factor_agent.llm.log#
The LLM log: every call of a run, in full, one JSON file per call under llm_log/.
A file <number>-<role>.json holds the request (role, model, prompt and
the parameters sent) and the reply (text, usage, the model that
answered and the provider’s raw reply); the numbers count the calls of the
run from 1, across every role, in the order they finish (calls sent side by
side are written one at a time); the calls of a round a resume dropped are
kept in dropped/, out of the sequence. No key is ever part of a request
or a reply, so none is written.
Classes
The call files of one run directory's |
|
A backend that writes every call it passes on to |