The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes git, gh, and LLM CLI engines (codex, claude, pi) via sub-processes for its primary functions. The optional --parallel-tests feature allows execution of arbitrary user-supplied test commands; risk is mitigated by running these in an isolated temporary directory with a sanitized environment allowlist.
[EXTERNAL_DOWNLOADS]: The skill transmits code review bundles to established LLM service providers (OpenAI, Anthropic) and interacts with GitHub. These network operations are intended and target trusted infrastructure.
[DATA_EXFILTRATION]: The skill includes a 'fail-closed' security mechanism that automatically scans patches and untracked repository files for secrets, credentials, and sensitive paths (e.g., .env, ~/.ssh), blocking transmission if risks are found.
[PROMPT_INJECTION]: The skill manages potential indirect prompt injection from repository data through several controls.
Ingestion points: Local and branch git diffs, untracked files (processed in SKILL.md), and prompt/dataset files (validated in scripts/autoreview_test.py).
Boundary markers: Present. The skill freezes a baseline scope and provides explicit agent instructions in SKILL.md to treat LLM outputs as advisory.
Capability inventory: Spawning of git, engine CLIs, and shell test commands (identified in scripts/test-review-harness.py and tests/test_autoreview_hardening.py).
Sanitization: Present. The tool redacts secrets and sensitive file paths from diffs and strips potentially dangerous environment variables like LD_PRELOAD and NODE_OPTIONS before engine invocation.

autoreview