Never claim untested things work.
CrashAtlas is a two-part diagnostic suite for Windows PCs that crash, run slow, or misbehave — built for non-technical users first, with a details mode and a WinDbg-style Debug Console for power users. When something can't be run in a given environment, that's said explicitly, rather than guessed at.
Design principles actually followed.
No unattended system changes
Every feature that modifies the system is one explicit user click away, shows its own UAC prompt, and never runs as a background auto-fix.
No custom kernel-mode driver
Sensor access rides on already-signed, already-trusted third-party drivers rather than a new EV code-signing + WHCP dependency.
Offline-first
Self-hosted fonts, inlined icons, no CDN scripts, no network calls the app doesn't strictly need.
Bilingual by design
Every user-facing string, dataset, and Wiki entry ships in English and Spanish from the same schema, not bolted on later.
Approximate is labeled as approximate
The call chain isn't a real symbolicated stack trace and says so; the GPU test isn't a real 3D stress test and says so too.
Never claim untested things work
Every verification claim distinguishes unit-tested, typechecked, and confirmed-on-real-hardware as different tiers of confidence.
Where each component actually stands.
| Component | What's tested | How |
|---|---|---|
| tools/minidump-cli | Byte-level parsing correctness | Hand-built synthetic .dmp fixtures matching the documented format, asserting correct extraction of bugcheck code, faulting module, system info. |
| Desktop app — services | Every core service, the report model, exporters, IPC-argument handling | node:test + ts-node, 805+ tests, no Electron binary needed. |
| Desktop app — renderer | Type correctness of every route and the typed API layer | tsc --noEmit on a separate renderer tsconfig — the root typecheck alone can't see it. |
| Desktop app — UI rendering | Every screen actually renders, Tailwind/font/i18n wiring works | Headless Chromium via Playwright, screenshotted across resolutions and both languages. |
| Desktop app — real hardware | Actual Electron launch, real PowerShell output, real sensor values | Manual verification passes on a real Windows machine — this is where the elevation bug was found. |
| boot-image / diagtool | Every test module's algorithm logic, the config and report models | make test (native host-gcc, assert-based) and make asan (AddressSanitizer/UBSan). |
| diagtool interactive UI | The ncurses menu, live progress bars, ASCII-art result banners | Driven programmatically over a real pty, asserting on actual terminal escape sequences. |
| boot-image boot pipeline | genimage.cfg / grub.cfg / post-build.sh / inittab actually produce a bootable image | QEMU + OVMF, using a host-package substitute for Buildroot's cross-compiled output. |
| boot-image on real hardware | Nothing yet | Needs a completed Buildroot cross-build and a physical machine. |
A large project, built in small, honest increments.
250+ numbered increments in the roadmap, every one with real, committed code behind it.
News & roadmap
Every phase shipped, in order, with an honest account of what's verified.