Skill Pilot

Adaptive skill scheduling engine with environment-aware routing, user preferences, and self-learning optimization. Use when: (1) auto-select best skill for t...

What to consider before installing: - Trust boundary: SkillPilot loads/imports other skill modules (via importlib) and executes skill scripts. Imported Python modules run in-process (they can execute arbitrary code), and subprocesses inherit the full environment by default. Only install SkillPilot if you trust the other skills present in your workspace. - Secrets exposure: Although SKILL.md states it does not exfiltrate secrets, the engine inherits environment variables and passes them to child processes. If you have high-value credentials in your environment (AWS keys, Cloud tokens, SSH keys, OPENCLAW_TOKEN, etc.), consider removing them or running SkillPilot in an isolated environment (container/VM) or using a minimal set of env vars. - Audit before use: Review the skill-pilot code paths that import and execute other skills (engine._load_skill, SkillExecutor/OpenClawCaller) and confirm input validation is active in your shipped version. The repository documents fixes for past shell-injection issues — verify those fixes are present in the actual code you will run. - Limit network surface: The skill performs active network probes (DNS, TCP connects, port scans). If you are in a sensitive environment, run it in a restricted network or review/disable probing. - Least privilege: Prefer running SkillPilot in a sandbox (Docker, VM) and avoid running it with elevated OS permissions. Remove or scope sensitive env vars before running, and periodically review the ~/.openclaw/workspace/skills/skill-pilot/ history/config files for unexpected activity. If you want, I can point to the exact code lines that import other skills and that pass os.environ to subprocesses so you (or an auditor) can inspect them further.

Type: OpenClaw Skill Name: skill-pilot Version: 0.4.6 SkillPilot is a sophisticated meta-scheduling framework designed to optimize the selection and execution of other OpenClaw skills. While it performs high-risk actions such as network probing (DNS/latency checks in `environment.py`), reading other skill configurations in the workspace (`registry.py`), and executing local scripts (`engine.py`), these behaviors are transparently documented and strictly aligned with its stated purpose of environment-aware routing. The codebase demonstrates security maturity by including an input validator (`_validate_args`) to prevent shell injection, utilizing list-based subprocess calls instead of `shell=True`, and providing detailed security audit documentation (`SECURITY_FIX.md`, `SECURITY_DECLARATION.md`).