The Agent Skills Directory

[COMMAND_EXECUTION]: The skill and its associated script (review-round.sh) execute several shell commands including git, gh, base64, jq, and bash to interact with the local repository and the GitHub API.
[DATA_EXFILTRATION]: The skill provides a workflow for frontend changes that involves taking screenshots and uploading them to a public repository (windmill-labs/agent-screenshots-internal). While the instructions explicitly warn against capturing sensitive information like secrets or workspace variables, the process inherently involves transmitting visual data to a public external destination.
[INDIRECT_PROMPT_INJECTION]: The automation script review-round.sh fetches and parses pull request comments to identify review verdicts. The logic for some reviewers (like Codex) relies on header matching rather than strictly validating the commenter's identity, which could allow an unauthorized user's comment to influence the agent's decision to promote a PR to 'ready' status.
Ingestion points: Pull request comments and review bodies fetched via gh api in review-round.sh.
Boundary markers: None; the script searches the entire body of fetched comments for specific substrings and regex patterns.
Capability inventory: The skill has the capability to modify pull request status (gh pr ready), post comments, and commit changes (ee-repo-ref.txt).
Sanitization: The script uses grep and jq to match specific status strings (e.g., "Good to merge", "Should address issues before merging") and performs partial validation on the commenter's login for certain bots (Claude, Pi, Cubic), but lacks strict validation for others.

pr