The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard development tools including git, gh (GitHub CLI), and mise. These tools are used for their intended purposes: managing branches, fetching issue metadata, running local pre-commit hooks, and creating pull requests.
[EXTERNAL_DOWNLOADS]: The skill fetches issue details and remediation plans from GitHub. Interactions with this well-known and trusted service for project management are considered safe behavior.
[PROMPT_INJECTION]: The skill identifies an indirect prompt injection surface by processing instructions (remediation plans) from issue comments. This risk is effectively mitigated by several architectural safeguards.
Ingestion points: Untrusted data enters the context via issue comments fetched through the gh issue view command in SKILL.md.
Boundary markers: The skill validates the presence of the security-review-agent marker string to ensure the instructions originate from a trusted automated process.
Capability inventory: The skill can write code, create commits/branches, push to the remote repository, and execute project hooks via mise (as described in Steps 5 and 8 of SKILL.md).
Sanitization: A strict manual gate is enforced; the agent will only proceed if the topic:security and state:agent-ready labels have been manually applied by a human operator, preventing the processing of unvetted content.

fix-security-issue