images`/`mask`, `ImageResponse.images`, and the `media` `LLMEvent` all share it. Protocols branch on `asset.source.type` and `asset.kind` and use `ProviderShared.requireInlineMedia` / `inlineRequired` / `mediaUrl` / `mediaReference` and `MediaInput.inlineBytes` / `refID` rather than re-deriving base64 or URL handling.

 `schema/messages.ts → media.ts → route/executor-service.ts` is an accepted runtime dependency from the schema layer on the executor service tag: `Media.Asset.bytes()` must be able to download `url` sources, and the tag lives in that leaf module precisely so the schema barrel never imports the executor implementation (which imports the schema barrel back). Do not move the tag into `route/executor.ts` or import `route/executor.ts` from `src/schema/*` or `src/media.ts`.

@@ -98,7 +98,7 @@ 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`, 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`. `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.

@@ -112,7 +112,7 @@ For providers where the URL is derived from typed inputs (Azure resource name, B

 ### Provider Facades

-Provider-facing APIs are configured facades over route values. Endpoint/auth/resource/API-version setup happens before model selection, and model selectors accept only a model or deployment id. Media models use per-modality selectors on the same facade (`openai.image(id)`, later `.video` / `.speech` / `.transcription`) that mirror `openai.responses(id)`; the one-word overlap with the request namespace is accepted over a second construction path:
+Provider-facing APIs are configured facades over route values. Endpoint/auth/resource/API-version setup happens before model selection, and model selectors accept only a model or deployment id. Media models use per-modality selectors on the same facade (`openai.image(id)`, `.speech(id)`, `.transcription(id)`, `google.video(id)`) that mirror `openai.responses(id)`; the one-word overlap with the request namespace is accepted over a second construction path:

 ```ts
 const openai = OpenAI.configure({ apiKey, baseURL })
@@ -475,18 +475,18 @@ const program = Effect.gen(function* () {
 Common fields are portable in shape, not in support. Unsupported fields fail with a typed `AIError` before any network
 call rather than being dropped, so check this table before swapping only the `model`:

-| Provider              | `n` | `size`    | `aspectRatio` | `seed` | `format` | `images`                  | `mask`              |
-| --------------------- | --- | --------- | ------------- | ------ | -------- | ------------------------- | ------------------- |
-| OpenAI                | ✓¹  | ✓         | ✗             | ✗      | ✓        | ✓                         | ✓                   |
-| Google (Gemini)       | 1   | ✗         | ✓             | ✓      | ✗        | ✓ (no public URLs)        | ✗                   |
-| xAI                   | ✓   | ✗         | ✓             | ✗      | ✗        | ✓                         | ✗                   |
-| Z.ai                  | ✗   | ✓         | ✗             | ✗      | ✗        | ✗                         | ✗                   |
-| Meta                  | ✓   | ✓ (hint)  | ✗             | ✗      | ✓        | ✓                         | ✗                   |
-| Black Forest Labs     | 1   | per model | per model     | ✓      | ✓        | per model (1–8)           | `flux-pro-1.0-fill` |
-| fal                   | ✓   | per model | per model     | ✓      | ✓        | 1 (several on `/edit`)    | ✓                   |
-| Replicate             | ✗   | ✗         | ✗             | ✗      | ✗        | ✗ (use `providerOptions`) | ✗                   |
-| Stability `image`     | 1   | ✗         | ✓             | ✓      | ✓        | 1 (not on `core`)         | ✗                   |
-| Stability `upscale()` | ✗   | ✗         | ✗             | ✓      | ✓        | exactly 1 (required)      | ✗                   |
+| Provider              | `n` | `size`    | `aspectRatio` | `seed` | `format` | `images`                         | `mask`              |
+| --------------------- | --- | --------- | ------------- | ------ | -------- | -------------------------------- | ------------------- |
+| OpenAI                | ✓¹  | ✓         | ✗             | ✗      | ✓        | ✓                                | ✓                   |
+| Google (Gemini)       | 1   | ✗         | ✓             | ✓      | ✗        | ✓ (no public URLs)               | ✗                   |
+| xAI                   | ✓   | ✗         | ✓             | ✗      | ✗        | ✓                                | ✗                   |
+| Z.ai                  | ✗   | ✓         | ✗             | ✗      | ✗        | ✗                                | ✗                   |
+| Meta                  | ✓   | ✓ (hint)  | ✗             | ✗      | ✓        | ✓                                | ✗                   |
+| Black Forest Labs     | 1   | per model | per model     | ✓      | ✓        | per model (1–8)                  | `flux-pro-1.0-fill` |
+| fal                   | ✓   | per model | per model     | ✓      | ✓        | 1 (several on `/edit`, `/multi`) | ✓                   |
+| Replicate             | ✗   | ✗         | ✗             | ✗      | ✗        | ✗ (use `providerOptions`)        | ✗                   |
+| Stability `image`     | 1   | ✗         | ✓             | ✓      | ✓        | 1 (not on `core`)                | ✗                   |
+| Stability `upscale()` | ✗   | ✗         | ✗             | ✓      | ✓        | exactly 1 (required)             | ✗                   |

 ✓ lowers natively; ✗ fails whenever the field is set (including `n: 1`); `1` means `n > 1` fails. ¹ `Image.stream` on OpenAI generates one image. fal
 rejects `size` and `aspectRatio` together; which one a fal or BFL model takes depends on the model.
@@ -621,8 +621,7 @@ persist the bytes promptly if they must remain available.
 ### Partial images

 OpenAI's GPT image models stream previews. `Image.stream` sends `stream: true` with `partialImages` (0–3, default 2)
-and emits `image-partial` events before each final `image`; `Image.generate` keeps the plain JSON request.
-`dall-e-*` models do not stream and fail typed:
+and emits `image-partial` events before each final `image`; `Image.generate` keeps the plain JSON request:

 ```ts
 import { Stream } from "effect"
@@ -699,7 +698,7 @@ const program = Effect.gen(function* () {
 })
 ```

-The hosted result is represented as a provider-executed tool call and tool result, and the generated image is also emitted as a first-class `media` `LLMEvent` (`response.message` then carries a `media` part). Gemini image-capable models emit the same `media` event for inline image output. Retaining `response.message` preserves the generated image for continuation on both routes.
+The hosted result is represented as a provider-executed tool call and a tool result whose content carries the generated image as a file. Gemini image-capable models instead emit a first-class `media` `LLMEvent` for inline image output (`response.message` then carries a `media` part). Retaining `response.message` preserves the generated image for continuation on both routes.

 ## Video generation

@@ -840,9 +839,10 @@ Provider notes:
 - **OpenAI** streams over SSE (`stream_format: "sse"`), which is also the only place it reports token usage; `tts-1`
  and `tts-1-hd` do not support SSE and stream the raw audio body instead. `pcm` is 24 kHz 16-bit mono. `language`
  and `timestamps` are not supported.
- **Gemini TTS** returns raw 16-bit PCM only (`audio/L16;codec=pcm;rate=24000`), so any `format` other than `pcm`
-  fails typed; wrap the samples yourself. Style is directed in the text, so `instructions` and `speed` fail typed.
-  Only `gemini-3.1-flash-tts-preview` and later support streaming. Two-speaker audio goes through
+- **Gemini TTS** returns the provider's default output: WAV for Gemini 3.8 TTS `generate`, raw 16-bit PCM
+  (`audio/L16;codec=pcm;rate=24000`) otherwise. `pcm` is the only explicit `format` it accepts, and it fails typed on
+  Gemini 3.8 `generate`; the route never wraps PCM as WAV. Style is directed in the text, so `instructions` and
+  `speed` fail typed. Only `gemini-3.1-flash-tts-preview` and later support streaming. Two-speaker audio goes through
  `providerOptions.speechConfig.multiSpeakerVoiceConfig`.
 - **ElevenLabs** requires `voice` (the path voice id) and authenticates with `xi-api-key`. `format` maps to the
  `output_format` query parameter (`mp3_44100_128`, `pcm_24000`, `wav_24000`, `opus_48000_64`);
@@ -944,6 +944,7 @@ const transcript = await generation.await({ poll: { interval: 3_000 } })
 - **`ImageClient`** — Effect service and layer for image execution, parallel to `LLMClient`.
 - **`Media`** — the shared asset type (`Media.Asset`, `Media.Source`) and constructors used by messages, tool results, and media requests.
 - **`Generation`** — provider-neutral handle for an in-flight media generation (`await`, `refresh`, `cancel`, `events`) used by queued media routes.
+- **`Video.request` / `generate` / `stream` / `start` / `resume`** — queued video generation through a provider-neutral request; `VideoClient` is its Effect service and layer.
 - **`Speech.request` / `Speech.generate` / `Speech.stream`** — text-to-speech through a provider-neutral request; `SpeechClient` is its Effect service and layer.
 - **`Transcription.request` / `generate` / `stream` / `start` / `resume`** — speech-to-text over inline, streaming, and queued routes; `TranscriptionClient` is its Effect service and layer.
 - **`AIClient.layer` / `AIClient.layerWith(executor)`** — every modality client plus the request executor in one layer.
@@ -1,6 +1,7 @@
 # Media generation in `@opencode/ai` — public API direction

-Status: phases 1–4 implemented (through Image queued routes and partial images); phase 5 proposal.
+Status: phases 1–4 implemented (through Image queued routes and partial images; ElevenLabs Scribe transcription
+pending); phase 5 proposal.

 ## Goal

@@ -40,7 +41,7 @@ The design below is derived from a survey of the raw provider APIs (OpenAI, Gemi

 ### Model selection

-A model value is built as `OpenAI.configure({ apiKey }).responses("gpt-5")` or `.image("gpt-image-2")`: `configure` fixes credentials, endpoint, and defaults; the selector fixes which of the provider's APIs to hit and binds the typed `providerOptions` generic. Media follows the same shape with one selector per modality — `openai.image(id)` today, `.video(id)` / `.speech(id)` / `.transcription(id)` as those modalities land — mirroring `openai.responses(id)`. `Image.request` accepts `ImageModel` only, exactly as `LLM.request` accepts `LanguageModel`.
+A model value is built as `OpenAI.configure({ apiKey }).responses("gpt-5")` or `.image("gpt-image-2")`: `configure` fixes credentials, endpoint, and defaults; the selector fixes which of the provider's APIs to hit and binds the typed `providerOptions` generic. Media follows the same shape with one selector per modality — `.image(id)`, `.video(id)`, `.speech(id)`, `.transcription(id)` on the facades that offer each — mirroring `openai.responses(id)`. `Image.request` accepts `ImageModel` only, exactly as `LLM.request` accepts `LanguageModel`.

 ```ts
 import { OpenAI, Google } from "@opencode/ai/providers"
@@ -135,8 +136,8 @@ Effect.gen(function* () {
 })
 ```

-`size` and `aspectRatio` are not interchangeable; each route rejects fields it cannot lower — see the README's Image
-portability matrix.
+`size` and `aspectRatio` are not interchangeable; each route rejects fields it cannot lower — see the portability table
+in the README's Image generation section.

 Editing is not a separate function; `images`/`mask` on the request select the edit path in the route (OpenAI `/images/edits`, Gemini multimodal parts, xAI `/images/edits`). Routes that cannot honor `mask` fail with `Unsupported`.

@@ -166,8 +167,8 @@ Effect.gen(function* () {

  // Simple: wait for it.
  const response = yield* Video.generate(request, { poll: { interval: "10 seconds", timeout: "10 minutes" } })
-  response.video                                     // Media.Asset: url with expiresAt (+ transient `headers` for Veo downloads)
-  response.usage                                     // credits on Runway; the other three report none
+  response.video                                     // Media.Asset: url (expiresAt on Veo and Runway; transient `headers` for Veo downloads)
+  response.usage                                     // credits on Runway; the other three report none (xAI's usage.cost_in_usd_ticks is not decoded)
  response.notices                                   // Veo raiMediaFilteredReasons → filtered, xAI respect_moderation → moderated
  yield* response.video.materialize()                // pull bytes before the URL expires

@@ -175,7 +176,7 @@ Effect.gen(function* () {
  const generation = yield* Video.start(request)     // Generation<VideoResponse>
  generation.id; generation.status; generation.progress; generation.position; generation.token
  yield* generation.await({ poll })                  // VideoResponse
-  yield* generation.cancel()                         // fal PUT cancel_url, Runway DELETE /tasks/{id}; no-op for Veo and xAI
+  yield* generation.cancel()                         // fal PUT cancel_url, Runway DELETE /tasks/{id}; Veo and xAI succeed without a request

  // Resume from another process. The token is validated against the route's codec and refreshed once. It carries no
  // route identity, so persist the provider and model ID alongside it: `resume` needs the model.
@@ -188,10 +189,11 @@ Effect.gen(function* () {

 Tokens are route-owned JSON: Veo `{ operation }`, xAI `{ requestID }`, Runway `{ taskID }`, fal
 `{ requestID, statusURL, responseURL, cancelURL }` (fal's follow-up URLs are authoritative and absolute). Common-field
-lowering per provider: Veo takes inline media only and rejects `audio: false` and `n > 1`; xAI rejects `seed` and
-`negativePrompt` and routes a `video` input to edits or (`providerOptions.mode: "extend"`) extensions; fal rejects
-`durationSeconds`, `references`, and `frames.last` because the field names and enums differ per model; Runway passes
-`aspectRatio` through as its pixel `ratio` and rejects `n`.
+lowering per provider: Veo takes inline media only, rejects `audio: false` and `n > 1`, and requires `frames.first`
+when `frames.last` is set; xAI rejects `n`, `seed`, and `negativePrompt` and routes a `video` input to edits or
+(`providerOptions.mode: "extend"`) extensions; fal rejects `n`, plus `durationSeconds`, `references`, and `frames.last`
+because the field names and enums differ per model; Runway passes `aspectRatio` through as its pixel `ratio` and
+rejects `n`.

 Deferred: `Video.complete(model, token, webhook)` (finish from a webhook payload without polling) and provider poll
 hints (none of the four providers emit one). Later providers: Luma, Kling, MiniMax, Replicate.
@@ -241,10 +243,12 @@ name→id resolution. Multi-speaker (Gemini `speechConfig.multiSpeakerVoiceConfi
 `opus_48000_64`, Cartesia `{ container, encoding, sample_rate }`, Deepgram `encoding`+`container`) and declares the
 asset's media type rather than sniffing, because headerless PCM can look like an MPEG frame sync. Headerless PCM
 always carries `info.encoding`, `info.sampleRate`, and `info.channels`; its media type is the provider's declaration
-(Gemini `audio/L16;codec=pcm;rate=24000`, Deepgram's `content-type`) or `audio/pcm`. Gemini returns PCM only, so any
-other `format` is rejected rather than wrapped as WAV by the route. Every `format` value a route cannot produce (unknown
-to it, a container on Cartesia SSE, WAV on an ElevenLabs stream, anything but PCM on Gemini) fails the same way as an
-unsupported field: `UnsupportedOperation` with `operation: "media.format"`.
+(Gemini `audio/L16;codec=pcm;rate=24000`, Deepgram's `content-type`) or `audio/pcm`. Gemini's asset follows the
+provider's declared type: WAV for Gemini 3.8 TTS `generate`, headerless PCM otherwise. The route never wraps PCM as WAV,
+so `pcm` is the only explicit `format` it accepts, and not on Gemini 3.8 `generate`. Every `format` value a route cannot
+produce (unknown to it, a container on Cartesia SSE, WAV on an ElevenLabs stream, anything but `pcm` on Gemini, `pcm` on
+Gemini 3.8 `generate`) fails the same way as an unsupported field: `UnsupportedOperation` with
+`operation: "media.format"`.

 **Timestamps.** `timestamps: true` on the request asks for alignment. ElevenLabs selects the `with-timestamps`
 endpoints (character-level, NDJSON when streaming); Cartesia sets `add_timestamps` on `/tts/sse` (word-level; a
@@ -277,7 +281,7 @@ const request = Transcription.request({
  language: "en",                                  // provider-native passthrough
  timestamps: "segment",                           // none | segment | word
  diarize: true,
-  speakers: 2,                                     // expected count, hint only (AssemblyAI)
+  speakers: 2,                                     // exact speaker count (AssemblyAI only)
  providerOptions: { known_speaker_names: ["agent"] },
 })

@@ -291,10 +295,11 @@ yield* Transcription.resume(model, token)
 Transcription is the first modality whose providers span all three protocol kinds, and it needed no fourth kind.
 Every `MediaRoute` now carries its `kind`; `TranscriptionRoute` is the union of the inline, stream, and queued routes;
 `TranscriptionModel.fromRoute` is overloaded per protocol kind (arity picks the overload: `<Options>`,
-`<Options, Frame, State>`, `<Options, Token>`) and composes through `MediaRoute.inline` / `stream` / `queued`; and
-`TranscriptionClient` dispatches on `route.kind`. `generate` on a queued route is `start` then `await`; `stream` on an
-inline route is the response as a single `finish`, and on a queued route it is the status observations followed by
-`finish`. `start` / `resume` on a non-queued route fail with `UnsupportedOperation` (`transcription.start`). The
+`<Options, Frame, State>`, `<Options, Token>`) and composes through the shared `composeRoute` (`src/media-model.ts`),
+which picks `MediaRoute.inline` / `stream` / `queued`; and `TranscriptionClient`, like every modality client, is
+`MediaClient.make` (`src/media-client.ts`), which dispatches on `route.kind`. `generate` on a queued route is `start`
+then `await`; `stream` on an inline route is the response as a single `finish`, and on a queued route it is the status
+observations followed by `finish`. `start` / `resume` on a non-queued route fail with `UnsupportedOperation` (`transcription.start`). The
 `finish` event carries the whole transcript (text, segments, words, language, duration, usage), so the stream route's
 `collect` is just "take `finish`".

@@ -309,11 +314,12 @@ Settled rules:
  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 a hint; only AssemblyAI (`speakers_expected`) accepts it.
+  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.
 - **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, `english` from whisper-1, `en_us` from AssemblyAI. Deepgram and AssemblyAI assume English
-  unless asked to detect, so a missing `language` enables their detection.
+  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.
 - **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.
@@ -323,7 +329,7 @@ Settled rules:

 | Provider | Kind | Audio input | `timestamps` | `diarize` | Unsupported | Usage |
 |---|---|---|---|---|---|---|
-| OpenAI | stream (`stream: true` in `stream` mode) | multipart `file` (inline only) | `whisper-1` (`verbose_json`); diarize model: `segment` | `gpt-4o-transcribe-diarize` (`diarized_json`) | `speakers`; `prompt` on the diarize model; streaming on `whisper-1` | `tokens` or `seconds` |
+| 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`) |
 | AssemblyAI | queued (upload → submit → poll) | `/v2/upload` then `audio_url`, or a URL | words always; `segment` → `speaker_labels` | `speaker_labels` | — | `seconds` (`audio_duration`) |
@@ -353,7 +359,7 @@ GenerationAwaitOptions = { poll?: Poll }
 Poll = { interval?: Duration; timeout?: Duration }
 ```

-`Generation` is not video-specific. Image routes on BFL, fal, and Replicate are queued; `Image.start` exists for them. A route declares itself `inline` or `queued`; `generate` on a queued route is `start` then `await`.
+`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`.

 ### Usage

@@ -400,14 +406,15 @@ Streams become `AsyncIterable` via `Stream.toAsyncIterable`. `AIError` is thrown

 ### Providers

-Existing facades gain per-modality selectors; the modality routes each facade provides:
+Existing facades gain per-modality selectors; the modality routes each facade provides (*italics* are not
+implemented):

 | Facade | llm | image | video | speech | transcription | other |
 |---|---|---|---|---|---|---|
-| `OpenAI` | responses (default), chat | Images API (stream) | Sora (deprecated 2026-09-24) | ✓ | ✓ | |
+| `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 | soundEffect, music |
+| `ElevenLabs` | | | | ✓ | *Scribe (pending)* | *soundEffect, music (phase 5)* |
 | `Cartesia` | | | | ✓ | | |
 | `Deepgram` | | | | Aura | ✓ | |
 | `Fal` | | ✓ (queued) | ✓ | | | |
@@ -416,7 +423,7 @@ Existing facades gain per-modality selectors; the modality routes each facade pr
 | `Replicate` | | ✓ (queued) | | | | |
 | `Stability` | | `image` (inline), `upscale()` (queued) | | | | |
 | `Runway` | | | ✓ | | | |
-| `Luma`, `Kling`, `MiniMax` | | per provider | | | | |
+| `Luma`, `Kling`, `MiniMax` | | *deferred* | *deferred* | | | |

 New facades follow the existing one-file-per-provider rule. The facade selector is the public path for media models; modality-specific package entrypoints (for example `@opencode/ai/providers/openai/images`) are deferred until Core has a modality-aware model resolver.

@@ -53,6 +53,8 @@ export type Event = Observation | { readonly type: "generation-finished"; readon

 const TERMINAL: ReadonlySet<Status> = new Set(["completed", "failed", "cancelled", "expired"])

+export const isTerminal = (status: Status) => TERMINAL.has(status)
+
 export class Generation<Response> {
  readonly id: string
  readonly status: Status
@@ -81,7 +83,7 @@ export class Generation<Response> {
  }

  get terminal() {
-    return TERMINAL.has(this.status)
+    return isTerminal(this.status)
  }

  refresh(): Effect.Effect<Generation<Response>, AIError> {
@@ -109,9 +111,10 @@ export class Generation<Response> {
  }

  /**
-   * Status observations as a stream, ending after the first terminal observation. Each poll is bounded by the time
-   * remaining until `poll.timeout`, so a hung status request fails the stream instead of stalling it. (`Stream.interruptWhen`
-   * would express this directly but deadlocks under `TestClock` when the source completes while the timer sleeps.)
+   * Status observations as a stream, ending after the first terminal observation. Each poll and each sleep between polls
+   * is bounded by the time remaining until `poll.timeout`, so a hung status request or a long interval fails the stream at
+   * the deadline instead of stalling it. (`Stream.interruptWhen` would express this directly but deadlocks under
+   * `TestClock` when the source completes while the timer sleeps.)
   */
  events(options?: AwaitOptions): Stream.Stream<Event, AIError> {
    if (this.terminal) return Stream.make(this.event())
@@ -120,17 +123,26 @@ 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) =>
-              this.refresh().pipe(
-                Effect.timeoutOrElse({
-                  duration: Duration.millis(Math.max(0, deadline - now)),
-                  orElse: () => this.timeoutError(timeout),
-                }),
-              ),
+              now >= deadline
+                ? this.timeoutError(timeout)
+                : this.refresh().pipe(
+                    Effect.timeoutOrElse({
+                      duration: Duration.millis(deadline - now),
+                      orElse: () => this.timeoutError(timeout),
+                    }),
+                  ),
            ),
          )
-          return Stream.fromEffectSchedule(refresh, this.schedule(options?.poll)).pipe(
+          const schedule = this.schedule(options?.poll).pipe(
+            Schedule.modifyDelay((meta) =>
+              Effect.succeed(Duration.min(meta.duration, Duration.millis(Math.max(0, deadline - meta.now)))),
+            ),
+          )
+          return Stream.fromEffectSchedule(refresh, schedule).pipe(
            Stream.takeUntil((generation) => generation.terminal),
            Stream.map((generation) => generation.event()),
          )
@@ -110,8 +110,11 @@ const fromRequest = Effect.fn("AssemblyAITranscription.fromRequest")(function* (
        language_code: request.language,
        language_detection: request.language === undefined ? true : undefined,
        prompt: request.prompt,
-        // Turn-level `utterances`, the only segments AssemblyAI returns, require speaker labels.
-        speaker_labels: request.diarize === true || request.timestamps === "segment" ? true : undefined,
+        // Turn-level `utterances`, the only segments AssemblyAI returns, and `speakers_expected` require speaker labels.
+        speaker_labels:
+          request.diarize === true || request.timestamps === "segment" || request.speakers !== undefined
+            ? true
+            : undefined,
        speakers_expected: request.speakers,
      },
      request.providerOptions,
@@ -155,8 +158,7 @@ const decodeResult = Effect.fn("AssemblyAITranscription.decodeResult")(function*
  const error = transcript.error ?? undefined
  if (status === "failed")
    return yield* output.ended("failed", `${route.name} transcription failed${error === undefined ? "" : `: ${error}`}`)
-  if (status !== "completed")
-    return yield* output.invalid(`${route.name} transcript ${context.token.transcriptID} has not finished`)
+  if (status !== "completed") return yield* output.pending(context.token.transcriptID)
  const duration = transcript.audio_duration ?? undefined
  return new TranscriptionResponse({
    text: transcript.text ?? "",
@@ -31,13 +31,21 @@ export type Request = ImageRequestFor<BlackForestLabsImageOptions>
 // 2. Token and response schemas
 // ---------------------------------------------------------------------------

-/** Regional clusters answer on different hosts, so the returned `polling_url` is followed verbatim. */
-export const Token = Schema.Struct({ id: Schema.String, pollingURL: Schema.String })
+/**
+ * Regional clusters answer on different hosts, so the returned `polling_url` is followed verbatim. BFL reports the
+ * credit cost on submit, so it rides on the token; it is optional so tokens persisted before it existed still decode.
+ */
+export const Token = Schema.Struct({
+  id: Schema.String,
+  pollingURL: Schema.String,
+  cost: Schema.optionalKey(Schema.Number),
+})
 export type Token = Schema.Schema.Type<typeof Token>

 const StartResponse = Schema.Struct({
  id: Schema.String,
  polling_url: Schema.String,
+  cost: optionalNull(Schema.Number),
 })

 const Result = Schema.Struct({
@@ -145,7 +153,11 @@ const fromRequest = Effect.fn("BlackForestLabsImages.fromRequest")(function* (re
 // ---------------------------------------------------------------------------

 const decodeStart = route.decodeStarted(StartResponse, (value) => ({
-  token: { id: value.id, pollingURL: value.polling_url },
+  token: {
+    id: value.id,
+    pollingURL: value.polling_url,
+    ...(value.cost === undefined || value.cost === null ? {} : { cost: value.cost }),
+  },
  snapshot: { id: value.id, status: "queued" },
 }))

@@ -169,14 +181,16 @@ const decodeResult = Effect.fn("BlackForestLabsImages.decodeResult")(function* (
  if (isModerated(document.status)) return yield* output.contentPolicy(`${route.name} moderated the generation`)
  if (status === "failed" || status === "expired")
    return yield* output.ended(status, `${route.name} generation ${context.token.id} ended with ${document.status}`)
-  if (status !== "completed" || document.result === undefined || document.result === null)
+  if (status !== "completed") return yield* output.pending(context.token.id)
+  if (document.result === undefined || document.result === null)
    return yield* output.invalid(`${route.name} generation ${context.token.id} has no result`)
  const { sample, seed, prompt, ...rest } = document.result
+  // A settled `cost` on the result supersedes the submit-time cost carried on the token.
+  const cost = document.cost ?? context.token.cost
  return new ImageResponse({
    // `sample` is a signed URL that expires 10 minutes after the result is ready, so it is downloaded now.
    images: [yield* context.materialize(Media.url(sample))],
-    usage:
-      document.cost === undefined || document.cost === null ? undefined : { type: "credits", credits: document.cost },
+    usage: cost === undefined ? undefined : { type: "credits", credits: cost },
    providerMetadata: {
      bfl: { id: context.token.id, seed: seed ?? undefined, prompt: prompt ?? undefined, ...rest },
    },
@@ -67,6 +67,9 @@ const queryParameters = (request: Request) => {
 }

 const fromRequest = Effect.fn("DeepgramSpeech.fromRequest")(function* (request: Request) {
+  // Not in `unsupported`: that list would also reject `timestamps: false`, which asks for nothing.
+  if (request.timestamps === true)
+    return yield* route.unsupported("media.timestamps", `${route.name} does not return timestamps`)
  if (
    request.format !== undefined &&
    FORMATS[request.format] === undefined &&
@@ -117,7 +120,7 @@ const finish = (state: State, context: MediaProtocol.ResponseContext<Request>) =
 // ---------------------------------------------------------------------------

 export const protocol = MediaProtocol.stream<Request, SpeechEvent, Uint8Array, State>(route, {
-  unsupported: ["voice", "language", "instructions", "timestamps"],
+  unsupported: ["voice", "language", "instructions"],
  body: { from: fromRequest },
  frames: (bytes) => bytes,
  initial: () => ({ chunks: [] }),
@@ -49,7 +49,7 @@ const QueueResult = Schema.StructWithRest(
 // ---------------------------------------------------------------------------

 const sizing = (model: string) => {
-  if (/^fal-ai\/(nano-banana|flux-pro\/v1\.1-ultra)/.test(model)) return "aspect_ratio"
+  if (/^fal-ai\/(nano-banana|flux-pro\/(v1\.1-ultra|kontext))/.test(model)) return "aspect_ratio"
  if (model.startsWith("fal-ai/flux")) return "image_size"
  return undefined
 }
@@ -63,20 +63,24 @@ const validate = (request: Request) => {
    return Effect.fail(route.unsupported("media.size", `${id} sizes by aspectRatio`))
  if (request.aspectRatio !== undefined && field === "image_size")
    return Effect.fail(route.unsupported("media.aspectRatio", `${id} sizes by size (image_size)`))
-  if ((request.images?.length ?? 0) > 1 && !isEdit(id))
+  if ((request.images?.length ?? 0) > 1 && !takesImageList(id))
    return Effect.fail(
-      route.unsupported("media.images", `${id} takes one image_url; use an /edit endpoint for several images`),
+      route.unsupported(
+        "media.images",
+        `${id} takes one image_url; use an /edit or /multi endpoint for several images`,
+      ),
    )
  return Effect.void
 }

-// `/edit` endpoints take an `image_urls` list; image-to-image, fill, and Ultra take one `image_url` (beside `mask_url`).
-const isEdit = (model: string) => model.endsWith("/edit")
+// `/edit` and `/multi` (Kontext) endpoints take an `image_urls` list; image-to-image, fill, and Ultra take one
+// `image_url` (beside `mask_url`).
+const takesImageList = (model: string) => model.endsWith("/edit") || model.endsWith("/multi")

 const fromRequest = Effect.fn("FalImages.fromRequest")(function* (request: Request) {
  yield* validate(request)
  const images = yield* Effect.forEach(request.images ?? [], (image) => FalQueue.mediaUrl(image, route.name))
-  const edit = isEdit(request.model.id)
+  const list = takesImageList(request.model.id)
  return MediaProtocol.json(
    mergeJsonRecords(
      {
@@ -86,8 +90,8 @@ const fromRequest = Effect.fn("FalImages.fromRequest")(function* (request: Reque
        image_size: request.size === undefined ? undefined : MediaInput.dimensions(request.size),
        aspect_ratio: request.aspectRatio,
        output_format: request.format,
-        image_urls: edit && images.length > 0 ? images : undefined,
-        image_url: edit ? undefined : images[0],
+        image_urls: list && images.length > 0 ? images : undefined,
+        image_url: list ? undefined : images[0],
        mask_url: request.mask === undefined ? undefined : yield* FalQueue.mediaUrl(request.mask, route.name),
      },
      request.providerOptions,
@@ -112,12 +116,14 @@ const decodeResult = Effect.fn("FalImages.decodeResult")(function* (
  // With the safety checker on, flagged images come back blacked out rather than omitted.
  const flagged = (has_nsfw_concepts ?? []).flatMap((value, index) => (value ? [index] : []))
  return new ImageResponse({
-    images: images.map((image) =>
-      Media.url(image.url, {
-        mediaType: image.content_type ?? undefined,
-        info: { width: image.width ?? undefined, height: image.height ?? undefined },
-      }),
-    ),
+    images: images.map((image) => {
+      const info = { width: image.width ?? undefined, height: image.height ?? undefined }
+      // `sync_mode: true` returns data URIs instead of hosted URLs.
+      return (
+        Media.parseDataUrl(image.url, { info }) ??
+        Media.url(image.url, { mediaType: image.content_type ?? undefined, info })
+      )
+    }),
    notices:
      flagged.length === 0
        ? undefined
@@ -101,7 +101,7 @@ const generationConfig = (request: Request) => {
 const fromRequest = Effect.fn("GoogleImages.fromRequest")(function* (request: Request) {
  if (request.n !== undefined && request.n > 1)
    return yield* route.unsupported(
-      "image.n",
+      "media.n",
      `${route.name} generates one image per request; call it once per image instead of n=${request.n}`,
    )
  const parts = yield* Effect.forEach(request.images ?? [], (image) =>
@@ -56,6 +56,9 @@ interface State extends SpeechStream.Audio, GeminiGenerateContent.Metadata {
 // ---------------------------------------------------------------------------

 const fromRequest = Effect.fn("GoogleSpeech.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  // Not in `unsupported`: that list would also reject `timestamps: false`, which asks for nothing.
+  if (request.timestamps === true)
+    return yield* route.unsupported("media.timestamps", `${route.name} does not return timestamps`)
  if (request.format === "pcm" && request.mode === "generate" && /^gemini-3\.8-.*-tts(?:-|$)/.test(request.model.id))
    return yield* route.unsupported(
      "media.format",
@@ -125,7 +128,7 @@ const finish = (state: State, context: MediaProtocol.ResponseContext<Request>) =
 // ---------------------------------------------------------------------------

 export const protocol = MediaProtocol.stream<Request, SpeechEvent, string, State>(route, {
-  unsupported: ["instructions", "speed", "timestamps"],
+  unsupported: ["instructions", "speed"],
  body: { from: fromRequest },
  frames: (bytes, context) => GeminiGenerateContent.frames(bytes, context.request.mode),
  initial: () => ({ chunks: [] }),
@@ -149,8 +149,7 @@ const decodeResult = Effect.fn("GoogleVideo.decodeResult")(function* (
  const output = yield* decodeOperation(response)
  const operation = output.value
  const status = statusOf(operation)
-  if (status === "running")
-    return yield* output.invalid(`${route.name} operation ${context.token.operation} has not finished`)
+  if (status === "running") return yield* output.pending(context.token.operation)
  if (status === "failed")
    return yield* output.ended(
      "failed",
@@ -48,15 +48,17 @@ const Usage = Schema.Struct({
  output_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
 })

-const OpenAIImageResponse = Schema.Struct({
-  data: Schema.Array(
-    Schema.Struct({
-      b64_json: Schema.optional(Schema.String),
-      url: Schema.optional(Schema.String),
-      revised_prompt: Schema.optional(Schema.String),
-    }),
-  ),
+/** What the provider actually rendered; it can differ from the request when `auto` or a default applied. */
+const Settings = {
  output_format: Schema.optional(Schema.String),
+  size: Schema.optional(Schema.String),
+  quality: Schema.optional(Schema.String),
+  background: Schema.optional(Schema.String),
+}
+
+const OpenAIImageResponse = Schema.Struct({
+  data: Schema.Array(Schema.Struct({ b64_json: Schema.String })),
+  ...Settings,
  usage: Schema.optional(Usage),
 })

@@ -69,11 +71,13 @@ const StreamEvent = Schema.Union([
    type: Schema.Literals(["image_generation.partial_image", "image_edit.partial_image"]),
    b64_json: Schema.String,
    partial_image_index: Schema.Number,
+    ...Settings,
    output_format: Schema.String,
  }),
  Schema.Struct({
    type: Schema.Literals(["image_generation.completed", "image_edit.completed"]),
    b64_json: Schema.String,
+    ...Settings,
    output_format: Schema.String,
    usage: Schema.optional(Usage),
  }),
@@ -92,6 +96,9 @@ type Frame = string | { readonly document: string; readonly requested: string |
 interface State {
  readonly completed: number
  readonly format?: string
+  readonly size?: string
+  readonly quality?: string
+  readonly background?: string
  readonly usage?: MediaUsage
 }

@@ -110,10 +117,6 @@ const nativeOptions = (options: OpenAIImageOptions | undefined) => {

 const streamOptions = (request: MediaProtocol.Addressed<Request>) => {
  if (request.mode !== "stream") return Effect.succeed(undefined)
-  if (request.model.id.startsWith("dall-e"))
-    return Effect.fail(
-      route.unsupported("media.stream", `${request.model.id} does not stream; use Image.generate or a GPT image model`),
-    )
  if (request.n !== undefined && request.n > 1)
    return Effect.fail(
      route.unsupported("media.n", `${route.name} streams one image; use Image.generate for n=${request.n}`),
@@ -194,21 +197,34 @@ const usage = (value: Schema.Schema.Type<typeof Usage> | undefined): MediaUsage
        details: { openai: value },
      }

-const eventImage = (frame: string, label: string, data: string, format: string) =>
+/** `size` echoes the rendered `WIDTHxHEIGHT`; `auto` or any other value leaves the dimensions unknown. */
+const info = (format: string, size: string | undefined): Media.Info => {
+  const match = size?.match(/^(\d+)x(\d+)$/)
+  return match ? { format, width: Number(match[1]), height: Number(match[2]) } : { format }
+}
+
+const eventImage = (frame: string, label: string, data: string, format: string, size: string | undefined) =>
  MediaInput.decodedAsset((message, cause) => route.frameError(message, frame, cause), label, data, `image/${format}`, {
-    info: { format },
+    info: info(format, size),
  })

 const onEvent = Effect.fn("OpenAIImages.onEvent")(function* (state: State, frame: string) {
  const event = yield* decodeEvent(frame)
  const format = event.output_format
  if ("partial_image_index" in event) {
-    const image = yield* eventImage(frame, `${route.name} partial image`, event.b64_json, format)
+    const image = yield* eventImage(frame, `${route.name} partial image`, event.b64_json, format, event.size)
    return [state, [ImagePartialEvent.make({ index: event.partial_image_index, image })]] as const
  }
-  const image = yield* eventImage(frame, `${route.name} result ${state.completed}`, event.b64_json, format)
+  const image = yield* eventImage(frame, `${route.name} result ${state.completed}`, event.b64_json, format, event.size)
  return [
-    { ...state, completed: state.completed + 1, format, usage: usage(event.usage) },
+    {
+      completed: state.completed + 1,
+      format,
+      size: event.size,
+      quality: event.quality,
+      background: event.background,
+      usage: usage(event.usage),
+    },
    [ImageOutputEvent.make({ index: state.completed, image })],
  ] as const
 })
@@ -219,16 +235,20 @@ const onDocument = Effect.fn("OpenAIImages.onDocument")(function* (frame: Exclud
    Effect.mapError((cause) => invalid(`${route.name} returned an invalid response`, cause)),
  )
  const format = decoded.output_format ?? frame.requested ?? "png"
-  const mediaType = `image/${format}`
  const images = yield* Effect.forEach(decoded.data, (item, index) =>
-    MediaInput.imageOutput(invalid, `${route.name} result ${index}`, item, mediaType, {
-      info: { format },
-      providerMetadata:
-        item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
+    MediaInput.decodedAsset(invalid, `${route.name} result ${index}`, item.b64_json, `image/${format}`, {
+      info: info(format, decoded.size),
    }),
  )
  if (images.length === 0) return yield* invalid(`${route.name} returned no images`)
-  const state: State = { completed: images.length, format, usage: usage(decoded.usage) }
+  const state: State = {
+    completed: images.length,
+    format,
+    size: decoded.size,
+    quality: decoded.quality,
+    background: decoded.background,
+    usage: usage(decoded.usage),
+  }
  return [state, images.map((image, index) => ImageOutputEvent.make({ index, image }))] as const
 })

@@ -237,7 +257,17 @@ const step = (state: State, frame: Frame) => (typeof frame === "string" ? onEven
 const finish = (state: State) => {
  if (state.completed === 0) return Effect.fail(route.incomplete())
  return Effect.succeed([
-    ImageFinishEvent.make({ usage: state.usage, providerMetadata: { openai: { outputFormat: state.format } } }),
+    ImageFinishEvent.make({
+      usage: state.usage,
+      providerMetadata: {
+        openai: {
+          outputFormat: state.format,
+          size: state.size,
+          quality: state.quality,
+          background: state.background,
+        },
+      },
+    }),
  ])
 }

@@ -143,12 +143,14 @@ const adapter = {
  restoreHostedToolItem: (item: unknown) => (Schema.is(OpenAIResponsesHostedToolItem)(item) ? item : undefined),
 } satisfies OpenResponses.ProviderAdapter

-// Only GPT-6 Astra accepts `configuration_update`, and never alongside automatic `context_management` compaction.
+// GPT-6 Astra, Sol, and Luna accept `configuration_update` only in standard mode (not `reasoning.mode: "pro"` or
+// `-pro` slugs), and never alongside automatic `context_management` compaction.
 const supportsEffortUpdates = (request: LLMRequest) => {
  if (request.providerOptions?.contextManagement !== undefined) return false
+  if (Schema.is(Schema.Struct({ mode: Schema.Literal("pro") }))(request.http?.body?.reasoning)) return false
  const override = request.model.compatibility?.supportsEffortUpdates
  if (override !== undefined) return override
-  return /(?:^|\/)gpt-6-astra$/i.test(request.model.id)
+  return /(?:^|\/)gpt-6-(?:astra|sol|luna)$/i.test(request.model.id)
 }

 const nativeImageToolInput = (tool: ToolDefinition) => {
@@ -61,6 +61,9 @@ interface State extends SpeechStream.Audio {
 const supportsSse = (model: string) => !/^tts-1(-hd)?(-|$)/.test(model)

 const fromRequest = Effect.fn("OpenAISpeech.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  // Not in `unsupported`: that list would also reject `timestamps: false`, which asks for nothing.
+  if (request.timestamps === true)
+    return yield* route.unsupported("media.timestamps", `${route.name} does not return timestamps`)
  return MediaProtocol.json(
    mergeJsonRecords(
      {
@@ -121,7 +124,7 @@ const finish = (state: State, context: MediaProtocol.ResponseContext<Request>) =
 // ---------------------------------------------------------------------------

 export const protocol = MediaProtocol.stream<Request, SpeechEvent, string | Uint8Array, State>(route, {
-  unsupported: ["language", "timestamps"],
+  unsupported: ["language"],
  body: { from: fromRequest },
  frames: (bytes, context) => (isSse(context.body) ? Framing.sse.frame(bytes) : bytes),
  initial: () => ({ chunks: [], done: false }),
@@ -1,8 +1,9 @@
 import { Effect, Schema, Stream } from "effect"
+import { classifyProviderFailure } from "../provider-error.js"
 import { Framing } from "../route/framing.js"
 import { MediaProtocol } from "../route/media-protocol.js"
 import { MediaRoute } from "../route/media.js"
-import { mergeJsonRecords, type MediaUsage } from "../schema/index.js"
+import { AIError, mergeJsonRecords, type MediaUsage } from "../schema/index.js"
 import {
  TranscriptionFinishEvent,
  TranscriptionModel,
@@ -59,6 +60,9 @@ const Usage = Schema.Union([
    input_tokens: Schema.optional(Schema.Number),
    output_tokens: Schema.optional(Schema.Number),
    total_tokens: Schema.optional(Schema.Number),
+    input_token_details: Schema.optional(
+      Schema.Struct({ audio_tokens: Schema.optional(Schema.Number), text_tokens: Schema.optional(Schema.Number) }),
+    ),
  }),
  Schema.Struct({ type: Schema.Literal("duration"), seconds: Schema.Number }),
 ])
@@ -75,14 +79,23 @@ const transcriptFields = {
  usage: Schema.optional(Usage),
 }

+/** OpenAI may add stream event types; frames outside `EVENT_TYPES` are ignored. */
+const EventType = Schema.Struct({ type: Schema.String })
 const Event = Schema.Union([
  Schema.Struct({ type: Schema.Literal("transcript.text.delta"), delta: Schema.String }),
  Schema.Struct({ type: Schema.Literal("transcript.text.segment"), ...Segment.fields }),
  Schema.Struct({ type: Schema.Literal("transcript.text.done"), ...transcriptFields }),
+  Schema.Struct({
+    type: Schema.Literal("error"),
+    message: Schema.optional(Schema.String),
+    error: Schema.optional(Schema.Struct({ message: Schema.optional(Schema.String) })),
+  }),
 ])
+const EVENT_TYPES = new Set(["transcript.text.delta", "transcript.text.segment", "transcript.text.done", "error"])
 const Transcript = Schema.Struct(transcriptFields)
 type Transcript = Schema.Schema.Type<typeof Transcript>

+const decodeEventType = route.decodeFrame(EventType)
 const decodeEvent = route.decodeFrame(Event)
 const decodeTranscript = route.decodeFrame(Transcript)

@@ -118,10 +131,12 @@ const capabilities = (model: string): Capabilities => {
  return TRANSCRIBE
 }

+/** whisper-1 ignores `stream`, so its `stream` mode sends a plain request and emits only `finish`. */
+const streamsEvents = (request: MediaProtocol.Addressed<Request>) =>
+  request.mode === "stream" && capabilities(request.model.id).stream
+
 const validate = (request: MediaProtocol.Addressed<Request>, model: Capabilities) => {
  const id = request.model.id
-  if (request.mode === "stream" && !model.stream)
-    return Effect.fail(route.unsupported("media.stream", `${id} does not stream; use Transcription.generate`))
  if (request.diarize === true && !model.diarize)
    return Effect.fail(route.unsupported("media.diarize", `${id} does not diarize; use gpt-4o-transcribe-diarize`))
  if (request.prompt !== undefined && model.diarize)
@@ -173,7 +188,7 @@ const fromRequest = Effect.fn("OpenAITranscription.fromRequest")(function* (requ
      timestamp_granularities: responseFormat === "verbose_json" ? [request.timestamps] : undefined,
      // Diarizing audio longer than 30 seconds requires a chunking strategy.
      chunking_strategy: model.diarize ? "auto" : undefined,
-      stream: request.mode === "stream" ? true : undefined,
+      stream: streamsEvents(request) ? true : undefined,
    },
    {
      overlay: mergeJsonRecords(request.providerOptions, request.http?.body),
@@ -196,7 +211,15 @@ const segment = (value: Schema.Schema.Type<typeof Segment>): TranscriptionSegmen
 })

 const onEvent = Effect.fn("OpenAITranscription.onEvent")(function* (state: State, frame: string) {
+  if (!EVENT_TYPES.has((yield* decodeEventType(frame)).type)) return [state, []] as const
  const event = yield* decodeEvent(frame)
+  if (event.type === "error")
+    return yield* new AIError({
+      reason: classifyProviderFailure({
+        message: `${route.name} stream failed: ${event.message ?? event.error?.message ?? "unknown error"}`,
+        rawBody: frame,
+      }),
+    })
  if (event.type === "transcript.text.done") return [{ ...state, transcript: event }, []] as const
  if (event.type === "transcript.text.delta")
    return [state, event.delta.length === 0 ? [] : [TranscriptionTextDeltaEvent.make({ delta: event.delta })]] as const
@@ -246,7 +269,7 @@ export const protocol = MediaProtocol.stream<Request, TranscriptionEvent, Frame,
  unsupported: ["speakers"],
  body: { from: fromRequest },
  frames: (bytes, context) =>
-    context.request.mode === "stream"
+    streamsEvents(context.request)
      ? Framing.sse.frame(bytes)
      : Framing.document.frame(bytes).pipe(Stream.map((document) => ({ document }))),
  initial: () => ({ segments: [] }),
@@ -132,8 +132,7 @@ const decodeResult = Effect.fn("ReplicateImages.decodeResult")(function* (
      status,
      `${route.name} prediction ${context.token.id} ${prediction.status}${typeof prediction.error === "string" ? `: ${prediction.error}` : ""}`,
    )
-  if (status !== "completed")
-    return yield* output.invalid(`${route.name} prediction ${context.token.id} has not finished`)
+  if (status !== "completed") return yield* output.pending(context.token.id)
  if (prediction.data_removed === true)
    return yield* output.ended("expired", `${route.name} removed the output of prediction ${context.token.id}`)
  if (!isOutput(prediction.output))
@@ -141,8 +141,7 @@ const decodeResult = Effect.fn("RunwayVideo.decodeResult")(function* (
  }
  if (status === "cancelled")
    return yield* output.ended("cancelled", `${route.name} task ${context.token.taskID} was cancelled`)
-  if (status !== "completed")
-    return yield* output.invalid(`${route.name} task ${context.token.taskID} has not finished`)
+  if (status !== "completed") return yield* output.pending(context.token.taskID)
  const urls = task.output ?? []
  if (urls.length === 0) return yield* output.invalid(`${route.name} task succeeded without any output`)
  return new VideoResponse({
@@ -171,7 +170,7 @@ export const protocol = MediaProtocol.queued<Request, VideoResponse, Token>(rout
  start: { body: { from: fromRequest }, decode: decodeStart },
  status: { path: taskPath, decode: decodeStatus },
  result: { path: taskPath, decode: decodeResult },
-  cancel: { method: "DELETE", path: taskPath },
+  cancel: { method: "DELETE", path: taskPath, activeOnly: true },
 })

 const startPath = (request: Request) => {
@@ -175,7 +175,7 @@ const decodeUpscaleResult = Effect.fn("StabilityImages.decodeUpscaleResult")(fun
 ) {
  if (response.status === 202) {
    const output = yield* upscaleRoute.text(response)
-    return yield* output.invalid(`${upscaleRoute.name} upscale ${context.token.id} has not finished`)
+    return yield* output.pending(context.token.id)
  }
  return yield* decodeUpscaleImage(response)
 })
@@ -101,7 +101,8 @@ const decodeResponse = Effect.fn("XAIImages.decodeResponse")(function* (
  )
  if (images.length === 0) return yield* output.invalid(`${route.name} returned no images`)
  const usage = ProviderShared.isRecord(decoded.usage) ? decoded.usage : undefined
-  // xAI reports image counts rather than tokens, seconds, or credits; the raw record stays in provider metadata.
+  // xAI reports a USD cost (`cost_in_usd_ticks`) rather than tokens, seconds, or credits; the raw record stays in
+  // provider metadata.
  return new ImageResponse({
    images,
    providerMetadata: usage === undefined ? undefined : { xai: { usage } },
@@ -136,8 +136,7 @@ const decodeResult = Effect.fn("XAIVideo.decodeResult")(function* (
  const output = yield* decodeVideoStatus(response)
  const decoded = output.value
  const status = yield* MediaProtocol.status(STATUS, decoded.status, output)
-  if (status === "running")
-    return yield* output.invalid(`${route.name} request ${context.token.requestID} has not finished`)
+  if (status === "running") return yield* output.pending(context.token.requestID)
  if (status === "failed") {
    const code = decoded.error?.code ?? undefined
    const message = decoded.error?.message ?? undefined
@@ -1,7 +1,6 @@
-import { Effect, Schema } from "effect"
+import { Duration, Effect, Schema } from "effect"
 import type { HttpClientResponse } from "effect/unstable/http"
 import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
-import { Media } from "../media.js"
 import { MediaProtocol } from "../route/media-protocol.js"
 import { MediaRoute } from "../route/media.js"
 import { mergeJsonRecords, type OpenString } from "../schema/index.js"
@@ -9,6 +8,7 @@ import { mergeJsonRecords, type OpenString } from "../schema/index.js"
 const route = MediaProtocol.identity({ id: "zai-images", name: "Z.ai Images", provider: "zai" })
 export const DEFAULT_BASE_URL = "https://api.z.ai/api/paas/v4"
 export const PATH = "/images/generations"
+const OUTPUT_RETENTION = Duration.days(30)

 // ---------------------------------------------------------------------------
 // 1. Public model input
@@ -76,7 +76,7 @@ const decodeResponse = Effect.fn("ZAIImages.decodeResponse")(function* (
  const filters = decoded.content_filter ?? []
  return new ImageResponse({
    // Z.ai returns only URLs and no content type; the media type resolves when the asset is materialized.
-    images: decoded.data.map((item) => Media.url(item.url)),
+    images: yield* Effect.forEach(decoded.data, (item) => MediaProtocol.expiringUrl(item.url, OUTPUT_RETENTION)),
    // Z.ai reports applied content filters alongside a successful result; surface them instead of dropping them.
    notices:
      filters.length === 0
@@ -137,6 +137,11 @@ export interface Queued<Request, Response, Token> {
  readonly cancel?: {
    readonly method: AuthInput["method"]
    readonly path: (token: Token) => string
+    /**
+     * Fetch a fresh status first and skip the call for terminal generations, for providers whose cancel endpoint
+     * destroys finished work (Runway's `DELETE /v1/tasks/{id}` deletes completed tasks and their outputs).
+     */
+    readonly activeOnly?: boolean
  }
 }

@@ -195,7 +200,8 @@ 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); `contentPolicy` is a moderated result.
+   * 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.
   */
  const text = Effect.fn("MediaProtocol.text")(function* (response: HttpClientResponse.HttpClientResponse) {
    const http = context(response)
@@ -224,6 +230,14 @@ export const identity = (input: { readonly id: string; readonly name: string; re
              ? new ProviderInternalError({ message, body, http })
              : new InvalidRequestError({ message, body, http }),
        }),
+      pending: (id: string) =>
+        new AIError({
+          reason: new InvalidRequestError({
+            message: `${input.name} generation ${id} has not finished; await it before reading the result`,
+            body,
+            http,
+          }),
+        }),
      contentPolicy: (message: string) => new AIError({ reason: new ContentPolicyError({ message, body, http }) }),
    }
  })
@@ -285,9 +299,8 @@ export const status = <Table extends Record<string, Status>>(
  raw: string,
  output: Output,
 ): Effect.Effect<Status, AIError> => {
-  const normalized: Status | undefined = table[raw]
-  if (normalized === undefined) return Effect.fail(output.invalid(`Unknown generation status "${raw}"`))
-  return Effect.succeed(normalized)
+  if (!Object.hasOwn(table, raw)) return Effect.fail(output.invalid(`Unknown generation status "${raw}"`))
+  return Effect.succeed(table[raw])
 }

 /** A `url` asset whose provider-declared retention window starts now. */
@@ -5,7 +5,7 @@ import { Endpoint } from "./endpoint.js"
 import { RequestExecutorService, type Interface } from "./executor-service.js"
 import { RequestExecutor } from "./executor.js"
 import { MediaProtocol } from "./media-protocol.js"
-import { Generation } from "../generation.js"
+import { Generation, isTerminal } from "../generation.js"
 import type { Media } from "../media.js"
 import {
  AIError,
@@ -164,14 +164,19 @@ export const queued = <Request extends MediaRequest, Response, Token>(
      transport
        .call("GET", operation.path(token), http, execute)
        .pipe(Effect.flatMap((sent) => operation.decode(sent.response, { token, auth: sent.auth, materialize })))
+    const status = poll(protocol.status)
    const cancel = protocol.cancel
+    const send =
+      cancel === undefined
+        ? undefined
+        : transport.call(cancel.method, cancel.path(token), http, execute).pipe(Effect.asVoid)
    return {
-      status: poll(protocol.status),
+      status,
      result: poll(protocol.result),
      cancel:
-        cancel === undefined
-          ? undefined
-          : transport.call(cancel.method, cancel.path(token), http, execute).pipe(Effect.asVoid),
+        send !== undefined && cancel?.activeOnly
+          ? status.pipe(Effect.flatMap((snapshot) => (isTerminal(snapshot.status) ? Effect.void : send)))
+          : send,
    }
  }

@@ -380,7 +385,11 @@ const encode = (body: MediaProtocol.Body | undefined, headers: Headers.Headers)
  }
 }

-/** Common fields are never silently dropped: a present field the protocol declared unsupported fails typed. */
+/**
+ * Common fields are never silently dropped: a present field the protocol declared unsupported fails typed. `false`
+ * counts as present because some booleans mean something when false (video `audio`); protocols reject opt-in
+ * booleans such as speech `timestamps` with `=== true` in `body.from` instead of listing them.
+ */
 const rejectUnsupported = <Request extends object>(
  route: string,
  provider: ProviderID,
@@ -351,10 +351,34 @@ describe("OpenAI Responses effort updates", () => {
    }),
  )

+  it.effect("strips markers when the body overlay selects pro reasoning mode", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: OpenAI.configure({ apiKey: "fixture", http: { body: { reasoning: { mode: "pro" } } } }).responses(
+            "gpt-6-sol",
+          ),
+          messages: conversation,
+          providerOptions: { reasoningEffort: "low" },
+        }),
+      )
+
+      expect(updates(prepared.body)).toEqual([])
+      expect(prepared.body.reasoning).toEqual({ effort: "low" })
+    }),
+  )
+
  for (const [id, supported] of [
    ["gpt-6-astra", true],
    ["openai/gpt-6-astra", true],
+    ["gpt-6-sol", true],
+    ["openai/gpt-6-sol", true],
+    ["gpt-6-luna", true],
+    ["openai/gpt-6-luna", true],
    ["gpt-6-astra-2026-09-01", false],
+    ["gpt-6-sol-pro", false],
+    ["gpt-6-luna-pro", false],
+    ["gpt-6-sol-fast", false],
    ["gpt-5.6-sol", false],
  ] as const) {
    it.effect(`${supported ? "lowers" : "strips"} markers for ${id}`, () =>
@@ -23,6 +23,7 @@ import { Provider as ProviderSubpath } from "@opencode/ai/provider"
 import {
  AssemblyAI,
  Baseten,
+  BlackForestLabs,
  Cartesia,
  CloudflareAIGateway,
  CloudflareWorkersAI,
@@ -32,14 +33,18 @@ import {
  Fal,
  Fireworks,
  Google,
+  Meta,
  OpenCodeZen,
  OpenAI,
  OpenAICompatible,
  OpenRouter,
+  Replicate,
  Runway,
+  Stability,
  TypeSafeAI,
  VercelAIGateway,
  XAI,
+  ZAI,
 } from "@opencode/ai/providers"
 import {
  OpenAIChat,
@@ -151,8 +156,34 @@ describe("public exports", () => {
    expect(XAI.provider.chat).toBe(XAI.chat)
    expect(XAI.configure({ apiKey: "fixture" }).responses("grok-4.3").route.id).toBe("openai-responses")
    expect(XAI.configure({ apiKey: "fixture" }).chat("grok-4.3").route.id).toBe("openai-compatible-chat")
+    expect(OpenAI.configure({ apiKey: "fixture" }).image("gpt-image-2").route.id).toBe("openai-images")
+    expect(OpenAI.provider.image).toBe(OpenAI.image)
+    expect(Google.configure({ apiKey: "fixture" }).image("imagen-4.0-generate-001").route.id).toBe("google-images")
+    expect(Google.provider.image).toBe(Google.image)
+    expect(XAI.configure({ apiKey: "fixture" }).image("grok-imagine-image").route.id).toBe("xai-images")
+    expect(XAI.provider.image).toBe(XAI.image)
+    expect(Fal.configure({ apiKey: "fixture" }).image("fal-ai/flux/dev").route.id).toBe("fal-images")
+    expect(Fal.provider.image).toBe(Fal.image)
+    expect(BlackForestLabs.configure({ apiKey: "fixture" }).image("flux-2-pro").route.id).toBe("bfl-images")
+    expect(BlackForestLabs.provider.image).toBe(BlackForestLabs.image)
+    expect(Replicate.configure({ apiKey: "fixture" }).image("black-forest-labs/flux-schnell").route.id).toBe(
+      "replicate-images",
+    )
+    expect(Replicate.provider.image).toBe(Replicate.image)
+    expect(Stability.configure({ apiKey: "fixture" }).image("sd3.5-large").route.id).toBe("stability-images")
+    expect(Stability.provider.image).toBe(Stability.image)
+    expect(Stability.configure({ apiKey: "fixture" }).upscale().route.id).toBe("stability-upscale")
+    expect(Stability.provider.upscale).toBe(Stability.upscale)
+    expect(Meta.configure({ apiKey: "fixture" }).image("muse-image").route.id).toBe("meta-images")
+    expect(Meta.provider.image).toBe(Meta.image)
+    expect(ZAI.configure({ apiKey: "fixture" }).image("glm-image").route.id).toBe("zai-images")
+    expect(ZAI.provider.image).toBe(ZAI.image)
    expect(XAI.configure({ apiKey: "fixture" }).video("grok-imagine-video-1.5").route.id).toBe("xai-video")
+    expect(XAI.provider.video).toBe(XAI.video)
+    expect(Google.configure({ apiKey: "fixture" }).video("veo-3.1-generate-preview").route.id).toBe("google-video")
+    expect(Google.provider.video).toBe(Google.video)
    expect(Fal.configure({ apiKey: "fixture" }).video("fal-ai/veo3.1").route.id).toBe("fal-video")
+    expect(Fal.provider.video).toBe(Fal.video)
    expect(Runway.configure({ apiKey: "fixture" }).video("gen4.5").route.id).toBe("runway-video")
    expect(Runway.provider.video).toBe(Runway.video)
    expect(OpenAI.configure({ apiKey: "fixture" }).speech("gpt-4o-mini-tts").route.id).toBe("openai-speech")
@@ -97,6 +97,26 @@ describe("Generation", () => {
    }),
  )

+  it.effect("fails an event stream at the deadline when the poll interval is longer than the timeout", () =>
+    Effect.gen(function* () {
+      const scripted = yield* scriptedRoute(["running"], "never")
+      const generation = new Generation(scripted.route, "t", { id: "gen_1", status: "queued" })
+
+      const fiber = yield* Effect.forkChild(
+        generation
+          .events({ poll: { interval: "30 seconds", timeout: "10 seconds" } })
+          .pipe(Stream.runCollect, Effect.flip),
+      )
+      yield* TestClock.adjust("9 seconds")
+      expect(fiber.pollUnsafe()).toBeUndefined()
+      yield* TestClock.adjust("1 second")
+      const error = yield* Fiber.join(fiber)
+
+      expect(error.reason._tag).toBe("Timeout")
+      expect(yield* Ref.get(scripted.polls)).toBe(1)
+    }),
+  )
+
  it.effect("surfaces the route failure body for failed generations", () =>
    Effect.gen(function* () {
      const scripted = yield* scriptedRoute(["running", "failed"], "unused")
@@ -78,8 +78,11 @@ describe("Image", () => {
        mediaType: "image/webp",
      })
      expect(yield* response.image.bytes()).toEqual(Uint8Array.from([1, 2, 3]))
-      expect(response.image.providerMetadata).toEqual({ openai: { revisedPrompt: "A precise robot" } })
+      expect(response.image.info).toEqual({ format: "webp", width: 2048, height: 2048 })
      expect(response.usage).toMatchObject({ type: "tokens", total: 12 })
+      expect(response.providerMetadata).toEqual({
+        openai: { outputFormat: "webp", size: "2048x2048", quality: "high", background: "opaque" },
+      })
    }).pipe(
      Effect.provide(
        ImageClient.layer.pipe(
@@ -107,8 +110,11 @@ describe("Image", () => {
                })
                return input.respond(
                  JSON.stringify({
-                    data: [{ b64_json: "AQID", revised_prompt: "A precise robot" }, { b64_json: "BAUG" }],
+                    data: [{ b64_json: "AQID" }, { b64_json: "BAUG" }],
                    output_format: "webp",
+                    size: "2048x2048",
+                    quality: "high",
+                    background: "opaque",
                    usage: { input_tokens: 4, output_tokens: 8, total_tokens: 12 },
                  }),
                  { headers: { "content-type": "application/json" } },
@@ -144,6 +150,7 @@ describe("Image", () => {
        ),
      )
      expect(response.image.source).toEqual({ type: "bytes", data: Uint8Array.from([1, 2, 3]), mediaType: "image/png" })
+      expect(response.image.info).toEqual({ format: "png" })
    }),
  )

@@ -725,6 +732,7 @@ describe("Image", () => {
      const errors = yield* Effect.all(
        [
          Image.start({ model: Google.configure({ apiKey: "test" }).image("gemini-3.1-flash-image"), prompt }),
+          Image.generate({ model: Google.configure({ apiKey: "test" }).image("gemini-3.1-flash-image"), prompt, n: 2 }),
          Image.start({
            model: BlackForestLabs.configure({ apiKey: "test" }).image("flux-2-pro"),
            prompt,
@@ -735,7 +743,6 @@ describe("Image", () => {
            prompt,
            size: "512x512",
          }),
-          Stream.runCollect(Image.stream({ model: openai.image("dall-e-3"), prompt })),
          Stream.runCollect(Image.stream({ model: openai.image("gpt-image-2"), prompt, n: 2 })),
          Image.start({ model: replicate, prompt, seed: 7 }),
          Image.start({
@@ -749,9 +756,9 @@ describe("Image", () => {
      expect(errors.map((error) => [error.reason._tag, "operation" in error.reason && error.reason.operation])).toEqual(
        [
          ["UnsupportedOperation", "image.start"],
+          ["UnsupportedOperation", "media.n"],
          ["UnsupportedOperation", "media.aspectRatio"],
          ["UnsupportedOperation", "media.size"],
-          ["UnsupportedOperation", "media.stream"],
          ["UnsupportedOperation", "media.n"],
          ["UnsupportedOperation", "media.seed"],
          ["InvalidRequest", false],
@@ -761,6 +768,80 @@ describe("Image", () => {
    }).pipe(Effect.provide(layer(() => Effect.die("an unsupported request reached the network")))),
  )

+  const falToken = {
+    requestID: "r1",
+    statusURL: "https://queue.fal.test/fal-ai/flux/requests/r1/status",
+    responseURL: "https://queue.fal.test/fal-ai/flux/requests/r1",
+    cancelURL: "https://queue.fal.test/fal-ai/flux/requests/r1/cancel",
+  }
+  const falSubmitted = {
+    request_id: falToken.requestID,
+    status_url: falToken.statusURL,
+    response_url: falToken.responseURL,
+    cancel_url: falToken.cancelURL,
+  }
+  const bodies: Array<unknown> = []
+  it.effect("sizes fal Kontext by aspect ratio and sends several images to /multi", () =>
+    Effect.gen(function* () {
+      const fal = Fal.configure({ apiKey: "test", baseURL: "https://queue.fal.test" })
+      const images = [Media.url("https://example.test/a.png"), Media.url("https://example.test/b.png")]
+      const rejected = yield* Image.start({
+        model: fal.image("fal-ai/flux-pro/kontext"),
+        prompt: "A lighthouse",
+        size: "512x512",
+      }).pipe(Effect.flip)
+      yield* Image.start({
+        model: fal.image("fal-ai/flux-pro/kontext"),
+        prompt: "A lighthouse",
+        images: images.slice(0, 1),
+        aspectRatio: "16:9",
+      })
+      yield* Image.start({ model: fal.image("fal-ai/flux-pro/kontext/max/multi"), prompt: "A lighthouse", images })
+
+      expect(rejected.reason).toMatchObject({ _tag: "UnsupportedOperation", operation: "media.size" })
+      expect(bodies).toEqual([
+        { prompt: "A lighthouse", aspect_ratio: "16:9", image_url: "https://example.test/a.png" },
+        { prompt: "A lighthouse", image_urls: ["https://example.test/a.png", "https://example.test/b.png"] },
+      ])
+    }).pipe(
+      Effect.provide(
+        layer((input) => {
+          bodies.push(JSON.parse(input.text))
+          return Effect.succeed(json(input, falSubmitted))
+        }),
+      ),
+    ),
+  )
+
+  it.effect("decodes fal sync_mode data URIs as inline images", () =>
+    Effect.gen(function* () {
+      const generation = yield* Image.resume(Fal.configure({ apiKey: "test" }).image("fal-ai/flux/schnell"), falToken)
+      const response = yield* generation.await()
+
+      expect(response.images.map((image) => image.source)).toEqual([
+        { type: "base64", data: "AQID", mediaType: "image/png" },
+        { type: "url", url: "https://v3.fal.media/out.jpg", mediaType: "image/jpeg" },
+      ])
+      expect(response.image.info).toEqual({ width: 512, height: 512 })
+      expect(yield* response.image.bytes()).toEqual(Uint8Array.from([1, 2, 3]))
+    }).pipe(
+      Effect.provide(
+        layer((input) =>
+          Effect.succeed(
+            input.request.url === falToken.statusURL
+              ? json(input, { status: "COMPLETED" })
+              : json(input, {
+                  images: [
+                    { url: "data:image/png;base64,AQID", width: 512, height: 512, content_type: "image/png" },
+                    { url: "https://v3.fal.media/out.jpg", width: 512, height: 512, content_type: "image/jpeg" },
+                  ],
+                }),
+          ),
+        ),
+      ),
+    ),
+  )
+
  const moderated = { id: "req_1", status: "Content Moderated" }
  const prediction = {
    id: "p_1",
@@ -768,6 +849,56 @@ describe("Image", () => {
    output: { text: "not an image" },
    urls: { get: "https://replicate.test/p_1", cancel: "https://replicate.test/p_1/cancel" },
  }
+  for (const pending of [
+    {
+      model: BlackForestLabs.configure({ apiKey: "test" }).image("flux-2-pro"),
+      token: { id: "req_1", pollingURL: "https://bfl.test/v1/get_result?id=req_1" },
+      status: 200,
+      body: { id: "req_1", status: "Pending" },
+      message: "Black Forest Labs generation req_1",
+    },
+    {
+      model: Replicate.configure({ apiKey: "test" }).image("owner/model"),
+      token: { id: "p_1", getURL: "https://replicate.test/p_1", cancelURL: "https://replicate.test/p_1/cancel" },
+      status: 200,
+      body: {
+        id: "p_1",
+        status: "processing",
+        urls: { get: "https://replicate.test/p_1", cancel: "https://replicate.test/p_1/cancel" },
+      },
+      message: "Replicate generation p_1",
+    },
+    {
+      model: Stability.configure({ apiKey: "test", baseURL: "https://stability.test" }).upscale(),
+      token: { id: "up_1" },
+      status: 202,
+      body: { id: "up_1", status: "in-progress" },
+      message: "Stability AI generation up_1",
+    },
+  ]) {
+    it.effect(`rejects reading a ${pending.model.provider} result before the generation finishes`, () =>
+      Effect.gen(function* () {
+        const generation = yield* Image.resume(pending.model, pending.token)
+        const error = yield* generation.result().pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidRequest")
+        expect(error.message).toBe(`${pending.message} has not finished; await it before reading the result`)
+        expect(error.reason.body).toBe(JSON.stringify(pending.body))
+        expect(error.reason.http?.status).toBe(pending.status)
+      }).pipe(
+        Effect.provide(
+          layer((input) =>
+            Effect.succeed(
+              input.respond(JSON.stringify(pending.body), {
+                status: pending.status,
+                headers: { "content-type": "application/json" },
+              }),
+            ),
+          ),
+        ),
+      ),
+    )
+  }
+
  it.effect("classifies terminal outcomes the recordings never saw", () =>
    Effect.gen(function* () {
      const bfl = yield* Image.resume(BlackForestLabs.configure({ apiKey: "test" }).image("flux-2-pro"), {
@@ -3,7 +3,7 @@ import { NodeFileSystem } from "@effect/platform-node"
 import { Effect, Ref, Schema } from "effect"
 import { FileSystem } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
-import { Media, Message } from "../src/index.js"
+import { AIError, Media, Message } from "../src/index.js"
 import { it } from "./lib/effect.js"
 import { dynamicResponse, scriptedResponses } from "./lib/http.js"

@@ -161,6 +161,56 @@ describe("Media", () => {
    }),
  )

+  it.effect("keeps transient url download headers out of toJSON and AssetSchema encoding", () =>
+    Effect.sync(() => {
+      const asset = Media.url("https://cdn.example.test/video.mp4", {
+        mediaType: "video/mp4",
+        expiresAt: 42,
+        headers: { "x-goog-api-key": "secret" },
+      })
+      expect(asset.headers).toEqual({ "x-goog-api-key": "secret" })
+      const source = { type: "url", url: "https://cdn.example.test/video.mp4", mediaType: "video/mp4", expiresAt: 42 }
+
+      expect(asset.toJSON()).not.toHaveProperty("headers")
+      expect(JSON.stringify(asset)).not.toContain("secret")
+      expect(asset.toJSON().source).toEqual(source)
+
+      const encoded = Schema.encodeSync(Media.AssetSchema)(asset)
+      expect(encoded).not.toHaveProperty("headers")
+      expect(encoded.source).toEqual(source)
+
+      const codec = Schema.fromJsonString(Media.AssetSchema)
+      const json = Schema.encodeSync(codec)(asset)
+      expect(json).not.toContain("secret")
+      const restored = Schema.decodeSync(codec)(json)
+      expect(restored).toBeInstanceOf(Media.Asset)
+      expect(restored.source).toEqual(source)
+      expect(restored.expiresAt).toBe(42)
+      expect(restored.headers).toBeUndefined()
+    }),
+  )
+
+  it.effect("fails url downloads with non-2xx status as a typed AIError keeping http and body", () =>
+    Effect.gen(function* () {
+      const body = JSON.stringify({ error: { message: "file expired" } })
+      const error = yield* Media.url("https://cdn.example.test/expired.png")
+        .bytes()
+        .pipe(
+          Effect.flip,
+          Effect.provide(
+            dynamicResponse((input) =>
+              Effect.succeed(input.respond(body, { status: 404, headers: { "content-type": "application/json" } })),
+            ),
+          ),
+        )
+      expect(error).toBeInstanceOf(AIError)
+      expect(error.message).toContain("file expired")
+      expect(error.reason.http?.status).toBe(404)
+      expect(error.reason.http?.url).toBe("https://cdn.example.test/expired.png")
+      expect(error.reason.body).toBe(body)
+    }),
+  )
+
  it.effect("reads files with sniffed media types and writes materialized assets", () =>
    Effect.gen(function* () {
      const fs = yield* FileSystem.FileSystem
@@ -33,6 +33,8 @@ describe("Black Forest Labs Images recorded", () => {

        expect(response.image.source.type).toBe("bytes")
        expect(dimensions(yield* response.image.bytes())).toEqual({ width: 512, height: 512 })
+        // BFL reports cost on submit only; the Ready result omits it.
+        expect(response.usage).toEqual({ type: "credits", credits: 1.4000000000000001 })
      }),
    { timeout: 15 * 60 * 1000 },
  )
@@ -29,7 +29,11 @@ describe("OpenAI Images recorded", () => {

      expect(response.images).toHaveLength(1)
      expect(response.image.mediaType).toBe("image/jpeg")
+      expect(response.image.info).toEqual({ format: "jpeg", width: 1024, height: 1024 })
      expect((yield* response.image.bytes()).length).toBeGreaterThan(0)
+      expect(response.providerMetadata).toEqual({
+        openai: { outputFormat: "jpeg", size: "1024x1024", quality: "low", background: "opaque" },
+      })
    }),
  )

@@ -76,8 +80,13 @@ describe("OpenAI Images recorded", () => {
      expect(events.map((event) => event.type)).toEqual(["image-partial", "image", "finish"])
      const image = events.find(ImageEvent.is.image)
      expect(image?.image.mediaType).toBe("image/jpeg")
+      expect(image?.image.info).toEqual({ format: "jpeg", width: 1024, height: 1024 })
      expect(dimensions(yield* image!.image.bytes())).toEqual({ width: 1024, height: 1024 })
-      expect(events.find(ImageEvent.is.finish)?.usage).toMatchObject({ type: "tokens" })
+      const finish = events.find(ImageEvent.is.finish)
+      expect(finish?.usage).toMatchObject({ type: "tokens" })
+      expect(finish?.providerMetadata).toEqual({
+        openai: { outputFormat: "jpeg", size: "1024x1024", quality: "low", background: "opaque" },
+      })
    }),
  )
 })
@@ -44,7 +44,12 @@ describe("OpenAI Transcription recorded", () => {
      expect(deltas.length).toBeGreaterThan(1)
      expect(deltas.join("")).toBe(finish.text)
      expect(finish.text).toMatch(TRANSCRIPT)
-      expect(finish.usage).toMatchObject({ type: "tokens", input: expect.any(Number), output: expect.any(Number) })
+      expect(finish.usage).toMatchObject({
+        type: "tokens",
+        input: expect.any(Number),
+        output: expect.any(Number),
+        details: { openai: { input_token_details: { audio_tokens: expect.any(Number) } } },
+      })
    }),
  )

@@ -32,7 +32,12 @@ describe("Z.ai Images", () => {

      expect(response.images).toHaveLength(1)
      expect(response.image.mediaType).toBe("application/octet-stream")
-      expect(response.image.source).toEqual({ type: "url", url: "https://cdn.z.ai/generated.png" })
+      // Z.ai documents that output URLs expire 30 days after generation; the test clock starts at 0.
+      expect(response.image.source).toEqual({
+        type: "url",
+        url: "https://cdn.z.ai/generated.png",
+        expiresAt: 30 * 24 * 60 * 60 * 1000,
+      })
      expect(response.notices).toEqual([
        {
          type: "moderated",
@@ -97,6 +97,42 @@ describe("Speech", () => {
    }).pipe(Effect.provide(layer(() => Effect.die("an unsupported request reached the network")))),
  )

+  it.effect("treats timestamps: false as not asking for timestamps on routes that cannot return them", () =>
+    Effect.gen(function* () {
+      const bytes = Uint8Array.from([1, 2, 3])
+      const gemini = JSON.stringify({
+        candidates: [
+          { content: { parts: [{ inlineData: { mimeType: "audio/L16;codec=pcm;rate=24000", data: "AQID" } }] } },
+        ],
+      })
+      const responses = yield* Effect.all([
+        Speech.generate({ model: openai, text: "Hi", timestamps: false }).pipe(
+          Effect.provide(respond(new Blob([bytes]).stream(), "audio/mpeg")),
+        ),
+        Speech.generate({ model: google, text: "Hi", timestamps: false }).pipe(
+          Effect.provide(respond(gemini, "application/json")),
+        ),
+        Speech.generate({ model: deepgram, text: "Hi", timestamps: false }).pipe(
+          Effect.provide(respond(new Blob([bytes]).stream(), "audio/mpeg")),
+        ),
+      ])
+      for (const response of responses) expect(yield* response.audio.bytes()).toEqual(bytes)
+
+      const errors = yield* Effect.all(
+        [openai, google, deepgram].map((model) =>
+          Speech.generate({ model, text: "Hi", timestamps: true }).pipe(Effect.flip),
+        ),
+      ).pipe(Effect.provide(layer(() => Effect.die("an unsupported request reached the network"))))
+      expect(errors.map((error) => [error.reason._tag, "operation" in error.reason && error.reason.operation])).toEqual(
+        [
+          ["UnsupportedOperation", "media.timestamps"],
+          ["UnsupportedOperation", "media.timestamps"],
+          ["UnsupportedOperation", "media.timestamps"],
+        ],
+      )
+    }),
+  )
+
  it.effect("classifies stream failures and keeps the provider payload and HTTP context", () =>
    Effect.gen(function* () {
      const badFrame = JSON.stringify({ type: "speech.audio.delta", audio: "not base64!" })
@@ -5,7 +5,7 @@ import { HttpClientRequest } from "effect/unstable/http"
 import { Media, Transcription, TranscriptionClient } from "../src/index.js"
 import { AssemblyAI, Deepgram, Google, OpenAI } from "../src/providers.js"
 import { it } from "./lib/effect.js"
-import { dynamicResponse } from "./lib/http.js"
+import { dynamicResponse, json, observe, type Call } from "./lib/http.js"

 const layer = (handler: Parameters<typeof dynamicResponse>[0]) =>
  TranscriptionClient.layer.pipe(Layer.provideMerge(dynamicResponse(handler)))
@@ -25,7 +25,6 @@ describe("Transcription", () => {
    Effect.gen(function* () {
      const errors = yield* Effect.all(
        [
-          Stream.runCollect(Transcription.stream({ model: openai.transcription("whisper-1"), audio })),
          Transcription.generate({ model: openai.transcription("gpt-4o-mini-transcribe"), audio, diarize: true }),
          Transcription.generate({ model: openai.transcription("gpt-4o-mini-transcribe"), audio, timestamps: "word" }),
          Transcription.generate({ model: openai.transcription("gpt-4o-transcribe-diarize"), audio, prompt: "Names" }),
@@ -53,7 +52,6 @@ describe("Transcription", () => {
      )
      expect(errors.map((error) => [error.reason._tag, "operation" in error.reason && error.reason.operation])).toEqual(
        [
-          ["UnsupportedOperation", "media.stream"],
          ["UnsupportedOperation", "media.diarize"],
          ["UnsupportedOperation", "media.timestamps"],
          ["UnsupportedOperation", "media.prompt"],
@@ -70,6 +68,67 @@ describe("Transcription", () => {
    }).pipe(Effect.provide(layer(() => Effect.die("an unsupported request reached the network")))),
  )

+  it.effect("ignores unknown OpenAI stream events and fails on an error event with the frame", () =>
+    Effect.gen(function* () {
+      const sse = (...frames: ReadonlyArray<string>) => frames.map((frame) => `data: ${frame}\n\n`).join("")
+      const failure = `{"type":"error","error":{"type":"server_error","code":"server_error","message":"The server had an error"}}`
+      const bodies = [
+        sse(
+          `{"type":"transcript.text.delta","delta":"Hi"}`,
+          `{"type":"transcript.text.future","payload":1}`,
+          `{"type":"transcript.text.done","text":"Hi"}`,
+          "[DONE]",
+        ),
+        sse(`{"type":"transcript.text.delta","delta":"Hi"}`, failure),
+      ]
+      const model = openai.transcription("gpt-4o-mini-transcribe")
+      const program = Effect.gen(function* () {
+        const events = Array.from(yield* Stream.runCollect(Transcription.stream({ model, audio })))
+        const error = yield* Stream.runCollect(Transcription.stream({ model, audio })).pipe(Effect.flip)
+        return { events, error }
+      })
+      const { events, error } = yield* program.pipe(
+        Effect.provide(
+          layer((input) =>
+            Effect.sync(() =>
+              input.respond(bodies.shift() ?? "", { headers: { "content-type": "text/event-stream" } }),
+            ),
+          ),
+        ),
+      )
+
+      expect(events.map((event) => event.type)).toEqual(["text-delta", "finish"])
+      expect(error.reason).toMatchObject({ _tag: "ProviderInternal", body: failure })
+      expect(error.message).toContain("The server had an error")
+    }),
+  )
+
+  it.effect("streams whisper-1 as a single finish from a plain request", () =>
+    Effect.gen(function* () {
+      const bodies: Array<string> = []
+      const events = Array.from(
+        yield* Stream.runCollect(Transcription.stream({ model: openai.transcription("whisper-1"), audio })).pipe(
+          Effect.provide(
+            layer((input) =>
+              Effect.sync(() => {
+                bodies.push(input.text)
+                return input.respond(
+                  JSON.stringify({ text: "Hello there.", usage: { type: "duration", seconds: 2 } }),
+                  { headers: { "content-type": "application/json" } },
+                )
+              }),
+            ),
+          ),
+        ),
+      )
+
+      expect(bodies[0]).not.toContain('name="stream"')
+      expect(events).toEqual([
+        expect.objectContaining({ type: "finish", text: "Hello there.", usage: { type: "seconds", seconds: 2 } }),
+      ])
+    }),
+  )
+
  it.effect(
    "uploads inline audio to AssemblyAI, resumes polling from a persisted token, and surfaces failed transcripts",
    () =>
@@ -171,4 +230,45 @@ describe("Transcription", () => {
        expect(failure.reason).toMatchObject({ _tag: "ProviderInternal", body: failed })
      }),
  )
+
+  it.effect("enables AssemblyAI speaker labels when only an expected speaker count is given", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      yield* Transcription.start({ model: assemblyai, audio: Media.url("https://a.test/call.mp3"), speakers: 2 }).pipe(
+        Effect.provide(
+          layer((input) => observe(calls, input).pipe(Effect.as(json(input, { id: "tr_1", status: "queued" })))),
+        ),
+      )
+      expect(calls.map((call) => JSON.parse(call.body))).toEqual([
+        {
+          audio_url: "https://a.test/call.mp3",
+          speech_models: ["universal-3-5-pro"],
+          language_detection: true,
+          speaker_labels: true,
+          speakers_expected: 2,
+        },
+      ])
+    }),
+  )
+
+  it.effect("rejects reading an AssemblyAI result before the transcript finishes", () =>
+    Effect.gen(function* () {
+      const generation = yield* Transcription.resume(assemblyai, { transcriptID: "tr_1" })
+      const error = yield* generation.result().pipe(Effect.flip)
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toBe("AssemblyAI generation tr_1 has not finished; await it before reading the result")
+      expect(error.reason.body).toBe(JSON.stringify({ id: "tr_1", status: "processing" }))
+      expect(error.reason.http?.status).toBe(200)
+    }).pipe(
+      Effect.provide(
+        layer((input) =>
+          Effect.succeed(
+            input.respond(JSON.stringify({ id: "tr_1", status: "processing" }), {
+              headers: { "content-type": "application/json" },
+            }),
+          ),
+        ),
+      ),
+    ),
+  )
 })
@@ -576,7 +576,7 @@ describe("Video / Runway", () => {
  const model = runway.video("gen4.5")
  const taskUrl = "https://runway.test/v1/tasks/task_1"

-  it.effect("submits image_to_video with the API version header, polls the task, and reports credits", () =>
+  it.effect("submits image_to_video, polls the task, reports credits, and keeps the finished task on cancel", () =>
    Effect.gen(function* () {
      const calls: Array<Call> = []
      const program = Effect.gen(function* () {
@@ -624,7 +624,7 @@ describe("Video / Runway", () => {
                return json(input, { id: "task_1", estimatedCost: { credits: 25 } })
              }
              expect(call.url).toBe(taskUrl)
-              if (call.method === "DELETE") return input.respond(null, { status: 204 })
+              if (call.method === "DELETE") return yield* Effect.die("cancel deleted a finished Runway task")
              if (nth === 1) return json(input, { id: "task_1", status: "PENDING", estimatedCost: { credits: 25 } })
              if (nth === 2) return json(input, { id: "task_1", status: "THROTTLED", estimatedCost: { credits: 25 } })
              if (nth === 3) return json(input, { id: "task_1", status: "RUNNING", progress: 0.5 })
@@ -653,6 +653,32 @@ describe("Video / Runway", () => {
        `GET ${taskUrl}`,
        `GET ${taskUrl}`,
        `GET ${taskUrl}`,
+        `GET ${taskUrl}`,
+      ])
+    }),
+  )
+
+  it.effect("cancels a task that is still running", () =>
+    Effect.gen(function* () {
+      const calls: Array<Call> = []
+      yield* Effect.gen(function* () {
+        const generation = yield* Video.start({ model, prompt: "x" })
+        yield* generation.cancel()
+      }).pipe(
+        Effect.provide(
+          layer((input) =>
+            Effect.gen(function* () {
+              const { call } = yield* observe(calls, input)
+              if (call.method === "POST") return json(input, { id: "task_1" })
+              if (call.method === "DELETE") return input.respond(null, { status: 204 })
+              return json(input, { id: "task_1", status: "RUNNING", progress: 0.2 })
+            }),
+          ),
+        ),
+      )
+      expect(calls.map((call) => `${call.method} ${call.url}`)).toEqual([
+        "POST https://runway.test/v1/text_to_video",
+        `GET ${taskUrl}`,
        `DELETE ${taskUrl}`,
      ])
    }),
@@ -812,3 +838,55 @@ describe("Video / Runway", () => {
    }),
  )
 })
+
+// ---------------------------------------------------------------------------
+// Shared queued behavior
+// ---------------------------------------------------------------------------
+
+describe("Video / queued result", () => {
+  for (const pending of [
+    {
+      model: Google.configure({ apiKey: "test", baseURL: "https://google.test/v1beta" }).video("veo-3.1"),
+      token: { operation: "models/veo-3.1/operations/op_1" },
+      body: { name: "models/veo-3.1/operations/op_1", done: false },
+      name: "Google Veo",
+    },
+    {
+      model: XAI.configure({ apiKey: "test", baseURL: "https://xai.test/v1" }).video("grok-imagine-video-1.5"),
+      token: { requestID: "req_1" },
+      body: { status: "pending", progress: 40 },
+      name: "xAI Video",
+    },
+    {
+      model: Runway.configure({ apiKey: "test", baseURL: "https://runway.test/v1" }).video("gen4.5"),
+      token: { taskID: "task_1" },
+      body: { status: "RUNNING", progress: 0.5 },
+      name: "Runway",
+    },
+  ]) {
+    it.effect(`rejects reading a ${pending.model.provider} result before the generation finishes`, () =>
+      Effect.gen(function* () {
+        const generation = yield* Video.resume(pending.model, pending.token)
+        const error = yield* generation.result().pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidRequest")
+        expect(error.message).toBe(
+          `${pending.name} generation ${generation.id} has not finished; await it before reading the result`,
+        )
+        expect(error.reason.body).toBe(JSON.stringify(pending.body))
+        expect(error.reason.http?.status).toBe(200)
+      }).pipe(Effect.provide(layer((input) => Effect.succeed(json(input, pending.body))))),
+    )
+  }
+
+  it.effect("rejects a status that only matches an inherited property", () =>
+    Effect.gen(function* () {
+      const error = yield* Video.resume(
+        XAI.configure({ apiKey: "test", baseURL: "https://xai.test/v1" }).video("grok-imagine-video-1.5"),
+        { requestID: "req_1" },
+      ).pipe(Effect.flip)
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      expect(error.message).toBe('Unknown generation status "constructor"')
+      expect(error.reason.body).toBe(JSON.stringify({ status: "constructor" }))
+    }).pipe(Effect.provide(layer((input) => Effect.succeed(json(input, { status: "constructor" }))))),
+  )
+})
@@ -55,7 +55,7 @@ ultimate source of truth. Upstream test262 files run verbatim from `test/test262
 - [x] Tagged templates: a tag applied to a template literal is called as `tag(strings, ...values)`, with the tag read
      like a callee so a member tag keeps its receiver. `strings` is an array of the cooked text with a read-only `raw`
      array of the source text; an invalid escape such as `\unicode` cooks to `undefined`. One template object per
-      site and both arrays are frozen, as in JS: `strings[0] = "x"` throws a `TypeError`.
+      site, as in JS, but it is not frozen: `strings[0] = "x"` succeeds here where JS throws.
 - [x] Regular-expression literals.
 - [x] `NaN` and `Infinity` globals.
 - [ ] BigInt literals and in-interpreter BigInt arithmetic; BigInt remains invalid at JSON-like host boundaries.
@@ -101,11 +101,9 @@ ultimate source of truth. Upstream test262 files run verbatim from `test/test262
 - [x] `for`, `while`, and `do...while`.
 - [x] `for...of` over arrays, strings, Maps, Sets, URLSearchParams, Headers, Uint8Arrays, built-in iterators, custom
      synchronous iterators, and confined synchronous generators. Abrupt completion invokes the iterator's optional `return()`.
- [x] `for...in` over the enumerable keys of plain objects, arrays, strings, and tool references, following the
-      prototype chain like JS (`for (k in Object.create({ a: 1 }))` visits `a`; built-in prototype methods are
-      non-enumerable so `for (k in [])` visits nothing). A key deleted before its turn is skipped and keys added during
-      the loop are not visited. `null`, `undefined`, and other non-objects iterate nothing. An un-awaited promise
-      throws rather than iterating.
+- [x] `for...in` over own keys of plain objects, arrays, strings, and tool references. `null`, `undefined`, and other
+      non-objects iterate nothing. An un-awaited promise throws rather than iterating.
+- [ ] `for...in` over inherited enumerable keys (`Object.create(proto)`), and skipping keys deleted during the loop.
 - [x] Unlabeled `break` and `continue`.
 - [x] `try`, `catch`, optional catch bindings, and `finally`.
 - [x] `throw` with arbitrary values.
@@ -172,11 +170,8 @@ Math.floor)` is `"3"`). A detached method loses its receiver, as in JS: `values.
 - [x] Redeclaring a function in the same scope, or alongside a `var`, is allowed: the last declaration wins.
 - [x] Generator functions have their own `prototype` (inheriting the shared generator prototype), so
      `g() instanceof g` holds. Plain functions have none, since they cannot construct.
- [x] Generator functions inherit from `GeneratorFunction.prototype` (async ones from
-      `AsyncGeneratorFunction.prototype`), an ordinary non-callable object under `Function.prototype` whose `prototype`
-      is the shared generator prototype and vice versa (`Object.getPrototypeOf(g).prototype.constructor`). Neither is
-      a global; async non-generator functions still inherit from `Function.prototype` directly. A generator whose
-      `prototype` was replaced by a non-object creates from the shared generator prototype.
+- [ ] `GeneratorFunction.prototype`: every function, generator or not, inherits directly from `Function.prototype`,
+      and a generator whose `prototype` was replaced by a non-object still creates from the shared generator prototype.
 - [x] Generator and async generator functions bind parameters (defaults, destructuring) at the call and defer only the
      body to the first `next()`, so a bad argument throws synchronously from the call site, as in JS.
 - [x] Synchronous and async generator declarations/expressions, `yield`, and `yield*`, including lazy bodies,
@@ -252,13 +247,9 @@ Math.floor)` is `"3"`). A detached method loses its receiver, as in JS: `values.
      hint (`"abc".indexOf({ toString() { return "b" } })` is `1`, `(255).toString({ valueOf() { return 16 } })` is
      `"ff"`, `String.prototype.trim.call({ toString() { return " a " } })` is `"a"`). Only consumed positions
      convert; a RegExp pattern is used as is, and `includes`/`startsWith`/`endsWith` reject one before converting.
- [x] `Error.prototype.toString` converts an object `name` or `message` through its own `toString` (`String(e)` with
-      `e.message = { toString() { return "m" } }` is `"Error: m"`); an uncaught error's report at the result boundary
-      still uses the built-in form.
- [x] `Array.prototype.toString` calls `this.join`, so `arr.join = () => "j"` makes `arr + ""`, `String(arr)`, and
-      `${arr}` all `"j"`; a non-callable `join` gives `"[object Array]"`.
- [ ] ToPrimitive elsewhere: numeric arguments of the Array and Uint8Array methods (`at`, `indexOf` start, `slice`)
-      still use the built-in form (`NaN`) and ignore own methods.
+- [ ] ToPrimitive elsewhere: `Error.prototype.toString` on an object `message` and numeric arguments of the Array and
+      Uint8Array methods (`at`, `indexOf` start, `slice`) still use the built-in form (`NaN`, `"[object Object]"`) and
+      ignore own methods.
 - [x] Property keys follow ToPropertyKey: `x[null]` and `x[true]` become string keys, and a data object key
      converts through its own `toString`/`valueOf` (string hint) exactly once per access, in reads, writes,
      compound assignment, `++`, `delete`, `in`, object literals, and destructuring:
@@ -268,8 +259,6 @@ Math.floor)` is `"3"`). A detached method loses its receiver, as in JS: `values.
 ## Promises and tools

 - [x] Tool calls start eagerly and return supervised, run-once CodeMode promises.
- [x] Tool references have identity: `tools.x === tools.x` and `tools.ns === tools.ns`, so they work as `Map`/`Set`
-      members and `switch` cases like any other object.
 - [x] Direct `await`, repeated awaits, and recursive thenable assimilation when a promise or thenable is returned from
      a function/program.
 - [x] `Promise.resolve` and `Promise.reject`.
@@ -346,7 +335,7 @@ reject }` object.
      `TypeError`.
 - [x] `Object.create(proto)` with an object or `null` prototype; any other prototype is a `TypeError`
      (`Object prototype may only be an Object or null`). Inherited reads, `in`, `hasOwnProperty`, and own-only
-      `Object.keys` follow the chain as in JS, and `for...in` enumerates inherited keys too. A second `properties`
+      `Object.keys` follow the chain as in JS, but `for...in` still enumerates own keys only. A second `properties`
      argument other than `undefined` throws a `TypeError`: property descriptors are not supported (there is no
      `Object.defineProperty` either).
 - [x] `Object.freeze`, `Object.seal`, and `Object.preventExtensions`, with `isFrozen`, `isSealed`, and `isExtensible`.
@@ -20,7 +20,7 @@ import {
  type Value,
 } from "./objects.js"
 import type { Interpreter } from "./interpreter.js"
-import { toPrimitiveString, withPrimitives } from "./callback.js"
+import { toPrimitiveString } from "./callback.js"
 import { formatValue } from "../stdlib/console.js"

 export const normalizeError = (error: unknown): Diagnostic => {
@@ -48,9 +48,7 @@ export const normalizeError = (error: unknown): Diagnostic => {
  if (error instanceof Throw) {
    const value = error.value
    if (value instanceof ErrorObj) {
-      return value.host
-        ? normalizeError(value.host)
-        : { kind: "ExecutionFailure", message: errorToString(get(value, "name"), get(value, "message")) }
+      return value.host ? normalizeError(value.host) : { kind: "ExecutionFailure", message: errorToString(value) }
    }
    let message: string
    if (containsRuntimeReference(value)) {
@@ -115,7 +113,9 @@ export const materialize = <R>(ctx: Interpreter<R>, thrown: unknown): Value => {
 }

 /** Error.prototype.toString: `name: message`, omitting whichever side is empty. */
-const errorToString = (name: Value, message: Value): string => {
+const errorToString = (self: Obj): string => {
+  const name = get(self, "name")
+  const message = get(self, "message")
  const shownName = name === undefined ? "Error" : coerceToString(name)
  const shownMessage = message === undefined ? "" : coerceToString(message)
  if (shownMessage === "") return shownName
@@ -179,16 +179,7 @@ export const errorGlobal = <R>(type: ErrorType, ctx: Interpreter<R>) => {
  })
  if (type === "Error") {
    methods(builtins, prototype, [
-      [
-        "toString",
-        0,
-        (thisValue) => {
-          const self = receiver(Obj, thisValue, "Error.prototype.toString")
-          return withPrimitives(ctx, "string", [get(self, "name"), get(self, "message")], ([name, message]) =>
-            errorToString(name, message),
-          )
-        },
-      ],
+      ["toString", 0, (thisValue) => errorToString(receiver(Obj, thisValue, "Error.prototype.toString"))],
    ])
    methods(builtins, ctor, [["isError", 1, (_, args) => args[0] instanceof ErrorObj]])
    return ctor
@@ -79,8 +79,7 @@ import {
  has,
  hidden,
  hasPrototype,
-  ownKeys,
-  enumerable,
+  keys,
  Native,
  parseArrayIndex,
  Arguments,
@@ -101,7 +100,7 @@ import { Pending, resolvePromise, resolvePromiseValue } from "./promises.js"
 import { describeValue, isOpaque, rejectCircularInsertion, typeofValue } from "./references.js"
 import { ScopeStack } from "./scope.js"
 import { constructRegExp } from "../stdlib/regexp.js"
-import { enumerableSource, restrict } from "../stdlib/object.js"
+import { enumerableSource } from "../stdlib/object.js"
 import { compoundOperators } from "../stdlib/value.js"

 /** The binary operators that convert object operands through ToPrimitive before acting on primitives. */
@@ -507,7 +506,7 @@ class Frame<R> {
  ): Fn {
    const builtins = this.ctx.builtins
    const fn = new Fn(
-      node.generator ? (node.async ? builtins.AsyncGeneratorFunction : builtins.GeneratorFunction) : builtins.Function,
+      builtins.Function,
      name,
      node.params,
      node.body,
@@ -956,24 +955,10 @@ class Frame<R> {
  }

  // for...in over null/undefined iterates nothing, like JS.
-  // EnumerateObjectProperties: own keys, then each prototype's, visiting a shadowed key once.
  private enumerableKeys(value: Value, node: AstNode): Array<string> {
    if (value instanceof ToolReference) return [...this.ctx.tools.keys(value.path)]
    if (value === null || value === undefined) return []
-    const seen = new Set<string>()
-    const result: Array<string> = []
-    for (
-      let current: Obj | null = enumerableSource(this.ctx, "for...in", value, node);
-      current !== null;
-      current = current.proto
-    ) {
-      for (const key of ownKeys(current)) {
-        if (typeof key !== "string" || seen.has(key)) continue
-        seen.add(key)
-        if (enumerable(current, key)) result.push(key)
-      }
-    }
-    return result
+    return keys(enumerableSource(this.ctx, "for...in", value, node))
  }

  private evaluateForInStatement(
@@ -997,8 +982,6 @@ class Frame<R> {
      const assignment = left.type === "VariableDeclaration" ? undefined : left

      for (const key of keys) {
-        // A key deleted before its turn is skipped, as in JS.
-        if (right instanceof Obj && !has(right, key)) continue
        const result = yield* Effect.gen(function* () {
          if (declared?.lexical) {
            self.scopes.push()
@@ -1727,15 +1710,12 @@ class Frame<R> {
    })
  }

-  // Built-ins throw without a location, synchronously or inside their Effect; the call site supplies it. A built-in
-  // reached through `ctx.call` runs on the root frame, so the deeper of the frame and the enclosing site counts.
+  // Built-ins throw without a location, synchronously or inside their Effect; the call site supplies it.
  private native(body: () => Effect.Effect<Value, unknown, R>, node?: AstNode): Effect.Effect<Value, unknown, R> {
-    return Effect.flatMap(CallSite, (site) =>
-      Effect.provideService(
-        Effect.catchDefect(Effect.suspend(body), (defect) => Effect.die(locate(defect, node))),
-        CallSite,
-        { node, depth: Math.max(this.depth, site.depth) },
-      ),
+    return Effect.provideService(
+      Effect.catchDefect(Effect.suspend(body), (defect) => Effect.die(locate(defect, node))),
+      CallSite,
+      { node, depth: this.depth },
    )
  }

@@ -2175,16 +2155,12 @@ class Frame<R> {
    define(
      strings,
      "raw",
-      restrict(
-        "freeze",
-        new Arr(
-          array,
-          node.quasi.quasis.map((quasi) => quasi.value.raw),
-        ),
+      new Arr(
+        array,
+        node.quasi.quasis.map((quasi) => quasi.value.raw),
      ),
      frozen,
    )
-    restrict("freeze", strings)
    this.ctx.templates.set(node, strings)
    return strings
  }
@@ -2232,7 +2208,7 @@ class Frame<R> {
      if (typeof key !== "string") {
        throw typeError("Tool paths must use string property names.", propertyNode)
      }
-      return objectValue.child(key)
+      return new ToolReference([...objectValue.path, key])
    }

    if (objectValue instanceof Obj) return { target: objectValue, key, receiver: objectValue }
@@ -1,5 +1,5 @@
 import { Effect } from "effect"
-import { Arr, define, ErrorObj, hidden, Native, Obj, readOnly, type Value } from "./objects.js"
+import { define, hidden, Native, Arr, ErrorObj, Obj, type Value } from "./objects.js"

 export const errorTypes = [
  "Error",
@@ -41,8 +41,6 @@ const builtins = [
  "AsyncIterator",
  "Generator",
  "AsyncGenerator",
-  "GeneratorFunction",
-  "AsyncGeneratorFunction",
 ] as const

 /**
@@ -81,15 +79,6 @@ export const createBuiltins = (): Builtins => {
  }
  const iterator = plain()
  const asyncIterator = plain()
-  // %GeneratorFunction.prototype% is an ordinary object linked both ways with %GeneratorPrototype%.
-  const generatorFunction = (generator: Obj) => {
-    const proto = new Obj(fn)
-    define(proto, "prototype", generator, readOnly)
-    define(generator, "constructor", proto, readOnly)
-    return proto
-  }
-  const generator = new Obj(iterator)
-  const asyncGenerator = new Obj(asyncIterator)
  return {
    Object: object,
    Function: fn,
@@ -113,10 +102,8 @@ export const createBuiltins = (): Builtins => {
    Iterator: iterator,
    IteratorHelper: new Obj(iterator),
    AsyncIterator: asyncIterator,
-    Generator: generator,
-    AsyncGenerator: asyncGenerator,
-    GeneratorFunction: generatorFunction(generator),
-    AsyncGeneratorFunction: generatorFunction(asyncGenerator),
+    Generator: new Obj(iterator),
+    AsyncGenerator: new Obj(asyncIterator),
    Error: error,
    TypeError: derived("TypeError"),
    RangeError: derived("RangeError"),
@@ -36,7 +36,6 @@ export const hidden: Attributes = { writable: true, enumerable: false, configura
 export const readonly: Attributes = { writable: false, enumerable: false, configurable: true }
 /** Constants such as `Math.PI` and a constructor's `prototype`. */
 export const frozen: Attributes = { writable: false, enumerable: false, configurable: false }
-export const readOnly: Attributes = { writable: false, enumerable: false, configurable: true }

 /**
 * An object owned by the program: own properties plus a prototype link. Subclasses answer, in one place, how a
@@ -646,7 +645,7 @@ export const ownKeys = (target: Obj): Array<string | symbol> => {
  ]
 }

-export const enumerable = (target: Obj, key: string | symbol): boolean => own(target, key)?.enumerable === true
+const enumerable = (target: Obj, key: string | symbol): boolean => own(target, key)?.enumerable === true

 /** Own enumerable keys, including the iterator symbols; what spread and `Object.assign` copy. */
 export const enumerableKeys = (target: Obj): Array<string | symbol> =>
@@ -5,7 +5,6 @@ import { invalidData, IteratorSymbol, rangeError, typeError } from "../interpret
 import {
  define,
  get,
-  Callable,
  hidden,
  Arr,
  GeneratorObj,
@@ -166,12 +165,13 @@ export const arrayGlobal = <R>(ctx: Interpreter<R>) => {
    [
      "toString",
      0,
-      (thisValue) => {
-        // Spec: delegate to this.join, so an overridden join shows up in `arr + ""` and String(arr).
-        const target = self(thisValue, "toString")
-        const join = get(target, "join")
-        return join instanceof Callable ? ctx.call(join, target, []) : `[object ${target.tag}]`
-      },
+      (thisValue) =>
+        withPrimitives(
+          ctx,
+          "string",
+          Array.from(self(thisValue, "toString").items, (item) => item ?? ""),
+          (items) => items.map(coerceToString).join(","),
+        ),
    ],
    [
      "includes",
@@ -104,7 +104,7 @@ const propertyKey = (value: Value): PropertyKey =>

 // SetIntegrityLevel: primitives pass through. A typed array's bytes cannot carry attributes, so JS throws after
 // already making it non-extensible.
-export const restrict = (level: "freeze" | "seal" | "preventExtensions", value: Value): Value => {
+const restrict = (level: "freeze" | "seal" | "preventExtensions", value: Value): Value => {
  if (!(value instanceof Obj)) return value
  value.extensible = false
  if (level === "preventExtensions") return value
@@ -113,16 +113,7 @@ export const toolExpression = (path: string) =>
    .join("")

 export class ToolReference {
-  private readonly children = new Map<string, ToolReference>()
  constructor(readonly path: ReadonlyArray<string>) {}
-  /** One reference per path, so `tools.a === tools.a` holds like any other member read. */
-  child(key: string): ToolReference {
-    const existing = this.children.get(key)
-    if (existing !== undefined) return existing
-    const created = new ToolReference([...this.path, key])
-    this.children.set(key, created)
-    return created
-  }
 }

 // Dots in tool names are namespace separators; the last tool for a canonical path wins.
@@ -178,25 +178,6 @@ describe("call depth", () => {
    expect(Date.now() - started).toBeLessThan(2000)
  })

-  test("recursion routed through nested built-ins keeps counting depth", async () => {
-    const started = Date.now()
-    expect(
-      await value(`
-        const a = [1]
-        a.join = () => a + ""
-        const o = { toString() { return [o].map(String)[0] } }
-        const e = new Error()
-        e.message = { toString() { return String(e) } }
-        const names = []
-        for (const run of [() => String(a), () => String(o), () => String(e)]) {
-          try { run() } catch (error) { names.push(error.name) }
-        }
-        return names
-      `),
-    ).toEqual(["RangeError", "RangeError", "RangeError"])
-    expect(Date.now() - started).toBeLessThan(2000)
-  })
-
  test("uncaught overflow reports the call that overflowed", async () => {
    const failure = await error(`const f = (n) => f(n + 1); return f(0)`)
    expect(failure.kind).toBe("ExecutionFailure")
@@ -1382,7 +1382,7 @@ describe("Object.getPrototypeOf and Object.create", () => {
    expect((await error(`Object.getPrototypeOf(Symbol.iterator)`)).message).toContain("cannot convert a symbol")
  })

-  test("Object.create links the prototype: inherited reads, in, own-only keys, and for...in", async () => {
+  test("Object.create links the prototype: inherited reads, in, and own-only keys", async () => {
    expect(
      await value(`
        const p = { greet(name) { return "hi " + name }, a: 1 }
@@ -1392,7 +1392,7 @@ describe("Object.getPrototypeOf and Object.create", () => {
        for (const key in c) seen.push(key)
        return ["greet" in c, Object.keys(c), c.hasOwnProperty("a"), c.a, c.greet(c.name), Object.getPrototypeOf(c) === p, Object.getPrototypeOf(Object.create(null)), seen]
      `),
-    ).toEqual([true, ["name"], false, 1, "hi x", true, null, ["name", "greet", "a"]])
+    ).toEqual([true, ["name"], false, 1, "hi x", true, null, ["name"]])
  })

  test("Object.create rejects non-object prototypes and property descriptors", async () => {
@@ -1984,79 +1984,3 @@ describe("WeakMap and WeakSet", () => {
    expect((await error(`structuredClone(new WeakSet())`)).message).toContain("DataCloneError")
  })
 })
-
-describe("small language leftovers", () => {
-  test("for...in walks the prototype chain and skips keys deleted before their turn", async () => {
-    expect(
-      await value(`
-        const o = Object.create({ a: 1, shadowed: 1 })
-        o.b = 2
-        o.shadowed = 3
-        const keys = []
-        for (const k in o) keys.push(k)
-        const live = { a: 1, b: 2, c: 3 }
-        const seen = []
-        for (const k in live) { seen.push(k); delete live.b; live.z = 1 }
-        const none = []
-        for (const k in []) none.push(k)
-        for (const k in new TypeError("x")) none.push(k)
-        return [keys, seen, none]
-      `),
-    ).toEqual([["b", "shadowed", "a"], ["a", "c"], []])
-  })
-
-  test("tagged template objects are frozen", async () => {
-    expect(
-      await value(`
-        const tag = (s) => s
-        const f = () => tag\`a\${1}b\`
-        return [f() === f(), Object.isFrozen(f()), Object.isFrozen(f().raw)]
-      `),
-    ).toEqual([true, true, true])
-    expect((await error("const tag = (s) => s; tag`a`[0] = 'x'")).message).toContain("read only")
-  })
-
-  test("Array.prototype.toString delegates to join", async () => {
-    expect(
-      await value(`
-        const a = [1, 2]
-        a.join = () => "j"
-        const b = [1]
-        b.join = 5
-        return [a + "", String(a), \`\${a}\`, b.toString(), Array.prototype.toString.call([3, [4]])]
-      `),
-    ).toEqual(["j", "j", "j", "[object Array]", "3,4"])
-  })
-
-  test("Error.prototype.toString converts object name and message", async () => {
-    expect(
-      await value(`
-        const e = new Error("m")
-        e.message = { toString() { return "obj" } }
-        e.name = { valueOf() { return "N" }, toString() { return "T" } }
-        return [String(e), Error.prototype.toString.call({ name: "", message: "m" }), Error.prototype.toString.call({})]
-      `),
-    ).toEqual(["T: obj", "m", "Error"])
-    expect(
-      (await error(`const e = new Error(); e.message = { toString() { throw new RangeError("r") } }; String(e)`))
-        .message,
-    ).toContain("r")
-  })
-
-  test("generator functions inherit from GeneratorFunction.prototype", async () => {
-    expect(
-      await value(`
-        function* g() {}
-        async function* ag() {}
-        const GFP = Object.getPrototypeOf(g)
-        g.prototype = null
-        return [
-          typeof GFP, GFP === Function.prototype, Object.getPrototypeOf(GFP) === Function.prototype,
-          GFP.prototype.constructor === GFP, Object.getPrototypeOf(ag) === GFP, typeof g.bind,
-          Object.getPrototypeOf(g()) === GFP.prototype,
-        ]
-      `),
-    ).toEqual(["object", false, true, true, false, "function", true])
-    expect((await error(`function* g() {} Object.getPrototypeOf(g)()`)).message).toContain("not a function")
-  })
-})
@@ -884,6 +884,7 @@ built-ins/AsyncGeneratorFunction/invoked-as-constructor-no-arguments.js  # TypeE
 built-ins/AsyncGeneratorFunction/invoked-as-function-multiple-arguments.js  # TypeError: The Function constructor is not supported; write the function inline.
 built-ins/AsyncGeneratorFunction/invoked-as-function-no-arguments.js  # TypeError: The Function constructor is not supported; write the function inline.
 built-ins/AsyncGeneratorFunction/invoked-as-function-single-argument.js  # TypeError: The Function constructor is not supported; write the function inline.
+built-ins/AsyncGeneratorFunction/prototype/not-callable.js  # Expected SameValue(«"function"», «"object"») to be true
 built-ins/AsyncGeneratorPrototype/next/this-val-not-async-generator.js  # TypeError: Cannot read properties of undefined (reading '…').
 built-ins/AsyncGeneratorPrototype/return/this-val-not-async-generator.js  # TypeError: Cannot read properties of undefined (reading '…').
 built-ins/AsyncGeneratorPrototype/throw/this-val-not-async-generator.js  # TypeError: Cannot read properties of undefined (reading '…').
@@ -899,6 +900,7 @@ built-ins/Date/prototype/toJSON/invoke-result.js  # TypeError: Date.prototype.to
 built-ins/Date/prototype/toJSON/to-primitive-value-of.js  # TypeError: Date.prototype.toJSON called on incompatible receiver a data object.
 built-ins/Date/prototype/toString/format.js  # Expected SameValue(«null», «null») to be false
 built-ins/Date/prototype/toString/negative-year.js  # Date.prototype.toString serializes year -1 to "-0001" Expected SameValue(«undefined», «"-0001"») to 
+built-ins/Error/prototype/toString/tostring-message-throws-toprimitive.js  # ToPrimitive(msg) called by ToString(msg) throws TypeError Expected a TypeError to be thrown but no e
 built-ins/Function/15.3.5-1gs.js  # Expected a TypeError to be thrown but no exception was thrown at all
 built-ins/Function/15.3.5-2gs.js  # Expected a TypeError to be thrown but no exception was thrown at all
 built-ins/Function/15.3.5.4_2-1gs.js  # Expected a TypeError to be thrown but no exception was thrown at all
@@ -929,6 +931,7 @@ built-ins/GeneratorFunction/invoked-as-constructor-no-arguments.js  # TypeError:
 built-ins/GeneratorFunction/invoked-as-function-multiple-arguments.js  # TypeError: The Function constructor is not supported; write the function inline.
 built-ins/GeneratorFunction/invoked-as-function-no-arguments.js  # TypeError: The Function constructor is not supported; write the function inline.
 built-ins/GeneratorFunction/invoked-as-function-single-argument.js  # TypeError: The Function constructor is not supported; write the function inline.
+built-ins/GeneratorFunction/prototype/not-callable.js  # Expected SameValue(«"function"», «"object"») to be true
 built-ins/GeneratorPrototype/throw/from-state-completed.js  # Expected a E but got a TypeError
 built-ins/GeneratorPrototype/throw/from-state-suspended-start.js  # Expected a E but got a TypeError
 built-ins/Iterator/from/return-method-returns-iterator-result.js  # Iterator next must be a function.
@@ -1465,10 +1468,12 @@ language/expressions/function/dstr/ary-init-iter-get-err-array-prototype.js  # E
 language/expressions/function/dstr/ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
 language/expressions/function/dstr/dflt-ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/expressions/function/dstr/dflt-ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
+language/expressions/generators/default-proto.js  # Expected SameValue(«object», «undefined») to be true
 language/expressions/generators/dstr/ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/expressions/generators/dstr/ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
 language/expressions/generators/dstr/dflt-ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/expressions/generators/dstr/dflt-ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
+language/expressions/generators/prototype-relation-to-function.js  # Expected SameValue(«object», «») to be true
 language/expressions/greater-than-or-equal/S11.8.4_A3.2_T1.2.js  # TypeError: Binary operators require data values.
 language/expressions/greater-than/S11.8.2_A3.2_T1.2.js  # TypeError: Binary operators require data values.
 language/expressions/in/S8.12.6_A2_T2.js  # TypeError: Robin cannot be constructed: user-defined constructors and classes are not supported. Cal
@@ -1554,6 +1559,7 @@ language/expressions/super/prop-expr-obj-key-err.js  # unsupported syntax Super
 language/expressions/super/prop-expr-obj-ref-strict.js  # unsupported syntax Super
 language/expressions/super/prop-expr-obj-unresolvable.js  # Expected SameValue(«SyntaxError», «ReferenceError») to be true
 language/expressions/tagged-template/constructor-invocation.js  # The called value cannot be constructed: user-defined constructors and classes are not supported.
+language/expressions/tagged-template/template-object-frozen-strict.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/expressions/this/11.1.1-1.js  # Expected SameValue(«undefined», «undefined») to be false
 language/expressions/unary-minus/S11.4.7_A3_T5.js  # TypeError: Unary operators require data values.
 language/expressions/unary-plus/S11.4.6_A3_T5.js  # TypeError: Unary operators require data values.
@@ -1582,6 +1588,7 @@ language/statements/async-generator/dstr/dflt-ary-ptrn-elem-id-iter-val-array-pr
 language/statements/async-generator/return-undefined-implicit-and-explicit.js  # Actual ["tick 1", "tick 2", "g1 ret", "g2 ret", "g3 ret", "g4 ret"] and expected ["tick 1", "g1 ret"
 language/statements/const/dstr/ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/statements/const/dstr/ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
+language/statements/for-in/S12.6.4_A7_T2.js  # for...in visits keys deleted during the loop
 language/statements/for-in/order-enumerable-shadowed.js  # Object.create property descriptors are not supported; assign the fields after creating the object.
 language/statements/for-in/S12.6.4_A6.1.js  # TypeError: FACTORY cannot be constructed: user-defined constructors and classes are not supported. C
 language/statements/for-in/S12.6.4_A6.js  # TypeError: FACTORY cannot be constructed: user-defined constructors and classes are not supported. C
@@ -1653,10 +1660,12 @@ language/statements/function/dstr/ary-init-iter-get-err-array-prototype.js  # Ex
 language/statements/function/dstr/ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
 language/statements/function/dstr/dflt-ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/statements/function/dstr/dflt-ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
+language/statements/generators/default-proto.js  # generator functions have no GeneratorFunction.prototype
 language/statements/generators/dstr/ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/statements/generators/dstr/ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
 language/statements/generators/dstr/dflt-ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/statements/generators/dstr/dflt-ary-ptrn-elem-id-iter-val-array-prototype.js  # Expected SameValue(«3», «42») to be true
+language/statements/generators/prototype-relation-to-function.js  # generator functions have no GeneratorFunction.prototype
 language/statements/generators/restricted-properties.js  # Expected a TypeError to be thrown but no exception was thrown at all
 language/statements/labeled/value-await-non-module.js  # Failed to parse TypeScript: Expression expected.
 language/statements/let/dstr/ary-init-iter-get-err-array-prototype.js  # Expected a TypeError to be thrown but no exception was thrown at all
@@ -368,16 +368,3 @@ describe("tool references under ==", () => {
    ).toEqual([false, false, false, 0])
  })
 })
-
-describe("tool reference identity", () => {
-  test("repeated member reads yield the same reference", async () => {
-    const runtime = CodeMode.make({ tools: { probe: echo("Probe", "ok"), "ns.inner": echo("Inner", "in") } })
-    expect(
-      await value(
-        runtime,
-        `return [tools.probe === tools.probe, tools.ns.inner === tools.ns.inner, tools.ns === tools.ns, tools["probe"] === tools.probe,
-                 tools.probe === tools.ns.inner, new Set([tools.probe, tools.probe]).size, await tools.ns.inner({})]`,
-      ),
-    ).toEqual([true, true, true, true, false, 1, "in"])
-  })
-})
@@ -54,8 +54,7 @@ const modelID = (model: Model.Info) => model.modelID ?? model.id

 function claudeInfo(model: Model.Info) {
  const id = modelID(model)
-  // Versions are at most two digits so snapshot dates such as `-20260901` are not read as versions.
-  const familyFirst = /(?:claude-)?(opus|sonnet|haiku|fable|mythos)-(\d{1,2})(?:[.-](\d{1,2}))?(?!\d)/i.exec(id)
+  const familyFirst = /(?:claude-)?(opus|sonnet|haiku|fable|mythos)-(\d+)(?:[.-](\d+))?/i.exec(id)
  const versionFirst = /claude-(\d+)(?:[.-](\d+))?-(opus|sonnet|haiku|fable|mythos)/i.exec(id)
  const family = (familyFirst?.[1] ?? versionFirst?.[3])?.toLowerCase()
  const major = Number(familyFirst?.[2] ?? versionFirst?.[1])
@@ -65,12 +64,7 @@ function claudeInfo(model: Model.Info) {
    major,
    minor,
    manual: (major === 3 && minor === 7) || (major === 4 && minor < 6),
-    // Opus 5.5 and later reject disabled thinking.
-    always:
-      family === "fable" ||
-      family === "mythos" ||
-      id.toLowerCase().includes("mythos-preview") ||
-      (family === "opus" && (major > 5 || (major === 5 && minor >= 5))),
+    always: family === "fable" || family === "mythos" || id.toLowerCase().includes("mythos-preview"),
  }
 }

@@ -396,7 +390,6 @@ const bedrockConverse: Protocol = (model, support) => {
        return fields({ reasoningConfig: { type: "enabled", maxReasoningEffort: effort } })
      })
    case "toggle":
-      if (claude && claudeInfo(model).always) return []
      return claude
        ? toggle(fields({ thinking: { type: "disabled" } }), fields({ thinking: ADAPTIVE_THINKING }))
        : toggle(fields({ reasoningConfig: { type: "disabled" } }), fields({ reasoningConfig: { type: "enabled" } }))
@@ -456,7 +449,6 @@ const bedrockAISDK: Protocol = (model, support) => {
          : { reasoningConfig: { type: "enabled", maxReasoningEffort: effort } },
      }))
    case "toggle":
-      if (claude && claudeInfo(model).always) return []
      return claude
        ? toggle(
            { settings: { additionalModelRequestFields: { thinking: { type: "disabled" } } } },
@@ -515,7 +507,6 @@ const sapAICore: Protocol = (model, support) => {
          sap({ additionalModelRequestFields: { thinking: { type: "disabled" } } }),
          sap({ additionalModelRequestFields: { thinking: { type: "enabled" } } }),
        )
-      if (id.includes("anthropic") && claudeInfo(model).always) return []
      if (id.includes("anthropic"))
        return toggle(
          sap({ additionalModelRequestFields: { thinking: { type: "disabled" } } }),
@@ -104,30 +104,6 @@ test("recognizes Claude version spellings and future models", () => {
  ])
 })

-test("keeps thinking on for Claude Opus 5.5 and later", () => {
-  const supports: Variant.Support[] = [{ type: "toggle" }, { type: "effort", values: ["low", "high"] }]
-  const adaptive = ["low", "high"].map((effort) => ({
-    id: effort,
-    settings: { effort, thinking: { type: "adaptive", display: "summarized" } },
-  }))
-  for (const input of [
-    model("@opencode/ai/providers/cloudflare-ai-gateway", "anthropic/claude-opus-5.5"),
-    model("@opencode/ai/providers/anthropic", "claude-opus-5-5"),
-    model("@opencode/ai/providers/anthropic", "claude-opus-6"),
-  ])
-    expect(resolve(input, supports)).toEqual(adaptive)
-
-  expect(
-    resolve(model("@opencode/ai/providers/amazon-bedrock", "us.anthropic.claude-opus-5-5-v1:0"), [{ type: "toggle" }]),
-  ).toEqual([])
-
-  for (const id of ["claude-opus-5", "claude-opus-5-20260901"])
-    expect(resolve(model("@opencode/ai/providers/anthropic", id), supports)).toEqual([
-      { id: "none", settings: { thinking: { type: "disabled" } } },
-      ...adaptive,
-    ])
-})
-
 test("spells Cloudflare AI Gateway variants for their upstream routes", () => {
  const pkg = "@opencode/ai/providers/cloudflare-ai-gateway"
  expect(resolve(model(pkg, "openai/gpt-5.4"), [{ type: "effort", values: ["low", "xhigh"] }])).toEqual([
@@ -128,5 +128,4 @@ workspace with `/connect`, and `/models` lists their enabled models.
 ## Usage and limits

 The provider bills you for the tokens. Console records every request and, using the pricing configured on the model,
-counts it toward workspace and member monthly limits. In [usage exports](/console/usage), BYOK rows have
-`billing_source` set to `byok` and a zero `cost_micro_cents`.
+counts it toward workspace and member monthly limits.
@@ -149,8 +149,8 @@ Token prices are per 1M tokens.
 | Qwen3.7 Plus (> 256K tokens)            | $1.20  | $4.80  | $0.12       | $1.50        | **$60**                                              |
 | Qwen3.6 Plus (≤ 256K tokens)            | $0.50  | $3.00  | $0.05       | $0.625       | **$60**                                              |
 | Qwen3.6 Plus (> 256K tokens)            | $2.00  | $6.00  | $0.20       | $2.50        | **$60**                                              |
-| DeepSeek V4.1 Flash (Off-Peak)          | $0.15  | $0.60  | $0.003      | -            | ~~$15~~ **$60**<br /><small>4x · Ends Sep 27</small> |
-| DeepSeek V4.1 Flash (Peak)              | $0.30  | $1.20  | $0.006      | -            | ~~$15~~ **$60**<br /><small>4x · Ends Sep 27</small> |
+| DeepSeek V4.1 Flash (Off-Peak)          | $0.15  | $0.60  | $0.003      | -            | **$60**                                              |
+| DeepSeek V4.1 Flash (Peak)              | $0.30  | $1.20  | $0.006      | -            | **$60**                                              |
 | DeepSeek V4 Pro (Off-Peak)              | $0.66  | $1.98  | $0.022      | -            | **$15**                                              |
 | DeepSeek V4 Pro (Peak)                  | $1.32  | $3.96  | $0.044      | -            | **$15**                                              |
 | DeepSeek V4 Flash (Off-Peak)            | $0.15  | $0.60  | $0.003      | -            | **$30**                                              |
@@ -206,7 +206,7 @@ The table below provides an estimated request count based on typical Go usage pa
 | Qwen3.7 Max                                              | 170                       | 420                        | 840                         |
 | Qwen3.7 Plus                                             | 4,300                     | 10,800                     | 21,600                      |
 | Qwen3.6 Plus                                             | 3,300                     | 8,200                      | 16,300                      |
-| DeepSeek V4.1 Flash<br /><small>4x · Ends Sep 27</small> | ~~6,500~~<br />**26,000** | ~~16,250~~<br />**65,000** | ~~32,500~~<br />**130,000** |
+| DeepSeek V4.1 Flash                                    | 26,000                   | 65,000                   | 130,000                   |
 | DeepSeek V4 Pro                                          | 1,050                     | 2,600                      | 5,200                       |
 | DeepSeek V4 Flash                                        | 13,000                    | 32,500                     | 65,000                      |
 | DeepSeek V4 Flash Vision Exp                             | 6,500                     | 16,250                     | 32,500                      |
@@ -1,146 +0,0 @@
---
-title: "Usage API"
-description: "Export usage records as CSV for a workspace, member, service account, or model."
---
-
-Export usage records as CSV for a workspace, member, service account, or model. Use the versioned endpoint in scripts,
-reporting jobs, and billing integrations.
-
-## Quick start
-
-Create a service account API key in the [Console](https://opencode.ai/console). This endpoint accepts service account
-keys only; user session tokens are rejected.
-
-```bash
-export CONSOLE_URL="https://opencode.ai/console"
-export SERVICE_API_KEY="oc_sk_..."
-```
-
-Export the last seven days of workspace usage:
-
-```bash
-curl --fail-with-body --get \
-  "${CONSOLE_URL}/api/v1/usage/export" \
-  --header "Authorization: Bearer ${SERVICE_API_KEY}" \
-  --header "Accept: text/csv" \
-  --data-urlencode "scope=organization" \
-  --data-urlencode "range=7d" \
-  --output "usage-organization-7d.csv"
-```
-
-A successful request writes `usage-organization-7d.csv`. Remove `--output` to print the CSV in your terminal instead.
-
-## Endpoint and scopes
-
-```text
-GET /api/v1/usage/export
-```
-
-Every request requires `scope` and `range`. Supported ranges are `24h`, `7d`, and `30d`. They start at midnight UTC
-rather than being rolling windows. The full range is returned as a single streamed CSV file, regardless of size.
-
-| Scope             | Additional parameters  | Result                                   |
-| ----------------- | ---------------------- | ---------------------------------------- |
-| `organization`    | None                   | All usage in the workspace.              |
-| `member`          | `user_email`           | Usage attributed to one member.          |
-| `service_account` | `service_account_id`   | Usage attributed to one service account. |
-| `model`           | `provider` and `model` | Usage for one provider and model pair.   |
-
-<Callout title="Use only the parameters for your scope">
-  Member, service account, and model filters are mutually exclusive. Extra filters return HTTP 400 instead of being
-  silently ignored.
-</Callout>
-
-## Export examples
-
-One member:
-
-```bash
-curl --fail-with-body --get \
-  "${CONSOLE_URL}/api/v1/usage/export" \
-  --header "Authorization: Bearer ${SERVICE_API_KEY}" \
-  --header "Accept: text/csv" \
-  --data-urlencode "scope=member" \
-  --data-urlencode "range=24h" \
-  --data-urlencode "user_email=alice@example.com" \
-  --output "usage-member-24h.csv"
-```
-
-One service account:
-
-```bash
-curl --fail-with-body --get \
-  "${CONSOLE_URL}/api/v1/usage/export" \
-  --header "Authorization: Bearer ${SERVICE_API_KEY}" \
-  --header "Accept: text/csv" \
-  --data-urlencode "scope=service_account" \
-  --data-urlencode "range=30d" \
-  --data-urlencode "service_account_id=svcacct_..." \
-  --output "usage-service-account-30d.csv"
-```
-
-One provider and model:
-
-```bash
-curl --fail-with-body --get \
-  "${CONSOLE_URL}/api/v1/usage/export" \
-  --header "Authorization: Bearer ${SERVICE_API_KEY}" \
-  --header "Accept: text/csv" \
-  --data-urlencode "scope=model" \
-  --data-urlencode "range=7d" \
-  --data-urlencode "provider=anthropic" \
-  --data-urlencode "model=claude-sonnet-4-5" \
-  --output "usage-model-7d.csv"
-```
-
-Provider is the provider catalog key, such as `anthropic`, `deepseek`, or `opencode`. A valid filter with no matching
-records still returns HTTP 200 with a header-only CSV file.
-
-## CSV response
-
-Results are newest first. Empty optional values are blank cells. `cost_micro_cents` is the amount charged by Console
-in microcents, where 100,000,000 equals one dollar. Managed-inference usage carries its charge; free, BYOK, and
-unclassified legacy usage have a zero charge.
-
-Organization and member exports include [Websearch](/console/websearch) charges. Search rows identify `web-search` in
-the `service` column and leave provider, model, and token fields blank. Model and provider filters select inference
-records only.
-
-<div class="docs-table-scroll" role="region" aria-label="Usage export CSV fields" tabIndex={0}>
-
-| Field                     | Description                                                                                              |
-| ------------------------- | -------------------------------------------------------------------------------------------------------- |
-| `id`                      | Unique record ID. New service records use a `service:` prefix; inference and historical IDs are numeric. |
-| `user_email`              | Current email of the member that generated the usage, when applicable.                                   |
-| `service_account_name`    | Current name of the service account that generated the usage, when applicable.                           |
-| `app`                     | Client app title, or its referrer when no title was reported.                                            |
-| `provider`                | Provider catalog key, such as `openai`, `anthropic`, `deepseek`, or `opencode`.                          |
-| `model`                   | Model identifier reported for the request.                                                               |
-| `input_tokens`            | Number of input tokens consumed.                                                                         |
-| `output_tokens`           | Number of output tokens generated.                                                                       |
-| `reasoning_tokens`        | Number of reasoning tokens reported by the provider.                                                     |
-| `cache_read_tokens`       | Number of input tokens read from the provider cache.                                                     |
-| `cache_write_5m_tokens`   | Number of tokens written to a five-minute cache.                                                         |
-| `cache_write_1h_tokens`   | Number of tokens written to a one-hour cache.                                                            |
-| `reasoning_mode`          | Reasoning mode used, such as `effort`, `adaptive`, `budget`, or `disabled`.                              |
-| `reasoning_effort`        | Provider-specific reasoning effort value, when supplied.                                                 |
-| `reasoning_budget_tokens` | Requested reasoning-token budget, when supplied.                                                         |
-| `reasoning_source`        | Source from which the reasoning configuration was derived.                                               |
-| `billing_source`          | How the request was funded, such as `managed-inference`, `credit`, `byok`, or `free`.                    |
-| `cost_micro_cents`        | Amount charged by Console in microcents. Divide by 100,000,000 to convert to USD.                        |
-| `created_at`              | When the usage was recorded, as an ISO 8601 UTC timestamp.                                               |
-| `service`                 | Service identifier, currently `web-search`. Blank for model requests.                                    |
-| `quantity`                | Number of billable operations represented by the record; normally one.                                   |
-
-</div>
-
-## Errors
-
-| Status | Meaning                                                                              |
-| ------ | ------------------------------------------------------------------------------------ |
-| `400`  | Missing parameters, invalid values, or filters that do not match the selected scope. |
-| `401`  | Missing, invalid, expired, or revoked service API key.                               |
-| `403`  | The authenticated service account is not allowed to read usage.                      |
-
-Curl uses `--fail-with-body` in these examples so authentication and validation failures produce a non-zero exit code
-while preserving any error response.
@@ -146,7 +146,6 @@ export const docsSections: DocsSection[] = [
        items: [
          { title: "Inference", slug: "console/inference" },
          { title: "BYOK", slug: "console/byok" },
-          { title: "Usage", slug: "console/usage" },
          { title: "Budgets", slug: "console/budgets" },
        ],
      },
Author	SHA1	Message	Date
Shoubhit Dash	6585bb7105	fix(ai): keep OpenAI image output settings and Z.ai URL expiry (#51380 )	2026-09-25 22:51:46 +05:30
Shoubhit Dash	f954688fbb	fix(ai): harden OpenAI transcription stream parsing (#51378 )	2026-09-25 22:51:22 +05:30
Shoubhit Dash	1463dabde9	fix(ai): tighten media error consistency (#51374 )	2026-09-25 22:46:04 +05:30
Shoubhit Dash	29ea6ee05b	test(ai): cover media facade selectors and url asset edges (#51376 )	2026-09-25 22:37:20 +05:30
Shoubhit Dash	b170904731	docs(ai): describe transcription speakers as a constraint (#51382 )	2026-09-25 22:36:02 +05:30
Shoubhit Dash	88e1fa9304	fix(ai): accept timestamps: false on speech routes without timestamps (#51375 )	2026-09-25 22:33:59 +05:30
opencode-agent[bot]andvimtor	ff1bf315ed	docs(www): hide Console Usage API documentation (#51358 ) Co-authored-by: vimtor <36263538+vimtor@users.noreply.github.com>	2026-09-25 19:01:48 +02:00
Shoubhit Dash	144ce00e00	fix(ai): bound generation event poll sleep by the deadline (#51372 )	2026-09-25 22:31:35 +05:30
Shoubhit Dash	ad504094f0	fix(ai): never delete a finished Runway task on cancel (#51373 )	2026-09-25 22:24:16 +05:30
Shoubhit Dash	bad6834a3e	fix(ai): size fal Kontext by aspect ratio and decode sync_mode data URIs (#51371 )	2026-09-25 22:23:14 +05:30
Shoubhit Dash	0c4bbc3cd1	feat(ai): keep prompt cache across effort switches on GPT-6 Sol and Luna (#51339 )	2026-09-25 22:15:23 +05:30
Shoubhit Dash	ae7dd82126	fix(ai): report Black Forest Labs submit cost as image usage (#51370 )	2026-09-25 22:13:55 +05:30
Shoubhit Dash	65d5123ead	fix(ai): enable AssemblyAI speaker labels when speakers is set (#51369 )	2026-09-25 22:11:44 +05:30
Shoubhit Dash	4eb46a8885	fix(core): revert always-thinking variants for Claude Opus 5.5 (#51359 )	2026-09-25 22:03:12 +05:30
Jack	1986e92842	docs(go): show permanent DeepSeek $60 allowance (#51363 )	2026-09-26 00:17:08 +08:00