Evolution Skill - Detailed Workflow Guide

This document describes the complete workflow of the evolution skill.

Phase 0: Trigger

Manual trigger: /evolution

Weekly summary integration: If you have a weekly summary skill (周汇报) configured, it will check evolution's last run timestamp. If evolution hasn't run this week, the weekly summary will prompt you:

## 进化状态
- 上次运行：从未运行
- 本周状态：⚠️ 未运行
- 推荐：运行 /evolution 进行一次工作区扫描

Phase 1: First Run vs Subsequent Runs

First Run Detection

Evolution detects first run by checking if ~/.openclaw/skills/evolution/BASELINE.md exists.

On first run:

Prompt: "首次运行建议扫描更长周期以建立基线。是否扫描过去90天？"
- Default: Yes (establish comprehensive baseline)
- User can configure: baseline_retention_days in config.yaml
Scan configured time span (90 days by default)
Record baseline: mtime + size for ALL scanned files in BASELINE.md
On successful completion: ask once "是否设置定时运行cron？"
Create LAST_RUN.md with first run timestamp

After first run:

Load LAST_RUN.md → previous scan timestamp
Load BASELINE.md → baseline mtime + size for each file
For each file: if mtime > last_run OR size ≠ baseline_size → include in scan
Update BASELINE.md with new mtime + size
Update LAST_RUN.md

Subsequent Run Modes

Mode	Command	Behavior
Incremental (default)	`/evolution`	Scan only changed files since last run
Full	`/evolution --full`	Rescan everything, reset baseline
Config	`/evolution --config`	Edit config.yaml

Phase 2: Workspace Selection

Default: All workspaces in ~/.openclaw/

workspace/           (main agent)
workspace-researcher/
workspace-engineer/
workspace-evaluator/
workspace-planner/
workspace-creator/
workspace-incognito/

User-configurable exclude list in config.yaml:

exclude_workspaces:
  - workspace-incognito  # Exclude by default (messaging mode)

Workspace scanning priority:

memory/ - Daily session logs (richest source of patterns)
knowledge/ - Curated reference material
skills/ - Already installed skills (to avoid duplication)
Root daily logs (workspace/*/YYYY-MM-DD.md)

Phase 3: Scanning

Scope

Included:
├── workspace*/memory/*.md (last N days, default 30)
├── workspace*/knowledge/** (all subdirectories)
├── workspace*/skills/** (installed skills list only)
└── workspace*/SOUL.md, IDENTITY.md, AGENTS.md (for context)

Excluded (security/privacy):
├── logs/
├── credentials/
├── agents/*/sessions/
└── any file with "secret", "password", "token" in path

Incremental Detection Logic

For each file in scope:
    stat(file) → mtime, size
    Load BASELINE.md entry for file (if exists)
    
    IF no BASELINE entry:
        ADD to scan (new file)
    ELSE IF mtime > last_run_timestamp:
        ADD to scan (modified)
    ELSE IF size ≠ baseline_size:
        ADD to scan (content changed despite mtime)
    ELSE:
        SKIP (unchanged)

Output

Raw scan data saved to ~/.openclaw/skills/evolution/reports/<date>-raw-scan.md:

# Raw Scan Data - 2026-04-29

## Scan Parameters
- mode: incremental
- scan_span_days: 30
- timestamp: 2026-04-29T18:30:00+08:00

## Files Scanned

| Workspace | File | Size | Modified |
|-----------|------|------|----------|
| workspace | memory/2026-04-28.md | 42618 | 1745932800 |
| workspace-researcher | knowledge/research/... | ... | ... |

Phase 4: Pattern Recognition

Category A: 工作边界类 (Cross-Workspace Issues)

What to detect:

Documents created in one workspace but referenced from another
Knowledge stored in researcher workspace but needed by main
Files that should be centralized but are scattered

Detection signals:

"放在workspace-researcher" in memory logs
Cross-workspace file references
"无法找到" / "找不到" patterns
Knowledge silo complaints

Example candidate:

smart-archive: 自动检查researcher创建的文档是否需要同步到主工作区

Category B: 经验复用类 (Reusable Patterns)

What to detect:

Repeated successful prompt templates
Consistent workflow steps across sessions
Debug paths that solved problems

Detection signals:

Similar instructions appearing ≥3 times
"成功" / "解决了" following consistent steps
Repeated tool use sequences

Example candidate:

research-briefing-template: 标准化的行业简报生成模板

Category C: 效率提升类 (Efficiency Gains)

What to detect:

Repeated manual operations
Commands that could be automated
Tedious multi-step processes

Detection signals:

"重复" / "每次都" / "每次都要"
Same bash command executed repeatedly
Copy-paste patterns in memory logs

Example candidate:

brand-research-flow: 品牌研究的标准工作流封装

Category D: 系统健康类 (System Health)

What to detect:

Memory files growing excessively
Knowledge contradictions
Configuration drift

Detection signals:

Memory file size > 100KB
"矛盾" / "冲突" in knowledge files
Frequent re-explaining of same concept

Example candidate:

memory-maintenance: 定期清理和压缩memory文件的机制

Phase 5: Quality Filtering

Each candidate must pass ALL gates:

Basic Gates (must pass)

Gate	Threshold	Rationale
Frequency	≥3 occurrences in scan period	Not random noise
Reward	Time savings OR error reduction identifiable	Worth the effort
Stability	Pattern stable across different contexts	Won't break soon
Boundary	Does not overlap with existing skills	No duplication

Execution Constraint Gate (must pass)

Gate	Requirement
Verifiable	Must include a concrete verification method
Trigger Clarity	Must specify exact trigger conditions, not vague
Actionable	Must define concrete actions
Feedback Loop	Must define how skill confirms correct execution

OpenClaw Compliance Gate (must pass)

Gate	Requirement
Architecture Fit	Follows OpenClaw best practices
Tool Alignment	Can be executed with OpenClaw's tool set
Context Appropriate	Scope appropriate for a skill

Auto-rejection criteria:

Frequency < 3 → "too rare"
No clear reward → "insufficient benefit"
Only happened once → "one-time issue"
Overlaps with existing skills → "duplicate functionality"
Cannot design test → "unverifiable"
Vague trigger conditions → "需要重构"
No feedback loop → "需要重构"

Phase 6: Distillation Report

Output to ~/.openclaw/skills/evolution/reports/<date>-report.md:

# 进化报告 - 2026-04-29

## 运行状态
- 上次运行：从未运行
- 本周工作区变动：workspace-researcher (+3), workspace-engineer (+5)

---

## Skill候选推荐

### 候选 #1：smart-archive

**类型**：工作边界类

**问题描述**：
researcher创建的文档经常存放在本工作区而非主工作区，导致知识孤岛。

**Pattern证据**：
| 时间 | 证据摘要 |
|------|----------|
| 2026-04-17 | researcher生成了品牌报告，放在workspace-researcher/knowledge/ |
| 2026-04-20 | feishu wiki文档位置错误，main无法发现 |
| 2026-04-25 | 类似的文档位置错误再次出现 |

**影响评估**：
- 频率：3次/30天 ✅
- 收益：减少手动整理，每年节省约2小时
- 稳定性：高（跨多个研究项目）
- 边界：与现有skill不重叠 ✅
- 可验证：可以设计简单测试 ✅

**建议的Skill结构**：

smart-archive/ ├── SKILL.md # 主指令 ├── references/ # 操作规范 └── scripts/ # 归档脚本（草稿）


**预估工作量**：2-3小时

---

## 你的操作

对于每个候选，请回复：
- **确认** → 我生成正式草稿到 `staging/`
- **待定** → 在下次evolution运行时再考虑，不遗忘
- **拒绝** → 记录拒绝理由（用于未来过滤）
- **修改** → 告诉我需要修改什么

[候选 #1] 确认/待定/拒绝/修改？
[候选 #2] ...

Phase 7: User Confirmation + Installation

Step 1: Confirm candidate

User replies "确认" for a candidate.

Step 2: Generate staging SKILL.md

Evolution generates full SKILL.md in staging/<candidate-name>/.

Step 3: Prompt installation

已生成 smart-archive 草稿于 ~/.openclaw/skills/evolution/staging/smart-archive/

是否帮你安装？
- 输入"是"：我将提供安装命令（你手动执行）
- 输入"否"：草稿保留，随时可以手动安装

Step 4: Strict confirmation for installation

⚠️ 安装确认
Skill: smart-archive
位置: ~/.openclaw/skills/evolution/staging/smart-archive/
操作: 复制到 ~/.openclaw/skills/smart-archive/

此操作将：
- 在 ~/.openclaw/skills/ 下创建 smart-archive/
- 安装后需要重启网关: openclaw gateway restart

确认安装？ (yes/no)

Step 5: Provide install command (NOT auto-install)

安装命令：
cp -r ~/.openclaw/skills/evolution/staging/smart-archive ~/.openclaw/skills/
openclaw gateway restart

复制后告诉我完成，我会验证安装是否成功。

Phase 8: Installed Skill Tracking

Purpose: Track evolution-generated skills and generate optimization candidates.

Registry File: evolution-installed.md

When user confirms installation of an evolution-generated skill:

Evolution asks: "Install complete?"
If yes → record to ~/.openclaw/skills/evolution/evolution-installed.md

Tracking During Scan

During Phase 1 scan, for each active skill in evolution-installed.md:

Search recent memory for skill references and trigger keywords
Check if constraint was followed or bypassed
Log trigger count and effectiveness
Generate optimization candidates for low-effectiveness skills

Optimization Trigger Conditions

Condition	Action
Skill rarely triggered	Mark as low value, consider removing
Skill triggered but ineffective	Generate optimization candidate
OpenClaw architecture changed	Mark as potentially outdated

Output in Distillation Report

## 已安装Skill优化候选

### authorization-norms-ref (需要优化)
- 触发次数：1次
- 有效性：低
- 问题：规则存在但无执行机制
- 建议：重构为P0-core hook集成或移除
- 操作：[优化/移除/保留]

Phase 9: Memory Log Generation

Purpose: Create a memory log after each evolution run for continuity.

When Generated

After Phase 7 (Manual Install) completes, before ending the evolution session.

Output Location

~/.openclaw/skills/evolution/memory/YYYY-MM-DD.md

Format

# Evolution Memory Log - 2026-04-29

## Run Summary
- mode: incremental
- scan_span_days: 30
- files_scanned: 47
- candidates_found: 2
- candidates_confirmed: 1
- candidates_deferred: 0
- candidates_rejected: 1

## Installed Skill Tracking
| Skill | Status | Last Triggered | Effectiveness |
|-------|--------|----------------|----------------|
| authorization-norms-ref | needs_optimization | 2026-04-29 | low |

## Notes
- First run with 90-day baseline
- Generated candidates: authorization-norms-ref, steve-commands-regression-check
- authorization-norms-ref installed, flagged for optimization

Phase 10: Weekly Summary Integration

The weekly summary skill (周汇报) should check evolution status:

## 进化状态
- 上次运行：2026-04-22（3天前）
- 本周运行状态：✅ 已运行

- 上次运行：从未运行
- 本周运行状态：⚠️ 未运行
- 推荐：运行 /evolution 进行一次工作区扫描

Implementation in weekly summary skill:

## 进化状态
- 上次运行：{{ evolution.last_run }}
- 本周状态：{{ if evolution.run_this_week then "✅ 已运行" else "⚠️ 未运行" }}
{{ if not evolution.run_this_week }}
- 推荐：运行 /evolution 进行一次工作区扫描
{{ end }}

Strictly Prohibited Actions

Action	Forbidden Because
Auto-creating skills	User must approve each
Auto-installing skills	Irreversible permission grant
Modifying system configs	Candidates may contain errors
Deleting/moving files	Only recommendations
Modifying other skills	Boundary pollution

File Structure

~/.openclaw/skills/evolution/
├── SKILL.md                    # Main skill instructions
├── evolution-guide.md          # This workflow doc
├── config.yaml                 # User preferences
├── config.yaml.example        # Config template
├── BASELINE.md                 # Baseline mtime+size for incremental scan
├── LAST_RUN.md                 # Timestamp of last run
├── evolution-installed.md      # Registry of evolution-generated installed skills
├── staging/                    # Confirmed candidates get generated here
│   └── .gitkeep
├── reports/                    # Historical distillation reports
│   └── 2026-04/
│       └── 2026-04-29-report.md
└── memory/                     # Evolution's own memory (runtime generated)
    └── 2026-04-29.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Evolution Skill - Detailed Workflow Guide

Phase 0: Trigger

Phase 1: First Run vs Subsequent Runs

First Run Detection

Subsequent Run Modes

Phase 2: Workspace Selection

Phase 3: Scanning

Scope

Incremental Detection Logic

Output

Phase 4: Pattern Recognition

Category A: 工作边界类 (Cross-Workspace Issues)

Category B: 经验复用类 (Reusable Patterns)

Category C: 效率提升类 (Efficiency Gains)

Category D: 系统健康类 (System Health)

Phase 5: Quality Filtering

Basic Gates (must pass)

Execution Constraint Gate (must pass)

OpenClaw Compliance Gate (must pass)

Phase 6: Distillation Report

Phase 7: User Confirmation + Installation

Phase 8: Installed Skill Tracking

Registry File: evolution-installed.md

Tracking During Scan

Optimization Trigger Conditions

Output in Distillation Report

Phase 9: Memory Log Generation

When Generated

Output Location

Format

Phase 10: Weekly Summary Integration

Strictly Prohibited Actions

File Structure

FilesExpand file tree

evolution-guide.md

Latest commit

History

evolution-guide.md

File metadata and controls

Evolution Skill - Detailed Workflow Guide

Phase 0: Trigger

Phase 1: First Run vs Subsequent Runs

First Run Detection

Subsequent Run Modes

Phase 2: Workspace Selection

Phase 3: Scanning

Scope

Incremental Detection Logic

Output

Phase 4: Pattern Recognition

Category A: 工作边界类 (Cross-Workspace Issues)

Category B: 经验复用类 (Reusable Patterns)

Category C: 效率提升类 (Efficiency Gains)

Category D: 系统健康类 (System Health)

Phase 5: Quality Filtering

Basic Gates (must pass)

Execution Constraint Gate (must pass)

OpenClaw Compliance Gate (must pass)

Phase 6: Distillation Report

Phase 7: User Confirmation + Installation

Phase 8: Installed Skill Tracking

Registry File: evolution-installed.md

Tracking During Scan

Optimization Trigger Conditions

Output in Distillation Report

Phase 9: Memory Log Generation

When Generated

Output Location

Format

Phase 10: Weekly Summary Integration

Strictly Prohibited Actions

File Structure