Expand description
What a run spent.
A run makes many agent calls at scheduled effort, and the terminal report
said nothing about them: no count per agent, no wall time, no retries or
hand-overs. max_rounds, absorb_new_issues, max_followups and the
effort schedule are all documented in terms of what they cost, and nothing
in the output let anybody see whether changing one did what they hoped.
Ambient on purpose. Threading a counter through every call site would put accounting in the signature of everything that asks a model a question, and the one thing this must not do is change what the loop decides.
Structs§
- Spent
- One call to one agent.
Functions§
- for_
subject - What one issue cost, for the line under its result.
- record
- Record one call, and say so as it happens.
- summary
- One line per agent, and one for the run, or nothing when nothing was asked.
- taken
- Everything recorded so far.