control_loop — the router that learns

The final systemic loop: every routing decision + outcome is recorded, and the effort→tier thresholds (read live by [[auto_router]]) are nudged from the data — keep more work local when local is reliable, escalate sooner when it isn't.

python -m skills.control_loop.cli analyze            # local %, savings, escalation/success
python -m skills.control_loop.cli adapt              # proposed threshold change + why
python -m skills.control_loop.cli adapt --apply      # write it; the router uses it next time
python -m skills.control_loop.cli reset              # back to defaults

How the loop closes

1. Measure — auto_router.run records each call to ~/.botte/control-ledger.jsonl (effort, tier, local/cloud, tokens saved, escalated, success).
2. Analyze — aggregate: local %, escalation rate, success rate, tokens saved.
3. Adapt — conservative rule, needs ≥10 samples, small steps, clamped:
   • local reliable (success ≥85%, escalation <15%) → raise the LOCAL boundary (keep more work local → more savings).
   • local insufficient (escalation >30%) → lower it (escalate borderline sooner).
4. Apply — write ~/.botte/routing-thresholds.json; auto_router.effort reads it live, so the next decisions reflect what actually worked.

This is the Karpathy/second-brain pattern applied to routing: the system gets smarter from its own outcomes. Exposed via [[llm_mcp]] as routing_stats. Related: [[auto_router]], [[conductor]], [[metrics]].