kurultai

name	kurultai
description	Convene the Kurultai council of rival AI models and act as its judge. Use when the user asks to "run the council", "stress-test this idea with other models", "let models debate/attack/estimate this", or when an important decision deserves adversarial multi-model review.

/kurultai - convene the council, then judge it

Kurultai is a CLI council of rival models (council.py, pure stdlib). You are the CHAIR: launch the run, wait for the panel, read the transcript, and fill in the judge section right here in chat - never via a paid API call.

Steps

1. Pick the mode from the user's intent (unless they named one):
   • contested question, positions should fight -> debate "<topic>" [debate|brainstorm|decision] [rounds]
   • independent panel review of a doc/question + web grounding -> fusion <file> [quality|budget]
   • stress-test one idea, hunt for fatal flaws -> redteam "<idea or file>"
   • choose among 2-6 options -> tournament <file> (goal text + a markdown list)
   • numeric estimate (how many, probability) -> delphi "<question>" [rounds]
2. Run it: python3 council.py <mode> ... from the repo root. For long questions, write them to a temp file first and pass the path.
3. Wait for the DONE line with the runs/<...>.md path. If some models were silent, run python3 council.py resume runs/<...>.json (writes a new -r2 pair).
4. Read the transcript and fill the judge template in chat: Consensus / Contradictions / Partial coverage / Unique insights / Blind spots (+ the mode-specific sections from the template at the end of the file), then a hard final synthesis. Blind spots are YOUR layer on top of the panel, not a summary.
5. Reply to the user with a tight verdict (consensus + blind spots + your call), transcript path at the end.

Notes

• The judge is never a paid API call - the judge is you, in this chat.
• Providers may be down; the ladder reroutes automatically. doctor shows who is alive.
• Never overwrite anything in runs/ - it is an append-only archive.