The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection by processing GitHub Actions logs, which can contain untrusted content from external contributors.
Ingestion points: Log content is retrieved in scripts/inspect_pr_checks.py using gh run view --log and the GitHub API.
Boundary markers: The skill does not use specific delimiters to isolate log data, increasing the risk that the AI may interpret parts of the logs as instructions.
Capability inventory: The agent is empowered to implement code fixes as per the instructions in SKILL.md, which includes file-writing and command execution capabilities.
Sanitization: No sanitization is performed on the log snippets to filter out natural language instructions before they are passed to the agent.
[COMMAND_EXECUTION]: The skill's operation requires executing local system commands using gh and git. This usage is consistent with the skill's purpose and is performed via the agent's Bash tool. The Python script scripts/inspect_pr_checks.py uses argument lists in subprocess.run to handle parameters safely, mitigating common shell injection risks, although the agent retains high capability to execute code when implementing fixes.

gh-fix-ci