본문으로 건너뛰기
cccskills
무료GitHub에 공개

skill-evolution

Use when running, reviewing or changing coco's self-evolution cycle: the 30-day loop that observes how skills are actually used, proposes evidence-backed edits to them as one branch and one pull request, and lands nothing without the owner's merge. Also use when asked why a skill changed, what the loop would propose next, or how to reject a proposal so it is not raised again.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

설치 방법 보기

포함된 파일(8)

  • SKILL.md4.8 KB
  • references/lanes.json3.2 KB
  • references/proposal-schema.json4.4 KB
  • references/signal-sources.md2.8 KB
  • scripts/fixtures/ledger/chain.jsonl2.4 KB
  • scripts/fixtures/ledger/malformed.jsonl684 B
  • scripts/ledger.py14.8 KB
  • state/README.md1.7 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

Skill Evolution

A cycle that improves this repository's own skills from observed use, and proposes every change for a human to merge. It never writes to main and never merges.

Status: the scripts this file documents, scripts/evolve.py and the steps it runs, are not on main yet, so the commands below do not run today. They land with plan tasks 2 to 9 in .metagpt/plan.md: task 2 ledger.py through task 8 evolve.py, and task 9 run_fixtures.sh with the CI join.

The one rule

The loop proposes. The owner decides. There is no confidence level, budget or green gate that lets it land a change on its own. This is a design invariant, not a setting: a system that edits the repository that measures it has no natural stopping point.

What a cycle does

Five steps, and only the second one calls a model.

StepScriptModel?Output
1. Observescripts/observe.pynosignals/<cycle>.json, pinned to one commit
2. Proposescripts/propose.pyyes, one laneproposals/<id>.json and proposals/<id>.patch
3. Validatescripts/validate.pynoaccept or refuse, with a reason
4. Openscripts/evolve.pynoone branch, one pull request
5. Recordscripts/ledger.pynoa ledger row per event, and CYCLE.md

Run it:

python3 scripts/evolve.py run --window 30d            # the real cycle, needs credentials
python3 scripts/evolve.py run --window 30d --dry-run   # stops before opening a branch
python3 scripts/evolve.py run --window 30d --lane stub # no network at all
python3 scripts/evolve.py status                       # last cycle, open proposals

What the guard refuses

validate.py is the only thing standing between a model's output and a branch. It refuses, with the reason named, when any of these hold:

  • Out of scope. v1 may change a skill's description frontmatter, its references/ and its SKILL.md prose, and may add or remove example blocks. A persona file, a command file, a count file, a workflow or this skill's own scripts are all out of bounds.
  • An unresolvable citation. Every proposal cites its evidence by file and line, or by receipt id, and the citation must resolve at the pinned commit.
  • Over a cap. Proposals per cycle and diff size per proposal are both capped in references/lanes.json. A cycle that hits the cap states the remainder rather than truncating silently.
  • Already rejected. A diff whose hash a ledger row rejected inside the two cycle window is not proposed again.
  • A count that moved alone. If a proposal changes a published number, the coupled files must move with it, checked by running the repository's own checkers rather than a private copy of their rules. See docs/rules.md R7.

Exit codes

Three valued, like every gate in this repository: 0 pass, 1 blocked by a guard, 2 unrunnable or unmeasured. An empty cycle is exit 0 with nothing to propose, which is a valid result and not a failure.

Cost

Every model call goes through scripts/lane.py, which reads references/lanes.json, checks the cycle budget before spending, obeys the model routing policy, and records the lane id and token count on the proposal it produced. The expensive review lane refuses unless the owner has explicitly authorised it for that cycle.

Where things live

  • references/lanes.json - the allowed lanes and their budgets.
  • references/proposal-schema.json - the proposal shape, and the only fields a proposal may carry.
  • references/signal-sources.md - what observe.py reads and what it deliberately does not.
  • state/ledger.jsonl - the committed ledger. Append only. One record per event, and the record of why something was rejected is the part that matters.
  • state/CYCLE.md - the rendered human artifact for the most recent cycle.

Testing

bash scripts/run_fixtures.sh runs every suite and prints N/N passed. Every case uses the stub lane, so the suite needs no network and no model. CI runs the same command, so a proposal pipeline that stops working fails the build rather than shipping quietly.

Changing this skill

Changing the loop's own scripts is out of scope for a loop proposal by design, which means a change here is an ordinary pull request from a human or an agent working under review. If a change alters a rule in this file, say so in the PR title and update docs/rules.md in the same commit.

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

Train and optimize AI agents using Microsoft's Agent Lightning framework with reinforcement learning. Use when setting up agent training, instrumenting agents with tracing, configuring LightningStore, implementing reward functions, or optimizing prompts with RL/APO algorithms.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

coco-research/coco7472026년 10월 18일 업데이트

Post-run self-evaluation system that scores agent output on correctness, clarity, actionability, and conciseness. Use after /team runs, skill executions, or when explicitly asked to evaluate output quality.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

coco-research/coco7472026년 10월 18일 업데이트

Create AI marketing videos for ads, promos, product launches, and brand content. Models: Veo, Seedance, Wan, FLUX for visuals, Kokoro for voiceover. Types: product demos, testimonials, explainers, social ads, brand videos. Use for: Facebook ads, YouTube ads, product launches, brand awareness. Triggers: marketing video, ad video, promo video, commercial, brand video, product video, explainer video, ad creative, video ad, facebook ad video, youtube ad, instagram ad, tiktok ad, promotional video, launch video

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

coco-research/coco7472026년 10월 18일 업데이트

Use when building AI features into a product: LLM integration, RAG pipelines, guardrails, streaming, AI UX, prompt engineering, or AI cost control. Treats prompts as code and validates every model output.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

coco-research/coco7472026년 10월 18일 업데이트

Your AI research and engineering brain trust. 59 named personas across 8 cells covering frontier labs, applied product, model architecture, reasoning/RL/agents, alignment and interpretability, theory and science of DL, multimodal and…

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

coco-research/coco7472026년 10월 18일 업데이트

Use when designing a new REST or GraphQL API, reviewing an API spec before implementation, setting team API standards, or migrating REST to GraphQL. Covers resources, HTTP semantics, pagination, error handling, and pitfalls.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

coco-research/coco7472026년 10월 18일 업데이트

coco-research의 스킬 모두 보기

이 스킬의 문제 신고