The Agent Skills Directory

[COMMAND_EXECUTION]: Reviewed subprocess.run calls in scripts/github-pr.py, scripts/fetch_comments.py, and scripts/inspect_pr_checks.py. All calls use shell=False and strictly validate arguments using regular expressions for repository slugs, branch names, and PR identifiers, effectively preventing command injection.\n- [PROMPT_INJECTION]: Evaluated references/evals.yaml which contains text patterns flagged as prompt injection. Analysis confirms these are adversarial test cases for benchmarking the agent's resistance to such attacks and do not constitute instructions for the agent.\n- [DATA_EXFILTRATION]: The skill instructions explicitly mandate the redaction of secrets, tokens, and credentials. Network activity is restricted to authenticated GitHub operations via the gh CLI.\n- [REMOTE_CODE_EXECUTION]: The test command in scripts/github-pr.py executes repository-local build and installation scripts. This functionality is restricted to the local project context and uses validated package manager identification.

gh-workflow