By name: orchestra description: "Round-robin multi-LLM consultation with full-context debate. GPT proposes, Gemini refines, Grok challenges — each sees the full chain including wheat/chaff verdicts and justifications. Nothing is filtered between rounds." version: 1.0.0
The Orchestra — Round-Robin Multi-LLM Debate
Ask the orchestra. Let them debate. Take what survives.
Usage
/orchestra "task description"
Mode: recipe (auto-detected from this repository)
Domain: Granny Hudson's recipe transcription and site
How It Works
Unlike /orchestrate (linear pipeline), /orchestra runs a round-robin debate where each model sees EVERYTHING from prior rounds — proposals, verdicts, justifications, and rejections. Nothing is filtered. What one model calls chaff, the next might rescue.
cd /home/user/ken/orchestrator && python3 orchestra.py recipe "task description"
The Rounds
Round 1: GPT Proposes
- Receives: task + relevant memories
- Produces: proposals (with confidence + justification), low-hanging fruit
Round 2: Gemini Refines
- Receives: task + GPT's FULL response
- For each GPT proposal: WHEAT | CHAFF | WHEAT_WITH_REFINEMENT (with justification)
- Adds own proposals + low-hanging fruit
- May rescue ideas it thinks GPT undervalued
Round 3: Grok Challenges
- Receives: task + GPT's FULL response + Gemini's FULL response (including verdicts)
- For EVERY prior proposal: verdict + justification
- Specifically looks for dismissed ideas worth rescuing
- Adds own proposals + identifies blind spots + low-hanging fruit
Blind Spot Check
- The most contrarian model (Grok) reviews the emerging synthesis
- "What is the single most important thing we're all still missing?"
Claude Synthesizes
- Sees the full debate chain: every proposal, every verdict, every justification
- Produces attributed plan: "GPT proposed X, Gemini refined to Y, Grok challenged Z"
- Encodes key decisions to cognitive-memory
Why Full Context Matters
If Claude filters between rounds, rejected ideas die silently. With full context:
- Gemini can disagree with GPT's reasoning, not just its conclusions
- Grok can rescue an idea Gemini dismissed, explaining WHY Gemini was wrong
- The final synthesis is informed by the debate, not just the survivors
Cost Reporting
Every run reports:
- Per-round costs (model, tokens, dollars)
- Total pipeline cost
- Idea survival rate (wheat vs chaff vs rescued)
- Low-hanging fruit from all models
Typical cost: $0.03–$0.08 per full orchestra run.
Integration
- cognitive-memory — recalls relevant memories before Round 1, encodes decisions after
- consult — for quick single-model questions (use /orchestra for the full debate)
- orchestrate — the older linear pipeline (still available for simpler tasks)
Soli Deo Gloria — Let the orchestra play. Take what glorifies God.