The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates the execution of local development tools (git, gh) and AI model CLIs (codex, claude-code, amp, pi, kimi) to perform its code review tasks. It resolves these binaries strictly from the system PATH and excludes entries from within the reviewed repository to prevent execution of malicious binaries.
[DATA_EXFILTRATION]: The skill is designed to send repository diffs and untracked files to external AI service providers for review. To mitigate the risk of secret leakage, it performs a mandatory TruffleHog scan of the outgoing review pack and omits known sensitive paths (e.g., .env, tokens/session.dat) during bundle construction.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code, which presents a surface for indirect prompt injection.
Ingestion points: Reads code changes and untracked files from the repository being reviewed as documented in scripts/autoreview.
Boundary markers: Uses clear section headers in the generated review bundle (e.g., '# Change Bundle') and enforces aggregate prompt limits to maintain context.
Capability inventory: Executes local CLI tools via subprocess calls to facilitate code analysis and review generation.
Sanitization: Implements environment isolation (stripping process-injection variables), process-group termination for timeouts, and secret scanning to prevent malicious content from compromising the agent environment.

autoreview