The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill orchestrates the review of external data from git diffs and project documentation, which presents a surface for indirect prompt injection.
Ingestion points: The skill reads output from git diff, git diff --cached, and local project files such as AGENTS.md and architecture documentation (SKILL.md).
Boundary markers: The instructions include explicit constraints requiring sub-agents to be "read-only" and specifically forbids them from executing state-mutating actions like apply_patch, staging changes, or committing (SKILL.md Step 2).
Capability inventory: The skill uses git diff for ingestion and identifies apply_patch as a forbidden tool for sub-agents.
Sanitization: The instructions do not specify sanitization or escaping for the ingested diff content before it is processed by the sub-agents.
[COMMAND_EXECUTION]: The skill utilizes standard, read-only git commands to determine the scope of changes for the review process.
Evidence: Instructions specify the use of git diff and git diff --cached (SKILL.md Step 1).

review-swarm