feat: native OS window screenshot for Tauri and Electron

[goosewobbler]

Summary

• Adds browser.tauri.nativeScreenshot() and browser.electron.nativeScreenshot() — captures the full OS window including title bar, close/minimize buttons, and window decorations (not just the webview content)
• Tauri: implemented via a new POST /wdio/native-screenshot endpoint in the embedded WebDriver server using the xcap crate (macOS + Windows; unsupported_operation on Linux)
• Electron: implemented via screencapture -R (macOS) and PowerShell PrintWindow (Windows) — no native Node module, no desktopCapturer (avoids macOS Screen Recording permission in CI)
• Returns Buffer of PNG bytes; Tauri requires the embedded driver provider

E2E verification — three-layer approach

Layer	Runs when	Catches
1. PNG structural + dimensions	every PR	broken capture / webview-only
2. OCR via tesseract.js	every PR	wrong window / missing fixture content
3. Vision LLM (Ollama, OLLAMA_API_KEY)	merge to main only	subtle rendering regressions tied to runtime state

Layer 3 requires OLLAMA_API_KEY set as a CI secret (gate via github.event_name == 'push' && github.ref == 'refs/heads/main'). The visionEnabled() guard ensures specs pass on PR runs without the key.

Test plan

• cargo check passes on macOS (aarch64-apple-darwin) — verified locally
• pnpm --filter @wdio/tauri-service typecheck — clean
• pnpm --filter @wdio/electron-service typecheck — clean
• pnpm --filter e2e typecheck — clean
• Run e2e/test/tauri/native-screenshot.spec.ts locally on macOS — layers 1+2 pass without API key
• Run e2e/test/electron/native-screenshot.spec.ts locally on macOS — layers 1+2 pass without API key
• Add OLLAMA_API_KEY + OLLAMA_BASE_URL to merge-to-main CI job for layer 3

🤖 Generated with Claude Code

[github-actions]

Release Preview — no release

No bump label detected.
Reason: No release labels found (need bump:* or release:stable)
Note: Add bump:patch, bump:minor, or bump:major to trigger a release.

---

Updated automatically by ReleaseKit

[greptile-apps]

Greptile Summary

Adds browser.tauri.nativeScreenshot() and browser.electron.nativeScreenshot() commands that capture the full OS window including title bar and decorations. All previously identified blocking issues — HWND little-endian byte order, GetHdc/ReleaseHdc pairing, PowerShell path escaping, spawnSync exit-code checking, incomplete visionEnabled() guard, and the ambiguous title-only window lookup — have been addressed in this revision.

• Tauri (Rust): new POST /wdio/native-screenshot endpoint using xcap, gated to macOS/Windows at the Cargo level; Linux returns unsupported_operation; window lookup now filters by PID first, then title.
• Electron (Node): screencapture -R on macOS and a PowerShell PrintWindow/BitBlt path on Windows, with correct little-endian HWND reading, ReleaseHdc before $b.Save, forward-slash path normalization, and bounded spawnSync timeouts.
• E2E tests: three-layer verification (PNG structure → OCR via tesseract.js → vision LLM behind visionEnabled()) with appropriate platform/CI skips.

Confidence Score: 5/5

Safe to merge; all previously identified blocking issues in the Electron Windows capture path have been fixed and the Tauri xcap path is well-scoped to macOS/Windows.

The prior round resolved every critical bug: HWND byte order, GetHdc pairing, path escaping, spawnSync error propagation, and the incomplete visionEnabled guard. What remains is a minor cleanup edge case in the finally block.

packages/electron-service/src/commands/nativeScreenshot.ts — the finally/unlinkSync edge case noted above; otherwise all changed files look correct.

Important Files Changed

Filename	Overview
packages/electron-service/src/commands/nativeScreenshot.ts	New command implementing native Electron window screenshot; previous issues with HWND byte order, GetHdc release, path escaping, and spawnSync error handling are addressed; minor cleanup concern in finally block
packages/tauri-plugin-webdriver/src/platform/macos.rs	Implements xcap-based native screenshot for macOS, correctly filtering by PID before title, with a fallback to first process-owned window
packages/tauri-plugin-webdriver/src/platform/windows.rs	Implements xcap-based native screenshot for Windows with PID-first matching, title-substring fallback for CI, and detailed diagnostic error output when no window is found
packages/tauri-service/src/commands/nativeScreenshot.ts	Routes nativeScreenshot to the embedded Tauri WebDriver server; enforces the embedded-provider requirement and caches the DirectEvalClient per browser instance
packages/tauri-plugin-webdriver/src/server/handlers/native_screenshot.rs	New Axum handler for POST /wdio/native-screenshot; window-not-found returns 404 with available labels, unsupported platform returns error via WebDriverErrorResponse::into_response
e2e/lib/visionAssert.ts	visionEnabled() now requires both the API key and OLLAMA_BASE_URL; regex tightened to /^(YES
e2e/lib/screenshotChecks.ts	Full 8-byte PNG magic validation, OCR helpers via tesseract.js, and assertCapturesChrome; OCR worker lazily initialised and correctly terminated

Sequence Diagram

sequenceDiagram
    participant Test as E2E Test
    participant Service as electron/tauri Service
    participant App as Electron/Tauri App
    participant OS as OS (screencapture / xcap)

    Test->>Service: browser.electron.nativeScreenshot()
    Service->>App: CDP execute: getBounds() + getNativeWindowHandle()
    App-->>Service: "{ bounds, nativeHandle, gpuCompositing }"
    alt macOS
        Service->>OS: spawnSync screencapture -R x,y,w,h out.png
        OS-->>Service: PNG file written
    else Windows (GPU)
        Service->>OS: spawnSync powershell PrintWindow(PW_RENDERFULLCONTENT)
        OS-->>Service: PNG file written
    else Windows (no GPU)
        Service->>OS: spawnSync powershell BitBlt(GetWindowDC)
        OS-->>Service: PNG file written (may be blank under WARP)
    end
    Service->>Service: readFileSync(out) + unlinkSync(out)
    Service-->>Test: Buffer (PNG bytes)

    Test->>Service: browser.tauri.nativeScreenshot()
    Service->>App: "POST /wdio/native-screenshot {window_label}"
    App->>OS: xcap::Window::capture_image()
    OS-->>App: RgbaImage
    App->>App: encode to PNG bytes
    App-->>Service: image/png response
    Service-->>Test: Buffer (PNG bytes)

Loading

_{Reviews (15): Last reviewed commit: "test(e2e): skip tauri native-screenshot ..." | Re-trigger Greptile}

[goosewobbler]

Closing in favour of using @wdio/visual-service — see VRT spike findings and PR #123 on the example repo.

The spike validated that @wdio/visual-service works out of the box across all three Electron variants and all three Tauri providers (with documented exclusions for crabnebula+macOS-CI and official+linux). It covers the in-app UI surface — which is what 99% of users actually want to test — and the existing mock APIs cover assertions about native UI behaviour (menu/tray/dialog creation). CrabNebula's macOS driver also captures the OS window with native chrome for free.

The remaining "native chrome only" use cases this PR was originally chasing (multi-window-in-one-capture, OS-level menu/tray pixel diffs) turned out to be either niche or already covered. Combined with the inability to make the capture work on Windows CI (Hyper-V/WARP can't satisfy any GDI / desktop-duplication path), the feature isn't worth the carrying cost.