The Executor Protocol - OpenClaw Task Workflow
Six-Character Mantra: Plan first, then execute; report as you go.
⚡ Core Prohibitions
- Proceeding without task classification → ❌
- Spawning subagents without a plan → ❌
- Abandoning tasks on timeout → ❌
- Failing to respond to TOM within 10s → ❌
- Guessing without reading local code → ❌
- Answering based on assumptions, not facts → ❌
Task Classification
| Type | Definition | Strategy |
|---|
| ------ | ------------ | ---------- |
| Simple | Single step, no research needed | Execute directly |
| Medium | Multi-step, requires some research | Spawn 1-2 subagents |
| Complex | Multiple directions, heavy research | One subagent per direction, parallel |
Detailed Workflow
Phase 1: PLAN
- Assess task complexity
- List execution plan (steps and what each step does)
- Report plan to TOM and wait for confirmation (unless TOM says otherwise)
Phase 2: Execute + Report
- Simple: Execute directly, return result
- Medium/Complex: Spawn subagents for parallel execution
- Report progress as it happens, don't wait for completion
- Auto-retry on timeout, never give up
- Monitor actively, check status regularly
Phase 3: Report Results
- Report content/results to TOM
- Explain next steps if any
Subagent Standards (Universal)
Parallel Strategy
- 5+ independent directions → one subagent each, parallel
- 2-4 directions → serial or small-scale parallel
- Simple tasks → no subagent
Key Rules
- Auto-retry on timeout: Never give up on timeout
- Report as you go: Every 30-60s or on progress
- 5 max parallel: Wait if limit exceeded
- Active monitoring: Check status regularly, don't just wait
Plan→Spawn→Monitor→Report
- Plan: List plan, report, wait for confirmation
- Spawn: sessions_spawn, mode=run, runtime=subagent
- Monitor: sessions_history/list to check status, sessions_send to correct direction
- Report: Forward results immediately when received
OpenClaw Code Review Standards
When analyzing OpenClaw's own capabilities:
- Read docs/tools/exec.md and docs/tools/exec-approvals.md
- Verify exec tool availability
- Check Python library installation
- Review openclaw.json config (disabled tools)
Notes
-遇到问题 → Analyze cause first, try alternatives
- Sensitive operations → Must get authorization first
- Execution failure → Record reason, adjust strategy and retry
Changelog
- 2026-04-15 v1.0: Initial release - The Executor Protocol