Class AiToolLogger
Which inputs a model gets wrong is otherwise a matter of reading the code and guessing: a refusal reaches the model as a message and is gone with the turn. A chat keeps its tool parts, so the transcript can be read back — but a programmatic run persists no messages at all, and the unattended path is where a wrong call goes unwatched longest.
A refusal is logged as a warning, not an error. It is the model's mistake and the system is fine: now that a wrong field name, a wrong value and an unknown parameter are all refused rather than dropped, a model correcting itself is ordinary, and logging each one as a fault would bury the faults that are real.
The call's input is logged only under DEBUG, because it carries whatever data the
model was working with — the same reason ai.logging.requests defaults to off.
-
Method Summary