skill-evolver

name	skill-evolver
description	evolution, evolver, evolve, skills, run, check, status, skill 進化, 跑進化, 進化報告
version	0.1.0
tools	Bash, Read
argument-hint	run [--max-skills N --max-rounds N] | status | ledger | dry-run
disable-model-invocation	true
io	{"input":[{"mime":"text/plain","description":"Subcommand and optional parameters"}],"output":[{"mime":"text/markdown","description":"Evolution report or status output"}]}

Skill Evolver

AutoResearch-inspired overnight skill evolution engine. Frozen eval metrics that agents cannot modify + keep/discard loop with git-backed experiment tracking.

Subcommands

run — Execute evolution loop

~/.local/bin/python3 ~/workshop/stations/skill-evolver/cli/skill_evolver.py run \
  --max-skills 5 --max-rounds 10

Options:

• --max-skills N — Max skills per run (default: 5)
• --max-rounds N — Max mutation rounds per skill (default: 10)
• --json — Output JSON results
• --config PATH — Custom config JSON

status — Show latest report

~/.local/bin/python3 ~/workshop/stations/skill-evolver/cli/skill_evolver.py status

ledger — View experiment history

~/.local/bin/python3 ~/workshop/stations/skill-evolver/cli/skill_evolver.py ledger --last 20

dry-run — Preview without executing

~/.local/bin/python3 ~/workshop/stations/skill-evolver/cli/skill_evolver.py dry-run

SDK Usage

from sdk_client.skill_evolver import SkillEvolverClient

client = SkillEvolverClient()
targets = client.dry_run()
results = client.run(max_skills=1, max_rounds=2)
report = client.status()

Architecture

• Frozen evals: stations/skill-evolver/frozen_evals/ (quality_judge + scoring_rubric)
• Golden cases: stations/skill-evolver/golden_cases/{skill}/cases.json
• Evolution directions: stations/skill-evolver/evolution.md (human-authored)
• Ledger: ~/.claude/data/skill-evolver/evolution_ledger.json
• Reports: stations/skill-evolver/reports/evolution-YYYY-MM-DD.md
• Cronicle: Nightly at 02:00 (event: ws-skill-evolution)

Adding Golden Cases for a Skill

Create stations/skill-evolver/golden_cases/{skill-name}/cases.json:

{
  "skill": "skill-name",
  "cases": [
    {
      "id": "case-01",
      "input": "test input for the skill",
      "expected_traits": ["trait 1", "trait 2"],
      "weight": 1.0
    }
  ]
}

Mutation Themes

Each round applies ONE theme: simplify | clarify | restructure | example_tune | constraint

Budget

Default: 5 skills x 10 rounds = 50 eval calls/night. LLM-as-Judge uses Haiku.