The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates code reviews by invoking local AI CLI binaries (e.g., claude, aider, codex) via spawnSync. This is an intended feature that allows for cross-model verification of local code changes.
[PROMPT_INJECTION]: To protect against malicious instructions embedded in reviewed code, the skill implements a 'prompt-shield' in scripts/prompt-shield/compose.mjs which wraps untrusted content in salted XML tags with an anti-injection preamble.
[DATA_EXFILTRATION]: While the skill transmits local code/diffs to other local CLIs (which may send data to cloud providers), it implements a 'secret-shield' in scripts/secret-shield/ that scans for and redacts a wide range of sensitive credentials (AWS keys, GitHub tokens, JWTs, private keys) before transmission.
[SAFE]: The implementation follows industry security best practices, including:
Path Hijacking Prevention: Resolving binaries to absolute paths using locateBinary before execution.
Environment Isolation: Using a fail-closed environment allowlist (buildChildEnv) to prevent sensitive environment variable leakage to child processes.
Transparency: Providing clear warnings about the visibility of data in the system process list for CLI configurations that do not support stdin.
Consent: Requiring explicit user confirmation before executing cross-model or parallel reviews.

code-reviewer