The Agent Skills Directory

[Command Execution]: The skill instructs the agent to execute shell commands like git diff and grep to identify changes and verify test coverage. While these are standard development workflows, they involve executing logic on the host machine.
[Indirect Prompt Injection]: The skill processes untrusted external data in the form of source code and git diffs. If this content contains malicious instructions designed to influence the AI's behavior, it could potentially affect the review's integrity.
Ingestion points: Reads git diff output and source code files during analysis steps (Step 1 and Step 2).
Boundary markers: Uses Markdown headers and specific prompt templates to delimit the input code from the agent's instructions.
Capability inventory: Includes file reading, execution of git/grep commands, and writing checkpoint files to the .copilot/review/ directory.
Sanitization: Relies on the agent's internal safety boundaries; no explicit sanitization of the input code is performed before analysis.
[Local Persistence]: The skill implements a checkpointing mechanism that writes markdown files to a hidden directory (.copilot/review/). This is used to maintain state across sessions but involves writing data to the local file system.

code-review