Web source: crop to the frame's ALPHA bounding box — the widget's true rendered extent, no DOM timing, can't break the render

- FindContentBounds scans the Bgra32 capture for the bounding box of non-transparent pixels
  (background is injected transparent, so alpha = widget extent). Crop to that rect + Stretch=Fill
  maps the widget flush under the selection box: full-canvas widgets fall through to the full frame
  (identical to the confirmed-good image); floating widgets get tight-fitted with no dead space.
- Removes QueryContentBoundsAsync + the JS script entirely — the DOM-union attempt (a5b9952) broke
  rendering by measuring an unsettled layout on NavigationCompleted; alpha measurement is immune.
- 3 new unit tests: tight box, no-alpha no-crop, full-alpha no-crop. 7/7 pass, 0 warnings.
This commit is contained in:
2026-08-28 11:53:36 -07:00
parent 5348b5c480
commit bccdb4892c
5 changed files with 117 additions and 80 deletions
+26 -23
View File
@@ -6,34 +6,36 @@
working). **NOT pushed — user rule: never push without explicit instruction (2026-08-24).** No feature
branches pre-1.0: all work lands on `main` per work unit.
## The web-source rendering work (2026-08-28) — CURRENT MODEL: canvas-size viewport + content crop
## The web-source rendering work (2026-08-28) — CURRENT MODEL: canvas-size viewport + ALPHA-BBOX crop
**Root causes finally isolated:** (1) hosting the page at the element-box size clips full-bleed
widgets to their upper-left chunk (fixed — page now renders at the master canvas 1920×1080,
stable). (2) The content-bounds crop that was supposed to kill right/bottom dead space never ran:
the script returned `JSON.stringify(...)` (a string), and `ExecuteScriptAsync` JSON-encodes the
result, so `JsonDocument.Parse` hit a string node and `GetProperty` threw → bounds stayed unset →
no crop. ALSO the intended measure was wrong: scroll extents ≠ widget rect (they include the
page's own margins). Fixed together: the script now returns a real OBJECT (union of every visible
`body *` `getBoundingClientRect`), giving the widget's true rect INCLUDING its top-left offset,
and `CaptureFrame` crops AT that offset — so the widget anchors at (0,0) and the box is tight on
all four sides. `Stretch="Fill"` maps it flush under the box. Creator confirmed render+resize are
solid at the canvas-size change; this commit delivers the tight/0,0 bounding box.
**History of failures (see MyMistakes.md for the full log):** box-size rendering clipped full-bleed
widgets to their upper-left chunk (fixed — page renders at the master canvas 1920×1080, stable, the
one render the creator confirmed solid). Then four crop attempts failed in different ways: (1) the
`JSON.stringify` return double-encoded under `ExecuteScriptAsync`, silently disabling the crop;
(2) `scrollWidth/scrollHeight` at canvas size returns the whole viewport for full-canvas wrapper
bodies → no trim (ghost boundary); (3) union-of-DOM-elements measured an unsettled layout on
NavigationCompleted and BROKE the image (creator's "don't break the rendering" boundary crossed).
(4) Reverting to f39ebd9's scroll-crop was another dead end (#2 again).
**The CURRENT MODEL:**
- `Services/WebView2Manager.cs` — page renders at 1920×1080 (stable); `overflow:hidden` +
transparent-bg injection; `QueryContentBoundsAsync` measures real widget bounds (object return
from JS, NOT JSON.stringify); `CaptureFrame` crops at the widget's (x,y) offset → anchored 0,0.
- `MainWindow.xaml` — web `<Image>` `Stretch="Fill"`.
- Invariant: **render at canvas size; crop to the widget's true bounds; Fill to the box.**
**The CURRENT MODEL (this commit):**
- `Services/WebView2Manager.cs` — page renders at 1920×1080 (stable, never tracked); `overflow:hidden`
+ transparent-bg injection. `FindContentBounds` scans the Bgra32 capture per frame and crops to the
bounding box of NON-TRANSPARENT pixels — the widget's true rendered extent, measured from the frame
itself. No DOM query, no layout timing: it can never truncate visible content (worst case = full
frame = the confirmed-good image). `Stretch="Fill"` maps the cropped frame flush under the selection
box → the box hugs the widget on all four sides. `QueryContentBoundsAsync` + the JS script are
DELETED.
- `MainWindow.xaml` — web `<Image>` Stretch set to Fill via the IsWebSource DataTrigger (unchanged).
- Invariant: **render at canvas size; crop to the alpha extent; Fill to the box.**
- Unit tests: 3 new for `FindContentBounds` (tight box, no-alpha no-crop, full-alpha no-crop).
**Property panel / sliders (2026-08-28):** web URI ✕ clears the textbox; web ✓/✕ icon buttons
`IsTabStop="False"` so Tab walks X→Y→W→H→URI; Slider style `IsMoveToPointEnabled="True"`
restores click-anywhere-on-bar. All three confirmed fixed by the creator.
Build: 0 warnings; tests 4/4. Rejected for the record: box-sized viewport (upper-left truncation),
fixed 1920×1080 + `Uniform` (letterbox dead space), CSS margin/100% reset (enlarged gap),
viewport==box (reflow/truncation/scrollbars).
Build: 0 warnings; tests 7/7 (WebView2Manager). Rejected for the record: box-sized viewport (upper-left
truncation), fixed 1920×1080 + `Uniform` (letterbox dead space), CSS margin/100% reset (enlarged gap),
viewport==box (reflow/truncation/scrollbars), DOM-union crop (broke the image).
## Key facts / landmines
@@ -50,7 +52,8 @@ viewport==box (reflow/truncation/scrollbars).
- No `CODING.md`; conventions live in `AGENTS.md` + `ai.md`.
## Next steps
1. AWAIT creator's verdict on viewport==box (geometry-agnostic) rendering — visual check with a
widget URL at several box sizes; the selection frame must snap flush with no padding.
1. AWAIT creator's verdict on the ALPHA-BBOX crop — reload a widget URL at several box sizes; the
selection frame must snap flush to the widget with no dead space, image must stay identical for
full-bleed widgets. If an animated widget causes visible jitter, add crop hysteresis as a follow-up.
2. Follow-up only if the creator asks: graceful handling of navigation errors / invalid URIs (TASK 17 ×9).
3. Continue queue: TASK 18 local recording, TASK 21 media source, TASK 22 audio sync offset.
+9 -7
View File
@@ -31,15 +31,17 @@ Stream widget bounding box: hug the widget on all four sides, pinned to canvas (
extents regardless. Lesson: verify declared mechanics actually execute; assume nothing ran.
## Current iteration (web bounding box — UNCOMMITTED, verification pending)
RENDER RESTORED to the last creator-verified-good state (`70fe2b8`): full-frame capture (no crop),
1920×1080 canvas-size viewport, plain `Stretch="Fill"`. Image rendering is back to "renders and
resizes beautifully". The bounding-box dead-space problem is STILL OPEN — do NOT reintroduce any
crop until the creator explicitly approves the approach.
ALPHA-BBOX CROP: `FindContentBounds` scans the captured Bgra32 frame and returns the bounding box
of non-transparent pixels (the background is injected transparent, so alpha = the widget's extent).
Capture is cropped to that rect and `Stretch="Fill"` maps it flush under the box. No DOM query, no
layout timing — the crop rect IS the rendered widget by construction, so it can NEVER truncate
visible content. Worst case (widget fills the canvas) = full frame = the confirmed-good render.
Canvas-size viewport + Fill untouched. `QueryContentBoundsAsync` + the JS script are DELETED.
## Iteration that just failed (`a5b9952` — wrapper-excluded content-union crop + image alignment)
## Iteration that just failed (`a5b9952` — wrapper-excluded DOM-union crop + image alignment)
Broke the image rendering — creator could not evaluate the box because the image was gone/ruined.
Why: the visible-element-union rect is measured ONCE on NavigationCompleted; widget layout isn't
settled then (async fonts/images, load animations, iframes), so the crop rect comes out tiny/wrong
and the Fill-stretched frame renders a broken/zoomed sliver. This was against the creator's explicit
"don't break the rendering" instruction. Lesson: the canvas-size render + Fill is the ONLY verified
good base; any crop must be built on it one tested step at a time, never shipped blind.
"don't break the rendering" instruction. Lesson: measure actual RENDERED PIXELS (alpha), never
transient DOM state; and never ship a crop blind.
+28 -42
View File
@@ -27,10 +27,6 @@ public sealed class WebView2Manager : IDisposable
public required WebView2 Control { get; init; }
public required DispatcherTimer CaptureTimer { get; init; }
public VideoFrame? LatestFrame;
public double ViewportCssW;
public double ViewportCssH;
public double ContentCssW;
public double ContentCssH;
public bool Disposed;
public bool Initialized;
}
@@ -48,9 +44,11 @@ public sealed class WebView2Manager : IDisposable
// OBS model: the page renders at the MASTER CANVAS size (1920×1080),
// stable, never tracked. Full-bleed widgets (designed for the canvas)
// render fully instead of being clipped to a small box-sized viewport;
// dragging the box never reflows the page. The capture is cropped to
// the widget's actual content bounds (QueryContentBoundsAsync) and the
// display (Stretch=Fill) maps it flush under the box, anchored at (0,0).
// dragging the box never reflows the page. The capture is tightly
// cropped to the widget's ALPHA bounding box (FindContentBounds) and
// the display (Stretch=Fill) maps it flush under the box, anchored at
// (0,0) — no DOM query, so no timing fragility; worst case is the
// full frame, which is the confirmed-good rendering.
var webView = new WebView2
{
Width = 1920,
@@ -132,7 +130,6 @@ public sealed class WebView2Manager : IDisposable
"document.documentElement.style.overflow='hidden';" +
"document.body.style.background='transparent';" +
"document.body.style.overflow='hidden';");
_ = QueryContentBoundsAsync(source.Id);
_ = CaptureFrame(source.Id);
};
@@ -159,33 +156,30 @@ public sealed class WebView2Manager : IDisposable
catch (Exception ex) { AppLog.Write($"WebView2Manager: navigate failed: {ex.Message}"); }
}
// The page's real content rect (scrollWidth/Height) can be smaller than the
// viewport for widgets that don't fill it — the leftover area is dead space
// that must not stretch across the selection box. Crop the capture to the
// content bounds (anchored at 0,0) so the box shows exactly the widget.
private async Task QueryContentBoundsAsync(string sourceId)
// The widget's extent is measured from the captured frame itself: the
// bounding box of non-transparent pixels (the background is injected
// transparent, so the alpha channel IS the widget). This is immune to
// DOM timing (fonts/images/animations), needs no JS round-trip, and
// CANNOT break rendering: if the widget fills the canvas the box equals
// the full frame (identical to the confirmed-good uncropped display); if
// it floats in transparent margin, the crop maps it flush under the box.
internal static (int X, int Y, int W, int H) FindContentBounds(byte[] pixels, int pixW, int pixH, int stride)
{
if (!_sessions.TryGetValue(sourceId, out var session) || session.Disposed) return;
try
int minX = pixW, minY = pixH, maxX = -1, maxY = -1;
for (var y = 0; y < pixH; y++)
{
if (session.Control.CoreWebView2 is not { } cws) return;
var json = await cws.ExecuteScriptAsync(
"JSON.stringify({w:document.documentElement.scrollWidth||0," +
"h:document.documentElement.scrollHeight||0," +
"vw:window.innerWidth||0,vh:window.innerHeight||0})");
if (string.IsNullOrEmpty(json) || json == "null") return;
using var doc = JsonDocument.Parse(json);
var root = doc.RootElement;
session.ViewportCssW = root.GetProperty("vw").GetDouble();
session.ViewportCssH = root.GetProperty("vh").GetDouble();
session.ContentCssW = root.GetProperty("w").GetDouble();
session.ContentCssH = root.GetProperty("h").GetDouble();
_ = CaptureFrame(sourceId);
var row = y * stride;
for (var x = 0; x < pixW; x++)
{
if (pixels[row + x * 4 + 3] == 0) continue;
if (x < minX) minX = x;
if (x > maxX) maxX = x;
if (y < minY) minY = y;
if (y > maxY) maxY = y;
}
catch (Exception ex)
{
AppLog.Write($"WebView2Manager: content bounds query failed: {ex.Message}");
}
if (maxX < minX || maxY < minY) return (0, 0, pixW, pixH);
return (minX, minY, maxX - minX + 1, maxY - minY + 1);
}
private async Task CaptureFrame(string sourceId)
@@ -223,23 +217,15 @@ public sealed class WebView2Manager : IDisposable
var stride = pixW * 4;
int cropW = pixW, cropH = pixH;
if (session.ViewportCssW >= 1 && session.ViewportCssH >= 1 &&
session.ContentCssW >= 1 && session.ContentCssH >= 1)
{
var sx = pixW / session.ViewportCssW;
var sy = pixH / session.ViewportCssH;
cropW = (int)Math.Clamp(Math.Round(session.ContentCssW * sx), 1, pixW);
cropH = (int)Math.Clamp(Math.Round(session.ContentCssH * sy), 1, pixH);
}
var (cropX, cropY, cropW, cropH) = FindContentBounds(pixels, pixW, pixH, stride);
var outPixels = pixels;
if (cropW != pixW || cropH != pixH)
if (cropW != pixW || cropH != pixH || cropX != 0 || cropY != 0)
{
var dstStride = cropW * 4;
outPixels = new byte[dstStride * cropH];
for (var r = 0; r < cropH; r++)
Buffer.BlockCopy(pixels, r * stride, outPixels, r * dstStride, dstStride);
Buffer.BlockCopy(pixels, (cropY + r) * stride + cropX * 4, outPixels, r * dstStride, dstStride);
}
var wb = new WriteableBitmap(cropW, cropH, 96, 96, PixelFormats.Bgra32, null);
+9 -7
View File
@@ -917,14 +917,16 @@ capturing the primary display (which causes the infinity mirror effect).
8. ✅ Preview shows live web content (not just a blank rectangle)
9. ☐ Handle navigation errors, invalid URIs, timeout gracefully
**Rendering model (2026-08-28, canvas-size viewport, NO crop — the only creator-verified state):**
**Rendering model (2026-08-28, canvas-size viewport + ALPHA-BBOX crop):**
the page renders at the MASTER CANVAS size (1920×1080), stable, never tracked (no reflow/truncation;
scrollbars suppressed via `overflow:hidden`). Capture is the FULL frame; `Stretch="Fill"` maps it
under the box. Image rendering verified good ("renders and resizes beautifully"). WEB BOX DEAD SPACE
STILL OPEN: previous cropping attempts all failed (JSON double-encode silently disabled the crop;
scrollWidth returns the full canvas for full-canvas wrappers; content-union measured an unsettled
layout on NavigationCompleted and BROKE the image). Any further crop needs creator approval first.
⚠️ `ExecuteScriptAsync` JSON-encodes the returned value — never return `JSON.stringify` from script.
scrollbars suppressed via `overflow:hidden`). `FindContentBounds` scans the Bgra32 capture and crops
to the bounding box of non-transparent pixels — the widget's true rendered extent, measured from the
frame itself (NO DOM query, immune to layout timing; can never truncate content; full-canvas widgets
fall through to the full frame = prior verified-good image). `Stretch="Fill"` maps the cropped frame
flush under the selection box → box hugs the widget on all four sides. `QueryContentBoundsAsync` +
the JS content-bounds script are DELETED. ⚠️ Lesson bank: never return `JSON.stringify` from
`ExecuteScriptAsync` (double-encodes); never measure DOM stuff on NavigationCompleted (unsettled
layout broke the image — `a5b9952`); measure rendered pixels instead.
**Properties panel (2026-08-28):** web URI ✓/✕ icon buttons are `IsTabStop="False"` so Tab flows
X→Y→W→H→URI; the ✕ button now clears the URI textbox (was reverting to the pre-accept snapshot).
+44
View File
@@ -110,4 +110,48 @@ public sealed class WebView2ManagerTests
Assert.Equal(System.Drawing.Color.Transparent, webView.DefaultBackgroundColor);
});
}
[Fact]
public void FindContentBounds_Returns_Bounding_Box_Of_NonTransparent_Pixels()
{
const int w = 10, h = 6;
var pixels = new byte[w * h * 4];
for (int yy = 2; yy <= 4; yy++)
for (int xx = 3; xx <= 6; xx++)
{
var i = (yy * w + xx) * 4;
pixels[i] = 255; // B
pixels[i + 1] = 0; // G
pixels[i + 2] = 0; // R
pixels[i + 3] = 255; // A
}
var (x, y, bw, bh) = WebView2Manager.FindContentBounds(pixels, w, h, w * 4);
Assert.Equal((3, 2, 4, 3), (x, y, bw, bh));
}
[Fact]
public void FindContentBounds_No_Alpha_Means_No_Crop()
{
const int w = 10, h = 6;
var pixels = new byte[w * h * 4];
var (x, y, bw, bh) = WebView2Manager.FindContentBounds(pixels, w, h, w * 4);
Assert.Equal((0, 0, w, h), (x, y, bw, bh));
}
[Fact]
public void FindContentBounds_Full_Opacity_Means_No_Crop()
{
const int w = 10, h = 6;
var pixels = new byte[w * h * 4];
for (var i = 3; i < pixels.Length; i += 4)
pixels[i] = 255;
var (x, y, bw, bh) = WebView2Manager.FindContentBounds(pixels, w, h, w * 4);
Assert.Equal((0, 0, w, h), (x, y, bw, bh));
}
}