| name | self-evolve |
| description | Drive the self-evolving-agent project (~/Documents/self-evolving-agent) — run a self-improvement generation, check genome status, refresh the user's profile, or gate an external proposal. Use when the user wants to evolve/improve the agent, run a loop/work/eval generation, inspect or roll back the genome (agent-config), refresh about-andy, ingest an orch branch, or manage the daily auto-loop automation. |
self-evolve
Operational driver for ~/Documents/self-evolving-agent — a general agent that self-evolves its own prompt/procedures/skills/lessons, every change gated by an objective fitness suite and persisted as a git commit (1 commit = 1 generation). All user-facing output in Japanese (ユーザーのポリシー).
Hard rules (read first)
Figure out where the user is, then act
| User intent | Command |
|---|
| 「今どんな状態?」ゲノム版・baseline・世代ログ | mise exec -- npm run status |
| 1 タスク実行して採点だけ（進化なし・安全） | mise exec -- npm run work -- "<task>" |
| 1 世代まるごと自己進化（work→eval→reflect→evolve→guard→commit/revert） | mise exec -- npm run loop -- "<task>" [--count N] |
| 直近 run の進化プランを見るだけ（LLM コストなし） | mise exec -- npm run evolve |
| golden スイート実行 / baseline 更新 | mise exec -- npm run eval [-- --set-baseline] |
| ユーザー プロフィール（知識）を work の最新から再生成 | mise exec -- npm run refresh-profile |
orch の orch/<id> ブランチ提案を eval ゲートで採否判定（merge はしない） | mise exec -- npm run ingest -- orch/<id> [--dry-run] |
| node パスのズレ点検（FDA 要付け替え検知） | mise exec -- npm run doctor |
<task> is a concrete work item the agent performs (e.g. "この議事録を要約して workspace/x.md に出せ"). Multi-line tasks must be quoted. loop reflects on the result and proposes genome edits; the fitness gate keeps the change only if post + 3 >= baseline, else auto-reverts.- After a
loop, report: accepted or reverted, the post score vs baseline, and the new genome version (from status). Inspect diffs with git -C ~/Documents/self-evolving-agent log/show. Roll back a bad generation with git revert.
Daily auto-loop (already wired)
A LaunchAgent (com.wada811.self-evolving-agent.autoloop) runs one gated generation every day at 08:35, but only if Claude Code was used since the last run (a global SessionEnd hook touches runs/.active; scripts/auto-loop.ts gates on it). Idle days cost nothing; never more than once per activity period.
- Standing task:
goals/auto-loop-task.md if present, else a built-in kaizen default. Edit that file to retune what the autonomous generation works on.
- Did it run?
tail runs/auto-loop.history.log — 1 line per fire (skip reason or run outcome: behavior evolved vX→vY / reverted / error). Full detail in runs/autoloop.out.log / runs/autoloop.err.log. Stamps: runs/.active, runs/.last-auto-loop.
- Manage:
launchctl list | grep autoloop, launchctl bootout gui/$(id -u)/com.wada811.self-evolving-agent.autoloop to disable, re-bootstrap the plist to re-enable. Force a run now: mise exec -- node --import tsx scripts/auto-loop.ts (still respects the activity gate). See docs/auto-loop.md.
genome → Claude Code bridge (already wired)
The genome compounds inside the SDK agent; ordinary Claude Code sessions don't read agent-config/ on their own. scripts/sync-to-claude.ts mirrors it across (idempotent, marker-bounded / se- namespaced — never clobbers hand-written content):
lessons.md → managed block in ~/.claude/CLAUDE.mdabout-andy.md → memory/se-about-andy.md (+ MEMORY.md index line)agent-config/skills/<name> → ~/.claude/skills/se-skill-<name>/ (prunes removed ones)agent-config/procedures/<name> → ~/.claude/skills/se-proc-<name>/ (on-demand, avoids context bloat)
It runs automatically at the end of every auto-loop (skip or run). After any manual genome change (manual loop, ingest merge, refresh-profile), re-sync:
cd ~/Documents/self-evolving-agent && mise exec -- node --import tsx scripts/sync-to-claude.ts
See docs/claude-code-bridge.md. Check SYNC lines in runs/auto-loop.history.log.
After the work
Per ユーザーのポリシー, when you've completed a step, propose ranked next actions with the AskUserQuestion tool rather than ending flatly.
