evolve

Evolution pass (prompt-space gradient descent)

You are updating this system's weights: the personalized rules injected into every prompt. Be conservative — a wrong learned rule costs the user on every message.

Procedure

1. Read ~/.claude/command-autopilot/events.jsonl. Also read learned.json if present (schema below).
2. Cluster events by task-type × command/skill. Look for consistent patterns, for example:
   • a command suggested repeatedly and consistently dismissed → candidate negative rule
   • a skill invoked on the same kind of task again and again → candidate positive rule ("invoke X early for Y-type tasks")
   • a habit self-used regularly → mastered, teaching for it should stop
3. Apply the discipline:
   • Promote only patterns with ≥3 consistent observations and no contradicting evidence → status: "in_force".
   • Demote existing in_force rules contradicted by new evidence (decrement evidence; at 0, set status: "candidate").
   • Decay: rules not reconfirmed for ~60 days → delete.
   • Max 5 in_force rules; each text ≤ 25 tokens, English, imperative, generic phrasing ("research-type tasks: offer /fork early").
   • NEVER write rules that override the safety net (/rewind) or the one-suggestion contract.
4. Write learned.json:

{
  "updated": "<ISO>",
  "rules": [
    {"text": "...", "evidence": 4, "first": "<ISO>", "last": "<ISO>", "status": "in_force"}
  ]
}

1. Archive processed events — rotate FIRST to avoid racing concurrent appends: rename events.jsonl to events-archive-<timestamp>.jsonl, and only then read the renamed file for the distillation. New events land in a fresh events.jsonl untouched. (If you already read before rotating, rotate anyway and accept the tiny overlap.)
2. Report to the user in their language: what was learned (each rule + its evidence), what was demoted or deleted, and one line on what will change. If nothing met the bar, say so plainly — no fabricated learnings.