The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted data sourced from repository trackers (e.g., issues, comments, PR bodies). It identifies this attack surface and mitigates it by instructing the agent to treat all external repository content as data to be analyzed rather than instructions to be followed. It requires strict regex validation of external identifiers and advocates for fresh execution contexts to prevent cross-contamination of instructions.
[DATA_EXFILTRATION]: The instructions explicitly prohibit the agent from exfiltrating data, accessing credential stores, or pasting sensitive information (like tokens or environment variables) into logs or tracker comments. It defines a 'Secrets hygiene' rule that applies to all operations.
[COMMAND_EXECUTION]: The skill executes commands for building and testing code but restricts these activities to an isolated worktree. It provides specific constraints against performing remote or tracker operations during the implementation phase to ensure local-only execution.
[PROMPT_INJECTION]: A static detector flagged a reference to 'ignore previous instructions' in the setup guidelines. However, context shows this is a defensive instruction teaching the agent to recognize and report malicious directives found in untrusted repository data, rather than an attempt to bypass security filters.

om-fix-issue