Repository navigation
Building from Source
DiffLens builds as a standard Tauri 2 workspace: a Rust workspace with the core/ library and src-tauri/ shell, plus a Vite/React frontend at the repo root.
-
Rust via rustup — stable channel. The MSRV is 1.88 (enforced by the workspace
rust-version; a current stable is fine). - Node.js 24 + npm (CI uses Node 24).
- Linux only: the Tauri system libraries — the same line CI uses (Ubuntu 24.04):
sudo apt-get update && sudo apt-get install -y \
libwebkit2gtk-4.1-dev build-essential curl wget file libxdo-dev \
libssl-dev libayatana-appindicator3-dev librsvg2-dev pkg-configAdd tesseract-ocr tesseract-ocr-vie if you want the OCR integration tests to run locally (they are #[ignore]d without it).
git clone https://github.com/vibecoder880/difflens
cd difflens
npm ci # frontend deps + @tauri-apps/cli
npm run tauri dev # desktop app with hot reload
npm run tauri build # full release bundle (per-OS installers)Note: npm run dev/test/build each run scripts/export-labels.mjs first — it exports the core's i18n label keys so the frontend parity tests compare against the live backend list, not a stale copy.
-
embeddings-ort(off by default): compiles the real ONNX e5 embedder (ortdownloads prebuilt ONNX Runtime at build time). With the feature on but model files absent, the pipeline logs a warning and falls back to the deterministic fake hash embedder — this keepscargo testhermetic. The choice is surfaced honestly viaDiffResult::embeddings_usedand the UI's embeddings indicator. - Model fetch scripts:
scripts/fetch-models.sh(--testfor the CI-sized e5 model,--bundlefor release resources) andscripts/fetch-pdfium.shstage the pinned PDFium build. CI setsDIFFLENS_PDFIUM_PATH+DIFFLENS_PDFIUM_STRICT=1so PDF tests must run, not silently skip.
cargo fmt --check -p difflens-core # formatting gate
cargo clippy -p difflens-core --all-targets -- -D warnings # lints gate (CI: with and without features)
cargo test -p difflens-core # core suite (fake embedder)
npm test # vitest + EN/VI i18n parity
npm run build # tsc -b && vite build
scripts/audit-egress.sh # egress invariantTest tiers, all exercised in CI:
| Tier | What | Locally |
|---|---|---|
| Core unit + integration | ~241 unit + ~53 integration #[test] fns (extraction, alignment, deltas, download/ archive red-team suites, store, report escaping goldens) |
runs |
| Gold alignment gates | precision ≥ 0.85 / recall ≥ 0.80 over the gold corpus (core/tests/gold/) |
runs (fake embedder in default tier) |
| PDF extraction | real pinned PDFium (fetch-pdfium.sh, DIFFLENS_PDFIUM_STRICT=1) |
#[ignore]d without PDFium |
| Real e5 embeddings |
embeddings-ort feature + ONNX model |
#[ignore]d |
| OCR e2e | real tesseract + vie (ocr-ci-ok assertion) |
#[ignore]d |
| LLM roundtrip | pinned llama-server + Qwen3-0.6B GGUF through the download manager (llm_real_roundtrip) |
#[ignore]d |
| Frontend | vitest (52 cases: filters, row building, formatting, EN/VI parity vs backend label keys) + typecheck/build | runs |
"Assert tests actually ran" greps in CI (e.g. pdf-extraction-ok, ocr-ci-ok) defeat silent skipping.
See Architecture for the annotated tree and the Tauri command/event surface table.
A Tauri/Rust target/ directory routinely grows to 10–20 GB across debug/release and feature combinations. On smaller machines: build one platform locally (release/3-OS builds belong in CI anyway), use cargo clean between long gaps, or relocate CARGO_TARGET_DIR. The ort feature adds a one-time ~100–300 MB build-time download of ONNX Runtime binaries.
- Conventional Commits (
feat:,fix:,docs:, …) — see CONTRIBUTING.md. -
The egress rule: any new network code must live in
core/src/download/and passscripts/audit-egress.sh; PRs that break the invariant will not merge. - Keep UI strings in both
src/i18n/en.jsonandsrc/i18n/vi.json— the parity test fails otherwise. - LLM code paths must never alter diff results or severity bands.