Skip to content

History / App Shell and Layout

Revisions

  • docs(app-shell): record the drawio viewer and the shared zoom/pan stage The editor's drawio viewer is now `@dskripchenko/drawio` behind a thin adapter, and its zoom/pan is `common/zoom-pan-stage` — the same surface an image gets. The entry names the measured reasons: the library's two page encodings, the three decisions the adapter owns, the 17.9 KB gzip cost against the GPL and maxgraph alternatives, and the two refit behaviours (stage resize while unzoomed, new content of the same size).

    @rajebdev rajebdev committed Nov 12, 2026
  • docs(editor): record the spreadsheet grid's interaction and save paths The grid is now a spreadsheet, and three decisions in it fail silently when re-broken: selection is anchor+focus (a single cell cannot express shift-click or a drag), CSV edits go through the text buffer while workbook edits go through bytes (a multipart text field corrupts a ZIP), and a workbook save is lossy in one direction (SheetJS writes values, formulas and number formats; styling, charts and images are not read). The entry also records why `.xls` needs SheetJS and what it costs.

    @rajebdev rajebdev committed Nov 12, 2026
  • docs(app-shell): record the backdrop-root cost of the glass plane and dialog The plane and the dialog tier each carried a `backdrop-filter` at window scale. Measured in Chromium at 1280x860 @2x: +211 MB for the plane, +209 MB for a 1280x810 dialog, +48 MB for a 400x300 popover, +325 MB for the set. Both window-sized filters were removed; the rendered frame differs by a mean of 2.58/255 and 0.97/255, so the material is unchanged. Also records that a WKWebView harness on the same page measured zero delta — a non-bundled binary never takes the accelerated path, so it is not a valid probe for backdrop cost.

    @rajebdev rajebdev committed Nov 12, 2026
  • docs(editor): record the content-hash disk watcher and the Reload control The editor page listed the toolbars' rules but not how an open tab follows a file written from outside. Adds the one bullet that names the measurement and the two failures it replaces: an mtime-only probe reloading a tab over bytes that did not change, and a clean buffer never being checked at all.

    @rajebdev rajebdev committed Nov 11, 2026
  • docs(toasts): record the three tones and their icon mapping The toast card is no longer success/error only: the tone also picks the glyph and its tint, and `warning` carries partial outcomes that used to be forced into one of the other two. The entry names the mapping, the budget each tone pays, and the two call sites that had the tone wrong — a settings save announced as an error, and a failed worktree creation painted as a success.

    @rajebdev rajebdev committed Nov 11, 2026
  • docs(sidebar): record the unified session tree, guides and child mark The sidebar's child sessions and subagent roster are one branch behind one chevron. Document the 22px slot-to-harness pitch, the stem's vertical clearance from the chevron, the 2px strokes that abut instead of overlapping (the translucent double-paint read as darker bands), and the GitBranch child mark for a leaf child. Include the measured Chromium geometry (harness centre x=42, child slot centre x=42, line centre x=20, stem start 4.7px below the chevron ink).

    @rajebdev rajebdev committed Nov 10, 2026
  • docs(wiki): record the phone dialog safe-area and single-scroller rules Measured on the About sheet at 390x844: full-bleed box `[0,0,500,913]` put the wordmark and close button under the status bar, a `46vh` cap on the harness list plus the card's own `overflow-y-auto` made two nested scrollers, and the desktop `9.5rem` version track starved the agent name at 324px of content width (168/112/28px after the phone pair). The stale-stylesheet trap is noted too: an arbitrary `grid-cols-[…]` utility is compiled by the build, so a dev server started before the edit serves a sheet with no track at all.

    @rajebdev rajebdev committed Nov 8, 2026
  • docs: record how a provider lands in each agent's registry Adds the measured detail for the two new surfaces: the shared landing path for "copy" and "type" (probe first, protocol from the measurement, refusal by name, env flattening at the host root, per-provider credential variable names), and the per-harness field plan that stopped drawing a `GEMINI_API_KEY` hint on a Claude Code card. Also corrects kimi's registry path to `~/.kimi-code/config.toml` — the retired Python CLI's `~/.kimi/` is left in place by `kimi migrate`, so the old row described a directory this build never writes — and records the searchable listbox's four rules in the app-shell page.

    @rajebdev rajebdev committed Nov 6, 2026
  • docs(icons): measure the light brightness ceiling and the four changes that raised it The report was "light mode kusam, dark cerah". Three causes, each with the number that identifies it: - One tone table solved against the darkest light paper, which cost every brighter palette 0.05 of OKLCH lightness. Six buckets lift the weighted mean 0.6198 to 0.6260; a table per palette would only reach 0.6335, so the granularity is stated along with why six was chosen. - A solver that preferred chroma over brightness, so it spent lightness on colour the hue could not use. - A muting rule that lost contrast while removing chroma: 2.96:1 on a 13px folder before the compensating darkening. The ceiling is recorded too, because it is the answer to the next request for "brighter": at 3.15:1 on `#f1f1f0` the brightest legal lightness averages 0.632, and L 0.78 on that paper is 1.96:1 — invisible. Light cannot match dark.

    @rajebdev rajebdev committed Nov 6, 2026
  • docs(app-shell): record the toast queue, its host and the measured delay Eleven toast hosts each drew a fixed bottom-right portal, so two surfaces toasting at once printed two cards in the same rectangle, and the card's countdown restarted on every parent render — a flat 4000 ms that was both too short for the sentences this app pushes (24-89 chars, median 54) and too long for a two-word confirmation. The page now has one queue and one host, and the delay is derived from tone and length. Records the measured numbers behind the rule (a 62-character card on screen for 5515 ms, two cards stacked with no overlap, hover and focus holding the countdown, the 390px clamp above an open drawer) so a later change that reintroduces a per-surface host is recognised as the regression it is.

    @rajebdev rajebdev committed Nov 6, 2026
  • docs(app-shell): record the composer's red Stop / blue Send pair Names the fill token and the measurement instead of the rule alone: the raw palette hex is a chip ground (2.73:1 error on `mono-light`, 1.94:1 info on `aura-light`), the solved `-text` slot is the fill a control can rest on, and the hover mix keeps 4.71:1 where `hover:opacity-90` fell to 3.77:1. Also records the half that is not visible in the component: the DEFAULT icon mode painted every glyph on a solid fill in the palette ink, measured 1.04:1 on `osaka-jade-light`, so the fill rule had to be stated for monochrome as well.

    @rajebdev rajebdev committed Nov 5, 2026
  • docs(shell): record the chrome row's one density module The navbar and the sidebar action row are the same surface and stated their own control plate and gap; measured, the phone header's glyphs sat 48px apart against the drawer's 38px. The new entry names the module that now owns the numbers, the measurement behind it, and why `.tap-target`'s 44px floor stays scoped to controls that stand alone.

    @rajebdev rajebdev committed Nov 5, 2026
  • docs(app-shell): record the backdrop-root finding behind opaque composer popovers A popover nested in a frosted surface has no working blur because its host is a backdrop root. Records the stripe-probe measurements (0/255 variance 16216 nested vs 221/224 variance 0.7 portalled), the tint sweep showing that raising opacity alone does not fix it, and the scope guards verified by the same probe: body-portalled menus, the Target Workspace picker above the card, a classic composer, and dialog dropdowns all keep 84% with a live blur.

    @rajebdev rajebdev committed Nov 5, 2026
  • docs(icons): measure the dark solve, the mute floor, and the two rules that outrank the mode Six facts, each with the number that justifies it: - Dark is solved per hue in a pinned lightness band; one `hsl()` pair spread its lightness over 0.25 and left the muddiest row at a quarter of the vividest row's chroma. - A mute keeps a chroma floor: `text-ink/40` painted 0.065 where the tone is 0.163, and 169 of 1139 coloured glyphs fell under 0.08. - An author's signal class outranks the generated hue, including its slashed forms. - A solid signal fill owns its icon's foreground: `Trash2` rendered 1.00:1 on `bg-error`. - A project accent is clamped for the glyph and exact for the swatch; nine of ten presets fail 3:1 on light paper. - No icon is painted by a Tailwind palette rung, enforced by a source scan. The two solved idle filters are corrected here too, since the shipped values changed.

    @rajebdev rajebdev committed Nov 5, 2026
  • docs(composer): record the armed goal and the dialog portal rule Two facts a future reader would otherwise re-break, each with its measurement. A Goal is ARMED: the modal records the intent, and the next prompt creates the record with `startTurn: false` so that prompt is turn one. The smoke on omp 18.16.0 recorded `tokensUsed: 0` before the prompt and `{user: 1, assistant: 4, toolResult: 3}` after it; the rejected design that opened its own turn produced `user: 0`. A guided interview is armed too, with the operator's message as the kickoff content, and `{{else}}` now renders (both branches used to land in the prompt). A dialog mounted from the composer must portal. Re-parenting the Goal dialog's real scrim into `[data-composer-style]` took it from 1440x900 at (0,0) to 710x164 at the card's box, so `inset-0` sized the dialog to the composer instead of the viewport. `backdrop-filter`, `transform` and `contain: layout` each trigger it; `container-type` alone does not, which corrects the first note.

    @rajebdev rajebdev committed Nov 5, 2026
  • docs(app-shell): record how an icon weight is muted `text-ink/60` is a weight, and the first answer reproduced it by mixing the hue toward the theme ink — which desaturates AND darkens. Measured on the settings sidebar, a muted icon read 7.48:1 where the active one read 3.96:1 and the monochrome ink it replaced read 4.74:1: darker than both, which is the "masih kusam di light mode" report. The page now records the rule (multiply chroma, leave lightness), the numbers (contrast holds at 3.87 mean / 3.07 min from /90 to /25 while chroma falls 0.146 -> 0.040), and the three selector facts that each cost a bug: the paint is stated once from a `--icon-color` source so a weight transforms whatever produced the colour; `[class~=...]` rather than `[class*=...]`, which also fired on the variant token and painted resting icons at their hover weight; and the paint rule covers provider and file-type marks too, which carry no `lucide` class and were inheriting their row's muted ink.

    @rajebdev rajebdev committed Nov 5, 2026
  • docs(app-shell): record why a light icon hue is solved per hue The color mode's light paint was one `hsl(H 60% 30%)` for all 36 generated hues, pinned to the 3:1 non-text floor, so its saturation was whatever the worst hue tolerated. Chroma is not evenly spread on the wheel: the pair spans OKLCH chroma 0.075 to 0.164 (mean 0.121) while dark's pair reaches 0.126 with 4.07:1 of headroom unspent, which is why light read muddy and dark read bright. Raising the pair is measured and rejected — 85% drops the worst hue to 2.66:1 and 100% to 2.32:1, both under the floor — so the tone is solved per hue in OKLCH instead, capped at chroma 0.19 and checked against every shipped light palette. Mean chroma 0.121 → 0.162, the muddiest row 0.075 → 0.102, every hue improves, worst contrast stays 3.15. The page also records what the change forced: the answer ships as a generated table because the solve needs the palette catalog and ~11k contrast checks the client must not pay for, `--icon-tone` is stamped at runtime by `PanelMark` and `PluginMark` because their hues cannot be generated rows, and the rail's light idle filter had to move from `brightness(1.25)` to `1.08 saturate(1.1)` once the paint started at the floor rather than 6.11:1 above it.

    @rajebdev rajebdev committed Nov 4, 2026
  • docs(app-shell): record why the switch knob exclusion lives in the well rule A same-property override is a specificity race and lost: `[role="switch"] .bg-canvas` (0,4,0) against the well's (0,5,0) meant every knob computed transparent under glass, on both layouts. The measure, the A/B (glass vs classic on one tab) and the reason the exclusion is a `:not()` inside the one rule that causes it are recorded so the next "…except X" rule is not written that way.

    @rajebdev rajebdev committed Nov 4, 2026
  • docs(app-shell): record why the phone's stuck header painted nothing The sticky folder header's material had two phone-only failures that both presented as "sticky active, background transparent", and neither is visible from the desktop column that the original measurement used. The scrollport search required an overflow, which reads 0/0 for a subtree mounted `hidden` — the drawer's own lifecycle — so the hook listened to `window.scrollY` instead of the list. The canvas well rule then outmatched the stuck fill on specificity because the touch header paints `bg-canvas`. Measured records and the resulting rules: `overflow-y` alone decides the scroll container, a `ResizeObserver` re-measures when the drawer appears, and the well selector excludes `.glass-sticky`.

    @rajebdev rajebdev committed Nov 4, 2026
  • docs(app-shell): record the panel Update buttons and the untracked-file defect Two entries with their measurements: which source decides what Update means (and that the bundled path keeps the sync's edit gate), and the clone "dirty" test that counted Kabmo's own build output as the user's edit — measured end to end as a 400 on a fixture repo with no .gitignore.

    @rajebdev rajebdev committed Nov 4, 2026
  • docs(app-shell): sweep the glass tiers over all 63 palettes The tier sweep recorded 43 palettes, which was the catalog's size when it was measured. It is 63 now that the nine newer OpenChamber families are ported, so the number in the rule had drifted from the catalog it describes. The measurement itself still holds: the worst case is `solarized-light`'s chrome at 4.89:1. What changed is how it is kept — `glass.test.ts` reads the alphas out of `glass.css` and composites every catalog palette, so a newly ported theme cannot land without clearing the 4.5:1 floor on the plane, the chrome or a card.

    @rajebdev rajebdev committed Nov 4, 2026
  • docs(app-shell): record the glass tiers, the icon paint contract and panel hues The rules a future reader would otherwise re-break, each with its measurement: the fallback blocks need `:root` to tie the dark-variant selector; a palette hue darkens paper, so light gets no bloom; `color-mix()` inside `background-image` drops the whole layer; `@layer utilities` needs the full order stated first; a weight does not apply to the user's own colour; a declared brand colour is not dimmed in the rail; and a panel's hue is keyed by the panel id, never by the glyph — measured, `Files` was `hsl(0) on the rail and `hsl(320)` in Settings.

    @rajebdev rajebdev committed Nov 4, 2026
  • docs(app-shell): record why the header halo is white and how strong The halo took three measured attempts before it looked right, and the two rejected ones are the shapes a future reader would reach for first: a brand-coloured glow at low alpha is invisible, and at high alpha it tints the whole bar. The numbers are in the bullet so the next change starts from a measurement rather than from taste.

    @rajebdev rajebdev committed Nov 4, 2026
  • docs(chat): record why the landing draft is memory-only The composer's draft used to persist under the provider's fallback session id when the URL named no chat, and nothing on that disabled landing could clear it. Measured: a browser-panel annotation block survived every launch and leaked into `?sessionId=1`, which is a real mock chat. Names the scope rule, where it lives (`use-draft.ts`), and the superseded symptom so the fallback bucket is not read as a draft store again.

    @rajebdev rajebdev committed Nov 3, 2026
  • docs(browser): record the agent panel's tab-close rules Measured on Chromium 1248 against a tab mid-screencast: closeTarget answers success, the browser socket survives, the attached session dies on its own and the remaining tabs stay attachable. Records why the strip is drawn from one tab upward and why a re-close reports alreadyClosed.

    @rajebdev rajebdev committed Nov 3, 2026
  • docs(layout): name the macOS UI delegate by what it now does `UploadPanelDelegate` became `ShellUIDelegate`: it answers `runOpenPanel` and `createWebViewWith` on the ONE `WKUIDelegate` the shell installs, because WebKit asks a single object for both and a web view with no `createWebViewWith` answers `window.open` / `target="_blank"` by doing nothing. The row keeps the measured reason the class lives in the library target — the window suite presses the shipped object instead of a stand-in — and adds the popup half, which a restated case would have hidden.

    @rajebdev rajebdev committed Nov 3, 2026
  • docs(layout): record the portalled-dropdown rule and the two menus it fixed

    @rajebdev rajebdev committed Nov 3, 2026
  • docs: record the file panel behind a file input and the Projects writes App-Shell-and-Layout: a `WKWebView` with no `uiDelegate` documents every file input as Cancel, which is why the macOS shell's upload buttons were dead while the same page worked in a browser. Also the element's `accept` never arrives in `WKOpenPanelParameters`, the dynamic-UTType trap that empties the panel, and why Windows, Linux, iOS and Android came out differently. Omp-Engine-and-Settings: the Projects pane writes one validated field at a time with named refusals, why the default model is `provider:modelId`, and why the custom accent's listeners are native rather than `onChange` props under `preact/compat`.

    @rajebdev rajebdev committed Nov 2, 2026
  • docs(app-shell): record why an SVG preview must own its ids DOMPurify deletes an `id` whose value the page already has (`body`, `title`) and leaves the `url(#id)` referencing it, so a gradient-painted shape silently adopts the wrong paint — measured rgb(9,77,61) against rgb(246,232,197), and 4648 vs 1003 differing pixels against the rasterised reference. Notes the prefix-plus-rewrite rule in svg-scope, why SANITIZE_NAMED_PROPS does not fix it, and that the happy-dom sanitizer skips the clobber check so only a browser measures this.

    @rajebdev rajebdev committed Nov 2, 2026
  • docs(editor): whitespace marks are drawn on the selection only The editor surface now paints a dot per space and an arrow per tab for the text a reader has selected, and nothing for a bare caret. Records why both always-on variants were rejected — the package's `highlightWhitespace()` marks every space in the viewport, and a leading-run-only mark is a permanent dotted layer over every block — and the two bounds that keep the plugin cheap: empty selection ranges are skipped, and the walk is the selection intersected with the viewport. Also drops the stale cross-reference to `indent-guide.ts`, which no longer exists: the per-level indent guides were deleted and the conflict-marker walk now describes itself.

    @rajebdev rajebdev committed Nov 2, 2026