The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources such as git history, CI logs, and GitHub issues/PRs, which could potentially contain malicious instructions intended to influence the agent's investigation.
Ingestion points: Git logs (git log), CI run output (gh run view --log-failed), PR/issue descriptions (gh pr view), and general project files.
Boundary markers: The skill does not define explicit delimiters for this data; however, it mandates an evidence-first approach where all claims must point to checkable sources (file:line, commit hash).
Capability inventory: The skill is restricted to read-only operations, including git and gh CLI tools, file reads, and the execution of side-effect-free tests.
Sanitization: There is no specific logic to sanitize or filter the content of logs or comments before they are analyzed by the agent.
[DYNAMIC_EXECUTION]: The skill instructions permit the agent to run existing project tests to verify behaviors or regressions.
Evidence: In the 'Gather evidence' section, the skill allows the agent to run tests specifically noted as 'only if they are side-effect free'. This involves executing code within the project context, though it is limited to pre-existing tests and scoped for read-only investigation.

ask-why