The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes command-line tools such as git and the GitHub CLI (gh) to interact with source control and pull request data.
The babysit-pr.md and worktree-cleanup.md playbooks involve viewing, pushing, and deleting repository state.
Mitigation strategies are defined in investigation.md (restricting to read-only commands) and worktree-cleanup.md (preventing force-deletion of dirty worktrees without user consent).
[INDIRECT_PROMPT_INJECTION]: The babysit-pr.md playbook is vulnerable to instructions embedded in external content.
Ingestion points: The skill retrieves review threads and PR metadata via gh api graphql and gh pr view in references/babysit-pr.md.
Boundary markers: There are no specified delimiters or "ignore" instructions to prevent the agent from following commands found within PR review comments.
Capability inventory: The agent can perform code modification via the implement skill and update remote repository state via land.
Sanitization: The skill lacks sanitization or validation of the ingested PR content.
[DYNAMIC_EXECUTION]: The autonomous-run.md playbook establishes an autonomous loop for unattended execution.
It automates the execution of multiple rounds of the implement skill (code generation) based on a specification.
While it forbids non-undoable actions and mandates limits, the loop facilitates the autonomous execution of generated code changes.

pecut