self-evolve
Skill9 starsChanged 8 months agoApache-2.0View on GitHub
触发词: /evolve retrospect, /evolve drill, /evolve benchmark 负责人: Wei (编排) + Alex (质量评估) 设计文档: evolution/self-evolution-design.md v2.0 白名单: evolution/changeset-whitelist.yaml 自演化引擎是 AI 科研军团的自我改进机制。每次 pipeline 完成后自动触发复盘,识别薄弱环节,生成改进提案。 核心原则: 自动进化的边界 = 可程序化验证的边界。 触发: pipeline 完成后自动触发(E2 约束),或手动调用 前置条件: outputs/ 目录下存在各模块的 trace.json 对照以下 T1 指标检查 trace 数据: 对每个诊断发现执行分类: 对 T1/T2 发现: - 读取 evolution/evidence-log.json - 检查是否有跨项目复现(同类问题在 ≥2 个独立项目中出现) - 若不满足归因门槛 → 仅记录证据,不触发进化 - 若满足 → 标记为 evolution_ready 输出文件: evolution/cycles/cycle-NNN/ 前置条件: 指定 finding 已标记 evolution_ready 变更域检查: 修复目标文件必须在 auto_safe 域内 从 evolution/benchmarks/frozen-core/ 加载对应基准数据集: - ref_accuracy → ref_golden_set.json - stat_method…
6 other projects use a version of this.
Adopted by
Most recently taken up by
- Echo-Scion/orioncortextheir own version, first seen Nov 8, 2026
- pillumina/ascend-sleuththeir own version, first seen Nov 2, 2026
- Echo-Scion/portable-brainvibingtheir own version, first seen Oct 15, 2026
- DaizeDong/self-evolvetheir own version, first seen Oct 9, 2026
- grdwng/claude-code-metatheir own version, first seen Oct 9, 2026
- avibebuilder/claude-primetheir own version, first seen May 10, 2026
Other versions6 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 6
- avibebuilder/claude-prime★ 119same name
- DaizeDong/self-evolve★ 26same name
- pillumina/ascend-sleuth★ 5same name
- Echo-Scion/orioncortex★ 2same name
- Echo-Scion/portable-brainvibing★ 2same name
- grdwng/claude-code-meta★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository29 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- academic-figure-engineSkill
- academic-figureSkill
- ai-radarSkill
- data-forensicsSkill
- data-profilerSkill
- deliverySkill
- diagram-generatorSkill
- journal-toolkitSkill
- manuscript-drafterSkill
- manuscript-draftSkill
- model-testerSkill
- paper-pipelineSkill
- paper-revisionSkill
- quality-gateSkill
- quality-reviewSkill
- reference-managerSkill
- ref-managerSkill
- research-designSkill
- research-frontierSkill
- research-strategySkill
- review-engineSkill
- review-thinkerSkill
- start-armySkill
- stat-analysisSkill
- statistical-analysisSkill
- submission-toolkitSkill
- submit-packageSkill
The file
Files
Self-Evolve — 自演化引擎
触发词:
/evolve retrospect,/evolve drill,/evolve benchmark负责人: Wei (编排) + Alex (质量评估) 设计文档:evolution/self-evolution-design.mdv2.0 白名单:evolution/changeset-whitelist.yaml
概述
自演化引擎是 AI 科研军团的自我改进机制。每次 pipeline 完成后自动触发复盘,识别薄弱环节,生成改进提案。
核心原则: 自动进化的边界 = 可程序化验证的边界。
子命令
/evolve retrospect — 自动复盘
触发: pipeline 完成后自动触发(E2 约束),或手动调用 前置条件: outputs/ 目录下存在各模块的 trace.json
执行流程:
Step 1: 收集 Trace
扫描 outputs/*/trace.json
汇总为 pipeline-trace-summary.json
若缺少必要 trace → 警告并记录缺失模块Step 2: 自动诊断
对照以下 T1 指标检查 trace 数据:
| 指标 | 来源 | 阈值 | 计算方式 |
| 引用验证通过率 | ref-manager/trace.json → metrics.ref_verified / metrics.ref_total | ≥0.90 | 精确除法 |
| 统计假设满足 | stat-analysis/trace.json → metrics.assumption_violations | = 0 | 精确计数 |
| 图表对账一致性 | academic-figure/trace.json → metrics.reconciliation_errors | = [] | 列表为空 |
| 格式合规性 | submit-package/trace.json → metrics.format_checks_passed / format_checks_total | = 1.0 | 精确除法 |
Step 3: 分层分类
对每个诊断发现执行分类:
for each finding:
if finding.has_ground_truth and finding.metric_type == "deterministic":
tier = "T1"
elif finding.has_rubric and finding.metric_type == "rule_based":
tier = "T2"
else:
tier = "T3"Step 4: 证据关联
对 T1/T2 发现:
- 读取
evolution/evidence-log.json - 检查是否有跨项目复现(同类问题在 ≥2 个独立项目中出现)
- 若不满足归因门槛 → 仅记录证据,不触发进化
- 若满足 → 标记为
evolution_ready
Step 5: 生成输出
输出文件: evolution/cycles/cycle-NNN/
- retrospection.md: 复盘报告
# Cycle NNN Retrospection Date: YYYY-MM-DD Project: project_id ## T1 发现(可验证) - [指标]: 当前值 vs 阈值, 证据数量, evolution_ready: yes/no ## T2 发现(可评估) - [指标]: 观察内容, 建议, 证据数量 ## T3 观察(主观判断) - [维度]: 观察内容 - classified_findings.json: 结构化分类结果
{ "cycle": "NNN", "project_id": "xxx", "findings": [ { "id": "F001", "tier": "T1", "metric": "ref_accuracy", "current_value": 0.85, "threshold": 0.90, "evidence_count": 2, "evolution_ready": true, "change_domain": "auto_safe" } ] } - 更新 evidence-log.json: 追加本次发现
- 更新 shadow-scores.csv: 追加本次各指标分数
- 追加 T3 到 observations.md: 记录主观观察
/evolve drill [finding_id] — 针对性训练
前置条件: 指定 finding 已标记 evolution_ready
变更域检查: 修复目标文件必须在 auto_safe 域内
执行流程:
Step 1: 加载 Drill 数据
从 evolution/benchmarks/frozen-core/ 加载对应基准数据集:
- ref_accuracy →
ref_golden_set.json - stat_method →
stat_scenarios.json - figure_recon → figure_recon_cases/
- data_quality → dirty_data_cases/
Step 2: 执行 Drill
只跑目标 skill(不跑全 pipeline),用基准数据:
input: frozen benchmark test cases
execute: target skill only
output: drill_results.json with per-case pass/failStep 3: 分析错误模式
对比: expected vs actual for each test case
归类: 错误模式分布(格式/匹配/逻辑/外部API)
定位: 根因在 skill 的哪个步骤Step 4: 生成修复提案
读取: 目标 skill 的 SKILL.md
定位: 需要修改的具体段落
生成: proposal.md 含:
- 根因分析
- 修改建议(精确到 SKILL.md 的行/段落)
- 预期效果注意: drill 阶段只生成提案,不自动修改文件。
Step 5: 输出
evolution/cycles/cycle-NNN/drill_results.jsonevolution/cycles/cycle-NNN/proposal.md
/evolve benchmark — 回归测试
触发: 进化提案被采纳(文件已修改)后手动或自动触发
执行流程:
Step 1: 加载 Frozen Core Suite
读取: evolution/benchmarks/frozen-core/ 下所有测试集Step 2: 执行测试
对每个 benchmark:
- 确定性 benchmark(ref_golden, figure_recon)→ 单次执行,硬门槛
- 非确定性 benchmark(stat_scenarios)→ 3 次执行取中位数
Step 3: 结果判定
for each benchmark:
if deterministic:
pass = (score >= threshold)
else:
scores = run 3 times
pass = (median(scores) >= threshold)Step 4: 输出
evolution/cycles/cycle-NNN/benchmark_results.json:{ "cycle": "NNN", "timestamp": "...", "results": [ {"benchmark": "B1_ref_golden", "score": 0.95, "threshold": 0.90, "pass": true}, {"benchmark": "B2_stat_scenarios", "scores": [1.0, 1.0, 1.0], "median": 1.0, "threshold": 1.0, "pass": true} ], "overall_pass": true }- 更新
evolution/scores.csv - 如果 overall_pass = false → 生成
failure.md说明哪个 benchmark 失败
Trace 输出标准
每个 pipeline 模块在完成时必须输出 outputs/<module>/trace.json。
通用 Schema
{
"module": "<module-name>",
"version": "1.0",
"timestamp": "<ISO 8601>",
"project_id": "<project-id>",
"duration_seconds": 0,
"metrics": {},
"errors": [],
"decisions": []
}模块特定 metrics
ref-manager: ref_total, ref_verified, ref_failed, ref_failed_reasons stat-analysis: methods_used, methods_considered, assumption_violations, analysis_chain_depth, analysis_chain quality-review: total_score, breakdown (8层), review_rounds, block_issues academic-figure: figures_total, reconciliation_passed, reconciliation_errors manuscript-draft: narrative_thread_versions, word_count, sections_complete submit-package: format_checks_passed, format_checks_total, files_generated
详细 schema 见 evolution/self-evolution-design.md §3.2。
安全约束
- 不得修改 Immutable 域: O1-O18, M1-M5, R1-R3, PROMPT.md, TEAM.md
- 不得自动合入: MVP 阶段所有提案需人类 approve (E10)
- 不得使用客户原始数据做 benchmark: frozen core 必须脱敏
- 跨项目归因: 单次项目失败不触发进化 (E5)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

