The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest a 'packet' containing code changes and evidence, and to include this text directly in the prompt of a reviewer agent. This creates a surface for indirect prompt injection if the reviewed code contains malicious instructions meant to override the reviewer's logic.
Ingestion points: The 'packet' output generated by gate.py brief (referenced in SKILL.md).
Boundary markers: Absent; instructions suggest including the raw text directly in the prompt.
Capability inventory: The skill can execute gate.py and invokes sub-agents (reviewer and refuter) which may have their own tool capabilities.
Sanitization: No sanitization or escaping of the packet content is described.
[COMMAND_EXECUTION]: The skill relies on the execution of local shell scripts and Python utilities to perform its core functions.
Evidence: SKILL.md invokes gate.py brief and gate.py record. The file scripts/review-guard.sh is a shell script that performs file system operations (cp, mkdir, rm) and executes git status to manage project snapshots.

second-opinion