本文へ移動
cccskills
無料GitHub で公開

continuous-learning

基于 instinct 的持续学习系统,通过 hooks 观察会话,创建带置信度评分的 atomic instincts, 并将高置信度 instinct 演进为 skills/commands/agents。v2.1 增加项目级 instincts 防止跨项目污染。

インストール方法を見る

含まれるファイル(11)

  • SKILL.md4.3 KB
  • agents/observer-loop.sh7.6 KB
  • agents/observer.md7.2 KB
  • agents/openai.yaml279 B
  • agents/session-guardian.sh6.2 KB
  • agents/start-observer.sh7.3 KB
  • config.json135 B
  • hooks/observe.sh15.1 KB
  • scripts/detect-project.sh7.6 KB
  • scripts/instinct-cli.py58.4 KB
  • scripts/test_parse_instinct.py32.3 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Continuous Learning v2.1 - Instinct-Based Architecture

一个高级学习系统,通过 atomic "instincts" 将 Claude Code 会话转化为可重用知识。

v2.1 增加 项目级 instincts — React 模式保留在 React 项目中,Python 约定保留在 Python 项目中,通用模式(如"始终验证输入")全局共享。

何时激活

  • 设置从 Claude Code 会话自动学习
  • 配置基于 hooks 的行为提取
  • 调优学习行为的置信度阈值
  • 审查、导出或导入 instinct 库
  • 将 instincts 演进为完整 skills、commands 或 agents
  • 管理项目级 vs 全局 instincts
  • 将 instincts 从项目级提升到全局级

Instinct 模型

一个 instinct 是一个小的学习行为:

---
id: prefer-component-colocation
trigger: "当多个组件使用相同状态时"
confidence: 0.7
domain: "frontend-architecture"
source: "session-observation"
scope: project
project_id: "a1b2c3d4e5f6"
project_name: "points-frontend"
---

# 组件状态共置

## 行为
当多个组件使用相同状态时,考虑将状态提升到共同祖先。

## 证据
- 在 5 个实例中观察到状态提升模式
- 用户在 2026-03-29 将组件本地状态改为共享状态

属性:

  • Atomic -- 一个 trigger,一个 action
  • Confidence-weighted -- 0.3 = 试探性,0.9 = 几乎确定
  • Domain-tagged -- code-style, testing, git, debugging, workflow 等
  • Evidence-backed -- 跟踪创建它的观察
  • Scope-aware -- project(默认)或 global

存储结构

~/.claude/homunculus/
+-- projects.json           # 项目注册表: hash -> name/path
+-- observations.jsonl      # 全局观察(fallback)
+-- instincts/
|   +-- personal/          # 全局自动学习的 instincts
|   +-- inherited/         # 全局导入的 instincts
+-- evolved/
|   +-- agents/           # 全局生成的 agents
|   +-- skills/           # 全局生成的 skills
|   +-- commands/         # 全局生成的 commands
+-- projects/
    +-- a1b2c3d4e5f6/    # 项目 hash(来自 git remote URL)
        +-- project.json    # 项目元数据
        +-- observations.jsonl
        +-- instincts/
        |   +-- personal/  # 项目特定自动学习
        |   +-- inherited/  # 项目特定导入
        +-- evolved/
            +-- skills/
            +-- commands/
            +-- agents/

项目检测

系统自动检测当前项目:

  1. CLAUDE_PROJECT_DIR env var(最高优先级)
  2. git remote get-url origin -- hash 生成项目 ID
  3. git rev-parse --show-toplevel -- fallback
  4. 全局 fallback -- 如果未检测到项目,instincts 进入全局范围

置信度评分

分数含义行为
0.3试探性建议但不强制
0.5中等相关时应用
0.7强自动批准应用
0.9几乎确定核心行为

置信度增加当:

  • 模式被重复观察
  • 用户不纠正建议的行为
  • 相似 instincts 同意

置信度减少当:

  • 用户明确纠正行为
  • 模式长时间未观察到
  • 出现矛盾证据

与 Skills 的区别

特性SkillsInstincts
粒度完整工作流Atomic 行为
触发手动调用自动观察
置信度无0.3-0.9
演化直接成为 skill先 instinct 再 cluster

相关工具

  • scripts/lib/memory_store.py - 底层存储接口
  • Error Experience Library - 错误模式的专门 instinct

命令接入

  • /instinct-status - 显示学习的 instincts
  • /evolve - 将 instincts 聚类为 skills/commands
  • /instinct-export - 导出 instincts 到文件
  • /instinct-import - 从文件导入 instincts
  • /promote - 将项目 instincts 提升到全局

最佳实践

  1. 观察 -- 启用 hooks 捕获所有工具调用
  2. 分析 -- 定期运行 observer agent 分析观察
  3. 演进 -- 将高置信度 instincts 演进为 skills
  4. 分享 -- 通过导出/导入在团队中共享 instincts

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Use AtomLane to compile and execute safe atomic parallel plans on macOS and native Windows Preview for worthwhile independent argv tasks, dependency DAGs, supported platform entrypoints, or Apple-silicon operators. Use at task start or an execution boundary when structured local work may contain two or more worthwhile units; skip plain answers, one quick command, and work whose effects cannot be safely bounded.

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2752026年10月11日 更新

add

無料

Register a deferred decision in the debt registry. Trigger by judgment, not a marker scan, whenever a future reader would ask "why this way?": an unmade decision, stub, loosened type, bypassed check, swallowed error, a default picked "for now", or a TODO/FIXME/HACK/XXX marker. Trigger immediately whenever you defer work, or when the user invokes $add. Over-register freely; the developer drops with "drop A", "drop A,C", or "drop all".

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2752026年10月11日 更新

ADK 框架适配层。为 LangChain / EINO / AutoGen / AgentScope / CrewAI 提供框架特定的 代码模板、惯用模式、API 映射和项目结构,供 agent-dev-workshop Phase 5 代码生成使用。 每个框架 reference 文件标注 verified_date 用于版本锁定。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2752026年10月11日 更新

中文调试修复技能。用于报错、测试失败、页面异常、功能不符合预期、需要定位根因并做最小修复时。触发语包括"进入调试模式""帮我修问题""报错了""测试失败""页面坏了""找根因"。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2752026年10月11日 更新

交互式 AI Agent 开发工作坊:通过 6 阶段深度协作对话,引导用户完成 Agent 需求分析、架构设计、 工具定义、Prompt 与编排设计、代码生成、验证迭代,产出可直接运行的 Agent 项目。 框架无关设计优先,支持 LangChain / EINO / AutoGen / AgentScope / CrewAI 等 ADK 框架。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2752026年10月11日 更新

中文漂移审计技能。用于项目或学习过程变乱、上下文漂移、任务分叉、多个方案冲突、命名不一致、Codex 可能顺手改多了时。触发语包括"漂移检查""感觉跑偏了""项目变乱了""检查是否失控""分叉太多""上下文漂移"。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2752026年10月11日 更新

hashgraph-online のスキルをすべて見る

このスキルの問題を報告する