OpenAI Responses provider worker with Chat Completions compatibility; implements provider::openai::stream and provider::openai::refresh_models behind llm-router.
iii trigger compose::add worker=provider-openai@1.2.20Run in your project with the iii engine running. Exact versions are immutable and digest-pinned.
Serve the provider-openai worker's injected console UI assets (content function for its console:script / console:style / console:module triggers).
The asset path from the trigger config (e.g. `state/page.js`).
The asset source, verbatim.
MIME type the console should serve the asset with.
Cancel the in-flight upstream stream for a request_id (router::abort fan-out), stopping billed generation immediately.
Count prompt tokens for {model, system_prompt?, tools?, messages} locally with the tiktoken tokenizers; never runs the model and costs nothing.
Wire agent messages, the same shape `provider::openai::stream` accepts. Must be non-empty.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Single-variant role tags: exact-match on deserialize, correct wire string on serialize, and they let `AgentMessage` be an untagged union.
Model id the prompt targets; selects the tokenizer (o200k_base for gpt-4o/gpt-5/o-series and anything unknown-modern, cl100k_base for gpt-3.5 and non-o gpt-4 families).
System prompt counted as its own wire message when present.
Function invocation schemas; each serialized schema counts toward the total.
Always `tiktoken`: a local tokenizer produced the estimate.
Estimated prompt tokens for the assembled request.
Batch text embeddings via the OpenAI embeddings endpoint, using the router-resolved credential. One vector per input, order preserved. Default model text-embedding-3-small.
Texts to embed, one vector returned per input, order preserved.
Embedding model id; the provider default when omitted.
One embedding per input, in input order.
Generate one image from a text prompt with gpt-image-2.5-sunburst or gpt-image-2.5-flare (OpenAI Images API, router-resolved credential). Returns viewable content blocks (image + caption) that render inline, and saves the file under data/provider-openai/images/ (details.path) for other processes to pick up.
`transparent`, `opaque`, or `auto`; transparency needs png or webp.
File stem for the persisted image (letters, digits, `.`, `_`, `-`; anything else becomes `-`). A timestamp and a random suffix are always appended so files never collide. Defaults to `image`.
Whether the result embeds image bytes: `none` (default — the file is saved, `details.path` names it, and the console fetches its own preview), `preview` (a JPEG at most 512px / ~80 KB, for an agent that must look at the picture; fits the harness result cap), or `full` (the original bytes — over a megabyte for a 1024px PNG).
Image model: `gpt-image-2.5-sunburst` or `gpt-image-2.5-flare`.
`png` (default), `jpeg`, or `webp`: the format of the file saved under data/ (and of the inline picture when `inline` is `full`).
What to draw (up to 32k characters).
`low`, `medium`, `high`, or `auto`; the model default when omitted.
`1024x1024`, `1536x1024` (landscape), `1024x1536` (portrait), or `auto`; the model default when omitted.
`[image block, text line]` (`[text line]` with `inline: none`) — render `content` and the picture shows.
Saved file size in bytes.
Pixel height of the saved file, when decodable.
`preview`, `full`, or `none`: what the inline image block carries.
Mime of the saved file: `image/png`, `image/jpeg`, or `image/webp`.
Image model that produced the picture.
Absolute path of the full-resolution file under `data/provider-openai/images/`.
Present when `inline` is `preview` and the preview could be built.
Requested quality (`auto` when left to the model).
The prompt as the model actually rendered it, when upstream reports one.
Requested size (`auto` when left to the model).
Pixel width of the saved file, when decodable.
Read back an image saved by image::generate (its details.path or bare file name under data/provider-openai/images/) as viewable content blocks: a JPEG preview (default, fits the harness result cap) or the full bytes. Refuses any path outside that folder.
Absolute path returned by `image::generate`, or the bare file name inside `data/provider-openai/images/`.
`preview` (default; a JPEG at most 512px / ~80 KB that fits the harness result cap) or `full` (the saved bytes as they are).
`[image block, text line]` — the same viewable envelope as generate.
Saved file size in bytes.
Mime of the saved file.
Absolute path of the file that was read.
`preview` or `full`: what the image block carries.
Internal: router::ready subscriber that re-declares this provider and refreshes its catalog.
Refresh the OpenAI catalog slice from GET /v1/models and reconcile it through the router; returns the model count written.
Stream an OpenAI response: resolve credentials, call the configured Responses or Chat Completions endpoint, and relay AssistantMessageEvent frames to writer_ref.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Optional link to the stored original in the session-manager attachment store. Transcript bookkeeping for lazy readers only: the harness clears it before `router::chat`, and no provider mapping reads it. Tolerated here so a block that still carries it deserializes instead of failing the request; omitted on the wire when absent.
Single-variant role tags: exact-match on deserialize, correct wire string on serialize, and they let `AgentMessage` be an untagged union.
Set on speech models only; see [`SpeechModel`].
BCP-47 tags of the languages the model handles; empty when the provider does not say.
What a speech model does. Chat models carry no `speech` block.
Realtime input (stt) or streamed audio output (tts) is available.
Stable conversation identity for provider cache affinity.
Ordered sections behind `system_prompt` (always forwarded alongside the flat string, which equals the sections joined with "\n\n").
Find workers, kits and executables, or jump to a page.