Skip to content
loopctl

`src/provider/openai.rs` — `OpenAiClient` [feature: openai]

The Chat Completions client — base of the whole compatible family. (~3,950 lines with tests.)

Key items

  • from_env() (OPENAI_API_KEY / BASE_URL / MODEL → gpt-4o) + the full builder (with_stream_usage true default).
  • Full RequestOptions support: response_format → native json_schema (tools suppressed); ToolConstraint::Strict → tightened schemas + strict: true; Grammar → guided_json.
  • The StreamEmitter — lane bookkeeping for interleaved text/thinking/tool content (text lane index 0, thinking 1); deferred MessageDelta until the usage chunk; [DONE] sentinel handling; no synthetic MessageStop on failure.
  • Mid-stream error classification: rate_limit_error / 429 / 503 / 529 / rate_limit_exceeded → ApiError::RateLimit.

Behavior notes

  • Tool-call arguments stream as string fragments; the lane-closing discipline keeps them out of the text slot (the subtle bug this prevents: tools executing with {} because their JSON landed in the wrong lane — pinned by tests).
  • Tool results: is_error not forwarded (no wire field); N results → N tool messages + trailing user text; images dropped.
  • set_model("") rejected (false); base-URL setter trims trailing slashes.
  • Non-streaming: non-JSON tool arguments are an ApiError (not silently {}); all-zero usage collapses to None.

Deep dive: OpenAI.