Chat completions

Which request fields are real

Only two.

Everything else — model, temperature, tools, top_p, max_tokens, response_format — is accepted and ignored. The call above sent "model":"gpt-4o", "temperature":0.2 and a tools array, and the execution target received only {"goal":"How did Q3 revenue go?","caller_id":"user_abc","project_id":"proj_abc"}.

Only the last user message is sent to the agent, as its goal. Earlier messages, and a system prompt sent alongside it, do not reach the agent, so the agent does not see the conversation the client holds. If there is no user message at all:

{ "error": { "message": "messages[] must include at least one message with role \"user\"",
             "type": "invalid_request_error", "code": "missing_messages" } }

Streaming is framing, not incremental tokens

"stream": true produces exactly the frames an OpenAI client expects, in the right order:

data: {"id":"chatcmpl-exec_abc_77","object":"chat.completion.chunk","created":1789469617,"model":"af_abc123","choices":[{"index":0,"delta":{"role":"assistant","content":"Q3 revenue was 4.2M, up 11% on Q2."},"finish_reason":null}]}

data: {"id":"chatcmpl-exec_abc_77","object":"chat.completion.chunk","created":1789469617,"model":"af_abc123","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: [DONE]

Content-Type becomes text/event-stream, and the whole answer arrives in the first chunk.

Rate limit: 60 per minute, per key

This route is the only place on the gateway where the per-key limiter is consulted.

A single fresh key was fired at the endpoint 65 times in one window. The first 60 answered 200; the 61st and everything after it answered:

HTTP/1.1 429 Too Many Requests
Retry-After: 1

{"error":{"message":"too many requests for this API key — try again shortly","type":"rate_limit_exceeded","code":"rate_limited"}}

A second key, minted on the same project, answered 200 at the same moment — the window is per key id, not per project or per user.

Three details that change how you should retry:

• The window is a fixed one minute, not a sliding one. The allowance resets wholesale rather than draining.
• Retry-After is a hard-coded 1, not the real seconds remaining. Back off on your own schedule; waiting one second immediately after the 61st call will usually be refused again.

The project and organization quota windows are separate, apply to the governed tool routes, and produce Wexa's own error body with a genuinely computed Retry-After. Quota and credits covers them.

A credential with no key id — a user token — is not subject to this limiter at all. That was confirmed: the same request with a user token answered 200.

Every failure shape

Six of the nine were produced live; the ledger at the foot of the page says which.

The 409 deserves a note, because its message is unusually long on purpose:

{ "error": { "type": "invalid_request_error", "code": "not_ready",
  "message": "agentflow has no promoted version promote-flow will NOT fix this: it snapshots the flow's shape, not the agent's config. If this agent already has versions, rollback-agent-version makes one live and this endpoint then works — list them with agent-versions. Creating the FIRST version needs the Fabric console (Agent Simulation, an Enterprise-tier feature). run-process-flow and run-agent require no promotion at all." } }

An agent must have a promoted version before this endpoint will run it. If yours does not, run-agent and run-process-flow need no promotion and are the way through.

Four compatibility limits

1. 401 does not speak OpenAI. Every other failure on this route uses the OpenAI error envelope; a missing or invalid credential is refused by the gateway's shared authentication layer before the handler runs, and comes back as Wexa's own shape — {"error":"invalid_token","error_description":"missing bearer credential"}. An OpenAI client that parses errors strictly will fail to parse that one. Handle 401 separately.
2. usage.prompt_tokens is always 0. The execution target reports a total; the gateway puts it in both total_tokens and completion_tokens and leaves the prompt count at zero. Do not compute cost from these fields.
3. n, logprobs, tools and function calling are not supported. There is always exactly one choice, finish_reason is always stop on success, and tool use happens inside the agent rather than being negotiated with the caller.
4. There is no GET /v1/models. The agent id in the URL is the model selector; a client that lists models at startup will not find one.

Per-route ledger

The status mapping, the request translation and the rate limit are gateway behaviour. The answer text and the token count come from whatever execution target your deployment runs.