review-pr
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [SAFE]: The skill demonstrates exceptional security maturity through its implementation of strict trust boundaries and hardening measures. Its helper scripts (
scripts/reviewctl.mjsandscripts/snapshot.mjs) use secure practices such as atomic file writes, constrained path resolution, and explicit permission management. - [INDIRECT_PROMPT_INJECTION]: The skill identifies pull-request content as untrusted data and implements multiple defensive layers to prevent prompt injection or accidental execution of malicious code.
- Ingestion points: Pull request diffs (
pr.diff), file snapshots from the HEAD and base commits, and metadata from GitHub and Linear tasks. - Boundary markers: The instructions explicitly define a 'Trust boundary,' categorizing all PR code as untrusted. It uses structured Markdown headers and fences to separate untrusted inputs from the agent's logic.
- Capability inventory: The skill uses
git,gh, anddocker. It allows runtime checks in a dedicated Docker sandbox but strictly forbids the execution of repository code on the host machine. - Sanitization: The
snapshot.mjsscript performs thorough sanitization of repository snapshots by stripping executable bits, setting strict read-only permissions (0o444), and converting symlinks and gitlinks into inert text files to prevent path traversal or symlink attacks. - [COMMAND_EXECUTION]: The skill executes local helper scripts and standard CLI tools to manage the review lifecycle.
- Evidence: It invokes
git,gh,docker,tar, and a context-retrieval tool calledpolygraphvianode:child_process. - Context: Command arguments are passed as arrays rather than shell strings, mitigating command injection risks. The operations are limited to data acquisition (
git fetch,gh api), resource cleanup (docker prune), and state management. - [EXTERNAL_DOWNLOADS]: The skill fetches repository data and pull-request metadata from official sources.
- Evidence: It performs
git fetchoperations fromgithub.com/nrwl/nx.gitand uses theghCLI to interact with GitHub's API. - Context: These downloads are restricted to the vendor's own repository and official API, which is standard for a PR review tool. The skill also allows downloading pinned third-party package archives for static inspection but explicitly forbids their installation or execution.
- [DYNAMIC_EXECUTION]: The skill facilitates runtime verification of PR behavior within a secure environment.
- Evidence: Scripts or commands requiring verification are written to a scratch directory and executed inside a Docker-backed sandbox using
sandbox exec. - Context: This execution is strictly isolated from the host system. The instructions mandate that no untrusted input from the PR or external tasks may appear in the host shell command string, ensuring the sandbox remains the only execution surface for potentially unsafe code.
Audit Metadata