The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes local shell commands (git diff, git show, git log, git blame) to extract code changes for analysis. Instructions explicitly restrict these operations to read-only information gathering and forbid any commands that would modify the repository state, branches, or stashes.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data in the form of code diffs and commit messages, which could potentially contain malicious instructions intended to influence the agent.
Ingestion points: External code changes retrieved via git commands (SKILL.md Step 1).
Boundary markers: The instructions provide a structured task template for reviewers using clear section headers (LENS, INTENT, SCOPE, RULES) to isolate instructions from the data being reviewed (SKILL.md Step 4).
Capability inventory: Subagents are strictly limited to read-only toolsets such as read_file, grep, and glob, or executed within platform-specific read-only sandboxes (SKILL.md Step 4).
Sanitization: The skill implements a mandatory verification phase where the main agent must manually open cited file locations to confirm subagent claims, effectively filtering out hallucinations or malicious instructions embedded in the diff (SKILL.md Step 5).

review-swarm