The Agent Skills Directory

[Prompt Injection Defense]: The skill incorporates explicit logic to detect and ignore imperative framing in external data (e.g., issue or PR bodies). By instructing the agent to treat external content solely as data and specifically providing examples like 'ignore previous instructions' to watch for, the skill minimizes the risk of indirect prompt injection. The static detection of these strings is a false positive related to this defensive instruction.
[Propose-then-Confirm Model]: A core security control in this skill is its read-only nature. It generates a text-based routing proposal and provides the maintainer with a manual command to execute if they agree. This prevents the agent from making unauthorized modifications to the project's tracker or repository state.
[Scoped Command Execution]: The skill utilizes standard tools like the GitHub CLI (gh) and git to collect metadata and contribution history. These commands are localized to the project environment and are used as intended for analysis without accessing sensitive system paths or credentials.
[Privacy and Configuration Validation]: The inclusion of a pre-flight privacy check (privacy-llm-check) ensures that the agent context remains within project-approved boundaries before processing potentially sensitive PII in contributor comments.

reviewer-routing