The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell scripts for routing logic and dispatching reviews. Specifically, it executes rubric-path.sh, resolve-route.py, and codex-dispatch.sh from a local plugin cache directory. These are internal tool operations designed to manage the review workflow and do not involve arbitrary user-supplied command execution.
[REMOTE_CODE_EXECUTION]: While the skill dispatches tasks to external 'Codex' executors or providers, it does so using a defined adapter pattern with a read-only sandbox (--sandbox read-only) and restricted approvals (--approval never). It interacts with these executors through structured JSON responses and exit codes rather than executing raw remote strings.
[DATA_EXFILTRATION]: The skill explicitly instructs to 'not include secrets or unbounded logs' in the prompts sent to external reviewers, mitigating the risk of credential exposure during the review process.
[PROMPT_INJECTION]: The skill implements integrity checks by requiring a fixed_target (commit SHA) and verifying that the target still matches its fixed point after the worker returns, protecting against manipulation of the code being reviewed.

review