The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill executes load-time shell commands to retrieve repository state, including branch names, worktree paths, and Pull Request metadata. It also identifies hosted environments (such as Codespaces or Conductor) to conditionally adjust its teardown logic. These are diagnostic actions intended to provide the agent with situational awareness.\n- [COMMAND_EXECUTION]: The skill automates Git and GitHub CLI operations to merge code, delete local/remote branches, and manage worktrees. It utilizes a bundled utility script (sidecar.sh) for project-specific configuration management. All destructive operations are protected by human-in-the-loop confirmation gates and safety precondition checks.\n- [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection via the following:\n
Ingestion points: Processes external data from GitHub PR bodies (gh pr view) and commit messages (git log).\n
Boundary markers: Utilizes SHA-based 'review currency' stamps and a 'footprint' scanner to distinguish between user content and agent-generated instructions.\n
Capability inventory: Includes destructive actions like branch deletion, worktree removal, and merging to GitHub, as well as writing to the ~/.claude/sidecars filesystem via sidecar.sh.\n
Sanitization: Employs jq for robust data extraction and regular expressions to filter out potential command keywords from ingested text.\n- [DATA_EXFILTRATION]: The skill manages project state in a 'sidecar' directory structure within the user's home directory. It reads and writes configuration and research metadata to these local paths (~/.claude/sidecars and ~/.claude/research). No unauthorized network exfiltration was found; network activity is restricted to standard Git and GitHub operations.

closeout