The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code diffs and repository metadata to perform AI-powered reviews. This creates a surface where malicious code in the repository could attempt to influence the agent's review findings or behavior.
Ingestion points: The skill reads local work, committed branches, and PR candidates as described in SKILL.md.
Boundary markers: The SKILL.md file contains explicit warnings that "Findings are advice to verify, not instructions to apply blindly."
Capability inventory: The skill interacts with the local file system and invokes external CLI tools such as git and trufflehog (evidenced in scripts/autoreview_test.py).
Sanitization: A mandatory TruffleHog scan is performed on all outgoing review data to prevent accidental leakage of secrets or sensitive data (documented in SKILL.md).
[COMMAND_EXECUTION]: The skill and its accompanying test utilities execute shell commands to manage Git repositories and interface with AI service CLIs.
Evidence: scripts/test-review-harness.py uses subprocess.run to initialize Git and execute the review script. tests/test_codex_sandbox.py executes the codex sandbox command to verify isolation.
[DYNAMIC_EXECUTION]: The test suite utilizes dynamic module loading to verify the functionality of the main review script.
Evidence: scripts/autoreview_test.py uses SourceFileLoader and importlib.util to execute the autoreview script for testing purposes.

openclaw-autoreview