Skip to content
cccskills
FreePublished on GitHub

self-evolve

Improve an existing skill or repository through isolated proposals, reproducible evaluation and evidence-based acceptance. Use for self-evolve or skill improvement.

The English summary is being prepared. Showing the original description.

DaizeDong/self-evolve
See how to install

Files included (111)

  • SKILL.md3.3 KB
  • .claude-plugin/plugin.json837 B
  • .dataclass.json1.7 KB
  • .githooks/pre-commit628 B
  • .githooks/pre-push337 B
  • .gitignore441 B
  • .pii-allow599 B
  • CHANGELOG.md6.9 KB
  • commands/self-evolve-resume.md862 B
  • commands/self-evolve-status.md750 B
  • commands/self-evolve.md897 B
  • config.contract.json645 B
  • DATA.md11.6 KB
  • docs/diagrams/render.py2.4 KB
  • docs/diagrams/workflow-cn.dot1.8 KB
  • docs/diagrams/workflow-cn.png111.3 KB
  • docs/diagrams/workflow-en.dot1.7 KB
  • docs/diagrams/workflow-en.png94.6 KB
  • examples/btarget/report.json14.7 KB
  • LICENSE1.0 KB
  • PHILOSOPHY.md6.8 KB
  • README_CN.md6.6 KB
  • README.md7.2 KB
  • reference/acceptor_math.md8.6 KB
  • reference/evaluation.md7.3 KB
  • reference/maintenance.md3.8 KB
  • reference/runtime.md9.8 KB
  • requirements.txt277 B
  • retention.json.example135 B
  • ROADMAP.md9.4 KB
  • schemas/history-catalog.schema.json3.5 KB
  • schemas/storage-manifest.schema.json2.5 KB
  • schemas/storage-retention.schema.json1.2 KB
  • scripts/setup_btarget_repo.py5.9 KB
  • scripts/validate_btarget.py5.1 KB
  • spikes/__init__.py0 B
  • spikes/confseq_spike.py2.6 KB
  • storage.contract.json44.5 KB
  • tools/__init__.py0 B
  • tools/llm_fixtures.py5.8 KB
  • tools/make_fixtures.py69.2 KB
  • tools/sie_cli.py265 B
  • tools/sie/__init__.py38 B
  • tools/sie/acceptor.py16.1 KB
  • tools/sie/agents.py3.5 KB
  • tools/sie/anchors.py9.7 KB
  • tools/sie/archive.py11.8 KB
  • tools/sie/backends/__init__.py0 B
  • tools/sie/backends/builtin.py570 B
  • tools/sie/backends/llm.py17.4 KB
  • tools/sie/business_tree.py9.8 KB
  • tools/sie/calibrate.py38.9 KB
  • tools/sie/calibration_defects/cl-oss-lane-key-mismatch.json2.1 KB
  • tools/sie/calibration_defects/cl-reward-floor-flattens-low-end.json2.8 KB
  • tools/sie/calibration_defects/cl-track-weight-double-charged.json4.0 KB
  • tools/sie/calibration_defects/cl-unseen-arm-inherits-best.json2.7 KB
  • tools/sie/calibration_defects/cp-report-failure-collapses-into-cannot-check.json3.4 KB
  • tools/sie/calibration_defects/cp-unparseable-digest-name-read-as-a-clean-day.json3.1 KB
  • tools/sie/calibration_defects/cp-unscanned-archive-publishes-zero-missing-days.json6.7 KB
  • tools/sie/calibration_defects/lb-demand-bar-measured-against-the-shipped-floor.json2.9 KB
  • tools/sie/calibration_defects/lb-origin-labels-lose-case-folding.json2.5 KB
  • tools/sie/calibration_defects/lb-quote-capped-at-the-push-limit.json2.4 KB
  • tools/sie/calibration_defects/lb-undated-age-scores-as-brand-new.json2.0 KB
  • tools/sie/calibration_defects/sc-age-floored-at-collection-cadence.json3.1 KB
  • tools/sie/calibration_defects/sc-crowdedness-damper-outside-the-weights.json3.8 KB
  • tools/sie/calibration_defects/sc-final-map-drops-lifecycle-stage.json3.3 KB
  • tools/sie/calibration_defects/sc-undated-evidence-counts-toward-the-red-line.json3.1 KB
  • tools/sie/calibration_defects/yl-lane-alias-not-folded.json3.2 KB
  • tools/sie/calibration_defects/yl-previral-key-list-drift.json4.0 KB
  • tools/sie/calibration_defects/yl-propose-add-silent-cap.json3.2 KB
  • tools/sie/calibration_defects/yl-unparsed-ts-reads-as-in-window.json3.8 KB
  • tools/sie/check_reflection.py3.0 KB
  • tools/sie/cli.py10.5 KB
  • tools/sie/edgar_cache.py1.1 KB
  • tools/sie/evaluate.py18.4 KB
  • tools/sie/events.py3.1 KB
  • tools/sie/gate_human.py2.9 KB
  • tools/sie/immutable.py6.9 KB
  • tools/sie/judge_claude.py392 B
  • tools/sie/judge_codex.py389 B
  • tools/sie/judges.py9.9 KB
  • tools/sie/llm_adapter.py6.4 KB
  • tools/sie/llm_agent_child.py962 B
  • tools/sie/patch.py43.8 KB
  • tools/sie/probes/__init__.py0 B
  • tools/sie/probes/exec_probe.py6.5 KB
  • tools/sie/probes/fact_probe.py3.7 KB
  • tools/sie/profile.py10.0 KB
  • tools/sie/propose.py1.6 KB
  • tools/sie/proxy.py10.9 KB
  • tools/sie/reflect.py5.0 KB
  • tools/sie/runtime_data.py14.7 KB
  • tools/sie/sandbox.py8.4 KB
  • tools/sie/selfboot.py4.1 KB
  • tools/sie/selfdeception.py7.6 KB
  • tools/sie/stage_records.py3.0 KB
  • tools/sie/state.py840 B
  • tools/sie/statemachine.py72.9 KB
  • tools/sie/storage.py5.7 KB
  • tools/sie/supervisor.py8.0 KB
  • tools/sie/verifiable.py18.3 KB
  • tools/storage_retention.py8.4 KB
  • workflows/_agent_launch.js434 B
  • workflows/_claude_launch.js434 B
  • workflows/agent.js434 B
  • workflows/claude-judge.js434 B
  • workflows/claude-propose-artifact.js434 B
  • workflows/claude-propose.js434 B
  • workflows/codex-judge.js434 B
  • workflows/reflect-fanout.js434 B
  • workflows/review-fanout.js434 B

SKILL.md (original)

Review the instructions the agent will receive before you install.

self-evolve

在 Git worktree 中提出修改、运行评测，再按证据决定是否采纳。 先明确目标和可执行判据；通过检查只证明所测范围。

开始前

  1. 确认目标仓库、基线 revision、允许修改的范围和期望改善的行为。 记录文档影响；无需修改的文档也说明原因。设计理由见 PHILOSOPHY.md。
  2. 确定评测信号：A 为测试结果，B 为独立事实核验，C 为已提供的主观证据。 主循环尚无 C 回归与一致性测量生产者，也不执行 A+B 组合。
  3. 按 DATA.md 配置 PRIVATE 伴生仓，保留真实运行证据。
  4. 运行 python <skill-path>/tools/sie_cli.py doctor --target <target>， 阅读配置和能力缺口。doctor 不调用模型，退出零不能单独证明就绪。

工作流

步骤要做的事
PROFILE实现前冻结评测策略、目标契约与留出集；续跑沿用原契约
REFLECT从已有记录形成诊断，核对引用；结构检查不等于事实核验
PROPOSE提出允许范围内的修改，保留实际 provider 和失败原因
PATCH在候选工作树应用提案，执行路径、AST 与 IMMUTABLE 检查
EVALUATE用相同口径测量父代和候选，保留任务身份及不可用原因
ACCEPT / REVIEW由代码处理证据与门控，归档采纳版本或记录拒绝、人审原因

agent 使用已安装的 llmcall.call(prompt, mode="agent")，文本 judge 使用默认模式。 继承其路由、模型、超时和回退策略，不另设 provider 链。 只有实际返回的已知不同家族才能支持独立评审结论。

运行与恢复

python <skill-path>/tools/sie_cli.py init --target <target>
python <skill-path>/tools/sie_cli.py run --target <target> --run-id <id> --base-ref HEAD
python <skill-path>/tools/sie_cli.py status --target <target> --run-id <id>

默认 builtin 将已提供的修复转成提案，不会自主发现修复。 模型提案使用 --live，JSON 产物使用 --proposer llm-artifact。 续跑复用 run ID；--max-rounds 是本次调用的额外轮数。 参数、自举和恢复语义见 runtime。

接受与交付

  • 沙箱与私有目录不等于操作系统权限隔离；真值不进入提案 prompt。
  • 缺少基线、回归记录、核验结果或独立 provider 时，明确报告缺口。
  • 没有回归或测试全绿不单独构成改进；按 evaluation 解释证据。
  • 人审队列记录待审动作；发布、合并和对外发送仍按已有授权单独执行。
  • 评审前完成实现、文档和已获授权的版本更新，再冻结准确的候选快照。 这不改变实现前的评测冻结；评审后改动须重新检查受影响部分。流程见 maintenance。
  • 交付候选差异、源码版本、实际测量、采纳原因、未覆盖范围和运行状态。 统计解释须满足 acceptor_math 的假设。

按需加载上述参考，不读取历史研究资料来代替当前实现和测量。

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

Report a problem with this skill