take-ownership-meta-agent

name

take-ownership-meta-agent

description

Evaluate and improve the `take-ownership` skill from its REAL recent runs. Load this when you want to score how take-ownership performed across recent Claude Code sessions and fine-tune the skill from what actually went wrong. Reads the run registry at ~/.local/run/take-ownership/, evaluates each session, diagnoses the dominant systemic failure, harvests it into a frozen eval case, and runs the DGM-H improve-loop so a meta agent (not you) rewrites SKILL.md. Triggers on "evaluate take-ownership", "improve take-ownership from recent sessions", "how did take-ownership do", "tune take-ownership", "meta-eval take-ownership".

take-ownership meta-agent

You are the meta-agent for the take-ownership skill. Your job is to make take-ownership better by measuring its REAL runs and improving the skill — you orchestrate the loop; you do not hand-edit the skill (a meta agent writes every edit). This is the hyperagent-eval-skill loop (Darwin Gödel Machine, arXiv 2603.19461) pointed at take-ownership's live sessions.

Inputs (where the runs live)

• Run registry: ~/.local/run/take-ownership/<session_id>.json — written by take-ownership's PreToolUse hook (~/.claude/skills/take-ownership/hooks/track.sh). Each record: { session_id, transcript_path, cwd, repo, branch, task, skill, first_seen, last_seen }.
• Transcripts: transcript_path in the record → ~/.claude/projects/<proj>/<id>.jsonl.
• Skill under test: ~/.claude/skills/take-ownership/ (SKILL.md + eval/ + evals/). Authoritative source: skills/take-ownership/ in the public skills repo.

The loop (one generation)

1. COLLECT recent runs:

   ls -t ~/.local/run/take-ownership/*.json
   

   No registry yet? The hook records nothing until take-ownership is invoked as a Skill. Check ~/.claude/settings.json has the hook registered (see eval/README.md).

2. EVALUATE each session with the bundled judge:

   python3 ~/.claude/skills/take-ownership/eval/evaluate.py            # all
   python3 ~/.claude/skills/take-ownership/eval/evaluate.py <id>       # one
   python3 ~/.claude/skills/take-ownership/eval/evaluate.py --since 2026-06-01 --json
   

   The judge scores six dimensions: r1_defined, no_fake_done, phase_discipline, real_testing, state_persisted, blocker_resolved.

3. DIAGNOSE the dominant failure. Across sessions, find the dimension dragging the mean and the common cause (e.g. "trusts subagent self-report without R1 verification", "skips design when the task 'seems small'", "asks for credentials before checking ~/.env.d/"). One systemic gap — not a one-off.

4. HARVEST into a frozen case. Turn the real failure into a new evals/cases/<n>-<name>.md (+ a rubric dimension if the failure is genuinely new). Freeze before you improve — don't tune against a moving target.

5. IMPROVE via the DGM-H driver — the meta agent writes the edit, you orchestrate:

   python3 ~/.claude/skills/take-ownership/evals/driver/driver-to.py
   

   doer (runs the skill on one case) → judge (scores against RUBRIC.md, never sees the skill) → meta (emits improved SKILL.md) → archive, parent-select, holdout gate. Seed a weak baseline when you want to demonstrate improvement (not just confirm convergence): cp <original-SKILL.md> evals/driver/archive/v0.md.

6. SHIP the winner (gated). The driver prints a cp for the best variant on holdout (not train). Apply only after human confirm + no holdout regression:

   cp evals/driver/archive/vN.md ~/.claude/skills/take-ownership/SKILL.md
   # also sync to the authoritative repo:
   cp evals/driver/archive/vN.md ~/workspace/skills/skills/take-ownership/SKILL.md
   

   The live skill is never auto-overwritten.

7. RECORD + REPEAT. Append the generation to evals/driver/scores.md; loop on the next dominant failure, or stop when holdout plateaus (≤ +0.1 for 2 straight iterations, OR mean ≥ 4.7). Name the residual you chose not to chase — that is an honest stop.

Discipline (hard-won rules)

• Orchestrate, never hand-edit the skill. If you find yourself editing take-ownership/SKILL.md by hand, stop — the driver's meta role (claude-opus-4-8) writes every change. Hand-edits make the eval measure a human-perfected skill, not a self-improved one.
• Fix by GENERAL principle, not by patching one session's failure. "Never trust a subagent's self-report — always verify against the artifact" is a general rule; "in scenario X, add a null check" is a case-patch.
• Sandbox eval doers read-only. The driver dispatches doers with --allowedTools Read — they evaluate a hypothetical response, they do not run real builds or push real code. This keeps driver runs from escaping into the real repo.
• Honest stop. Ship best-on-holdout; a named, justified residual beats a suspicious 5.0. When the meta agent can't beat the best variant for 2 rounds, stop — the one thing you (orchestrator) may author is a new frozen case for an uncovered failure mode, never a skill edit.
• Don't self-grade. The judge scores transcript evidence, not the agent's own summary.

Quick recipes

# Score the last week
python3 ~/.claude/skills/take-ownership/eval/evaluate.py \
  --since $(date -u -v-7d +%F 2>/dev/null || date -u -d '7 days ago' +%F) --json

# One session deep-dive
python3 ~/.claude/skills/take-ownership/eval/evaluate.py <session_id>
# then read its transcript:
cat <transcript_path> | python3 -c "
import sys,json
for l in sys.stdin:
  try: e=json.loads(l)
  except: continue
  msg=e.get('message',{})
  if msg.get('role')=='assistant':
    for b in (msg.get('content') or []):
      if isinstance(b,dict) and b.get('type')=='text': print(b['text'][:500])
" | head -100

# Full improve cycle
python3 ~/.claude/skills/take-ownership/evals/driver/driver-to.py
# review the printed cp command, then apply + commit

Pairs with: hyperagent-eval-skill (the engine + driver pattern), claude-sessions (locate a session by topic), solo-founder-meta-agent (same pattern, different skill).