结构化任务规划与分步执行 V2（异步子代理架构）

结构化任务规划与分步执行 V2（异步子代理架构）。支持将每个步骤通过 session_spawn 创建子代理异步执行，主会话保持非阻塞。功能包括：步骤分解、子代理执行、子代理检验（LLM判断）、状态跟踪、Heartbeat 监控、任务中断。触发词：/stp、任务规划、步骤执行。

This skill appears to do what it says: orchestrate asynchronous subagents, track status, and create heartbeat cron jobs. Before you run it: (1) review scripts/stp_orchestrator.py to confirm you understand what files it will read/write and how it determines subagent state; (2) be aware it reads platform session logs (~/.openclaw/agents/.../sessions.json and session jsonl files) — those may contain recent tool outputs or messages you consider sensitive; (3) note it will create cron jobs under your account for periodic heartbeat checks — ensure you're comfortable with that and know how to remove them (the skill provides interrupt commands); (4) address the small metadata mismatch (registry showed no required binaries while _meta.json lists 'openclaw'); (5) run first in a non-production or isolated environment if you want to observe behavior before trusting it with sensitive workflows.

Type: OpenClaw Skill Name: stp Version: 1.0.7 The skill is classified as suspicious due to a critical shell injection vulnerability identified in the `SKILL.md` instructions. Specifically, the agent is instructed to extract a Process ID (`details.pid`) from a subagent's execution history (via `sessions_history`) and then directly use this value in a `kill <PID>` command. If a malicious subagent could manipulate its output to inject arbitrary commands into the `details.pid` field (e.g., `1; rm -rf /`), the main agent would execute these commands on the host system, leading to Remote Code Execution (RCE). While the intent is for cleanup, this lack of input sanitization for a value sourced from a potentially untrusted subagent constitutes a significant vulnerability. Other risky capabilities, such as executing user-provided prompts in subagents, are inherent to the skill's purpose and are mitigated by user confirmation and verification steps.