The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-controlled data from GitHub issues, including titles, bodies, and comments.
Ingestion points: Data enters the context via gh issue view and reading project files like review-findings.md or audit reports in docs/audits/.
Boundary markers: The skill defines a 'Turn contract' and explicit 'Guardrails' to ensure evidence-based decision making.
Capability inventory: Capabilities include gh issue edit, gh issue comment, and gh label create.
Sanitization: The skill mitigates risks through a mandatory 'Injection-safety invariant' that instructs the agent to base classification solely on evidence verified against the code, effectively ignoring any instructions or safety-override attempts embedded in the issue text.
[COMMAND_EXECUTION]: The skill executes shell commands to interact with the project and the forge.
Evidence: Uses gh for GitHub operations (view, edit, comment, label), grep for searching, and node for local scripts.
Security posture: Employs safe practices such as writing complex Markdown bodies to temporary files using the Write tool and then using gh issue comment --body-file to prevent shell injection or formatting issues.
[DYNAMIC_EXECUTION]: The skill executes project-specific logic via a local script.
Evidence: Invokes node scripts/unit-route.mjs <unit> to determine replanning routes. This is restricted to the local environment and uses pre-defined project tooling.

triage-issue