跳转至

name: openclaw-task-delegation description: Deploy and delegate complex tasks to OpenClaw agent — wrapper scripts + task guide + handoff protocol, respecting OpenClaw's sandbox constraints and 8-tool-call limit. tags: - openclaw - delegation - agent-pairing - workflow

OpenClaw Task Delegation Protocol

Context

You (Hermes) work alongside OpenClaw（口水虾）, a subordinate agent with constraints: - 8 tool calls per session max — complex tasks get truncated - Sandbox-first — must write drafts to /tmp/openclaw_draft/, cannot modify Hermes files - Silent execution — doesn't ask questions, just does what it's told - Review-gate — any changes to server/Wiki must go through Hermes approval - 飞书 channel only — communicates via Feishu messages - Cannot create cron jobs — that's Hermes's domain

Your role: Hermes = supervisor/approver. OpenClaw = worker that executes well-defined tasks.

When to Delegate to OpenClaw

Any task where OpenClaw can be given a clear, single-line command and read back the result. Good candidates: - Running pre-installed tools/scripts - Reading output files and summarizing - Generating reports from structured data - Executing batch analyses

Deployment Steps (do this ONCE per tool)

1. Install the tool in a shared location

s

## Deployment Steps (do this ONCE per tool)

### 1. Install the tool in a shared location
```bash# Clone/install outside Hermes's directory, where OpenClaw can access
git clone <repo> /root/data/disk/<project-name>/
cd /root/data/disk/<project-name>/
pip install -e .

2. Write a wrapper script (critical!)

OpenClaw's 8-tool-call limit means it cannot do multi-step workflows. You must package the entire operation into a single bash command.

The wrapper script should: - Take simple arguments (stock code, date, mode) - Handle all setup internally (API keys, paths, error handling) - Output structured results to a predictable directory - Return a clear exit code

Script location: /root/data/disk/<project-name>/run_<task>.sh Output convention: /root/data/disk/<project-name>-results/<task-id>/

3. Write a task guide document

Place at /root/openclaw-learning/<project>-task-guide.md containing: - Project location and purpose - Exact commands OpenClaw should run - What each output file contains - Common issues and resolutions - Safety boundaries (what NOT to do)

4. Notify OpenClaw via Feishu

Send a message to OpenClaw's Feishu chat with: 

📢 **新任务上线通知**

你有了一个新能力：<description>

项目：/root/data/disk/<project-name>/
任务指南（必读）：/root/openclaw-learning/<project>-task-guide.md

你有两种模式：
- <模式1>：bash run_...sh <arg1> <arg2>
- <模式2>：bash run_...sh recommend <date>

请先读一遍任务指南熟悉项目。
模式： - <模式1>：bash run_...sh - <模式2>：bash run_...sh recommend

请先读一遍任务指南熟悉项目。 ``### 5. Update AGENTS.md (optional) If OpenClaw's AGENTS.md at/root/.openclaw/workspace/AGENTS.mdneeds a permanent role update, write a draft to/tmp/openclaw_draft/` and let OpenClaw handle it.

Handoff Protocol (per task)

User -> Hermes: "分析 688017"
Hermes -> OpenClaw (Feishu): "请分析 688017，执行：
bash /root/data/disk/<project>/run_analysis.sh 688017 [date]
读结果发报告给我审核"
OpenClaw -> (runs, reads) -> Hermes (Feishu): 报告内容
Hermes -> User: 审核后转发给用户

实际验证有效的派发方式：interagent.py API 调用

已验证可用（2026-05-14 Wiki 审查任务）： python3 /root/.hermes/scripts/interagent.py --to niouma "<任务描述>" --max-tokens XXXX

  • OpenClaw（niouma-doufu）通过 API Server（localhost:8643）接收任务
  • 任务消息必须自包含——OpenClaw 没有当前对话上下文
  • 较长的任务（grep 搜索 275 个文件等）容易 timeout（120s 限制）
  • 成功模式：写任务指南 → 写执行脚本 → 通过 interagent.py 派发单个明确命令

派发策略（经验总结）

任务复杂度 方式 示例
简单（读文件+报告） 直接在 interagent.py 消息里描述 "读取 X 文件，汇总清单"
中等（按规则执行操作） 先写执行脚本到 /tmp/openclaw_draft/，然后派发"读取并执行" archive_execute.sh
复杂（需要判断标准） 先写任务指南到 /root/openclaw-learning/，再派发"按指南执行" wiki-vrchat-audit-guide.md

已知限制

  • interagent.py 有 120-300s timeout——大数据集任务（275个文件 grep）容易超时
  • 超时后任务可能实际上已完成（只是没等到返回），可以依赖文件系统检查结果
  • 重复派发不影响结果——幂等操作可以安全重试
  • 不要一条消息塞太多任务——贴 40 个 mv 命令的脚本比在消息里列出更容易成功
  • 用文件传数据比在消息里传更可靠——长清单/脚本写到文件再让 OpenClaw 读取 以安全重试
  • 不要一条消息塞太多任务——贴 40 个 mv 命令的脚本比在消息里列出更容易成功
  • 用文件传数据比在消息里传更可靠——长清单/脚本写到文件再让 OpenClaw 读取### 旧方案对比（飞书群发/Control UI）
  • 飞书群发消息给 OpenClaw 不保证触发执行（不同 app_id）
  • 用户通过飞书私聊 @OpenClaw bot 是可行的，但 Hermes 无法自动触发
  • Control UI（localhost:18381/0qtaye/）可靠但需手动操作
  • interagent.py API 调用是目前最可靠的自动派发方式

Common Pitfalls

  • pyproject.toml format bugs — upstream projects often have malformed builds. The dependencies array must be under [project.dependencies], not under [project.urls]. Check before installing.
  • env vars — always check how the project loads API keys. The TradingAgents project reads MINIMAX_API_KEY from env directly (no dotenv in the main code path).
  • script syntax — OpenClaw's silent execution means it won't report minor shell errors. Always test the wrapper script yourself before delegating.
  • 8-call limit — OpenClaw cannot iterate, retry, or debug. Everything must work first-try. Test the full pipeline before telling OpenClaw.
  • OpenClaw cannot modify /root/.hermes/ — all project files must live under /root/data/disk/ or another non-Hermes path.