Module chat_api 

HTTP client for the OpenRouter chat-completions API (direct HTTP, no CLI subprocess). HTTP client for the OpenRouter chat-completions API.

Sends structured-output chat requests to the OpenAI-compatible endpoint at openrouter.ai/api/v1/chat/completions and returns the parsed JSON object the model produced under a strict json_schema response_format.

This mirrors crate::embedding_api for the embeddings endpoint: same retry/backoff policy (immediate abort on 401/400/404, retry-after on 429, exponential backoff + jitter on 5xx) and the same minimal headers (only Authorization: Bearer, no HTTP-Referer/X-Title). The shared error envelope and backoff helper live in crate::openrouter_http (GAP-SG-74).

v1.0.95 (ADR-0054): adds an OpenRouter REST transport for the enrich JUDGE so structured extraction no longer requires a locally installed claude / codex / opencode CLI subprocess.

v1.1.00 (GAP-SG-70/72-chat): the OpenAI-compatible contract surfaces choices[].finish_reason and usage.{prompt_tokens,completion_tokens}. finish_reason == "length" means the response was truncated because max_tokens was too small — not a malformed generation. OpenRouterChatClient::complete now detects this BEFORE attempting JSON repair, grows max_tokens and re-issues the request (bounded by crate::constants::ENRICH_MAX_LENGTH_RETRIES), and always reports the diagnostics (finish_reason, token counts) to the caller via ChatCompletion on success or ChatError on failure.

Structs

ChatCompletion

Successful OpenRouterChatClient::complete result (GAP-SG-72-chat).

ChatError

OpenRouterChatClient::complete failure (GAP-SG-72-chat / GAP-SG-72 reauditor addendum).

OpenRouterChatClient

Process-wide OpenRouter chat client. Holds the model name so that callers only thread the per-item prompt/schema/input through Self::complete.