The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the gh CLI to interact with GitHub pull requests. Execution is handled via child_process.execFile in scripts/generate_review_batch.mjs (for ingestion) and scripts/execute_decisions.mjs (for review submission). This is the primary purpose of the skill and is implemented using secure argument passing (arrays rather than shell strings) to prevent command injection.
[EXTERNAL_DOWNLOADS]: The skill interacts with external services including GitHub (via the CLI) and Busabase (via busabase-sdk). Both are established platforms, and the interaction is restricted to fetching pull request data and storing review metadata. The skill follows best practices by using environment variables (BUSABASE_API_KEY, BUSABASE_SPACE_ID) for authentication rather than hardcoded secrets.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it ingests untrusted data from GitHub pull requests (titles, bodies, and diffs).
Ingestion points: scripts/generate_review_batch.mjs fetches PR metadata and diffs using gh search prs and gh pr diff.
Boundary markers: None identified in the ingestion scripts.
Capability inventory: The skill can submit PR reviews (gh pr review) and write to the Busabase database.
Sanitization: The truncateText function in scripts/generate_review_batch.mjs limits the ingested PR body to 2000 characters and the summary to 600 characters, reducing the attack surface. Furthermore, the mandatory human-in-the-loop approval process in the AirApp UI serves as a significant mitigation against automated execution of injected instructions.

kelly-pr-review