Multi-agent reflection (diverse critic perspectives)

[CasterlyGit]

What this adds

Multi-agent reflection — the L6 self-critique loop (#23) uses one Sonnet pass which has "cognitive entrenchment" (confirms its own initial framing). Multi-agent reflection runs N agents with diverse system prompts and synthesizes their independent critiques.

Based on Dec 2025 research: single-agent reflection plateaus, multi-agent improves continuously.

Mechanism

1. After L6 self-critique runs (Cross-session pattern memory (time / cwd / sequence patterns) #23), spawn 3 parallel agents with distinct lenses:
   • Skeptic: "What did shed do that was probably wrong?"
   • Optimizer: "What did shed do well that should be amplified?"
   • Pattern-spotter: "What pattern in this session has shed missed entirely?"
2. Run them in parallel via Anthropic batch API (cheap)
3. A 4th synthesis agent merges the 3 critiques into actionable proposals
4. Output goes to same ~/.shed/self-lessons/ as L6
5. Per-agent confidence weighting: if skeptic and pattern-spotter agree, higher confidence

Acceptance

• Diversity metric: critiques from different agents have <0.7 cosine similarity (proves they're not collapsing to same view)
• Self-tuning: track which agent's critiques get accepted most, adjust weights
• Cost: ~$5/month additional (batch API + Sonnet pricing)

Opt-in

• Default off until v1.0 because cost and complexity. Enable via corrections.multi_agent_reflection: true