ACP 线程代理实操：在同一线程管理 coding harness 生命周期
问题/场景：希望把“run this in Codex”落地为线程内可控执行，避免主会话被长任务污染。前置条件：ACP runtime 已启用，且可管理子会话。实施步骤：1) 在目标线程 spawn 子代理；2) 用 send 持续派发子任务；3) 启用合并回复减少噪音；4) 记录历史做审计；5) 完成后清理 runtime。关键命令：`openclaw sessions_spawn`、`openclaw sessions_send`、`openclaw sessions_list`、`openclaw sessions_history`。验证方法：执行与回包均在同线程，清理后无僵尸会话。风险与边界：线程隔离会丢失跨线程上下文，需显式传参。来源归因：官方 X 发布。
XDiscovered 2026-02-27Author @openclaw
Prerequisites
- ACP-thread runtime is enabled and healthy.
- Operators can inspect session lifecycle and logs.
Steps
- Spawn an ACP sub-agent from the active thread with clear task scope.
- Dispatch iterative subtasks through send calls in the same thread.
- Enable coalesced replies to keep output readable and compact.
- Review session history for checkpoints and failure diagnostics.
- Cleanup runtime resources after completion.
Commands
openclaw sessions_spawnopenclaw sessions_sendopenclaw sessions_listopenclaw sessions_historyVerify
All sub-agent activity stays thread-bound and idle sessions are cleaned up.
Caveats
- Thread isolation requires explicit context forwarding for cross-thread dependencies.
- Aggressive spawning without cleanup can increase queue latency（需验证）.
Source attribution
This tip is aggregated from community/public sources and preserved with attribution.
Open original source ↗