The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub pull requests, including diffs, review comments, and issue comments. This represents a vulnerability surface where malicious instructions could be embedded in the code or comments to influence the agent's behavior.
Ingestion points: The skill reads PR content via gh pr diff, review bodies from repos/{owner}/{repo}/pulls/N/reviews, and comment text from pulls/N/comments and issues/N/comments.
Boundary markers: The instructions explicitly require re-anchoring and re-verifying findings against the PR head to ensure claims are grounded in the actual code.
Capability inventory: The agent has permissions to execute gh CLI commands, manage git worktrees, commit changes, and merge pull requests.
Sanitization: Mitigation is implemented through Step 4, which uses sub-agents (leek-skills:finding-verifier) to verify findings in parallel and requires the model to re-verify findings deemed "LEGIT" before taking action.
[COMMAND_EXECUTION]: The skill uses the Bash tool extensively to execute GitHub CLI (gh) and git commands. These commands are used to view metadata, retrieve diffs, check CI status, manage local worktrees, commit fixes, and merge branches.
[DYNAMIC_EXECUTION]: The skill dynamically dispatches tasks to other vendor-specific sub-agents, such as leek-skills:pr-reviewer and leek-skills:finding-verifier, to parallelize the review and verification processes.

triage-github-pr