model-council

Multi-perspective AI deliberation for the pi coding agent. Inspired by Perplexity's Model Council.

Convene a council of analytical perspectives, cross-validate findings, and surface consensus, unique discoveries, and conflicts — so you know when to move faster with confidence, and when to dig deeper.

Install

pi install npm:@demi8-patch/model-council

Requires pi-subagents (auto-installed as peer dependency on first use).

Usage

Invoke the skill when facing high-stakes questions where single-perspective blind spots are costly:

/skill:model-council

Or let the agent fire it automatically — trigger words: "council", "deliberate", "multi-perspective", "cross-validate", "second opinion".

What you get

A consolidated answer in four sections:

• Consensus — what the council independently agreed on (with citations)
• Unique discoveries — insights only one perspective caught
• Conflicts — where perspectives disagree, and why
• Confidence — whether to move faster or dig deeper

Cost transparency

Cost: approximately $0.50–$3.00 per council depending on perspective count and model choices. Full breakdown:

| Path | Calls | | ------------------------ | ------------------------------------------------------- | | 3-perspective happy path | 5 subagent calls (Convene + 3 Collect + Synthesize) | | 5-perspective deep mode | 7 subagent calls (Convene + 5 Collect + Synthesize) | | Worst-case with retries | up to 8 subagent calls |

Each subagent() call expects 2k-8k output tokens. Suggested timeout: 120 seconds per call. The full 4-section synthesis output is capped at ~1200 words.

License

MIT — see LICENSE.