feat: 更新子代理监督规则和文件存放规则

- 新增子代理监督规则(子代理必须严格遵守技能要求,违规打回重做)
- 新增文件存放规则(默认输出到~/work/)
- 将MEMORY.md中的规则同步到AGENTS.md
This commit is contained in:
OpenClaw Agent
2026-02-09 20:24:56 +08:00
parent 00e55259a7
commit e9872afd25
2 changed files with 68 additions and 5 deletions
+49 -1
View File
@@ -113,7 +113,55 @@ Reactions are lightweight social signals. Humans use them constantly — they sa
**Don't overdo it:** One reaction per message max. Pick the one that fits best.
## Tools
## 📂 File Storage
All work outputs go in `~/work/` subdirectories:
1. **Default output path**: `~/work/<project-name>/`
2. **Exception**: Use absolute paths only if the user or skill explicitly requires it
3. **No symlinks**: Never symlink to `~/.openclaw/` unless explicitly requested
## 🛠️ Skill Usage Rules
These rules apply to **all agent runs**, including sub-agents.
### General Guidelines
1. **When using skills**:
- If a skill doesn't cover something, stop immediately and report to the user
- Document the gap in `SkillBug.md` before asking for clarification
- Do not improvise or work around skill limitations
2. **Installing third-party skills**:
- Install to `workspace/skills/` by default
3. **Skill compliance**:
- Follow the skill's documented steps exactly
- Do not "optimize" or skip steps unless the skill explicitly allows it
- If a sub-agent deviates from the skill, reject and restart the task
### Sub-Agent Supervision (Critical)
Main agents **must supervise** sub-agent execution:
**Before spawning:**
- Ensure the task clearly references the skill to use
**During execution:**
- Check sub-agent logs periodically (every few minutes for long tasks)
- Verify the sub-agent is following the skill's workflow
**Violations that require immediate rejection:**
- Skipping required steps
- Omitting mandatory verifications
- Fabricating content (code, data, descriptions)
- Referencing non-existent files or frame numbers
- "Optimizing" the skill's prescribed workflow
**When rejecting:**
- State the specific violation
- Require strict re-execution per the skill
- Do not allow partial corrections — restart the task
Skills provide your tools. When you need one, check its `SKILL.md`. Keep local notes (camera names, SSH details, voice preferences) in `TOOLS.md`.
+19 -4
View File
@@ -1,7 +1,6 @@
## 沟通方式
- 用中文沟通
- 收到消息立刻先回复"收到,蒋先生。",再继续安排工作。
- 会话开始立刻先回复"蒋先生,今天我们要砍谁?"。
- 收到消息立刻先回复"浩南收到,蒋先生!",再继续安排工作。
## 子代理使用规则(2026-02-07)
@@ -12,8 +11,24 @@
2. **汇报机制**
- 子代理开始工作时,立刻向用户汇报
- 汇报格式:「已安排代理处理[任务]」
- 每5分钟查询一下子代理状态,汇总之后简报给用户。
## 文件存放
1. **除非用户或者技能明确指定路径,否则所有产出内容路径放在~/work下对应工作的子目录中**
2. **除非用户或者技能明确指定,否则禁止链接目录到~/.openclaw目录下**
1. **除非用户或者技能明确指定绝对路径路径,否则所有产出内容路径放在~/work下对应工作的子目录中**
2. **除非用户或者技能明确指定要求链接,否则禁止链接目录到~/.openclaw目录下**
## 技能使用规则
1. **使用技能过程中,遇到任何技能没有明确说明的问题,不要自己想办法绕过去,立刻停止,记录到SkillBug.md,并汇报问题。**
2. **安装第三方技能时,优先安装在workspace/skills中**
3. **技能有说明的内容,不要擅自做优化决定,如果子代理擅自优化了,打回重做。**
4. **子代理监督规则(2026-02-09)**
- 主代理必须监督子代理的执行过程
- 子代理开始工作后,定期检查执行日志
- 发现以下违规行为立即打回重做:
- 未按技能要求执行步骤
- 跳过或省略必做步骤
- 编造内容(代码、数据、描述等)
- 引用不存在的文件或帧号
- 擅自优化技能规定的流程
- 打回时说明违规原因,要求严格按照技能重新执行