Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Drive and verify an Electron or Tauri desktop app from the inside, including the main-process and Rust IPC calls a browser tool cannot see. Use when a desktop app needs testing, when a feature works in the browser but not in the packaged app, when an IPC or invoke call needs proving, when a desktop screenshot or visual diff is wanted, or when you need a headless run of a desktop UI in CI.
.claude/skills/reticlehq-drive-desktop-app/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 7% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 12% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -13% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 27% | 0% |
A desktop app reaches its backend over IPC, not HTTP. Patching fetch/XHR cannot see that, so a browser-shaped tool is blind to every backend call the app makes: the network log reads empty, an action has no in-flight request to settle on, and asserting on the network is vacuously true. That is a false green by construction.
Reticle observes the renderer _and_ the IPC boundary, so a desktop verdict means what a web one does. Not installed? RETICLE_INSTALL_SOURCE=npx_skill npx @reticlehq/server@latest init, then the install-and-verify skill.
ts// vite.config.ts — desktop:true also runs the plugin for `vite build`, because a packaged // renderer is a production build with no dev server export default defineConfig({ base: './', // file:// needs relative asset paths plugins: [react(), reticle({ desktop: true })], });
js// electron/preload.cjs — FIRST line. This is what makes main-process IPC visible. require('@reticlehq/electron/preload');
It must be in the preload and it must be first. contextBridge.exposeInMainWorld hands the renderer a deeply frozen object, so nothing in the page can instrument it afterwards. The preload is the last point where ipcRenderer.invoke is still writable, and the shim has to run before your preload captures its own reference.
A sandboxed preload cannot resolve node_modules, so the bare require fails. Either bundle the preload (electron-vite and Forge do by default) or set sandbox: false.
The frontend is the same as any web app. The part people miss is that Tauri's default CSP blocks the bridge WebSocket before it opens, so the app runs perfectly and simply never connects:
json{ "app": { "security": { "csp": "default-src 'self' ipc: http://ipc.localhost; connect-src 'self' ipc: http://ipc.localhost ws://localhost:4400 ws://127.0.0.1:4400" } } }
Keep ipc: http://ipc.localhost: Tauri v2 needs it for invoke itself. Dev-only; drop the ws:// entries from your release config.
IPC observation needs nothing on the Rust side: an invoke('load_todos') already reaches Reticle as ipc://load_todos. The reticle-tauri crate is only for screenshots and headless, and it is versioned independently of the npm packages.
Also: use a hash router. A packaged renderer is served from file://, where history-based routing does not resolve.
Same loop as the web, with IPC in the predicates:
reticle_act_and_wait({ sessionId, ref, action: "click", until: { kind: "allOf", predicates: [
{ kind: "net", urlContains: "ipc://todos:archive", status: 200 },
{ kind: "element", query: { testid: "..." } },
{ kind: "console", level: "error", absent: true },
]}})IPC has no status code. 200/500 are synthetic, mapped from whether the command succeeded, precisely so the same predicates keep working. On Tauri you will see status: 500 next to statusText: "OK". That is not a bug: the transport answered fine and the 500 is the command's own verdict. ok is authoritative.
reticle_state reads the live store exactly as on the web. reticle_screenshot and reticle_visual_diff work once the platform's capture step is wired. Electron needs nothing extra; Tauri needs the crate. Headless on Tauri is RETICLE_HEADLESS=1, and screenshots keep working because the capture renders the webview rather than the screen.
A missing Electron preload is declared, not silent: verdicts come back with coverage: partial naming the line you did not add, instead of reading clean over a blind spot. If you see that, add the preload line before trusting anything.
If IPC calls never appear while the app works fine: on Electron, the shim's require is not first. On Tauri, invoke from @tauri-apps/api/core is observed, but a hand-rolled postMessage protocol is not.
unknown is not a pass on the desktop either. And do not weaken an IPC assertion to make a red verdict green: a desktop false green is the exact failure this wiring exists to remove.
Full desktop reference: curl https://docs.reticle.sh/desktop.md. Everything else: curl https://docs.reticle.sh/llms.txt.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 18,168 | 21,759 | +20% | 1 | 1 | 0% | 2,898 | 3,115 | +7% | 0 | 0 | — |
case-02 | fail→pass | 21,437 | 16,150 | -25% | 1 | 1 | 0% | 3,356 | 3,754 | +12% | 0 | 0 | — |
case-03 | fail→pass | 19,249 | 11,978 | -38% | 1 | 1 | 0% | 2,740 | 3,275 | +20% | 0 | 0 | — |
case-04 | pass→pass | 19,170 | 17,608 | -8% | 1 | 1 | 0% | 2,902 | 3,662 | +26% | 0 | 0 | — |
case-05 | fail→pass | 22,288 | 11,607 | -48% | 1 | 1 | 0% | 3,412 | 2,979 | -13% | 0 | 0 | — |
case-06 | pass→pass | 20,161 | 13,769 | -32% | 1 | 1 | 0% | 2,857 | 3,030 | +6% | 0 | 0 | — |
case-07 | fail→fail | 15,820 | 11,132 | -30% | 1 | 1 | 0% | 2,277 | 2,654 | +17% | 0 | 0 | — |
case-08 | fail→pass | 17,516 | 2,581 | -85% | 1 | 1 | 0% | 1,224 | 1,549 | +27% | 0 | 0 | — |
case-09 | pass→pass | 12,699 | 4,363 | -66% | 1 | 1 | 0% | 1,866 | 1,777 | -5% | 0 | 0 | — |
case-10 | pass→pass | 14,010 | 6,121 | -56% | 1 | 1 | 0% | 2,294 | 2,089 | -9% | 0 | 0 | — |
case-11 | pass→pass | 6,460 | 3,990 | -38% | 1 | 1 | 0% | 935 | 1,672 | +79% | 0 | 0 | — |
case-12 | fail→pass | 12,359 | 6,394 | -48% | 1 | 1 | 0% | 1,908 | 2,183 | +14% | 0 | 0 | — |
case-13 | fail→fail | 14,333 | 7,563 | -47% | 1 | 1 | 0% | 1,909 | 2,297 | +20% | 0 | 0 | — |
case-14 | pass→pass | 8,157 | 4,594 | -44% | 1 | 1 | 0% | 1,134 | 1,774 | +56% | 0 | 0 | — |
case-15 | fail→pass | 22,374 | 5,969 | -73% | 1 | 1 | 0% | 1,733 | 2,103 | +21% | 0 | 0 | — |
case-16 | pass→pass | 14,959 | 4,154 | -72% | 1 | 1 | 0% | 2,191 | 1,856 | -15% | 0 | 0 | — |
case-17 | pass→pass | 13,869 | 7,229 | -48% | 1 | 1 | 0% | 1,858 | 2,174 | +17% | 0 | 0 | — |
case-18 | fail→pass | 8,834 | 2,123 | -76% | 1 | 1 | 0% | 1,246 | 1,404 | +13% | 0 | 0 | — |
case-19 | fail→pass | 10,415 | 2,049 | -80% | 1 | 1 | 0% | 1,340 | 1,436 | +7% | 0 | 0 | — |
case-20 | pass→pass | 15,714 | 15,760 | +0% | 1 | 1 | 0% | 2,667 | 3,673 | +38% | 0 | 0 | — |
case-21 | pass→pass | 13,419 | 9,075 | -32% | 1 | 1 | 0% | 2,106 | 2,598 | +23% | 0 | 0 | — |
case-22 | pass→pass | 14,577 | 14,609 | +0% | 1 | 1 | 0% | 2,462 | 3,470 | +41% | 0 | 0 | — |
case-23 | pass→pass | 12,006 | 9,128 | -24% | 1 | 1 | 0% | 2,095 | 2,763 | +32% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 22 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +39 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.