The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs Git operations and executes user-defined test suites using subprocess.run (e.g., in scripts/test-review-harness.py and tests/test_autoreview_hardening.py). It implements environment isolation and temporary isolated roots to maintain system integrity during these operations.\n- [DATA_EXFILTRATION]: The skill manages authentication for various AI engines and provides a mechanism to stage "Blacksmith" credentials for remote test delegation. It includes robust internal logic, verified in tests/test_autoreview_hardening.py, to prevent secrets and sensitive configuration files from being passed to external analysis engines.\n- [REMOTE_CODE_EXECUTION]: The skill interfaces with external AI engine CLIs, such as Codex and Claude Code, to perform code analysis. It enforces isolation through specific configuration and flags (e.g., --safe-mode for Claude, isolated workspaces for Codex) to ensure analysis remains confined to the provided code bundle.\n- [PROMPT_INJECTION]: As a tool that processes untrusted code diffs, the skill is architected to resist indirect prompt injection. It instructs the agent to treat review findings as advisory and uses structural validation (e.g., extract_json in scripts/autoreview_test.py) to ensure the reviewer engine's output is correctly formatted and safe to process. Every indirect prompt injection surface follows a strict evidence chain:\n
Ingestion points: Reads git diffs and untrusted code changes via local, branch, or commit modes.\n
Boundary markers: Explicit instructions define review findings as advisory only and mandate human verification.\n
Capability inventory: Uses subprocess calls for Git and testing, and writes atomic JSON reports.\n
Sanitization: Implements secret scanning for diffs and terminal control sequence escaping for all output.

autoreview