The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes standard Git commands (including git status, git diff, git rev-parse, and git remote show origin) to establish the repository state for review purposes.
[COMMAND_EXECUTION]: The diff review procedure in references/diff-review.md instructs the agent to identify and execute verification commands defined in project-specific files such as CLAUDE.md, package.json, or Makefile. This represents an execution path for commands sourced from the project environment.
[INDIRECT_PROMPT_INJECTION]: The skill exposes an attack surface for indirect prompt injection by ingesting and processing data from the repository:
Ingestion points: Plan files (docs/plans/*), code diffs, and project memory files (CLAUDE.md, AGENTS.md).
Boundary markers: Absent; no delimiters are specified to isolate untrusted project content.
Capability inventory: File system write access (to memory and rules files) and shell command execution.
Sanitization: None; content from the repository is processed directly without filtering or escaping.
[INDIRECT_PROMPT_INJECTION]: The hygiene pass in references/memory-hygiene.md automates the distillation of "lessons" and user corrections into persistent rules stored in the agent's rules directory (e.g., .claude/rules/). This allows potentially untrusted input from a development cycle to influence the agent's long-term behavior and constraints.

stage-gate