The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple local bash scripts (ghchk, wtinit, wtadd, wcp, wtcmt, wtpush, wtpr, wtciwait, wtstep) to automate complex git and GitHub CLI workflows. While these are essential to the skill's function, they represent a significant execution surface.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection during Step 4 (Committing) and Step 6 (Creating Pull Request).
Ingestion points: The agent reads full file diffs via git diff <start-branch>...HEAD and the user's original request string.
Boundary markers: Absent. There are no delimiters or instructions used when reading code changes to prevent the LLM from obeying instructions embedded in code comments or strings.
Capability inventory: The agent can write persistent data to the repository via wtcmt (commits) and wtpr (PR creation), and push those changes to a remote server via wtpush.
Sanitization: None. The agent is instructed to produce an objective representation of the diff, which could be manipulated by adversarial content within the diff itself.
[COMMAND_EXECUTION]: The wtpush script temporarily modifies the repository's origin remote URL to an SSH format to facilitate pushing in different environments. Although it uses a trap to restore the original URL upon exit or interruption, this behavior modifies the repository's global configuration during the operation.

worktree-pr