Code Review — Gnomad Desktop Assistant

Reviewed: 2026-05-31 (updated after hiring-feedback implementation)
Scope: src/, src-tauri/src/, workflows, agent/security paths

Summary

The app is a Tauri v2 + React 19 desktop assistant: tray/panel, global shortcut, OS context (active window, clipboard), cloud + local LLM chat with an agent tool loop, persistent PTY shell session, filesystem agent tools, Sudo Gate / Path Gate HITL, and structured error payloads end-to-end.

Architecture (high level)

Tray + Global Shortcut (lib.rs)
        │
        ▼
React App shell (App.tsx) ──hooks──► useAgentExecution, useChatSubmit, …
        │                              │
        ▼                              ▼
ChatView / SettingsPanel ──invoke──► Tauri commands
                                     ├── agent_runtime / agent_fs
                                     ├── shell_session (PTY)
                                     ├── privilege.rs (safety + elevation)
                                     ├── error.rs (GnomadError payloads)
                                     └── context, keychain, llm, …

Agent error payload contract

Tauri commands still return Result<T, String>. During migration, error strings are JSON matching:

{
  "code": "safety_blocked",
  "message": "Human-readable summary",
  "detail": "Optional technical detail",
  "hint": "Optional remediation",
  "retryable": false
}

Frontend: src/lib/errors.ts — parseInvokeError, executionFailedLabel, formatErrorForUser.
UI: src/components/AgentErrorBanner.tsx on messages with errorPayload.

Stable code values are covered by error::tests::payload_codes_are_stable.

Strengths

1. Module split — Rust agent, shell, privilege, FS; React hooks mirror execution concerns.
2. HITL — Sudo Gate and Path Gate with explicit approve/deny.
3. Defense in depth — Server-side safety before shell; elevation rejects injection patterns.
4. Cross-platform awareness — platformInfo drives labels and capability flags.

Remaining gaps (priority)

P1 — Security / product

1. Wave B error migration — ✓ Shipped on LLM, planner, and chat history paths.
2. Windows elevation — Structured elevation_unsupported; user must use elevated terminal for admin ops.
3. Path Gate tokens — ✓ Shipped — path_token.rs; boolean IPC bypass rejected.

P2 — Engineering

4. Typed invoke errors — Optional future: Result<T, AgentErrorPayload> at Tauri boundary once JSON-in-string is stable everywhere.
5. Vitest — ✓ parseInvokeError tests in CI (npm run test).
6. GGUF planner + local chat — In-process inference via optional embedded-llm feature; Ollama remains fallback.

Suggested next steps

1. Generate updater signing keys per UPDATER.md and replace the placeholder pubkey in tauri.conf.json.
2. End-to-end updater test once signing keys are configured.
3. Snap / Flatpak manifests (community packaging).

File reference