The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/inspect_pr_checks.py invokes the GitHub CLI (gh) using subprocess.run. The command arguments are constructed as a list, which prevents shell injection. Furthermore, dynamic inputs like PR numbers are converted to integers, and run IDs are validated against a numeric regex pattern before being passed to the command.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it retrieves and processes GitHub Action logs, which are external and potentially attacker-controlled (e.g., in a public repository PR).
Ingestion points: The extract_failure_snippet function in scripts/inspect_pr_checks.py reads logs via gh run view --log.
Boundary markers: The extracted log snippets are presented to the agent in the FailureSummary without explicit boundary delimiters or "ignore embedded instructions" warnings.
Capability inventory: The skill instructions in SKILL.md grant the agent the capability to propose fix plans and implement code changes based on these logs.
Sanitization: The script performs no sanitization or instruction-filtering on the logs; it only uses a regex to locate error lines for snippet extraction.

github-fix-ci