e2e-reviewer
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill's internal logic utilizes dynamic execution patterns to optimize its analysis engine. In
scripts/scope-worker.pyandscripts/scope-graph.py, the skill uses Python'simportlib.utilto load its own sub-modules from the file system. Additionally,scripts/scope-watch.pyusesctypes.CDLLto interface with the macOS system library/usr/lib/libSystem.B.dylibfor local APFS filesystem probing (fstatfs64). These operations are used to manage complex graph traversals and file monitoring during the review process. - [COMMAND_EXECUTION]: The skill automates complex analysis tasks by executing shell commands. The script
scripts/scope-graph.pyspawns a subprocess to run the bundledscripts/scope-source.shhelper. Furthermore, theSKILL.mdinstructions guide the AI agent to use standard system tools such asrg,sed, andnlfor targeted reading and filtering of the code being reviewed. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code from target repositories, which inherently presents an indirect prompt injection surface. However, the skill implements a comprehensive defense-in-depth approach:
- Ingestion points: Repository files (Playwright/Cypress specs, Page Objects, and configurations) are read into the agent's context as instructed in
SKILL.mdusingrgandsed. - Boundary markers: The
SKILL.mdcontains a mandatory "Untrusted-input boundary" section that explicitly instructs the agent to treat all target repository content as data, never as instructions, and to ignore any commands embedded in comments or strings. - Capability inventory: The skill uses its own bundled analysis scripts (
scan.sh), system utilities, and internal Python scripts for lexing and scope tracking. - Sanitization: The instructions prohibit the agent from quoting or propagating suspected prompt-injection text in its generated findings.
- Ingestion points: Repository files (Playwright/Cypress specs, Page Objects, and configurations) are read into the agent's context as instructed in
Audit Metadata