The Agent Skills Directory

[COMMAND_EXECUTION]: The bundled shell script scripts/peer-review.sh allows for the execution of arbitrary binaries. While intended to wrap specific LLM CLIs, the execute_cli function includes a fallback case that executes any command name passed as an argument if it is found in the system's execution path.
Evidence: In scripts/peer-review.sh, the execute_cli function contains a *) case that runs run_cmd "$cli" "$prompt". The $cli variable is derived directly from the arguments passed to the script's execute command by the agent.
[DATA_EXFILTRATION]: The skill transmits repository context, including git diffs and file contents, to well-known external LLM services (Gemini, Codex, and Claude) via their respective CLI tools for the purpose of code review.
Evidence: The workflow in SKILL.md and the prompt templates in reference/prompt-template.md explicitly instruct the agent to provide the external tools with access to the repository's git history and file contents.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and processing "raw responses" from external LLM tools. These responses are synthesized into a final report which then influences subsequent agent actions, such as auto-fixing code via other skills.
Ingestion points: The agent reads raw output from /tmp/*.txt files generated by external tools as described in SKILL.md (Step 3) and reference/cli-commands.md.
Boundary markers: Absent. There are no delimiters or instructions provided to the synthesis logic to ignore potential command-like instructions embedded within the peer LLM reviews.
Capability inventory: The agent has access to the Bash tool to execute scripts/peer-review.sh and perform repository modifications.
Sanitization: Absent. The skill catenates and processes raw external output for synthesis without filtering or escaping content.

consulting-peer-llms