The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from GitHub PR review threads, discussion comments, and CI logs to automatically address feedback and fix code. * Ingestion points: Data is fetched from GitHub APIs and Actions logs via scripts/fetch-feedback.sh and scripts/fetch-checks.sh. * Boundary markers: No explicit delimiters or 'ignore' instructions are used when passing this data to the agent. * Capability inventory: The agent can write to the filesystem, push commits to remote repositories, resolve PR threads via API, and execute local shell commands. * Sanitization: The instructions explicitly advise using files or stdin for shell commands to prevent the agent from accidentally executing backticks or variables contained in the comment text.
[COMMAND_EXECUTION]: The skill instructs the agent to execute repository-specific tools and test suites, including pytest, vitest, jest, and prek. These commands are run locally to verify fixes.
[EXTERNAL_DOWNLOADS]: The skill's documentation suggests fetching third-party CI logs directly from external URLs if they are public. This involves network requests to URLs potentially provided by untrusted external services.

address-pr-feedback