code-review-loop
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from git diffs and untracked files in the repository under review. This data is interpolated directly into the prompts for both the Codex reviewer and the Claude author models, creating a surface for malicious code to influence model behavior.
- Ingestion points: Found in
resources/review-code.mjswithin thesnapshot()function which callsgit diff, and inresources/prompts/codex-review.mdwhere the{{DIFF}}variable is interpolated. - Boundary markers: Content is enclosed within
<diff>tags in the prompt templates (e.g., incodex-review.mdandclaude-address.md). - Capability inventory: The agent has access to
Bash(for running the check command),Edit,Write,Read,Glob,Grep, andWebFetchtools. - Sanitization: No explicit filtering or sanitization of the diff content is performed before it is added to the model context.
- [COMMAND_EXECUTION]: The skill executes a command specified by the user through the
--checkargument. This command is executed viash -cand is also available to the Claude model to run during the review loop. While the skill documentation mentions sandboxing (usingsandbox-execorbwrap), the integrity of this confinement depends on the externalagent-orchestratelibrary. - [EXTERNAL_DOWNLOADS]: As described in
SKILL.md, the skill may usepnpm dlxto download and execute version-pinned CLI tools if they are not available on the host system, which involves fetching external code from the npm registry at runtime.
Audit Metadata