OpenAI transport

For a walkthrough, see Use the OpenAI client.

sealing_http_client()Copy Anchor Link

http_client = vorq.sealing_http_client(
    *,
    base_url: str = "https://api.vorq.co",
    signer: Signer | None = None,
    cipher: Cipher | None = None,
    verifier: Verifier | None = None,
    timeout: float = 30.0,
    inner_transport: httpx.AsyncBaseTransport | None = None,
) -> httpx.Client

Returns an httpx.Client for openai.OpenAI(http_client=...), whose transport is a vorq.SealingTransport built from the same arguments.

Sync-only. Each call runs a fresh native vorq.Client, with its own session handshake, on a private event loop. Called inside a running event loop, it raises RuntimeError, which the openai package reports as openai.APIConnectionError.

Intercepted routesCopy Anchor Link

CreateCopy Anchor Link

The request body becomes the model input, minus model, background, vorq and metadata. A list-valued input is sealed under the key messages. The same local param validation as submit runs, and can't be skipped.

The vorq block (sent through extra_body={"vorq": {...}}):

Other keys in the block are ignored.

• background=True: returns at once with a queued Response whose id is the job id.
• Without background: blocks until the job settles or its sla window elapses. A timeout surfaces as a 400.

Refused with a 400 before anything is sealed:

Response objectsCopy Anchor Link

Every Response carries a top-level vorq block echoing the job's terms as the coordinator reports them.

• A text result becomes one message output item with output_text, plus usage.
• A media result becomes one image_generation_call item per frame, with the frame's base64 in result. Video uses the same item type; read content_type from the native MediaResult to tell them apart.
• An embeddings result renders no output items.
• A job that failed or was cancelled carries an error object whose code is the end cause.

Forwarded routesCopy Anchor Link

These routes carry no prompt and are forwarded to the coordinator over the transport's own wallet session. Each is matched on method and path.

The optional metadata on a batch create is stored in plaintext. The Authorization header your openai client sends is dropped. A successful forwarded response keeps its status, body, content type and x-request-id; an error is returned as described in Errors and retries.

Refused routesCopy Anchor Link

Every other route gets a 400 from the transport itself, before the request body is read, so nothing leaves your process. That includes chat.completions, files.create (a batch input file would go up in plaintext; use client.batches.submit()), POST /v1/jobs, and any route the openai package adds later.

Both the intercepts and the forward list match /v1/... exactly. An OpenAI(base_url=...) under a sub-path such as https://host/api/v1 matches nothing, Responses included, and every call is refused.

Errors and retriesCopy Anchor Link

A coordinator error keeps its HTTP status, so the openai package raises its usual classes:

Errors the transport raises itself, including refusals and a synchronous wait that timed out, are 400 / openai.BadRequestError.

POST /v1/responses and POST /v1/batches failures are marked not retryable, so the openai package doesn't retry them: a retry would create a second job or batch and could bill twice. Reads and cancels keep the package's normal retry behavior.

Differences from OpenAICopy Anchor Link