Performance, measured honestly
These are real timings from the moments you feel every day in an editor: opening the app, opening a project, opening a file, searching everything. Median of 5 cold runs on an ordinary consumer laptop. No lab hardware, no cherry-picking, and the harness that produced them ships in the repository.
| What you feel | Median | All 5 runs |
|---|---|---|
| Cold start to interactive window | 3.1 s | 2.9 / 2.9 / 3.1 / 3.3 / 4.4 s |
| Full code editor engine ready | 3.1 s | loads with the window, not after it |
| Open a real project (this repo's source tree) | 1.7 s | 1.6 / 1.7 / 1.7 / 1.7 / 1.9 s |
| Open a large TypeScript file into the editor | 249 ms | 225 / 243 / 249 / 265 / 281 ms |
| Search the whole source tree (first 200 hits) | 207 ms | 142 / 198 / 207 / 254 / 273 ms |
Timings include everything the user waits for: process launch, window paint, editor engine, file tree, editor model, and search results rendered to the UI thread.
Method (reproduce it yourself)
The harness is scripts/bench-ide.mjs in the LoWisa repository. It drives a real app instance over the Chrome DevTools Protocol and only stops each clock when the actual UI state exists, not when an internal event fires:
- Cold start: process spawn to the welcome UI present in the DOM and a JavaScript round trip completing.
- Editor engine: spawn to the Monaco engine reporting ready. LoWisa serves the engine from a local loopback server, so it loads in parallel with the window even under heavy disk load.
- Project open: workspace command to the file tree rendered with real entries.
- File open: open call to a live editor model for that exact file.
- Search: a plain-text query across the repository source tree (hundreds of files), first 200 results returned.
Machine: Windows 11 Home consumer laptop with OneDrive active and antivirus on, because that is the machine real learners have. Each run kills the app fully and starts cold. Results are written to release-evidence/ with every individual run, not just the flattering ones.
What we do not claim
We do not publish comparison numbers for other editors we did not measure under identical conditions, and we do not average away slow runs. When a number regresses, this page regresses with it: the same harness runs against every release.