src/provider/openai.rs

This file holds OpenAiClient, the Chat Completions client, which is the base of the whole compatible family. It runs about 3,950 lines with tests.

Key items

from_env() reads OPENAI_API_KEY, BASE_URL, and MODEL, defaulting to gpt-4o. The full builder gives you with_stream_usage and replay_reasoning, both defaulting to true.

Since 0.4.0 replay_reasoning returns prior thinking blocks on continuation requests, which keeps your reasoning visible across turns.

Full RequestOptions support covers three mappings. Your response_format becomes native json_schema with tools suppressed, ToolConstraint::Strict becomes tightened schemas plus strict: true, and Grammar becomes guided_json.

model_context_window() shipped in 0.4.0 and is the one client that discloses what it finds. It reads llama.cpp /props for n_ctx / total_slots and Ollama /api/show for the architecture-keyed context_length.

Both are server-root probes under a 5-second deadline, cached per model epoch, cleared when you call set_model, and fail-soft to None.

The StreamEmitter keeps lane bookkeeping for interleaved text, thinking, and tool content, with text at lane index 0 and thinking at 1.

It defers MessageDelta until the usage chunk, handles the [DONE] sentinel, and emits no synthetic MessageStop on failure.

Mid-stream error classification maps rate_limit_error, 429, 503, 529, and rate_limit_exceeded to ApiError::RateLimit.

Behavior notes

Tool-call arguments stream as string fragments, and the lane-closing discipline keeps them out of the text slot.

That discipline prevents a subtle bug where tools executed with {} because their JSON landed in the wrong lane, and tests pin it.

For tool results, is_error is not forwarded, since there is no wire field for it. Your N results become N tool messages plus trailing user text, and images are dropped.

For usage, prompt_tokens_details.cached_tokens fills Usage::cached_input_tokens. This reads leniently, so a non-numeric placeholder reads as zero without dropping the usage object.

set_model("") is rejected and returns false, and your base-URL setter trims trailing slashes.

In non-streaming mode, non-JSON tool arguments are an ApiError rather than a silent {}, and all-zero usage collapses to None.

For the deep dive, see OpenAI.