The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted GitHub issue titles, bodies, and comments. It proactively addresses this risk with a dedicated 'Security policy' that instructs the agent to treat issue content strictly as data, not instructions. It mandates a silent 'Security abort' if prompt injection is suspected, preventing the agent from being manipulated into performing unauthorized actions or posting exfiltrated data.
[DATA_EXFILTRATION]: The skill strictly prohibits sending data to external services and forbids the use or modification of API keys and secrets. All network-related operations are confined to official GitHub repository interactions using gh and git for the intended purpose of issue resolution and PR creation.
[COMMAND_EXECUTION]: Bash usage is heavily restricted to prevent abuse. The skill prohibits command chaining (pipes, redirects, &&, ;), the use of non-allowlisted utilities like cat or find, and any attempts to bypass tool restrictions. It requires the use of specific platform-provided tools for file operations and workspace inspection.
[SAFE]: The skill demonstrates excellent security hygiene by establishing clear boundaries for data ingestion (GitHub issues and CI logs) and providing explicit guidance on least-privilege tool usage. The use of --body-file for GitHub commands is a notable best practice to prevent shell mangling and injection.

fix-issue