Openclaw Skill Gastown

Multi-agent coding orchestrator using Gas Town (gt) and Claude Code. Use for ANY non-trivial coding task — multi-file changes, new features, refactors, bug fixes, anything involving code that needs to compile/run/test. Delegates work to parallel Claude Code agents (polecats) with git-backed persistent state, work tracking (beads), and coordination. Use when a task involves more than a single file edit or quick script.

This skill is coherent with being a Gas Town orchestrator, but it contains several red flags you should consider before installing: - Missing declarations: The metadata lists no required binaries or env vars, yet SKILL.md and scripts require gt, bd, Go, claude CLI, tmux, wget, and access to ~/gt and git — confirm these dependencies and that you are comfortable granting filesystem/git access. - Arbitrary execution risk: The instructions explicitly tell the agent to execute any work it finds on hooks (the 'GUPP' rule). That means an agent using this skill can run build/test/merge steps and arbitrary repo code without further human approval. If you have sensitive repositories or secrets in your environment, this is risky. - Install action: setup.sh fetches Go from the official site and uses `go install` to pull gt and bd from GitHub. If you prefer control, run these steps yourself in an isolated environment (VM/container) instead of letting the agent run them. Recommended mitigations before enabling: 1. Review SKILL.md and scripts/setup.sh line-by-line and inspect any formulas or .beads files in repos the skill will touch. 2. Run setup steps manually in a disposable VM/container; do not allow the agent to run them autonomously. 3. Do not grant this skill access to repositories with secrets or production credentials. Create a dedicated test workspace (~/gt) for trial runs. 4. Consider disabling autonomous invocation (if your platform allows) so the agent prompts you before running any gt/bd commands. 5. Verify git credential usage (use least-privilege deploy keys or read-only clones where possible). 6. If you proceed, monitor the first runs closely and restrict the skill to user-invoked usage until you’re confident in its behavior. If you want, I can list the exact places in SKILL.md and scripts where commands will execute and what filesystem paths they touch so you can audit them more quickly.

Type: OpenClaw Skill Name: openclaw-skill-gastown Version: 0.1.4 The skill bundle is classified as suspicious due to the broad system access granted to the AI agent and the explicit instructions for immediate, autonomous execution of 'work on your Hook' (GUPP principle) found in `SKILL.md`. This design makes the agent highly susceptible to prompt injection if a malicious 'bead' or 'molecule' (work item) is introduced into the Gas Town system. While the `scripts/setup.sh` performs standard installations from legitimate sources (go.dev, github.com), the reliance on `go install @latest` introduces a supply chain risk. The skill itself does not demonstrate intentional malicious behavior like data exfiltration or backdoor installation, but its operational model presents significant security risks.