[elsehow]

Lets AIRO's published instrument run through pi with the same loop rules as its own litellm loop (redlines.llm.call_tools), for a shadow comparison before any cut-over.

All options are off by default, so the ForecastBench loop is unchanged (its four tests pass untouched; the registered names still match the code).

option	mirrors in call_tools
floor_counts="returned"	min_evidence counts every returned research call
coverage=[keys]	the coverage rule: a recent_days search with cause per group
withhold_final=True	the final tool not offered until floor and coverage are met
warn_missing_keys=False	the final answer accepted as sent
max_turns=N	max_iters, the forced last turn as for models that refuse a forced tool_choice, + 2 turns; prose replies nudged
debrief={...}	_debrief_turn: merged deliveries, validator, retries
tool_parameters	AIRO's tool schemas verbatim (e.g. the cause enum)

Known differences, recorded for the shadow comparison: pi cannot force tool_choice, so every model gets the ask-in-words last turn (AIRO forces it for OpenAI models); after an accepted answer or debrief pi makes one more short request; the debrief's text-only fallback for a model that returns nothing (Fable 5.1) is not ported; per-turn max_tokens is pi's model default rather than 64k/48k.

Tests: each option through the real pi with the scripted model, each checked by breaking it in pi_tools.ts.

🤖 Generated with Claude Code