The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, specifically repository source code and pull requests, which could contain malicious or misleading instructions.\n
Ingestion points: The coordinator agent reads diffs, PR metadata, and surrounding code as defined in SKILL.md (Step 1).\n
Capability inventory: The skill utilizes the herdr CLI for environment management and agent coordination.\n
Boundary markers: Instructions enforce a read-only constraint and a specific coordinator marker (TEAM_CODE_REVIEW_COORDINATOR=1) to prevent recursive execution.\n
Sanitization: A dedicated verification phase (Step 3) requires the main agent to independently reproduce and verify all specialist findings, mitigating the risk of following instructions embedded in the analyzed code.\n- [COMMAND_EXECUTION]: The skill uses the herdr CLI to orchestrate the review workspace.\n
Evidence: references/herdr.md lists commands for pane management (herdr pane split), agent discovery (herdr agent get), and cross-pane communication (herdr agent prompt). These are used for internal coordination rather than executing arbitrary external scripts.\n- [DYNAMIC_EXECUTION]: The skill dynamically generates prompts for delegated agents at runtime.\n
Evidence: The coordinator constructs self-contained prompt strings incorporating dynamic environment variables, pane IDs, and repository paths to drive the multi-agent review logic.\n- [SAFE]: The skill promotes standard industry security and testing practices for application audits.\n
Evidence: references/rails.md recommends using established tools such as Brakeman for security scanning and RSpec for behavioral verification, which are standard developer tools for ensuring code quality.

team-code-review