Auto-improve loop refinement and enhancement

[akashgit]

Problem

The continuous improvement loop (--loop) works but has several rough edges that limit its effectiveness over long runs:

1. Stuck detection is too coarse — only triggers after 3 consecutive reverts in the same FEEC category. Three reverts across different categories doesn't trigger it, even though you're clearly stuck
2. FEEC heuristic doesn't learn — uses static keyword matching to categorize hypotheses. Doesn't incorporate outcome history (e.g., "FIX hypotheses succeed 80% of the time here, EXPLORE only 40%")
3. Backlog is passive — items are plain text with no scoring, prioritization, or dependency tracking. Loop mode ignores the backlog entirely
4. ACE playbooks are injected once at agent startup — no mid-cycle adaptation. If Builder violates playbook rules 3 times in a row, CEO doesn't know until the cycle ends
5. Cycle state expires after 24 hours — legitimate long cycles get abandoned. No cost tracking across cycles
6. No loop-level telemetry — can't see keep rate, cost efficiency, or hypothesis completion rate across cycles

What's needed

Smarter stuck detection

• Trigger on N consecutive reverts regardless of category (the current same-category requirement is too narrow)
• Add velocity metric: if revert rate > 60% over last 10 experiments, pause and suggest course correction
• Make stuck signals visible in CEO prompt so it can adapt

FEEC learning

• Track which FEEC categories actually succeed for this project
• Weight recent experiments more heavily (temporal decay)
• Feed success rates back into strategy: "FIX has 75% keep rate, EXPLORE has 40%" → soft bias toward what works
• Don't disable categories, just reorder by learned effectiveness

Active backlog management

• Store metadata with backlog items: category, estimated scope, tried count, dependencies
• In loop mode, process 1-2 backlog items per cycle alongside auto-detected work
• Graduate completed items to archive. Mark items tried 3x and reverted 3x as "needs research first"

Tighter feedback loops

• After each agent completes, check alignment with playbook bullets in real-time
• Mid-cycle cost tracking: if cost_spent > 70% of budget, skip Explore, prioritize FIX only
• Replace 24h cycle TTL with task-based TTL (cycle lives until all planned hypotheses have verdicts)

Loop telemetry

• Per-cycle metrics: hypotheses planned/completed, keep rate, avg score delta, cost, duration
• Rolling health indicators on the dashboard
• Cost efficiency metric: score_delta per dollar spent