Maxaubert · Maxaubert · Aug 20, 2026 · Aug 20, 2026 · Aug 20, 2026 · Aug 20, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -0,0 +1,24 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  check:
+    # windows-latest, not ubuntu: the app is Windows-first and several suites
+    # exercise Windows path semantics (drive letters, backslashes, reserved
+    # names) that a Linux runner would never catch.
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          cache: npm
+      - run: npm ci
+      - run: npm run typecheck
+      - run: npm run lint
+      - run: npx prettier --check .
+      - run: npm test
diff --git a/.gitignore b/.gitignore
@@ -27,3 +27,7 @@ build/logo-mockups/
 # Channel signing key — never commit. See scripts/registry-channel.mjs
 channel-private.pem
 channel.json
+
+# Playwright artifacts
+test-results
+playwright-report
diff --git a/.prettierignore b/.prettierignore
@@ -1,5 +1,7 @@
 node_modules
 out
 dist
-resources/bin
+resources
 package-lock.json
+# Design mockups are point-in-time artifacts, not living code.
+docs/mockups
diff --git a/.prettierrc.yaml b/.prettierrc.yaml
@@ -2,3 +2,6 @@ semi: false
 singleQuote: true
 printWidth: 100
 trailingComma: none
+# Match what git checks out on Windows (core.autocrlf) so `format` never
+# rewrites the whole tree over line endings.
+endOfLine: auto
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -62,7 +62,8 @@ resources/bin/ bundled CLI binaries (gitignored; fetched by scripts, packed by e
 - `npm test` — Vitest unit tests (arg-builders, format catalogs, output collision-safety).
 - `npm run lint` / `npm run format` — eslint (flat config) / prettier.
 - `npm run package` — electron-vite build + electron-builder NSIS installer to `dist/`.
-- UI end-to-end: Playwright (scaffolded once the real UI exists — per the global testing rule).
+- `npm run test:e2e` — Playwright end-to-end (launches the built app via `_electron`; run
+  `npm run build` first). Covers the preload/IPC/engine chain unit tests can't reach.
 
 ## Conventions
 

diff --git a/README.md b/README.md
@@ -1,15 +1,15 @@
 <div align="center">
   <img src="docs/logo.png" alt="Filesmith" width="116">
 
-  # Filesmith
+# Filesmith
 
-  A local file toolkit for Windows. Drop files, pick a tool, get results.
+A local file toolkit for Windows. Drop files, pick a tool, get results.
 
-  [![Latest release](https://img.shields.io/github/v/release/Maxaubert/Filesmith?style=flat-square&color=5b5bd6&cacheSeconds=1800)](https://github.com/Maxaubert/Filesmith/releases/latest)
-  [![Downloads](https://img.shields.io/github/downloads/Maxaubert/Filesmith/total?style=flat-square&color=5b5bd6&cacheSeconds=1800)](https://github.com/Maxaubert/Filesmith/releases)
-  [![Windows](https://img.shields.io/badge/Windows-10%20%7C%2011-0078D4?style=flat-square)](https://github.com/Maxaubert/Filesmith/releases/latest)
-  [![Built with](https://img.shields.io/badge/Electron%20·%20React%20·%20TypeScript-2b2e3a?style=flat-square)](#build-from-source)
-  [![License: MIT](https://img.shields.io/badge/License-MIT-22b364?style=flat-square)](LICENSE)
+[![Latest release](https://img.shields.io/github/v/release/Maxaubert/Filesmith?style=flat-square&color=5b5bd6&cacheSeconds=1800)](https://github.com/Maxaubert/Filesmith/releases/latest)
+[![Downloads](https://img.shields.io/github/downloads/Maxaubert/Filesmith/total?style=flat-square&color=5b5bd6&cacheSeconds=1800)](https://github.com/Maxaubert/Filesmith/releases)
+[![Windows](https://img.shields.io/badge/Windows-10%20%7C%2011-0078D4?style=flat-square)](https://github.com/Maxaubert/Filesmith/releases/latest)
+[![Built with](https://img.shields.io/badge/Electron%20·%20React%20·%20TypeScript-2b2e3a?style=flat-square)](#build-from-source)
+[![License: MIT](https://img.shields.io/badge/License-MIT-22b364?style=flat-square)](LICENSE)
 </div>
 
 ---
@@ -48,11 +48,11 @@ Plus batch queues, thumbnails for every kind (images, video frames, audio cover
 
 The everyday tools (convert, compress, resize, and all PDF / video / audio / document operations) run **fully offline** from bundled binaries, with no AI and no downloads. The AI features are **opt-in**, and nothing AI runs unless you choose it:
 
-| Feature | What it needs |
-|---|---|
-| **Remove Background** | The free [`uv`](https://docs.astral.sh/uv/) tool; a small AI model is downloaded on first use (once, then offline). The panel tells you before you commit files. |
-| **Upscale** | An upscaling model, downloaded on first use. NVIDIA (PiD) mode needs an NVIDIA GPU. |
-| **Generate** | An existing [ComfyUI](https://github.com/comfyanonymous/ComfyUI) install, which Filesmith drives headlessly. Supports SDXL, Flux 1, Flux 2 (klein), Z-Image, and Krea 2 models; missing text-encoders / VAEs can be downloaded from the panel. If something is missing, the panel says exactly what to do. |
+| Feature               | What it needs                                                                                                                                                                                                                                                                                              |
+| --------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Remove Background** | The free [`uv`](https://docs.astral.sh/uv/) tool; a small AI model is downloaded on first use (once, then offline). The panel tells you before you commit files.                                                                                                                                           |
+| **Upscale**           | An upscaling model, downloaded on first use. NVIDIA (PiD) mode needs an NVIDIA GPU.                                                                                                                                                                                                                        |
+| **Generate**          | An existing [ComfyUI](https://github.com/comfyanonymous/ComfyUI) install, which Filesmith drives headlessly. Supports SDXL, Flux 1, Flux 2 (klein), Z-Image, and Krea 2 models; missing text-encoders / VAEs can be downloaded from the panel. If something is missing, the panel says exactly what to do. |
 
 Don't want AI? Just use the core tools, and the AI features stay out of your way.
 

diff --git a/docs/adding-models.md b/docs/adding-models.md
@@ -1,18 +1,18 @@
 # Adding a model to Filesmith
 
-Filesmith does not bake models in. What a model *is* — how to recognize it, what files it needs,
+Filesmith does not bake models in. What a model _is_ — how to recognize it, what files it needs,
 which ComfyUI graph runs it — is **data on disk**, so you can add a model family that did not exist
 when your copy of Filesmith was built, without waiting for a release.
 
 ## Where the registry lives
 
 Three layers, merged by `id`, later layers winning **field by field**:
 
-| Layer | Path | Who writes it |
-|---|---|---|
-| 1. Built-in | `<install>/resources/registry/*.json` | ships in the installer, read-only |
-| 2. Channel | `%APPDATA%/Filesmith/registry/channel/` | refreshed over the network (signed) |
-| 3. **Yours** | `%APPDATA%/Filesmith/registry/user/` | you |
+| Layer        | Path                                    | Who writes it                       |
+| ------------ | --------------------------------------- | ----------------------------------- |
+| 1. Built-in  | `<install>/resources/registry/*.json`   | ships in the installer, read-only   |
+| 2. Channel   | `%APPDATA%/Filesmith/registry/channel/` | refreshed over the network (signed) |
+| 3. **Yours** | `%APPDATA%/Filesmith/registry/user/`    | you                                 |
 
 Two rules that will not change:
 
@@ -129,7 +129,7 @@ Available placeholders:
 `${unet}` `${clip}` `${clip2}` `${vae}` `${model}` · `${prompt}` `${negative}` `${seed}` `${steps}`
 `${cfg}` `${guidance}` `${sampler}` `${scheduler}` `${width}` `${height}` `${batch}` `${prefix}`
 
-A value that is *exactly* one placeholder is replaced with the raw value, so `"seed": "${seed}"`
+A value that is _exactly_ one placeholder is replaced with the raw value, so `"seed": "${seed}"`
 yields a number, not a string. Use `workflow` for a bare diffusion model (UNETLoader + separate
 encoders) and `checkpointWorkflow` for an all-in-one single-file checkpoint. An entry can have both.
 
@@ -165,7 +165,7 @@ always tell whether Filesmith saw the file at all.
 Every download in `resources/registry/gen-archs.json` carries a real `sha256` and a URL pinned to
 an immutable commit revision, with the `resolve/main` branch URL kept after it as a fallback mirror.
 
-The hashes are not invented: Hugging Face stores large files in git-LFS, and an LFS object id *is*
+The hashes are not invented: Hugging Face stores large files in git-LFS, and an LFS object id _is_
 the sha256 of the content, exposed per file by the repo tree API.
 
 ```
@@ -183,7 +183,7 @@ trust-on-first-use, which is where everything was before.
 
 ## Publishing a channel update
 
-Every companion URL points into someone else's repo. When one is reorganized, every *installed*
+Every companion URL points into someone else's repo. When one is reorganized, every _installed_
 copy of Filesmith gets a 404 and stays broken until a new release ships. The channel fixes that
 without a release: publish a signed pack, and every install picks it up within a day.
 

diff --git a/docs/comfyui-upscaler-import.md b/docs/comfyui-upscaler-import.md
@@ -11,7 +11,7 @@ uses) and run through our own tiled upscaler.
 ## Why this over bundling
 
 - **Licensing**: most good community upscalers are non-commercial (CC-BY-NC-SA),
-  so we can't ship them. Running the user's *own* files is distribution-free.
+  so we can't ship them. Running the user's _own_ files is distribution-free.
 - **VRAM**: tiled ESRGAN inference is bounded, so it avoids the large-output
   blow-ups the diffusion path (PiD) hits.
 - **Coverage**: spandrel auto-detects architecture, so one code path handles the
@@ -22,7 +22,7 @@ uses) and run through our own tiled upscaler.
 1. **Import policy** — load anything spandrel accepts; badge known-good models
    **Verified**, the rest **Experimental**; diffusion/unloadable files are shown
    greyed as **Unsupported** with the reason.
-2. **Placement** — imported models sit *alongside* the existing options
+2. **Placement** — imported models sit _alongside_ the existing options
    (Photo / Anime bundled Real-ESRGAN, PiD). Nothing is removed. PiD keeps its
    own diffusion path (and its pending input-resize fix).
 3. **Storage** — **reference in place**: read models straight from the ComfyUI

diff --git a/docs/compression-upgrade-plan.md b/docs/compression-upgrade-plan.md
@@ -4,11 +4,13 @@ A running design doc for expanding the **Compress** tab, decided category by
 category with the user. Status per section: DECIDED / discussing / TODO.
 
 **STATUS: IMPLEMENTED** (images, video, audio, PDF). Engine + per-kind options UI
-+ live video-resolution preview + Ghostscript/ffprobe bundling all landed and
-verified end-to-end against the real binaries. Possible follow-up: allow WAV/FLAC
-into audio Compress (wav→opus is a big win), currently excluded by canCompress.
+
+- live video-resolution preview + Ghostscript/ffprobe bundling all landed and
+  verified end-to-end against the real binaries. Possible follow-up: allow WAV/FLAC
+  into audio Compress (wav→opus is a big win), currently excluded by canCompress.
 
 Current baseline (what the Compress tab does today):
+
 - Images: CaesiumCLT (jpg/png/webp/gif/tiff) + ImageMagick fallback, re-encode at
   a quality value, **same format in/out**.
 - Video: ffmpeg CRF, x264 (VP9 for webm), fixed 128k AAC audio, one quality slider.
@@ -31,16 +33,18 @@ Add a **Format** choice to the image Compress options. Three options:
    narrower compatibility. The "squeeze it as hard as possible" option.
 
 Key decisions:
+
 - **Both new formats run in LOSSY mode**, driven by the existing quality slider.
   Rationale: a Compress button should reliably make files smaller; lossy is
   where the big wins are for photos (the common case). Lossless WebP/AVIF only
-  helps graphics and can make photos *larger*, so it's not offered here.
+  helps graphics and can make photos _larger_, so it's not offered here.
 - **No "→ JPEG" option in Compress.** Converting to JPEG is a footgun on
   transparency / sharp edges; that belongs in the Convert tool.
 - Perfect-quality (lossless) format conversion, if ever wanted, also belongs in
   Convert, not Compress.
 
 Implementation notes:
+
 - ImageMagick (already bundled) can already encode WebP and AVIF at `-quality`
   (verified in the conversion matrix), so an **MVP needs no new binary** — route
   the WebP/AVIF options through magick with the mapped quality.
@@ -51,6 +55,7 @@ Implementation notes:
 - Transparency is preserved in lossy WebP/AVIF (unlike JPEG).
 
 Second-wave / not now (nice-to-have specialists):
+
 - mozjpeg (better JPEG encoder, 10-30% smaller), oxipng / pngquant (PNG lossless
   / palette), gifsicle (GIF). Bigger, format-specific wins; revisit later.
 - "Strip metadata" toggle (EXIF/ICC/XMP) — cheap, always helps a little.
@@ -62,6 +67,7 @@ Second-wave / not now (nice-to-have specialists):
 Three independent controls in the video Compress options:
 
 ### a. Codec (the biggest lever) — same "compatible / smaller / smallest" framing as images
+
 1. **H.264 (compatible)** — universal, plays everywhere, least efficient. Today's
    default. ffmpeg `libx264`.
 2. **H.265 / HEVC (smaller)** — ~30-50% smaller than H.264 at equal quality,
@@ -70,16 +76,19 @@ Three independent controls in the video Compress options:
 3. **AV1 (smallest)** — another ~20-35% smaller than H.265, slowest encode,
    narrower support. ffmpeg `libsvtav1` (fast AV1 encoder). The bundled gyan.dev
    essentials ffmpeg includes SVT-AV1 (verified).
+
 - **Output is always .mp4**, driven by the chosen codec (H.264/H.265/AV1 all mux
   there). The codec picker replaces the old auto-VP9-for-WebM behavior: a `.webm`
   source is re-encoded to `.mp4` with the selected codec. (No VP9 option is
   offered; VP9/WebM output would be a Convert-tool concern if ever wanted.)
 
 ### b. Quality — the existing slider
+
 - Maps to CRF per codec (the CRF scale differs by codec: ~18-32 for x264/x265,
   higher numbers for AV1, so use a per-codec mapping, not one shared range).
 
 ### c. Resolution — presets + live output list
+
 - Presets: **Original / 1440p / 1080p / 720p / 480p / 360p / 240p**.
 - Each preset means "fit within that box, **preserve aspect ratio, never
   upscale**" — so it works for any shape (landscape, portrait, ultrawide) and any
@@ -99,6 +108,7 @@ Three independent controls in the video Compress options:
 - Optional later: a percentage / custom-dimensions mode for power users.
 
 ### Audio track (inside video)
+
 - MVP: keep the current fixed AAC 128k. Optionally expose audio bitrate / "remove
   audio" later, reusing the audio-category decisions below.
 
@@ -109,33 +119,38 @@ lossy formats — mp3/m4a/aac/ogg/opus/wma; lossless flac/wav are excluded by
 `canCompress`, as today.) All via bundled ffmpeg, no new tools.
 
 ### a. Codec / format picker — parenthetical labels for the tradeoff
+
 1. **MP3 (compatible)** — universal, least efficient. `libmp3lame`.
 2. **AAC (balanced)** — better quality per byte, widely supported (`.m4a`). ffmpeg `aac`.
 3. **Opus (smallest)** — most efficient (~40-60% smaller at equal quality),
    especially at low bitrates; fine on modern devices. `libopus`.
+
 - "Keep format" behavior can remain the default (re-encode in the source codec),
   with MP3/AAC/Opus as explicit targets.
 
 ### b. Bitrate picker (replaces the quality slider for audio)
+
 - User picks the **bitrate directly**, shown in kbps (not an abstract quality
   slider). Preset values: **320 / 256 / 192 / 128 / 96 / 64 kbps** (default ~192).
 - ffmpeg `-b:a <N>k`. Lower bitrate = smaller file; Opus stays clean lower than
   MP3/AAC.
 
 ### Explicitly NOT doing (this round)
+
 - **No "already-compressed" warning.** (User decision.)
 - Sample-rate reduction / stereo→mono downmix / Voice-Podcast-Music presets:
   deferred, revisit later if wanted.
 
 ## 4. PDF — DECIDED
 
 Bundle **Ghostscript** (`gs`) — the only mature offline single-binary tool that
-downsamples the *images inside* a PDF (mutool/qpdf only touch structure, which is
+downsamples the _images inside_ a PDF (mutool/qpdf only touch structure, which is
 why today's compress does almost nothing on scanned / image-heavy PDFs). Worth
 the ~60 MB. License: AGPL, compatible since Filesmith is open-source and already
 ships GPL ffmpeg (no network service, so AGPL's extra clause doesn't bite).
 
 ### a. Compression level picker — default Balanced
+
 - **Lossless** — today's `mutool clean -gggg -z`. No image changes, safe, small
   savings. Best for text-only PDFs / when you don't want to risk altering render.
 - **High quality** (~300 dpi) — Ghostscript `-dPDFSETTINGS=/printer`. Downsamples
@@ -149,9 +164,10 @@ ships GPL ffmpeg (no network service, so AGPL's extra clause doesn't bite).
   mutool is the conservative "shrink without altering" pass.
 
 ### b. Grayscale toggle — optional, OFF by default
+
 - Converts color → grayscale for docs that don't need color (scanned documents to
   email). ~30-50% extra. Ghostscript `-sColorConversionStrategy=Gray
-  -sProcessColorModel=DeviceGray`. Off by default so it never surprises anyone.
+-sProcessColorModel=DeviceGray`. Off by default so it never surprises anyone.
 
 Base gs invocation:
 `gs -sDEVICE=pdfwrite -dPDFSETTINGS=/<tier> -dNOPAUSE -dBATCH -dQUIET -o out.pdf in.pdf`