菜单
Run self-evolution cycle — score agents, detect drift, propose instruction improvements. Use at session end, weekly cadence, or monthly deep review.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Hard-learned development rules by domain — Python, FastAPI, React, Testing. Prevents the specific mistakes that caused 70+ audit findings across v0.21.0-v0.21.3.
Mandatory pre-merge quality checks for production code — security scan, packaging test, API consistency, no silent failures. Prevents the issues that caused 4 emergency releases in v0.21.x.
Pre-release validation — packaging verification, user acceptance testing, security scan, API drift detection. Prevents the recurring pattern of shipping broken wheels and missing files.
Test the library from an end-user perspective — pip install, import, run all workflows. Catches packaging, import, and API issues that internal tests miss.
Discover exact API function signatures, parameter names, and types from structural_lib. Use BEFORE wrapping or calling any API function. Prevents the #1 agent mistake: guessing parameter names (b_mm not width, fck not concrete_grade).
Run a structured innovation research cycle — discover gaps, propose novel capabilities, prototype and validate breakthrough ideas for structural engineering.
| name | agent-evolution |
| description | Run self-evolution cycle — score agents, detect drift, propose instruction improvements. Use at session end, weekly cadence, or monthly deep review. |
Run the self-evolving agent improvement cycle — from data collection through scoring, drift detection, and instruction evolution.
./run.sh evolve --status
./run.sh feedback log --agent <name>
Run these 5 steps in order:
Collect session data:
.venv/bin/python scripts/agent_session_collector.py
Score agents:
.venv/bin/python scripts/agent_scorer.py --all
Detect drift:
.venv/bin/python scripts/agent_drift_detector.py
Check compliance:
.venv/bin/python scripts/agent_compliance_checker.py
Analyze trends:
.venv/bin/python scripts/agent_trends.py --weekly
Or one command: ./run.sh evolve --review weekly
Additional steps beyond the weekly cycle:
Run weekly cycle first (steps 1-5 above)
Export paper data:
.venv/bin/python scripts/export_paper_data.py
Propose evolutions:
.venv/bin/python scripts/agent_evolve_instructions.py --propose
Review proposals:
.venv/bin/python scripts/agent_evolve_instructions.py --list
Apply approved:
.venv/bin/python scripts/agent_evolve_instructions.py --apply --confirm
Or: ./run.sh evolve --review monthly
Emergency rollback procedure:
.venv/bin/python scripts/agent_evolve_instructions.py --rollback <agent-name>
| File | Purpose |
|---|---|
logs/agent-performance/scorecard_index.json | Latest scores per agent |
logs/agent-performance/evolution-log.json | Applied evolution history |
logs/agent-performance/pending-evolutions.json | Proposed changes |
logs/agent-performance/sessions/*.json | Raw session data |
logs/agent-performance/trends/*.json | Trend analyses |
logs/agent-performance/backups/*.bak | Agent file backups |
--dry-run or --propose before --apply--rollbackThe system requires 15-20 sessions of data collection before evolution proposals become meaningful. During burn-in, focus on OBSERVE and MEASURE only — do not EVOLVE.