multi-model

Multi-Model Collaboration Skill

Orchestrate multi-model AI collaboration using external LLM CLIs. Each council member is a different model provider for genuine perspective diversity.

How It Works

The multi-model system uses external LLM CLIs discovered at runtime:

1. CLI Discovery — which codex copilot gemini opencode pi detects installed CLIs
2. Quorum Check — Council requires 2+ external CLIs; 0 = refuse, 1 = warn
3. Question Scaffold — All models answer the same fixed 4-question set
4. Parallel Dispatch — Each CLI receives the scaffold via stdin pipe, runs independently
5. Synthesis — Claude synthesizes all perspectives with model attribution

Quick Start

# Council mode — dispatches to external CLIs in parallel
/jam:council "Should we use JWT or sessions for auth?"

# Quick jam — single model, 4 personas, fast
/jam:quick "How should we improve the visual design?"

# Full brainstorm — single model, 4-6 personas, 2-3 rounds
/jam:brainstorm "Architecture for the notification system"

Supported CLIs

CLI	Install	Model
codex	brew install codex	OpenAI Codex
copilot	brew install copilot-cli	GitHub Copilot
gemini	npm i -g @google/gemini-cli	Google Gemini
opencode	brew install opencode	Configurable
pi	brew install pi-mono	Pi AI

Claude always participates as a council member alongside the external CLIs.

Architecture

┌─────────────────────────────────────────────────────┐
│  /jam:council Command                                │
│  - Parses topic, options, criteria                   │
│  - Dispatches to council agent                       │
├─────────────────────────────────────────────────────┤
│  Council Agent (agents/jam/council.md)               │
│  - Detects CLIs via `which`                          │
│  - Builds question scaffold (4 fixed questions)      │
│  - Pipes scaffold to each CLI in parallel            │
│  - Claude answers the same scaffold independently    │
├─────────────────────────────────────────────────────┤
│  External CLI Dispatch                               │
│  - cat scaffold.md | codex exec "..."                │
│  - cat scaffold.md | gemini "..."                    │
│  - cat scaffold.md | opencode run "..."              │
│  - cat scaffold.md | pi exec "..."                   │
├─────────────────────────────────────────────────────┤
│  Synthesis (3-stage)                                 │
│  - Stage 1: Raw responses per model                  │
│  - Stage 2: Synthesis matrix + risk convergence      │
│  - Stage 3: Verdict (consensus or fault lines)       │
└─────────────────────────────────────────────────────┘

Synthesis Framework

After gathering perspectives from different models, synthesize using:

Signal	Meaning	Action
Consensus (2+ models agree)	High confidence issue	Address immediately
Unique insight	One model caught it	Evaluate carefully
Disagreement	Genuine tradeoff	Human decides
Silence	No model flagged it	Lower priority

Persistence

Automatic (via save_transcript.py)

Council responses are persisted as transcript entries:

• Each model's response stored with persona_type: council
• Synthesis appended as entry_type: synthesis
• Retrievable via /jam:transcript and /jam:thinking

Manual (via wicked-garden:mem)

Store decisions with full attribution:

/wicked-garden:mem:store "Auth: JWT with 15min/7day expiry.
Consensus: Claude, Gemini, Codex (idempotency critical).
Unique: Gemini flagged session store scaling concern.
Dissent: none." --type decision --tags auth,multi-model-review

When to Use Multi-Model

Situation	Recommendation
Architecture decisions	Yes — high impact, catch blind spots
Security review	Yes — different models flag different risks
Important PRs	Yes — diverse review perspectives
Visual/UX design	Yes — different aesthetic sensibilities
Quick bug fix	No — overhead not worth it
Routine code	No — single AI sufficient

Fallback Behavior

If no external CLIs are detected, council refuses and suggests /jam:brainstorm (single-model, multi-persona) as an alternative. With only 1 CLI, it runs as "brainstorm with external guest" with a warning.

References

Orchestration:

• Orchestration Patterns — CLI dispatch, parallel execution, synthesis
• Context Management — Session state, cross-AI handoffs, context windows

CLI Providers:

• Codex | Copilot | Gemini | OpenCode | Pi

Quality:

• Auditability — Audit trails, compliance, decision tracking
• Examples — ADR templates, synthesis patterns, review templates