Swarm Sprint

Parallel multi-agent coding sprints using git worktree isolation. Use when running 2+ coding tasks on a repository that touch different parts of the codebase...

This skill appears to do what it says (create isolated git worktrees, generate per-task agent packages, and coordinate merges), but take these precautions before using it on important repositories: - Review the full, untruncated scripts/swarm.js file before running. The provided copy is truncated in the manifest; any hidden code could change the risk profile. - Run in --dry-run or --plan-only first to see planned worktrees and groups without touching disk or branches. Inspect generated swarm-packages.json before spawning agents. - Only run on a local clone or a disposable environment (not directly on a production checkout). Prefer a sandbox or CI workspace and ensure you have a backup/clean branch to recover. - Validate and sanitize task inputs. The script builds branch names and filesystem paths using task.id and repo path; a malicious or malformed task id could cause unexpected branch names or file paths. Do not accept tasks from untrusted sources without validation. - Be cautious about cleanup fallback: the script attempts rm -rf on worktree paths if git removal fails. Do not run as root and confirm the computed worktree paths are safe before allowing cleanup to run. - Consider agent/network privileges: spawning multiple subagents increases the number of processes with access to repository content and any network egress those agents have. Limit agent network and secret access if possible. - Confirm the coordinator (human or automated) reviews diffs before merge; the skill's safety relies on that step. Use git push/policy controls rather than allowing automated pushes from worker environments. If you need higher assurance, ask the publisher for the full, untruncated script and a short security note describing path sanitization and any safety checks they perform.

Type: OpenClaw Skill Name: swarm-sprint Version: 1.0.1 The skill implements a parallel multi-agent coding workflow using git worktrees but contains a critical shell injection vulnerability in `scripts/swarm.js`. The script executes shell commands via `bash -c` using unsanitized input from the `tasks.json` file (specifically the `id` field) when creating and cleaning up worktrees. While the tool's logic appears intended for legitimate coordination and includes safety instructions for sub-agents in `SKILL.md`, the lack of input validation in command construction allows for arbitrary command execution if a task ID contains shell metacharacters.