create-principles-disciple 1.112.3 → 1.113.0

package/templates/langs/en/skills/pd-mentor/SKILL.md

@@ -1,211 +0,0 @@
----
-name: pd-mentor
-description: Spicy Mentor - Interactive command guidance and scenario-based recommendations
-disable-model-invocation: true
----
-
-# /pd-mentor: Spicy Mentor
-
-I'm your intelligent mentor, helping you understand and use all features of Principles Disciple.
-
-**Remember: Pain is the fuel of evolution, and I'm the one who lights the fire.**
-
----
-
-## Execution Principles
-
-1. **Scenario-based Guidance**: Recommend the most appropriate commands based on user's current task
-2. **Interactive Q&A**: Use AskUserQuestion to understand user intent
-3. **Flowchart Display**: Visualize SOP processes
-4. **Options First**: Provide preset options to reduce typing burden
-
----
-
-## Command Index
-
-| Command | Purpose | Use Case |
-|---------|---------|----------|
-| `/pd-init` | Initialize strategy and OKRs | New project startup |
-| `/pd-bootstrap` | Environment tool scan and upgrade | Tool upgrade |
-| `/pd-research` | Initiate tool upgrade research | Deep research |
-| `/pd-thinking` | Manage mental models and candidates | Metacognition |
-| `/pd-status` | View system status (GFI and Pain Dictionary) | Health check |
-| `/pd-help` | Get interactive command guidance | This skill |
-
----
-
-## Scenario Matching
-
-### Scenario 1: New Project Initialization
-
-**Trigger**: User says "I just created a new project" or similar
-
-**Recommended Flow**:
-1. `/pd-init` - Establish strategic vision and OKR framework
-2. `/pd-bootstrap` - Scan environment tools, get capability list
-3. `/pd-thinking` - Establish project's mental model baseline
-
-**Script**: "A new project is like a blank canvas. Let me help you lay the foundation: strategy first, then equipment, finally mental framework."
-
----
-
-### Scenario 2: Bug Fix Needed
-
-**Trigger**: User says "there's a bug", "got an error", "something's wrong"
-
-**Recommended Flow**:
-1. `/pd-status` - Check system status (GFI and Pain Dictionary)
-2. `/pd-evolution-status` - Start full evolution loop
-
-**Script**: "Problems are the fuel of evolution. Let me help you diagnose and fix systematically."
-
----
-
-### Scenario 3: Daily Maintenance & Review
-
-**Trigger**: User says "what did I do today", "check progress", "give me a report"
-
-**Recommended Flow**:
-1. `/pd-evolution-status` - Send today's evolution report
-2. `/pd-evolution-status` - View current trust score
-
-**Script**: "Daily report in hand, evolution I command. Let me help you review today's achievements."
-
----
-
-### Scenario 4: Permission or Security Related
-
-**Trigger**: User says "not enough permissions", "blocked", "security level"
-
-**Recommended Flow**:
-1. `/pd-evolution-status` - View trust score and security stage
-2. Explain current stage's capability boundaries
-
-**Script**: "Trust is earned, not given. Let me help you understand your current security level."
-
----
-
-### Scenario 5: Tool Upgrade Needs
-
-**Trigger**: User says "want to upgrade tools", "research new version", "tech stack update"
-
-**Recommended Flow**:
-1. `/pd-bootstrap` - Scan current environment
-2. `/pd-research` - Initiate research for specific tools
-
-**Script**: "Tool upgrades are the guarantee of combat capability. Let me help you scan and plan."
-
----
-
-## SOP Flowchart
-
-```
-┌─────────────────────────────────────────────────────────┐
-│              Evolution Loop SOP                         │
-├─────────────────────────────────────────────────────────┤
-│                                                         │
-│  Step 1: TRIAGE                                         │
-│     └─→ Collect evidence, reproduction steps, risk      │
-│                                                         │
-│  Step 2: DIAGNOSIS                                      │
-│     └─→ Root cause analysis, impact assessment          │
-│                                                         │
-│  Step 3: AUDIT                                          │
-│     └─→ Security check, logic verification             │
-│                                                         │
-│  Step 4: PLAN                                           │
-│     └─→ Target files, execution steps, rollback plan   │
-│                                                         │
-│  Step 5: EXECUTE                                        │
-│     └─→ Modify code according to plan                   │
-│                                                         │
-│  Step 6: VERIFY                                         │
-│     └─→ Run tests, check metrics                        │
-│                                                         │
-│  Step 7: REVIEW                                         │
-│     └─→ Code review, quality assessment                 │
-│                                                         │
-│  Step 8: LOG                                            │
-│     └─→ Record lessons learned, update principles       │
-│                                                         │
-└─────────────────────────────────────────────────────────┘
-```
-
----
-
-## Execution Flow
-
-### Step 1: Intent Recognition
-
-Use `AskUserQuestion` to ask user about current task scenario:
-
-**Options**:
-- 🆕 New project initialization
-- 🐛 Bug fix
-- 📊 Daily review
-- 🔐 Permission/Security
-- 🔧 Tool upgrade
-- ❓ Other
-
-### Step 2: Scenario Matching
-
-Based on user's selected scenario, recommend corresponding command flow.
-
-### Step 3: Confirm Execution
-
-Ask user if they want to execute a command:
-- "Execute `/pd-xxx` command now?"
-- "Need me to explain this command in detail?"
-
----
-
-## Advanced Usage
-
-### Combined Skills
-
-For complex scenarios, combine multiple skills:
-
-| Scenario | Combined Flow |
-|----------|---------------|
-| Major refactor | `/pd-evolution-status` → `deductive-audit` → execute |
-| System optimization | `/pd-status` → `evolve-system` → `root-cause` |
-| Project review | `/pd-evolution-status` → `/pd-okr` → `reflection-log` |
-
-### Internal Skill Calls
-
-These skills are usually called automatically by the system, but advanced users can also trigger them manually:
-
-- `triage` - Issue triage
-- `root-cause` - Root cause analysis
-- `deductive-audit` - Deductive audit
-- `reflection` - Metacognitive reflection
-- `reflection-log` - Reflection logging
-
----
-
-## FAQ
-
-**Q: What is GFI (Friction Index)?**
-A: GFI (Global Friction Index) measures the system's "pain level", range 0-100. Higher values indicate more friction, requiring attention.
-
-**Q: How is trust score calculated?**
-A: Based on successful operations accumulating points, failed operations deducting points. Reaching thresholds automatically upgrades security stage, unlocking more permissions.
-
-**Q: What is Pain Signal?**
-A: Pain Signal is a problem signal detected by the system, stored in `.state/.pain_flag`. When triggered, the system starts the evolution loop.
-
-**Q: How to view installed tool capabilities?**
-A: Check `.state/SYSTEM_CAPABILITIES.json` or run `/pd-bootstrap` to rescan.
-
-**Q: What's the difference between `/pd-thinking` and `/pd-research`?**
-A: `/pd-thinking` manages mental models and candidate solutions for metacognition and decision support; `/pd-research` initiates tool upgrade research for technical investigation.
-
----
-
-## Closing
-
-**Remember our creed**:
-
-> "Pain is the fuel of evolution, reflection is the ladder of progress."
-
-For any questions, feel free to call `/pd-help` or ask me directly. I'm your Spicy Mentor, evolving with you.

package/templates/langs/en/skills/profile/SKILL.md

@@ -1,24 +0,0 @@
----
-name: profile
-description: Manually correct or update the user's expertise profile. Use to tell the system "I am an expert in X" or "I am a novice in Y".
-disable-model-invocation: true
----
-
-# Profile Corrector
-
-You are now the "Manual Intervention Profile" component.
-
-**Task**:
-1. Parse user input `$ARGUMENTS` (format like "Frontend: Expert").
-2. Generate an **incremental evaluation** file `.state/.user_verdict.json`, forcing that domain's weight to extremely high (or low, depending on user description).
-3. Inform user: Changes will take effect after task completion (Stop).
-
-**JSON Template**:
-```json
-{
-  "updates": [
-    {"domain": "<EXTRACTED_DOMAIN>", "delta": 10, "reason": "User self-declared expertise"}
-  ]
-}
-```
-*(Note: If Novice, set delta to -10)*

package/templates/langs/en/skills/reflection/SKILL.md

@@ -1,40 +0,0 @@
----
-name: reflection
-description: Perform a deep metacognitive reflection on the current task status, user sentiment, and systemic issues. Use this before context compaction or when stuck.
-disable-model-invocation: false
----
-
-# Metacognitive Reflection
-
-**Trigger Scenario**: Context about to be compacted (Memory Loss Imminent) or Task long-term stalled.
-**Goal**: Before forgetting the detailed process, extract "painful lessons" and solidify them as principles.
-
-Please execute the following reflection steps:
-
-## 1. Status Scan
-- **Goal**: What was our original objective? (Check early conversation context)
-- **Status**: How much is completed now? Where are we stuck?
-- **Cost**: We've consumed significant tokens. Is the output matching the cost?
-
-## 2. Pain Detection
-Please honestly answer the following questions (Yes/No):
-- [ ] **Task Stalled**: No实质性 code progress for 3+ consecutive rounds?
-- [ ] **Repeated Errors**: Bug fixed then reappeared, or same error reported twice?
-- [ ] **User Frustration**: User used negative words like "wrong", "no", "stop", or tone became impatient?
-- [ ] **Blind Action**: Modified code directly without PLAN or AUDIT?
-- [ ] **Architecture Degradation**: Is code more chaotic than when we started?
-
-## 3. Root Cause Analysis (If Pain Detected)
-If any of above is Yes, must perform deep attribution:
-- **Direct Cause**: What did we do (or not do) that led to current situation?
-- **Root Cause**: Where did our mental model go wrong? (Too eager? Ignored existing files? Disregarded tests?)
-
-## 4. Evolution Logging
-If pain detected, must execute:
-1. **Record**: Write analysis results to `memory/ISSUE_LOG.md`.
-2. **Extract**: If this is the second occurrence, extract a **Prohibitive Principle (Must NOT)**.
-3. **Reinforce**: Suggest a specific Hook or Test to prevent future recurrence.
-
-## 5. Recovery Plan
-- Since we're about to compact context, how should we continue with the "cleanest" state?
-- Update `memory/.scratchpad.md`, mark current progress, ensure seamless continuation after compaction.

package/templates/langs/en/skills/reflection-log/SKILL.md

@@ -1,37 +0,0 @@
----
-name: reflection-log
-description: Final task reflection and evolution logging. Use to capture pain signals, update profiles, and propose new principles.
-disable-model-invocation: true
----
-
-# Reflection & Evolution
-
-**Goal**: Transform single-task experience into system's permanent memory.
-
-Please execute the following closing operations:
-
-## 1. Pain Summary
-- Briefly describe the most painful, time-consuming, or failure-causing points in this task.
-
-## 2. Issue Logging
-- **Action**: Append detailed Pain Signal and diagnosis results to `memory/ISSUE_LOG.md`.
-
-## 3. Evolution Candidates
-- **Principle**: Propose a new principle (P-XX).
-- **Guardrail**: Suggest a specific Hook, Rule, or Test.
-  - **Path Interception**: Suggest adding sensitive directories to `risk_paths` in `.principles/PROFILE.json`.
-  - **Behavior Interception**: Suggest adding regex to `custom_guards` in `.principles/PROFILE.json` to intercept dangerous calls of specific tools (e.g., `Edit.*SYSTEM`).
-
-## 4. Positive Reinforcement
-- **Check Excellence Signals**: 
-  1. User's explicit praise (Quote user).
-  2. Objective performance/quality metric improvements (Cite data).
-  3. Reviewer's high evaluation (Excellent/Elegant).
-- **Extract Pattern**: If above signals exist, record `achievement` field in `.state/.user_verdict.json` describing the successful behavioral pattern.
-
-## 5. Attribution
-- **Agent Scorecard**: Evaluate sub-agent performance used in this task, write to `.state/.verdict.json`. Format follows `@.principles/schemas/agent_verdict_schema.json`.
-- **User Profile**: Evaluate user instruction quality and preferences, write to `.state/.user_verdict.json`. Format follows `@.principles/schemas/user_verdict_schema.json`.
-
-## 6. Cleanup
-- Clean up all intermediate marker files (e.g., `.pain_flag`, `.verdict.json`, etc.).

package/templates/langs/en/skills/report/SKILL.md

@@ -1,13 +0,0 @@
----
-name: report
-description: Manually request a formal status report from the Reporter agent.
-disable-model-invocation: true
----
-
-# /report: Get Work Report
-
-User (the boss) requests an immediate work report.
-
-## Execution Action
-1. Immediately delegate ``pd_spawn_agent(reporter)``.
-2. Task description: "The boss wants to know the current situation. Please analyze current conversation context and recent `memory/ISSUE_LOG.md` to write an elegant report for the boss. Remember to check their profile first!"

package/templates/langs/en/skills/root-cause/SKILL.md

@@ -1,33 +0,0 @@
----
-name: root-cause
-description: Deep dive analysis into why a problem occurred. Uses the 5 Whys method and classifies the cause.
-disable-model-invocation: true
----
-
-# Root Cause Analysis
-
-**Goal**: See through phenomena to essence, prevent repeated mistakes.
-
-Please execute diagnosis in the following format:
-
-## 1. Proximal Cause
-- **Verb-driven**: Describe what specific operation or lack thereof caused the immediate failure.
-
-## 2. 5 Whys (Deep Inquiry)
-1. Why did the proximal cause occur?
-2. ...
-3. ...
-4. ...
-5. Trace back to systemic or cognitive root.
-
-## 3. Root Cause
-- **Adjective/Design-driven**: Describe system architecture, process defects, or wrong assumptions.
-- **Guardrail Failure Analysis**: Why didn't existing gates (Hooks) or rules (Rules) intercept this error? Is it missing rules, loose matching, or logic loopholes?
-
-## 4. Category
-- [ ] **People**: Capability blind spots, habit issues.
-- [ ] **Design**: Process/tool defects, insufficient gates, architecture loopholes.
-- [ ] **Assumption**: Wrong assumptions about environment, versions, or dependencies.
-
-## 5. Principle Candidate
-- If we fix this issue, what principle should be added to prevent it from happening again?

package/templates/langs/en/skills/triage/SKILL.md

@@ -1,29 +0,0 @@
----
-name: triage
-description: Initial problem definition and risk assessment. Use to collect environment info, reproduction steps, and logs.
-disable-model-invocation: true
----
-
-# Triage (Problem Triage)
-
-**Goal**: Collect sufficient evidence to define the problem and assess change risk.
-
-Please output in the following structure:
-
-## 1. Goal
-- One sentence describing the final success criteria for this task.
-
-## 2. Problem
-- What is happening now? What is the expected behavior?
-- **Evidence**: List relevant log snippets, error codes, or observed anomalies.
-
-## 3. Reproduction
-- Provide clear reproduction commands or operation sequence.
-- Note key variables in current environment.
-
-## 4. Scope
-- Preview of involved files or modules.
-- **Risk Level**: Low / Medium / High (based on `.principles/PROFILE.json` definition).
-
-## 5. Next Step
-- Recommend which sub-agent to delegate (usually Explorer) for further evidence collection.

package/templates/langs/en/skills/watch-evolution/SKILL.md

@@ -1,33 +0,0 @@
----
-name: watch-evolution
-description: Start the background evolution daemon to process queued tasks from EVOLUTION_QUEUE.json.
----
-
-# /watch-evolution: Background Evolution Guardian
-
-**Goal**: Start a continuously running process to automatically analyze and fix backlog code issues.
-
-## 1. Environment Preparation
-- Ensure `evolution_mode: "async"` is configured in `.principles/PROFILE.json`.
-- Ensure `.state/evolution_queue.json` exists.
-
-## 2. Startup Command
-Run the following command to start the daemon:
-
-```bash
-python scripts/evolution_daemon.py
-```
-
-## 3. Workflow
-1. **Auto Scan**: Scan `.state/evolution_queue.json` every 30 seconds.
-2. **Scheduling Strategy**:
-   - Process tasks by `priority` from high to low (high priority first).
-   - Tasks with `status: retrying` are only re-executed after `next_retry_at` expires.
-   - Failed tasks enter exponential backoff retry, marked as `failed` after max attempts.
-3. **Diagnosis**: Load `root-cause` skill, locate error root cause in Headless mode.
-4. **Fix**: Load fix skill, automatically modify code and run tests.
-5. **Log**: Call `reflection-log` to store experience in `PRINCIPLES.md` and `ISSUE_LOG.md`.
-6. **Dashboard**: Real-time update of `memory/EVOLUTION_PRD.md` showing progress.
-
-## 4. Exit
-- Press `Ctrl+C` to stop background task. Task state will be preserved in JSON, can continue on next startup.

package/templates/langs/zh/skills/admin/SKILL.md

@@ -1,41 +0,0 @@
----
-name: admin
-description: System administration and recovery tool for humans. Use to init, repair, or reset the evolutionary agent framework.
-disable-model-invocation: true
----
-
-# Admin Console (管理员控制台)
-
-你现在扮演的是“可进化系统管理员”。你的职责是根据用户提供的参数 `$ARGUMENTS` 维护、修复或初始化系统的“毛坯房”架构。
-
----
-
-## 核心功能
-
-### 1. `diagnose` (系统诊断)
-**动作**: 检查“毛坯房”架构的完整性。
-- **核心组件**: 检查 `.claude/hooks/hook_runner.py` 是否存在且可执行。
-- **文档完整性**: 检查 `.principles/PROFILE.json` 等是否存在。
-- **工具感知**: 检查 `.state/SYSTEM_CAPABILITIES.json`。若缺失，提示用户："⚠️ 尚未进行工具链升级。建议运行 `/bootstrap-tools` 以大幅提升系统能力。"
-- **记忆挂载**: 检查 `CLAUDE.md` 是否包含 `System Integration` 章节。
-- **输出**: 生成一份健康报告，列出缺失或异常的项目。
-
-### 2. `repair` (系统修复)
-**动作**: 
-- **配置恢复**: 如果 `PROFILE.json` 缺失或损坏，尝试从 `.claude/templates/PROFILE.json` 恢复。
-- **结构补全**: 确保工作区结构完整。
-- **强制清理**: 删除 `.pain_flag`, `.verdict.json`, `.user_verdict.json`, `.pending_reflection` 等临时标记。
-
-### 3. `reset` (强制重置)
-
-### 3. `reset` (强制重置)
-**动作**: 在得到用户明确确认后，将 `USER_PROFILE.json` 和 `AGENT_SCORECARD.json` 归零。
-
-### 4. `status` (状态报告)
-**动作**: 汇报当前 Risk Paths、用户最高/最低分领域、Agent 排名。
-
----
-
-## 执行准则
-- 只有在人类用户输入 `/admin` 时，你才会看到此指令。
-- 执行前简述计划，执行后输出“✅ 系统已加固/已初始化”。

package/templates/langs/zh/skills/bootstrap-tools/SKILL.md

@@ -1,52 +0,0 @@
----
-name: bootstrap-tools
-description: Scans project tech stack and searches the web for the latest, most effective CLI tools to augment agent capabilities. Suggests and installs tools upon user approval.
-disable-model-invocation: true
----
-
-# /bootstrap-tools: 装备升级官
-
-你的目标是让智能体团队拥有最先进的武器。通过分析当前项目栈，并**实时联网搜索**，寻找能提升开发、重构、测试效率的最佳 CLI 工具。
-
-## 执行流程
-
-### 1. 侦察 (Recon)
-- **分析技术栈**: 读取 `package.json`, `Cargo.toml`, `requirements.txt` 等。确定核心框架（如 Next.js, FastAPI）。
-- **盘点现状**: 运行 `npm list -g --depth=0` 和 `command -v` 检查已安装的工具。
-
-### 2. 寻宝 (Hunt)
-- **联网搜索**: 针对当前栈，搜索最新的 CLI 神器。
-  - *Query 示例*: "best CLI tools for Next.js 15 development 2025", "fastest rust-based grep alternative", "modern linter for python".
-- **筛选标准**:
-  - **Headless**: 必须是 CLI 工具。
-  - **Performance**: 优先推荐 Rust/Go 编写的高性能工具 (e.g., `ripgrep`, `ast-grep`, `oxc`).
-  - **Relevance**: 能解决实际痛点（如 `knip` 查死代码，`depcheck` 查依赖）。
-
-### 3. 提案 (Pitch)
-- 使用 `AskUserQuestion` 向用户展示推荐清单。
-- **格式**:
-  - **工具名**: [Name]
-  - **推荐理由**: [Why it helps the agent/project]
-  - **安装命令**: `npm i -g ...` 或 `apt-get ...`
-  - **Demo**: 给出一个简单的用法示例。
-
-### 4. 部署与登记 (Deploy & Register)
-- 获得批准后，执行安装命令。
-    - **验证 (Verification - 强制)**:
-      - 安装完成后，**必须**运行 `<tool> --version` 或 `command -v <tool>` 来验证是否真的安装成功。
-      - **若失败**: 告知用户（可能是权限问题），请求用户手动安装，**不要**更新能力文件。
-      - **若成功**: 
-        - 更新 `.state/SYSTEM_CAPABILITIES.json`。记录新工具的路径。
-        - **全员广播**: 
-          - 扫描 `.claude/agents/*.md`。
-          - 检查每个文件是否包含 `@.state/SYSTEM_CAPABILITIES.json`。
-          - 若未包含，在文件末尾追加：
-            ```markdown
-            
-            ## Environment Capabilities
-            Check @.state/SYSTEM_CAPABILITIES.json for high-performance tools (e.g., ripgrep, ast-grep) available in this environment. Use them!
-            ```
-          - 提示用户运行 `/manage-okr` 或 `/admin diagnose` 以让 Agent 感知新能力。
-## 核心原则
-- **喜新厌旧**: 敢于推荐新工具替代旧工具（如推荐 `pnpm` 替 `npm`，推荐 `vitest` 替 `jest`），但要说明理由。
-- **安全第一**: 在安装前必须获得用户明确授权。

package/templates/langs/zh/skills/deductive-audit/SKILL.md

@@ -1,36 +0,0 @@
----
-name: deductive-audit
-description: Rigorous safety and logic check of a proposed solution. Evaluates against axioms, system impact, and edge cases.
-disable-model-invocation: true
----
-
-# Deductive Audit (演绎审计)
-
-**目标**: 在执行前拦截逻辑错误和潜在风险。
-
-请对当前提议的修复方案进行三审：
-
-## 1. Axiom Test (公理审计)
-- 语言规范、库的 API 契约、项目既定约定是否被违反？
-- 输入输出类型是否对齐？
-
-## 2. System Test (系统审计)
-- 是否引入了新的技术债？
-- 是否影响了性能、延迟或稳定性？
-- 是否引入了循环依赖？
-
-## 3. Entropy Audit (熵减审计) - *New*
-- **必要性**: 这个修改是绝对必要的吗？有没有更简单的解法？
-- **最小化**: 是否触碰了不该触碰的文件？（最小接触面原则）
-- **反博弈**: 这是一个真实的修复，还是为了过门禁而做的“形式主义”？
-
-## 4. Via Negativa (否定审计)
-- 最坏的情况下（网络断开、磁盘满、恶意输入）会发生什么？
-- 是否有安全红线（Token 泄露、越权）风险？
-
-## 5. Verdict (裁决)
-- **RESULT**: PASS | FAIL
-- **Must Fix**: 如果失败，列出必须修正的点。
-
----
-**动作**: 请将以上结果更新至 `AUDIT.md`。

package/templates/langs/zh/skills/evolution-framework-update/SKILL.md

@@ -1,31 +0,0 @@
----
-name: evolution-framework-update
-description: 拉取原则信徒进化框架的最新更新（包含 Orchestrator 模式、异步队列及地图优先协议）。
----
-
-# /evolution-framework-update: 进化框架自更新
-
-**目标**: 同步上游框架的最新代码（Hooks, Skills, Agents, Daemon），保持系统进化能力。
-
-## 1. 执行更新
-运行以下脚本拉取最新代码：
-
-```bash
-bash scripts/update_agent_framework.sh
-```
-
-## 2. 冲突处理 (Smart Merge)
-脚本运行后，请检查输出：
-- **无冲突**: 如果显示 "✅ Update complete"，则无需操作。
-- **有冲突**: 如果显示 "⚠️ Updates found with conflicts"：
-  1. 查找所有 `.update` 文件：
-     ```bash
-     find .claude -name "*.update"
-     ```
-  2. 对于每一个冲突文件（例如 `THINKING_OS.md` vs `THINKING_OS.md.update`）：
-     - **读取** 原文件和 `.update` 文件。
-     - **分析** 差异：合入上游的新功能，保留本地的个性化配置。
-     - **清理**：合并完成后删除 `.update` 文件。
-
-## 3. 重启生效
-更新完成后，建议重启 Session 以加载最新的神经中枢逻辑。

package/templates/langs/zh/skills/evolve-system/SKILL.md

@@ -1,46 +0,0 @@
----
-name: evolve-system
-description: Second-order observation and system-level evolution. Analyzes performance metrics and issue logs to propose optimizations for agents, hooks, and rules.
-disable-model-invocation: true
----
-
-# /evolve-system: 数字架构师 (二阶观察)
-
-你现在的身份是本系统的 **数字化架构师 (The Architect)**。你的职责不是修复业务代码，而是通过分析系统运行数据，优化系统自身的“基因”（Prompt、Hook、规则）。
-
-## 1. 现状度量 (Metrics Analysis)
-- **读取数据**:
-  - `.state/AGENT_SCORECARD.json`: 计算每个 Agent 的胜率 (wins / (wins + losses))。
-  - `memory/ISSUE_LOG.md`: 识别最近 10 条记录中的重复模式（Pain Patterns）。
-- **识别异常**:
-  - **低效 Agent**: 胜率低于 50% 且样本量 >= 3 的 Agent。
-  - **系统顽疾**: 在 Issue Log 中出现超过 2 次的同类系统性错误。
-
-## 2. 根因诊断 (Systemic Diagnosis)
-- 针对识别出的异常，分析其在 `.claude/agents/` 或 `.claude/hooks/` 中的定义。
-- **思考**: 
-  - 是 Prompt 描述太模糊导致幻觉？
-  - 是 Hook 逻辑存在边界死角？
-  - 是缺失了某个关键的 Guardrail？
-
-## 2.5 临床实验 (Clinical Trial) - *Optional*
-**如果根因不明确**，需进行实证：
-- **征询**: 使用 `AskUserQuestion` 询问：“为确诊问题，我需要对 [Agent] 进行一次自动诊断任务，这可能会消耗一些 Token，是否继续？”
-- **静默执行**:
-  - 若用户同意，直接调用 ``pd_spawn_agent` 工具` 发起测试。
-  - **指令**: "你正在被进行诊断测试。请执行以下任务：[Test Scenario]。请保持输出极其精简，只返回最终结果或错误信息。"
-  - **观察**: 检查其工具调用链是否符合预期（例如：是否使用了正确的 Search 工具）。
-- **确诊**: 基于测试表现，锁定病灶。
-
-## 3. 进化提案 (Optimization Proposal)
-**如果根因已确诊**，生成 `SYSTEM_OPTIMIZATION_PLAN.md`，内容包括：
-- **诊断结论**: 明确指出系统哪一部分“病了”。
-- **修改建议**: 提供具体的代码/Prompt 修改 Diff。
-- **预期收益**: 解释这次修改如何提升胜率或减少痛苦。
-
-## 4. 安全执行 (Safety Gate)
-- **强制确认**: 在修改任何系统文件 (`.claude/` 目录下) 之前，必须使用 `AskUserQuestion` 展示提案并获得用户明确授权。
-- **原子性**: 每次只建议一个高杠杆的优化点，不要试图一次性重构整个系统。
-
-## 结项
-输出：“✅ 系统自诊完成。提案已提交，等待老板决策。”