1. summarizeCanvas() walked only node.nodes, so everything inside Section/ColumnLayout/forms
(linkedNodes) was missing and live self-reviews could only name top-level sections. It now
walks nodes + linked slots (stable key order) and lists id, type and key props (text first,
links, main style values as style.<key>, colour-like props) as key="value", within a
documented budget (150 nodes, 6000 chars, 300 chars of props per line; detail is dropped
before nodes; cuts happen at line boundaries -- a mid-line cut used to make the server
refuse the whole review). HtmlBlock code is no longer summarised.
applyPatch no longer counts an op with nothing left to set as a change; ops that did nothing
are counted per reason. The self-review note and console say why ops were dropped (server
and editor reasons, counts only). Real-browser test 8n builds a nested page, sends the real
summary through the real WHP review lockdown (php) and checks the nested edit takes effect
while link/text edits, unknown ids and deletes are dropped.
3. Background placeholders use ?text=%E2%80%8B (measured: placehold.co then draws no glyphs;
the same-colour label left a 4-colour ghost in its PNG); the capture draws no <text> for a
zero-width or same-colour label. Pixel test on the case-4 hero markup in both browsers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0185ukXgk6UunFCKyXqLe2Cq
The capture stopped at 3000 CSS px, so the reviewer never saw the bottom or footer. Now the full page
height H is rendered at its normal width (1280). H <= 3000: one image, as before. Otherwise
N = min(3, ceil(H/3000)) equal-height strips (each <= 3000 px) are laid out side by side, left to
right (strip 1 = page top), with 24 px #808080 gutters, in one JPEG; H > 9000 shows the top 9000 px
(cropped). Geometry is the pure stripGeometry() (unit-tested); JPEG quality steps 0.8 -> 0.5 before
"too large" (3 MB cap, server limit 8 MB). CaptureResult gains strips/pageHeight; runSelfReview
returns them and says when only the top of a long page was reviewed. Real eval pages: 2 strips,
~0.8-1.1 s, 150-270 KB in Chromium and Firefox; e2e checks strip order/alignment/gutters/cropping.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured in Chromium and Firefox: the CSSOM exposes `background: var(--x)` and
`border: 1px solid var(--x)` only as pending-substitution longhands whose values
read back as "", so the capture sanitiser (which re-serialises the parsed
declarations) dropped them. The self-review screenshot then showed dark heroes,
image placeholders and card borders as white, and the model "fixed" things that
were fine on the real page.
The sanitiser now emits the declared shorthand once, at its first longhand,
through the same value/URL gate as every other declaration. Sandbox, CSP and
the single URL allow-list are unchanged.
- scripts/e2e-capture-fidelity.mjs: real-browser test through the whole capture
path (sanitise -> inline -> sandboxed frame -> html-to-image), decodes the
JPEG and samples pixels. Failed 9/12 checks per browser before, passes after.
- e2e-capture-security: hostile var() shorthand / var() fallback url() probes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>