The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted source code and diffs which are passed to sub-agents for analysis. An attacker could embed malicious instructions in code comments to influence sub-agent behavior. \n
Ingestion points: SKILL.md (Step 1: Determine the Scope). \n
Boundary markers: The skill instructs sub-agents to confine prompts to the review scope. \n
Capability inventory: spawn_agent, git worktree, npm install. \n
Sanitization: Isolated $TMPDIR worktrees and process groups with timeouts. \n- [REMOTE_CODE_EXECUTION]: The skill instructions include running project-specific build and test commands (e.g., npm install, build the project) which executes code provided in the untrusted repository being reviewed. \n- [COMMAND_EXECUTION]: The skill utilizes shell commands (git, gh, spawn_agent) to orchestrate the review process, including repository metadata extraction and parallel agent management. \n- [DYNAMIC_EXECUTION]: The skill dynamically generates environments using git worktree and git checkout-index to perform mutations and execute tests on code branches.

review-code