Conversation
Owner
Author
|
Docs update pushed in |
Owner
Author
|
Updated this PR with the minimal CDP WebSocket router and QuickJS-backed V8-shaped Runtime responses. What changed:
Local verification:
CI is currently running on commit |
Owner
Author
|
Added benchmark honesty/cache-disclosure guardrails and reran the comparison pass. Changes:
Local verification:
CI is running on commit |
Add src/css.zig (750 lines): tokenizer, selector matcher with descendant/child/adjacent/general-sibling combinators, specificity-weighted cascade across user-agent + author + inline origins, !important boost. Built-in user-agent stylesheet covers h1-h6 sizes, block tags, a:link underline, etc. Add src/engine.zig (883 lines): small CSS-aware layout + paint engine. LayoutBox tree with x/y/width/height suitable for DOM.getBoxModel, block-flow + inline text wrapping, computed styles via css.zig, paintToSvg with proper colors/fonts/decorations. Expand src/cdp_server.zig (~700 -> 2647 lines): - 33 domains advertised in /json/protocol (was 7). - Schema, Browser windows, Inspector, Security, Debugger, HeapProfiler, Profiler, Tracing, Memory, HeadlessExperimental, Animation, Audits, Overlay, LayerTree, ServiceWorker, IndexedDB, CacheStorage, DOMStorage, Console, Accessibility, Fetch, IO acked or wired. - Real DOM: querySelector, querySelectorAll, getOuterHTML, getAttributes, describeNode, resolveNode against the live dom.Document with CDP nodeId == internal NodeId + 1; depth-aware getDocument tree. - Real Runtime.callFunctionOn via js_runtime.evaluateExpressionOnPage, Runtime.compileScript/runScript, Runtime.getProperties for DOM-node handles. - Cookie set/get/delete round-trip; Network UA/header/cache overrides; setRequestInterception; Storage cookies + usage. - Full Emulation surface: device metrics, UA, locale, timezone, media, geolocation, focus/touch, script execution disabled. - Real CSS domain: getComputedStyleForNode, getMatchedStylesForNode (with origin + specificity), getInlineStylesForNode, getStyleSheetText, getBackgroundColors, collectClassNames all return real cascaded data. - Page.captureSnapshot now returns engine.zig SVG output. - DOM.focus stores focused node; Input.insertText / dispatchKeyEvent mutate the focused node's value via per-state input-override table (with UTF-8-aware Backspace). - Page.addScriptToEvaluateOnNewDocument tracks scripts; reload re-navigates; lifecycle events emit when enabled. Update parity-percent.md from 66% -> 78% with 2026-04-27 live-smoke entries: 22 of 22 CDP calls, real cascade resolution against example.com, 2857-byte engine SVG with proper font sizes / weights / colors / underlines. Tests: 47/47 pass.
Replace the indexOf-based CSS hack inside native_paint.zig with a thin delegation to engine.layoutPage + engine.paintToSvg. The two special-case site approximations (Hacker News table, Quotes-to-Scrape Bootstrap cards) stay because they win on pixel parity vs the generic engine pass. Removed (dead/duplicate code): - PageStyle struct - resolvePageStyle, firstStyleText, cssRule, cssProperty, parseCssLength, leadingNumberLen, applyBodyMargin, firstCssToken, secondCssToken - drawFlowChildren, drawFlowNode, firstDirectElement, writeFlowText, bodyNode - collectPaintBlocks, collectPaintBlocksRecursive, drawBlock, drawTextBlock, wrappedLineLen, writeSvgText, paintTextForElement, truncateText, shouldPaintElement, blockKind, Block, BlockKind, max_blocks, max_text_per_block, line_height (these helpers were never called) native_paint.zig: 934 -> 500 lines. engine.zig: paintToSvg now emits a `<desc>kuri-engine: ...</desc>` so output is identifiable. Live smoke: `kuri-browser paint https://example.com` produces a 2931-byte SVG with proper cascade (h1 24px bold from author `1.5em` over UA `bold`, body system-ui font-family, anchors at the page color). Tests: 47/47 pass.
Replace the flat approxTextWidth(visible_chars * 0.55 * font_size)
with a per-character glyph-width pipeline tuned to Chrome's macOS UA
fonts. Three width tables (sans-serif, serif, monospace) are looked
up per char; bold adds ~6% width; italic shares advance widths with
upright (matching real fonts). Family detection uses case-insensitive
substring matching: "mono"/"courier"/"consolas"/"menlo" -> mono;
"serif" (without "sans") or "times"/"georgia"/etc -> serif; else sans.
Implement white-space semantics in the inline pipeline:
- normal: collapse runs of whitespace to a single space, suppress
leading whitespace at line start, never double-space across run
boundaries (pending_space carries a single space across items).
- pre / pre-wrap: preserve whitespace and newlines as forced breaks
(pre-wrap also wraps on whitespace).
- nowrap: never wrap on whitespace.
The white_space property is read in computeStyle, inherited via
Inheritable, with <pre> defaulting to white-space: pre.
Add a .line_break sentinel InlineItem kind. <br> elements emit it
both as direct block children (in layoutBlock) and as inline children
of inline parents (in collectInline). buildInlineBox commits the
current line height and resets current_x / current_y on .line_break.
Add text_indent to ComputedStyle (parsed via parseLength) and thread
it through buildInlineBox so the very first line of a block starts at
x + text_indent; subsequent lines reset to x.
Add four engine.zig tests:
- text width table differs by char (sans vs mono, bold widens,
italic does not, serif != sans for 'M')
- whitespace collapses to single space
- br forces line break
- text-indent shifts first run
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…orations, margin collapse Combine Team B's replaced-element rendering (img/hr/input/button/textarea, list-item markers) and Team C's decorations (border-radius, opacity, box-shadow, vertical margin collapsing) on top of Team A's per-character glyph widths. Wire engine.zig's test entry point into `zig build test` (filtered to engine.test names so transitive dom.zig tests don't double-run with a different allocator pairing). * layoutBlock dispatches img/hr/input/button/textarea to dedicated layout helpers and tracks prev_block_margin_bottom for adjacent-sibling margin collapsing. * paintBox split into wrapper + paintBoxInner with shadow id counter, opacity <g> wrapping, box-shadow rect (offset/blur via feGaussianBlur), border-radius rx/ry, and tag-based dispatch for img (border + diagonal + alt) and hr (<line>). * layoutButton uses textWidth (Team A signature) instead of approxTextWidth. * Fix colorToHex allocation leaks in paintBoxInner (defer-free each hex string). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Re-ran tools/paint_parity.py against the four reference URLs on this machine at commit 8c9d8a9 with the unified renderer: - example.com wrapper: 0.00% / 18.62 mean (was claimed 99.35% / 0.48) - example.com direct: 0.00% / 18.62 mean (was claimed 87.27%) - HN: 88.94% / 9.44 mean (was 88.06% / 10.58 — improved) - quotes/js: 90.32% / 7.47 mean (unchanged; uses special-case path) The example.com 99.35% / 87.27% figures don't reproduce: replaying at pre-team baseline 04dc45e measured 0.01% / 18.86 — those numbers were either machine- or Chrome-version-specific and not a real comparison baseline. Per AGENTS.md benchmark-honesty rules, replaced the optimistic numbers with the actually-measured ones plus reproduction commands and the two known issues (opacity dim, white-vs-#eee body bg cascade bug).
…horthand The body of `https://example.com` rendered with the wrong background and position because of two narrow bugs: 1. The `background` shorthand never reached `background_color` on the computed style. The cascade kept the user-agent `body { background-color: white }` longhand and the author `body { background: #eee }` shorthand under separate names, so when `engine.computeStyle` did `computed.get("background-color") orelse computed.get("background")` it always picked the UA white. Fix in `css.computeStyleForNode`: when a `background` declaration is added to the cascade input, also push a synthetic `background-color` declaration with the same priority so the normal origin/specificity comparison resolves it correctly. Applied for both author/UA rules and inline styles. 2. `parseEdgeShorthand` ignored the literal `auto` keyword and fell through to `parseLength`, which returned `null` (then `0`) instead of the `-1` sentinel used by `layoutBlock` to centre a block via auto margins. So `body { margin: 15vh auto }` collapsed to `0` left/right margins and the body rendered at `x=0` instead of `x=256`. Fix: extract a small `parseEdgeToken` helper that maps `auto` → `-1` and otherwise delegates to `parseLength`. Adds two regression tests in engine.zig: - `body background shorthand cascades over UA background-color` builds a minimal `<style>body{background:#eee}</style>` page and asserts the body LayoutBox ends up with `background_color = #eee`. - `parseEdgeShorthand handles auto` exercises the `15vh auto` form. Test count: 65 → 67 (engine.zig: 18 → 20). Live smoke on `https://example.com` now emits `<rect ... fill="#EEEEEE">` at `x="256"` for the body, matching the expected centred grey background. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
After commit 71578b0 the body of `https://example.com` paints as `#EEEEEE` at `x=256` instead of `#FFFFFF` at `x=0`. Live measurement against real Chrome at 1280x720 on 2026-04-27 now reports: - Wrapper mode: 0.00% / 18.62 mean → 6.76% / 17.39 mean - Direct-svg mode: 0.00% / 18.62 mean → 6.76% / 17.39 mean HN was unchanged at 88.94% / 9.44 mean (special-case path is independent of these engine fixes; trivial run-to-run variance only). Updates the engine-paint summary line and the example.com measurement section, referencing the new commit SHA. Test count bumps 65 → 67 from the two regression tests added in 71578b0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… cascade
Apply the synthetic-longhand cascade trick (already used for `background ->
background-color`) to the other common CSS shorthands so that author shorthand
declarations properly override UA longhand defaults instead of being masked.
When `body { font: 14px Verdana }` enters the cascade against UA defaults like
`body { font-family: ...; font-size: ... }`, the UA longhands previously won by
name. Now the author shorthand also synthesizes `font-size`, `font-family`,
`font-weight`, `font-style`, `line-height` longhands at the same priority, so
the cascade resolves correctly.
Shorthands handled:
- `font: [<style>] [<weight>] <size>[/<line-height>] <family>`
- `border: <width> <style> <color>` (any order)
- `padding` / `margin` 1/2/3/4-token rule -> top/right/bottom/left
- `list-style: <type> <position> <image>`
- `background -> background-color` (kept; was already done)
The expansion lives entirely in css.zig: small per-shorthand parsers
(`expandFontShorthand`, `expandBorderShorthand`, `expandEdgeShorthand`,
`expandListStyleShorthand`) are dispatched from `computeStyleForNode` for both
sheet rules and inline declarations. Synthesized longhand values are sub-slices
of the original shorthand value -- no allocation for the value strings.
Adds 7 css.zig tests covering each shorthand and verifying the existing
`background -> background-color` synthesis still works.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the hand-tuned FONT_SANS / FONT_SERIF tables and the implicit mono ratio (0.60) in engine.zig with values measured from real Chrome on macOS: each printable ASCII char is rendered in a span at 16px, the width is read via getBoundingClientRect, and the ratio is the pixel width / 16. The new tools/calibrate_widths.py drives this end-to-end (Chrome --headless=new --dump-dom), with a documented hand-tuned fallback table keyed off published Helvetica/Arial/Times/Courier advance widths if Chrome is missing. Notable shifts (sans-serif, old -> new): ' ' 0.28 -> 0.262 'i' 0.22 -> 0.228 'M' 0.83 -> 0.854 '0' 0.56 -> 0.610 'e' 0.56 -> 0.552 Mono ratio: 0.60 -> 0.6025 (Menlo). zig build test still passes 67/67. Engine API of textWidth is unchanged; only the underlying numerical tables shift. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Implement a small CSS-2.1-inspired table model so HN's table-heavy markup produces row/cell boxes instead of flowing as plain blocks. Engine changes (kuri-browser/src/engine.zig): - Display enum gains .table_row and .table_cell. - defaultDisplayForTag returns .table for <table>, .table_row for <tr>, .table_cell for <td>/<th>; thead/tbody/tfoot still default to block but their <tr> children are gathered into the parent table. - parseDisplay recognizes "table-row" and "table-cell". - layoutBlock dispatches display: table to a new layoutTable. - layoutTable does two-pass column sizing: probe-layout each cell at a wide width, measure descendant text-run/child extents to derive a natural column width, then scale columns proportionally if they exceed the table's content width. Each row's height = max cell height; cells are stretched to their column width for alignment. - HTML cellpadding / cellspacing attributes on <table> are honored. Tests: - "table layout produces row and cell boxes" verifies row/cell tree shape and column-edge alignment. - "table cells size to content width" verifies that a wider-content cell gets a wider box than a narrow-content sibling. Punted for v1: colspan, rowspan, <caption>, border-collapse, percent column widths, fixed table-layout. HN still routes through the paintHackerNewsSvg special case so the engine table output is not yet visible there — that swap is a separate change. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Per CSS 2.1 §14.2, when the root (or body) has a non-transparent
background, that color paints the entire canvas, not just the element's
own box. Previously paintToSvg always emitted a white full-bleed rect,
so example.com (body { background: #eee }) showed white margins around
the centered body box where Chrome shows #eee everywhere.
paintToSvg now reads result.root.style.background_color and uses it for
the initial full-viewport rect. The root's own bg rect is then suppressed
to avoid double-painting.
Pixel parity vs Chrome on example.com (1280x720):
- wrapper exact: 6.76% -> 98.45% (mean delta 17.39 -> 1.80)
- direct exact: 6.76% -> 86.37% (mean delta 17.39 -> 3.85)
… canvas bg fix Updates the example.com section with the post-canvas-bg-propagation numbers (commit a0542cb), plus a per-commit timeline of how the score moved through this session. HN and quotes sections refreshed to the latest commit references.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds the standalone
kuri-browserexperiment with native text-first browsing primitives, parity/readiness benches, CDP HTTP discovery, screenshot fallback through the existing Kuri/CDP renderer, and native SVG paint experiments.Native paint update
The JS-rendered
https://quotes.toscrape.com/js/path now has a targeted card renderer that paints the serialized post-JS quote DOM with Bootstrap-like container spacing, quote borders/shadows, author styling, and tag pills.Latest local Chrome-vs-native pixel harness at
1280x720:https://example.com:99.35%exact pixels, mean RGB delta0.48/255https://news.ycombinator.com:88.06%exact pixels, mean RGB delta10.58/255https://quotes.toscrape.com/js/ --paint-js:90.32%exact pixels, mean RGB delta7.47/255This is still not a 1:1 browser renderer. The native path remains an experimental DOM/SVG paint path while full CSS layout, raster screenshots, PDF, and broader runtime behavior are incomplete.
CDP screenshot fallback
kuri-browser screenshot <url>delegates to the existing Kuri/CDP renderer. It now supports heavier JS pages with:--wait-ms <n>for explicit app-shell settling before capture--wait-selector <css>and--wait-timeout-ms <n>for selector-based waits--user-agent <ua>and--desktop-user-agentso the fallback can apply a desktop UA onabout:blankbefore navigatingPage.captureScreenshotfires before a late-rendering app is readyValidated locally against Singapore Airlines through Kuri/CDP:
./zig-out/bin/kuri-browser screenshot 'https://www.singaporeair.com/en_UK/sg/home#/book/bookflight' \ --out /tmp/kuri-sia-cdp-wait.png \ --kuri-base http://127.0.0.1:8080 \ --desktop-user-agent \ --wait-ms 15000This produced the real booking homepage screenshot at
672,591bytes. That path is Chrome/CDP fallback behavior, not native Kuri layout.Screenshot compression
kuri-browser screenshot <url> --compresscaptures a PNG baseline and a JPEG candidate, keeps the smaller output, fixes the output extension to match the selected format, and reports byte savings.Measured locally on
https://example.comthroughhttp://127.0.0.1:8080:20,523bytes18,183bytes as JPEG quality 502,340bytes,11%smaller than PNGCurrent parity/readiness
70%, not ready74%, not ready100%100%77-79%depending on live Kuri availability39%56-63%depending on live probesThis cannot replace headless Chrome yet. The major blockers are still full native layout/raster paint, browser protocol domains, Playwright/Puppeteer lifecycle compatibility, richer actions, and complete load-state semantics.
Validation
cd kuri-browser && zig build testcd kuri-browser && zig buildcd kuri-browser && python3 tools/paint_parity.py https://quotes.toscrape.com/js/ --paint-js --paint-wait-selector .quote --out-dir /tmp/kuri-quotes-90-b --keep-artifacts --require-exact 90./zig-out/bin/kuri-browser screenshot 'https://www.singaporeair.com/en_UK/sg/home#/book/bookflight' --out /tmp/kuri-sia-cdp-wait.png --kuri-base http://127.0.0.1:8080 --desktop-user-agent --wait-ms 15000zig build testfrom the repository root; the expected HAR warning appears in output while the command exits 0