agentic-tdd
Test-driven development for behaviour a unit test cannot reach, by writing the expectation…
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
$ npx -y skills add reticlehq/reticle --skill drive-desktop-app --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/drive-desktop-appContext preview
The summary Claude sees to decide when to auto-load this skill.
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
name: drive-desktop-app description: 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. license: Apache-2.0 metadata: version: 3.7.0 homepage: https://www.reticle.sh repository: https://github.com/reticlehq/reticle
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`](https://github.com/reticlehq/reticle/blob/main/skills/install-and-verify/SKILL.md) skill.
// vite.config.ts — desktop:true lets connect() start in a packaged renderer (NODE_ENV=production).
// A default (production-mode) `vite build` ships no Reticle code; to drive a packaged
// renderer with no dev server, build it with `vite build --mode development`
export default defineConfig({
base: './', // file:// needs relative asset paths
plugins: [react(), reticle({ desktop: true })],
});// 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**:
{
"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`](https://crates.io/crates/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_look { action: "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`.
AI agents can generate code, but still struggle to understand what they build. Reticle brings Jev-style machine-native runtime perception to web & desktop applications.
Repo: reticlehq/reticle
Test-driven development for behaviour a unit test cannot reach, by writing the expectation…
Sweep a whole running web app for what is broken, without writing a script or knowing the…
Find out why something in a running web app does not work, when the console is empty and the…
Check that the UI you actually rendered uses the design system, by reading computed styles in…
Find out why the tests pass but the app is broken. Catches false greens: a green suite over a…
Pick up the bugs a human flagged by pointing at them in the running app, each arriving with…