The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the wt, git, and gh command-line interfaces. Security is maintained through mandatory user confirmation steps before the agent performs merges, pushes, or configuration updates. It specifically prohibits the use of 'force' or 'auto-approve' flags unless explicitly requested by the user.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted data by deriving automation hooks from project files such as package.json, Makefile, pyproject.toml, or Cargo.toml.
Ingestion points: package.json, Makefile, pyproject.toml, Cargo.toml (detected in SKILL.md).
Boundary markers: Absent (instructions rely on LLM interpretation).
Capability inventory: Shell command execution via wt hooks and standard CLI tools.
Sanitization: Absent, but the risk is mitigated by instructing the agent to explain proposed commands and requiring manual user approval (wt config approvals add) before execution.
[SAFE]: The skill incorporates defensive instructions, such as avoiding hardcoded paths by using tool-provided metadata and ensuring that checks/tests are run before any code is proposed for merging.

worktree-workflow