plan-devex-review

Pass

Audited by Gen Agent Trust Hub on Aug 25, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses eval and source <(...) to execute local binaries located in ~/.claude/skills/gstack/bin/. These scripts manage internal state, configuration, and project-specific metadata (slugs, repo modes) required for the gstack framework to function correctly.
  • [REMOTE_CODE_EXECUTION]: The skill implements an 'Outside Voice' review phase that sends plan data to OpenAI Codex via a CLI tool or to a Claude subagent via the Agent tool. This feature provides independent critiques and is documented as a standard step that users can disable.
  • [DATA_EXFILTRATION]: The skill incorporates optional telemetry for usage tracking and a feature to sync design artifacts to a private GitHub repository. Both features require explicit user consent via the AskUserQuestion tool and are handled through official gstack logging binaries.
  • [EXTERNAL_DOWNLOADS]: The skill prompts the user to install the @openai/codex NPM package to enable cross-model plan reviews. It also performs periodic update checks for the gstack skill suite.
  • [INDIRECT_PROMPT_INJECTION]: The skill reads untrusted project data (README, plan files, docs) to perform its review. While this creates a surface for indirect prompt injection from repository content, the skill's reliance on human-in-the-loop decisions through AskUserQuestion significantly mitigates the risk of unauthorized actions.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 25, 2026, 01:58 AM
Security Audit — agent-trust-hub — plan-devex-review