Drag a widget in your running Flutter app and Pinpoint edits the Dart source that put it there — right file, right line, right literal — then hot-reloads.
One real drag on Wonderous, a third-party app I did not write.
- A value follows the drag — an inset opens continuously, through fractional values, with the number live on a HUD chip.
- The source shape is preserved — dragging one edge of
EdgeInsets.symmetric(horizontal:)moves both sides and stays.symmetric. - One line, every widget it builds — where one source line builds three cards, all three change, and all three are checked after the reload.
- Undo puts the file back byte-for-byte, and the layout snaps back with it.
- The round trip is real and measured. Gate G1: 52/52 scripted gestures bound, landed after reload and undid byte-identically. Gate G5: 128/128 preview cycles converged, 0 diverged.
- The gated edit family is bare
EdgeInsetsliterals — on Wonderous roughly 1% of what is on screen — and there is no product around it: no release, and a VS Code extension that has never been run inside an editor. - Ten families are wired and one of them is gated. Size, alignment, colour, radius,
opacity, elevation, gap, flex factor and text content joined
EdgeInsetson 2026-08-19 — planners, binders, knob tables and override maps, with unit and widget tests. Not one of the nine has been driven end to end against a running app, so nothing here claims they work. Typography is built and no gesture can select it.docs/CAPABILITIES.md§2a is the per-family status;docs/KNOWN-ISSUES.md§1 and §18–§22 list what they do not do. - Coverage is measured, both ways, and the two numbers are different questions. Over
50,149 style-value sites in fourteen real apps, the share a shipped gesture can act on went
21.5% → 40.9% against a 78.4% ceiling (
tools/census/BASELINE-2026-08-19.md). Over 223 real taps on two running apps, 3.1% produced an action and the rest a stated reason — and until 2026-08-19, 96.9% of them produced nothing at all, because the daemon had no way to say no (docs/KNOWN-ISSUES.md§36, §37). - Long sessions fall over (the 500-reload soak failed at reload 283) and smoothness is claimed at 60 Hz and nothing stronger.
Not installable yet — nothing is published and there is no release, so watch the repo, or
watch the full four-scene cut: demo/clip.mp4 (27 s), audited frame by frame
against the run logs in demo/AUDIT.md. One capability at a time:
demo/scenes/ — the same frames, one file per scene.
The pointer in the demo is script-driven — synthesised events through the same binding
entry point a real touch uses; hit test, binding, preview, write and reload underneath are the
real path (demo/DISCLOSURES.md).
| docs/OVERVIEW.md | the full picture — what each capability really does, every known gap in detail, how the claims are verified, the licence and business-model stance, the repository layout |
| CHANGELOG.md | what has landed, and what evidence each thing has behind it |
| demo/AUDIT.md | the reel's honesty audit: every splice, every on-screen step matched against its run log |
| docs/DESIGN.md | system design, interaction architecture, language rulings, v1 cut, risk register |
| CONTRIBUTING.md | what kind of contribution is actually useful right now |
MIT (LICENSE). It runs entirely on your machine, sends nothing anywhere, and works offline; what it will not become is in docs/OVERVIEW.md.
Demo app: Wonderous by gskinner, MIT licensed (© 2022 gskinner.com),
unmodified except for one dev_dependency and one added entrypoint file. Not affiliated with
or endorsed by gskinner or Google.