Expand description
LLM / embedding HTTP call tracing.
Every chat (distill) and embedding request flows through llm::post_json_retry,
which calls record once per call with the final outcome. Traces are appended
as JSONL to ~/.innate/logs/llm_trace.log (size-capped with single-file rotation)
so any process — MCP, CLI, evolve — contributes automatically, failures included.
Design notes:
- Never logs the API key. Only the request body is captured; the
Authorizationheader is not passed in and is never recorded. - Request/response bodies are truncated to keep the log bounded.
- Disable entirely with
INNATE_LLM_TRACE=0.
Functions§
- read_
recent - Read the most recent trace entries (newest first) for the web viewer. Optional
exact-match filters on
kind(“chat”/“embedding”) andstatus. Read-only. - record
- Record one LLM/embedding call outcome. Best-effort: tracing failures never affect the caller (the LLM call result is returned regardless).