The Agent Skills Directory

[COMMAND_EXECUTION]: The codex-review.sh script manages the review process by invoking the codex CLI tool.
[PROMPT_INJECTION]: The skill uses specific instructions to define the reviewer's persona and requires the local agent to adhere to the resulting verdict mechanically. These are architectural controls designed for quality assurance rather than malicious bypass attempts.
[PROMPT_INJECTION]: The skill processes untrusted input in the form of code diffs and implementation plans, presenting a potential surface for indirect prompt injection.
Ingestion points: The skill ingests repository code and implementation plans through codex-review.sh arguments and git diff commands.
Boundary markers: The prompt construction in the shell script uses specific section headers (e.g., 'DIFF UNDER REVIEW:') to isolate external data from the instructions.
Capability inventory: The skill provides the external model with the ability to execute workspace commands (such as tsc or vitest) via the codex tool.
Sanitization: The script implements a 'judge immutability' check (tree_fingerprint) that validates the integrity of the workspace using git status and HEAD hashes before and after execution to prevent unauthorized modifications.
[COMMAND_EXECUTION]: The external reviewer is explicitly instructed to execute local verification tools to provide evidence-based findings. The skill documentation states that this execution relies on sandboxing (bubblewrap) provided by the external codex tool environment.

codex-review