skills/archastro/archdev/reviews/Gen Agent Trust Hub

reviews

Pass

Audited by Gen Agent Trust Hub on Sep 11, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The bootstrap scripts (bootstrap.sh, bootstrap.ps1) download the ArchDev installer and binary from the official ArchAstro GitHub repository (github.com/ArchAstro/archdev). These downloads are directed to specific, pinned commit hashes, following standard installation patterns for vendor-provided tooling.
  • [REMOTE_CODE_EXECUTION]: The installation process for the archdev CLI utilizes a common 'curl to bash' pattern for Unix-like systems and a similar download-and-execute pattern in PowerShell. These actions are directed towards the vendor's own verified releases and are used to ensure the execution environment is up to date.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data, specifically source code and human feedback from local snapshots or GitHub Pull Requests, which could contain malicious instructions designed to influence the agent.
  • Ingestion points: Local working tree files, Git diffs, and inline comments received via a local loopback server or GitHub API.
  • Boundary markers: Instructions explicitly guide the agent to ground its findings in captured changes and establish author intent. The documentation also provides warnings to human reviewers regarding the privacy of their notes.
  • Capability inventory: The toolset includes file system reads/writes (via archdev and git), network operations (GitHub API, local server), and shell command execution.
  • Sanitization: Workflows utilize synthesis nodes to deduplicate and rank evidence, rejecting unsupported claims before final presentation to the user.
  • [DYNAMIC_EXECUTION]: The review workflows support bash and javascript nodes which execute logic defined in project-level configuration files (.archdev/reviews/*.json).
  • Evidence: references/workflows.md describes how these nodes run with host authority and can modify the active workflow directory.
  • Mitigation: The documentation advises checking executable code before running it and suggests using isolated worktrees for writer agents to prevent unintended modification of the parent checkout.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 11, 2026, 05:11 AM