plan-design-review

Pass

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted design plans and project documentation (e.g., DESIGN.md, TODOS.md) which can influence the agent's behavior or decision-making during the review process.
  • Ingestion points: Untrusted content enters the context via the Read tool as specified in SKILL.md and sections/review-sections.md.
  • Boundary markers: The skill implements a 'Scope gate' and uses the AskUserQuestion tool to ensure the user confirms the review target and approves all proposed modifications before they are applied to the plan file.
  • Capability inventory: The skill utilizes Bash, Edit, and Read tools to analyze files and modify project plans.
  • Sanitization: The instructions mandate writing untrusted content to temporary files and passing them as arguments to binaries rather than interpolating them directly into shell command strings, reducing the risk of command injection.
  • [COMMAND_EXECUTION]: The skill makes extensive use of shell commands and local binaries to manage project state, perform git operations, and run internal utilities.
  • All observed commands (e.g., git, gh, jq, bun, and various gstack-* binaries in the home directory) are consistent with the skill's documented purpose as a development and design tool.
  • [DATA_EXFILTRATION]: The skill can send plan data to external AI models (Codex and Claude subagents) to obtain independent design critiques.
  • This operation involves network communication with external services and is clearly documented as an optional step. The skill requires the user to explicitly opt-in through a choice-based AskUserQuestion prompt before any plan content is transmitted to these external providers.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 16, 2026, 08:17 PM