The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is centered around executing external CLI tools to perform its primary function.
Evidence: instructions in SKILL.md direct the agent to invoke codex review, codex exec, and the claude CLI with various flags like -p and --allowedTools.
Evidence: The skill utilizes git diff and git show to extract code context for the review process.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by processing untrusted external data (code changes) and passing it to an LLM.
Ingestion points: git diff main...HEAD, git show <SHA>, and git diff output are ingested in SKILL.md patterns 1, 2, and 3.
Boundary markers: The prompts in references/prompts.md do not utilize robust delimiters or instructions to ignore embedded commands within the code under review.
Capability inventory: The commands suggested in SKILL.md grant the reviewer models capabilities such as Read, Glob, Grep, and Bash(git *) via the --allowedTools flag.
Sanitization: No sanitization or filtering of the code content is performed before it is sent to the reviewer CLI.

cross-model-review