Evolution Core — 自进化资产模型与闭环

SkillMonitoring & ops

Unified evolution asset model and self-evolution closed-loop core. Defines a three-tier asset model, Gene (strategy DNA), Capsule (validated experience snapshot), and EvolutionEvent (append-only event log), implementing a Detect→Select→Replay→Validate→Solidify→Reuse closed loop, so that when the a

Use Evolution Core — 自进化资产模型与闭环 in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add Evolution Core — 自进化资产模型与闭环 and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the Evolution Core skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Evolution Core — 自进化资产模型与闭环Start free

What this skill tells your AI

The instructions your AI receives, as published by hashgraph-online/awesome-codex-plugins in plugins/Colin4k1024/tsp/skills/evolution-core/SKILL.md and read by ahel’s review.

用途

  • 即时复用:agent 遇到错误时,先从进化存储中查找匹配的已验证经验(Capsule),尝试重放,成功则跳过 LLM 推理
  • 经验沉淀:每次成功解决问题后,自动提取信号、生成策略 Gene 和执行快照 Capsule
  • 跨 agent 共享:不同 agent/项目的验证过经验可以发布到共享网络,其他 agent 直接复用
  • 治理安全:通过 Governor 控制进化速率、爆炸半径和回归检测,防止低质量策略扩散

核心概念

Gene — 策略 DNA

Gene 是一条可复用的问题解决策略,从一次或多次成功经验中提取并泛化:

{
  "gene_id": "g-a1b2c3d4",
  "signals": [
    {
      "signal_type": "compiler_error",
      "source": "compiler",
      "pattern": "Cannot find module '(.+)'",
      "context_hints": ["typescript", "import"]
    }
  ],
  "strategy": {
    "description": "检查 tsconfig.json paths 配置和 package.json dependencies,确认模块存在后重新安装",
    "steps": [
      "检查 tsconfig.json 中 paths 和 baseUrl 配置",
      "检查 package.json 中是否包含该依赖",
      "运行 npm install 或对应包管理器安装命令",
      "若为路径别名问题,更新 tsconfig paths"
    ],
    "constraints": ["不要直接删除 node_modules 重装,先定位根因"]
  },
  "validation_rules": {
    "success_indicator": "tsc 编译通过,无 Cannot find module 错误",
    "timeout_seconds": 120
  },
  "confidence": 0.75,
  "state": "promoted",
  "tags": ["typescript", "module-resolution", "build-error"],
  "env_fingerprint": {
    "languages": ["typescript"],
    "frameworks": ["next.js"],
    "tools": ["npm"]
  },
  "stats": {
    "total_uses": 12,
    "successes": 10,
    "failures": 2,
    "last_used": "2026-04-01T10:30:00Z"
  },
  "created_at": "2026-03-15T08:00:00Z",
  "updated_at": "2026-04-01T10:30:00Z"
}

Gene 状态生命周期:

candidate → promoted → archived
    ↓           ↓
  revoked    quarantined → (re-promoted | revoked)
状态含义
candidate新提取,尚未达到晋升门槛
promoted经过多次验证,高置信度可用
quarantined从外部导入或连续失败,需本地验证
revoked多次失败或人工标记无效
archived长期未使用,自动归档

Capsule — 已验证经验快照

Capsule 是 Gene 的一次具体执行记录,包含完整的上下文和结果:

{
  "capsule_id": "c-e5f6g7h8",
  "gene_ref": "g-a1b2c3d4",
  "execution_context": {
    "project_id": "proj-xyz",
    "agent_id": "agent-session-001",
    "trigger_signal": {
      "signal_type": "compiler_error",
      "source": "compiler",
      "raw_message": "Cannot find module '@/utils/helpers'",
      "file_path": "src/pages/index.tsx"
    }
  },
  "outcome": {
    "success": true,
    "changes_summary": "更新 tsconfig.json paths 配置,添加 @/ → src/ 映射",
    "diff_hash": "sha256:abc123...",
    "files_changed": ["tsconfig.json"],
    "lines_modified": 3,
    "duration_seconds": 15
  },
  "env_fingerprint": {
    "os": "darwin",
    "node_version": "20.x",
    "typescript_version": "5.4",
    "package_manager": "npm"
  },
  "confidence": 0.85,
  "created_at": "2026-04-01T10:30:00Z"
}

EvolutionEvent — 追加式事件日志

EvolutionEvent 是不可变的事件记录,构成进化存储的唯一事实源(single source of truth):

{
  "event_id": "ev-20260401-001",
  "type": "capsule_reused",
  "timestamp": "2026-04-01T10:30:00Z",
  "agent_id": "agent-session-001",
  "payload": {
    "gene_id": "g-a1b2c3d4",
    "capsule_id": "c-e5f6g7h8",
    "success": true,
    "confidence_delta": 0.05
  }
}

事件类型:

类型触发时机
signal_detected检测到错误/问题信号
gene_created新 Gene 从成功经验中提取
gene_updatedGene 置信度、统计或状态更新
gene_promotedGene 从 candidate 晋升为 promoted
gene_revokedGene 被标记为无效
gene_quarantinedGene 进入隔离区
capsule_created新的执行快照产生
capsule_reused已有 Capsule 被重放使用
replay_attempted尝试重放但未成功
governor_blockedGovernor 阻止了某个操作
network_publishedGene/Capsule 发布到共享网络
network_fetched从共享网络获取了资产

进化闭环流程

┌──────────────────────────────────────────────────────┐
│                  Evolution Loop                       │
│                                                       │
│  ① Detect  ──→  ② Select  ──→  ③ Replay             │
│     │              │              │                   │
│     │              │              ├─ 成功 → ⑥ Reuse  │
│     │              │              │                   │
│     │              │              └─ 失败 ↓           │
│     │              │                                  │
│     │              └──→  ④ Execute (LLM)              │
│     │                       │                         │
│     │                       ↓                         │
│     │               ⑤ Validate                        │
│     │                  │                              │
│     │                  ├─ 通过 → Solidify → Gene+     │
│     │                  │              Capsule          │
│     │                  └─ 失败 → 记录失败事件          │
│     │                                                 │
│     └─── Governor 全程监控速率/半径/回归 ──────────────│
└──────────────────────────────────────────────────────┘
  1. Detect: 从工具调用结果(编译错误、测试失败、运行时异常)中提取结构化信号
  2. Select: 用信号在 Evolution Store 中查询匹配的 Gene 候选
  3. Replay: 取置信度最高的 Capsule 尝试重放其策略
  4. Execute: 若无匹配或重放失败,走正常 LLM 推理解决问题
  5. Validate: 验证解决方案的正确性(编译通过、测试通过等)
  6. Solidify: 验证通过后提取 Gene + Capsule 写入进化存储
  7. Reuse: 成功重放时记录复用事件,提升 Gene 置信度

存储布局

~/.claude/evolution/
├── events.jsonl          # 追加式事件日志(唯一事实源)
├── genes.json            # Gene 投影缓存(从 events 重建)
├── capsules.json         # Capsule 投影缓存(从 events 重建)
├── LOCK                  # 并发写入保护
├── config.json           # 本地配置(Governor 参数等)
└── shared/               # 跨 agent 共享区
    ├── published/        # 本地发布的资产
    └── quarantine/       # 外部导入的待验证资产

置信度模型

初始置信度

来源初始值
首次成功解决0.40
从 error-experience 迁移根据 success_rate 映射
从 instinct 迁移保留原有 confidence
从外部网络导入0.30(进入 quarantine)

置信度变更

成功复用: confidence = min(0.95, confidence + 0.05)
失败复用: confidence = max(0.10, confidence - 0.10)
时间衰减: confidence = confidence × e^(-0.001 × days_since_last_use)
手动标记: confidence = 指定值

状态晋升门槛

转换条件
candidate → promotedconfidence ≥ 0.60 且 successes ≥ 3
promoted → quarantined连续失败 ≥ 3 次
quarantined → promoted本地验证通过 2 次以上
* → revokedconfidence < 0.15 或手动标记
promoted → archived90 天未使用

信号匹配算法

信号匹配分三级:

  1. 精确匹配 (score=1.0): signal_type + source + pattern 完全匹配
  2. 模糊匹配 (score=0.7): signal_type + source 匹配,pattern 相似度 > 0.8
  3. 上下文匹配 (score=0.4): 仅 tags 和 env_fingerprint 部分重叠

最终匹配分 = match_score × gene_confidence × env_similarity

命令接入

命令用途
/replay <error>手动查询匹配的 Gene 并尝试重放
/evolution-status查看进化存储状态面板
/evolve将 instinct 聚类为 Gene(升级版)
/instinct-export导出进化资产(升级版)
/instinct-import导入进化资产(升级版)

与现有系统的关系

现有系统关系
error-experience-libraryGene 的数据来源之一;迁移后 /error-lookup 仍可用,底层查询 Evolution Store
continuous-learningInstinct 可晋升为 Gene;观察流程继续保留
instinct-export/import升级为 Evolution Asset 导入导出
session-end hook增加 solidify 调用,自动从 git diff 提取 Gene
session-start hook增加进化上下文注入

相关工具

  • scripts/evolution/store.py — Evolution Store 读写接口
  • scripts/evolution/replay.py — Replay-First 执行器
  • scripts/evolution/governor.py — Governor 治理层
  • scripts/evolution/network.py — 跨 agent 共享网络
  • scripts/evolution/migrate.py — 从旧系统迁移数据

配合约束

  1. 所有进化操作必须产生 EvolutionEvent,事件日志是唯一事实源
  2. Gene/Capsule JSON 缓存可随时从 events.jsonl 重建
  3. 外部导入的资产必须先进 quarantine,本地验证通过后才能 promote
  4. Governor 可全局禁用(用于调试),但生产环境必须启用
  5. 进化能力默认 opt-in,通过 hooks.json 的 enabled flag 控制

Signals

GitHub stars
1k
Forks
316
Last commit
Oct 2026

ahel review

  • K1binfo
    installs-packages

Automated review, not a security audit. Ruleset v1+k2.

Advanced
Item type
skill
Key
evolution-core
Source
github.com/hashgraph-online/awesome-codex-plugins