fix(mobile): chat and composer chips stay readable in light mode#22585

open27

main

GitHub fork: fix/mobile-chip-light-contrast

opened2d agobyOoliver

18 files changed

View on GitHubTwo-way sync enabledPR last imported 42m ago

Note

🤖 Claude Opus 5.5 on behalf of Oliver

Update: dark themes and vivid accents

Colored dark themes put chips on a mid-tone user bubble where the shared accents nearly vanish. Ember's video chip read at 2.0:1 and Grove's terminal chip at 1.9:1. Composer and some reply chips in those themes were also below 4.5:1. The light-mode rule blended accents toward black, which greyed them out and overshot to about 6:1.

Each chip now gets one color per theme, the same in messages, agent replies and the composer:

themeAccentWithContrast keeps the accent's OKLCH hue and the most chroma sRGB allows, and moves only lightness, lighter on dark themes and darker on light ones, until the color reads at 5:1 on every chip surface (user bubble, thread canvas, composer). That leaves a margin over the 4.5:1 that normal text needs, because chip labels are small.
The user bubble is usually the hardest surface, so it decides the color. In colored dark themes, replies get the same pastel as the bubble, at about 9–11:1.
A surface whose text is the opposite shade can't share a color with the canvas, so it gets its own match. That's Material You's light theme, where the bubble is the primary color with white text.
Results are cached per accent and theme, so composer keystrokes don't redo the search.
In the default T3 Code dark theme, most accents brighten slightly; mention, terminal and thread are unchanged. Review-comment cards and PR markdown stay at 5.1:1 or better.

Evidence is the full theme matrix rendered from the real getMobileThemeVariables values and the shipped helper, as before vs after. Ratios are shown against each surface: red is under 4.5:1. All 360 pairs reach 5:1 after, versus 128 below 4.5:1 before.

Dark themes, before and after: user message, agent reply and composer per theme

Light themes, before and after: user message, agent reply and composer per theme

iOS simulator

iPhone 18 Pro, iOS 27.0. Before is main at 9a391eb3f5, after is 598d97f929. Both run on the same dev client and the same offline chat fixture, which has a user message with every chip kind, an agent reply and the composer. In the after shots, the agent reply's "Linked thread" renders as a chip because of the thread-link layer earlier in this stack, not because of the contrast change.

Ember dark, before and after

Grove dark, before and after

T3 Code dark, before and after

T3 Code light, before and after

Ember light, before and after

Validation: composerChipPresentation.test.ts checks every chip accent (including PR states) in every built-in theme and both modes. Each must be identical across the three surfaces and reach 5:1 on all of them, plus a Material You primary bubble. Also passing: mobileTheme, nativeTheme and enrichedLinkPresentation tests, mobile tsc-rs --noEmit, and lint on the changed files. useEnrichedLinkVariants.ts joins the allowUniwindTheme interop allowlist, because it hands raw colors to the native markdown renderer.

Made with Claude Opus 5.5 via Claude Code in T3 Code.

Original description by @flamboh

Note

🤖 Claude Opus 5.5 on behalf of Oliver

Important

Stacked PR 4 of 4. Review and merge in this order:

https://git.cafe/pingdotgg/t3code/pulls/22237: Thread links in chat render as chips
https://git.cafe/pingdotgg/t3code/pulls/22243: Thread chips show the thread's project icon
https://git.cafe/pingdotgg/t3code/pulls/22481: Projects without a favicon show a monogram, like desktop
https://git.cafe/pingdotgg/t3code/pulls/22585: Chat and composer chips stay readable in light mode ← this PR

Every PR in the stack targets main because the branches live on a fork, so this diff also includes https://git.cafe/pingdotgg/t3code/pulls/22237, https://git.cafe/pingdotgg/t3code/pulls/22243 and https://git.cafe/pingdotgg/t3code/pulls/22481. To review only this PR's changes, use this comparison.

When a PR earlier in the stack merges, I rebase the ones after it onto main.

Problem

Mobile chat chips have no fill or border. Their text and icon use the chip kind's full accent color, the sRGB form of web's OKLCH set. Those accents read fine in dark mode. On light surfaces most of them fall to about 3:1, below the 4.5:1 that normal-size text needs, and teal and green are the worst. Desktop avoids this because its chip text mixes only a little accent into the foreground, on top of a tinted pill. Mobile has no pill.

Fix

On a light surface, each accent is drawn in a darker shade of the same hue. The shade is the accent with its OKLCH lightness lowered just enough to reach 4.5:1 on every built-in light theme's user bubble and thread canvas. The values are precomputed in a small table in composerChipPresentation.ts, so rendering does no color math beyond one luminance check.
contextChipAccent(accent, textColor) returns the form to draw. It picks by the color of the surrounding text, not by the theme's appearance: dark text means the chip sits on a light surface. That matters for Material You on Android, where the light theme's user bubble is the saturated primary color with white text. Darkening chips there would make them worse, so they keep the shared accent.
Dark surfaces are unchanged.
This covers every chat chip: context records, thread links, skills, file links, and thread chips that show a project icon. Favicons, emoji and project monogram tiles are untouched. Monograms use the project's Tailwind color on their own tinted tile, not these accents.

Chip	Before	Bubble	Canvas	After	Bubble	Canvas	Worst built-in light theme
file	`#0090cd`	3.11	3.49	`#016795`	5.42	6.06	2.61 → 4.54
mention	`#0096af`	3.06	3.42	`#016b7d`	5.38	6.02	2.56 → 4.50
terminal, open PR	`#009f6e`	2.96	3.31	`#016f4c`	5.41	6.06	2.48 → 4.53
thread	`#009c96`	2.95	3.30	`#036d69`	5.39	6.03	2.47 → 4.51
skill	`#b261be`	3.42	3.83	`#90409c`	5.38	6.02	2.87 → 4.51
review, merged PR	`#8a70dd`	3.36	3.76	`#6a4eb8`	5.39	6.03	2.81 → 4.51
pull request	`#7079e4`	3.30	3.69	`#5155bd`	5.40	6.05	2.76 → 4.53
draft PR	`#7f8793`	3.16	3.54	`#5a626d`	5.37	6.02	2.65 → 4.50
image, closed PR	`#d55665`	3.42	3.83	`#af3346`	5.40	6.05	2.86 → 4.53
video	`#d06217`	3.36	3.76	`#a04701`	5.38	6.03	2.82 → 4.51
element, annotation	`#b87501`	3.28	3.67	`#895601`	5.39	6.03	2.74 → 4.51

Bubble and canvas are the default T3 Code light theme (#efeff1, #fcfcfc). The last column is the lowest ratio across every built-in light theme's bubble and canvas; the Grove bubble is the darkest of these surfaces.

Composer

#22518, now merged, gave composer chips these same accents. This PR applies the readable form there too: both composer editors (T3ComposerEditor.ios.tsx and T3ComposerEditor.native.tsx) pass contextChipAccent(presentation.accent, textColor) with the composer's own text color, so composer and message chips stay identical on light surfaces.

Evidence

The fixture thread has a user message holding every chip kind plus an assistant reply with thread and file chips. The installed dev clients ran this branch's JS over Metro.

iOS light mode, before and after

Android light mode, before and after

iOS dark mode, before and after: unchanged

Validation

composerChipPresentation.test.ts: every chip accent, including each PR state, reaches 4.5:1 on each built-in light theme's bubble and canvas when drawn with that surface's text color. Dark themes and a Material You primary bubble keep the shared accents.
reactCompilerCoverage.test.ts, mobile tsc --noEmit, and lint on the changed files pass.

Made with Claude Opus 5.5 via Claude Code in T3 Code.

Takeover

Continues @flamboh’s original PR on its fork branch, preserving the contributor commits. Replaces the manually maintained light-accent table with the existing theme contrast helper; rewrites the two regression tests without growing the suite. Keeps the offline fixture replay tooling in linked, hashed artifacts so it does not ship. Dependency: #22237 → #22243 → #22481, all open when inspected. These feature layers are inherited, not part of the contrast fix, and must land before this PR. Julius’s land label is required. Status: blocked on incomplete required iOS evidence; green CI and the source reviews do not clear the visual-proof gate.

Evidence

Before main and after takeover, t3-code light user-message chips

Overview is downscaled; full native surfaces and both diffs follow.

Verification scope: mobile verified on static fixture only, for the delivered states. Required gaps remain unverified and block evidence acceptance.

Current-head iOS capture set — INCOMPLETE; required gaps block evidence acceptance.

Before 955b4cc801; after bfd730292a; fixture 7d1650ea51. Baseline-to-head includes inherited thread/PR chips, project icons/monograms and palette behavior beyond the contrast change; Julius review remains required. Linked fixture tooling is separate from production scope.

397 paired v3 full native surfaces across 28 matrix row IDs. Independent09:23 audit accepts10/29 iOS rows for their stated combinations: U1,A1,E1,E3,E4,T1,T2,T3,T5,G1;18 other iOS rows remain partial, and M1 is the expected native-platform limit. Z1 Android is separately unverified. V3 inventory:475 BEFORE PNGs,397 AFTER PNGs,397 exact matched filenames;78 BEFORE-only extras remain unpaired and do not count as proof. Separate v4 delivers8 exact pairs. Counts describe delivered pages, not acceptance or offscreen coverage.

Environment: EAS iPhone 17e, iOS 26.5, 390×844 logical points portrait, DPR 3 (1170×2532 PNG), app base size 16pt, OS text size large except named W2 accessibility-medium captures. Native safe areas; keyboard/scroll state is named per page. Static message dates 2026-01-01; PNG status bar normalized to9:41; recordings retain the actual simulator status time/battery. The OS clock itself was not frozen.

Diff0: absolute RGB channel delta multiplied by 32 and clamped, with no threshold. Diff8: retain pixels whose maximum RGB channel delta is >8/255, then amplify by 32. Full surfaces appear at native width in each sheet.

Row · portrait / theme states	Before	After	Diff0	Diff8	Coverage / remaining limit
U1	U1 before full native surfaces 955b4cc801	U1 after full native surfaces bfd730292a	U1 diff0 full native surfaces 955b4cc801 → bfd730292a	U1 diff8 full native surfaces 955b4cc801 → bfd730292a	24 pages. CAPTURED P×T, both atlas pages (24 paired surfaces). All ten kinds visible; dark native iris retains legacy low contrast2.42–2.80:1. Capture coverage is not an all-theme readability pass.
A1	A1 before full native surfaces 955b4cc801	A1 after full native surfaces bfd730292a	A1 diff0 full native surfaces 955b4cc801 → bfd730292a	A1 diff8 full native surfaces 955b4cc801 → bfd730292a	24 pages. CAPTURED P×T, both assistant pages (24 paired surfaces), with no-record defaults, skill and file paths visible. Auditor sampled t3-code dark: zero product-region delta, status pixels only.
C1	C1 before full native surfaces 955b4cc801	C1 after full native surfaces bfd730292a	C1 diff0 full native surfaces 955b4cc801 → bfd730292a	C1 diff8 full native surfaces 955b4cc801 → bfd730292a	36 pages. INCOMPLETE: collapsed initial states in all six themes; focused thread composer is covered by keyboard. Native dismiss unsupported (no exposed hide key). Full thread-composer atlas, mention/skill and empty control unreadable/unverified.
P1	P1 before full native surfaces 955b4cc801	P1 after full native surfaces bfd730292a	P1 diff0 full native surfaces 955b4cc801 → bfd730292a	P1 diff8 full native surfaces 955b4cc801 → bfd730292a	26 pages. INCOMPLETE: six-theme user/composer initial states. Collapsed composer exposes first status only; all five composer statuses across T not proved. Actual NewTask PR pages are B-only supplemental, not a T substitute.
P2	P2 before full native surfaces 955b4cc801	P2 after full native surfaces bfd730292a	P2 diff0 full native surfaces 955b4cc801 → bfd730292a	P2 diff8 full native surfaces 955b4cc801 → bfd730292a	6 pages. INCOMPLETE: user/composer initial review and PR states. Expanded thread-composer review detail remains unverified.
P3	P3 before full native surfaces 955b4cc801	P3 after full native surfaces bfd730292a	P3 diff0 full native surfaces 955b4cc801 → bfd730292a	P3 diff8 full native surfaces 955b4cc801 → bfd730292a	6 pages. INCOMPLETE: user/assistant and streaming initial scenes captured. Scene replacement does not prove actual incremental streaming→settled transition.
F1	F1 before full native surfaces 955b4cc801	F1 after full native surfaces bfd730292a	F1 diff0 full native surfaces 955b4cc801 → bfd730292a	F1 diff8 full native surfaces 955b4cc801 → bfd730292a	12 pages. INCOMPLETE: seven MIME/extension distinctions in user/assistant/composer initial pages; expanded composer full atlas remains occluded.
F2	F2 before full native surfaces 955b4cc801	F2 after full native surfaces bfd730292a	F2 diff0 full native surfaces 955b4cc801 → bfd730292a	F2 diff8 full native surfaces 955b4cc801 → bfd730292a	4 pages. INCOMPLETE: file sizes and missing-record examples captured. Known image absent/zero/normal/large and long labels per size are not seeded; missing record is not a known attachment with absent size. Contract requires sizeBytes.
N1	N1 before full native surfaces 955b4cc801	N1 after full native surfaces bfd730292a	N1 diff0 full native surfaces 955b4cc801 → bfd730292a	N1 diff8 full native surfaces 955b4cc801 → bfd730292a	6 pages. INCOMPLETE: unknown/missing records in three initial paths captured. Supplied-empty-list distinction and high/low/unparseable custom accents not captured.
M1	Not captured	Not captured	Not captured	Not captured	0 pages. UNVERIFIED expected platform limit: actual Android Material You cannot be proved on iOS. Any frozen-palette page is supplemental only.
E1	E1 before full native surfaces 955b4cc801	E1 after full native surfaces bfd730292a	E1 diff0 full native surfaces 955b4cc801 → bfd730292a	E1 diff8 full native surfaces 955b4cc801 → bfd730292a	6 pages. COVERED per independent09:23 audit for stated combinations: Actual FileMarkdownPreview light/dark initial and pagination delivered. Native outer scroll can report no movement; no claim that a failed scroll exposes additional hidden content.
E2	E2 before full native surfaces 955b4cc801	E2 after full native surfaces bfd730292a	E2 diff0 full native surfaces 955b4cc801 → bfd730292a	E2 diff8 full native surfaces 955b4cc801 → bfd730292a	32 pages. INCOMPLETE: valid repository-aware screen and grouped-card attention fixtures plus actual description/timeline expand-collapse sequences delivered in light/dark. Distinct narrow-note width and complete caller-equivalence mapping remain unproved; failed nested scroll does not establish full body visibility.
E3	E3 before full native surfaces 955b4cc801	E3 after full native surfaces bfd730292a	E3 diff0 full native surfaces 955b4cc801 → bfd730292a	E3 diff8 full native surfaces 955b4cc801 → bfd730292a	10 pages. CAPTURED: Actual native expand/collapse sequence delivered in light/dark, with initial expanded/collapsed states and choices. Native nested-scroll limitations are logged.
E4	E4 before full native surfaces 955b4cc801	E4 after full native surfaces bfd730292a	E4 diff0 full native surfaces 955b4cc801 → bfd730292a	E4 diff8 full native surfaces 955b4cc801 → bfd730292a	12 pages. COVERED per independent09:23 audit for stated combinations: Nonempty and empty review-card initial and pagination states in light/dark. Retries used relaunch per scene to align navigator shell; preliminary AFTER Project overview versus BEFORE New thread title residue was fixture navigation history, unrelated to chip code. Failed nested scrolling limits any offscreen prose claim.
C2	C2 before full native surfaces 955b4cc801	C2 after full native surfaces bfd730292a	C2 diff0 full native surfaces 955b4cc801 → bfd730292a	C2 diff8 full native surfaces 955b4cc801 → bfd730292a	52 pages. INCOMPLETE: actual NewTask focused native editor and six pan pages per scene (C2-1/C2-2/C2-pr), both appearances, expose chips above keyboard, including PR states. These focused-visible sequences are diagnostic partial pairs: AFTER C2-1 light first focus failed and navigator/scroll conditions may differ. Relaunch-aligned retry was curtailed for v4 deadline; only matching completed subset is delivered. Thread composer expands but keyboard hides all contents; P×T full composer atlas remains unproved.
W1	W1 before full native surfaces 955b4cc801	W1 after full native surfaces bfd730292a	W1 diff0 full native surfaces 955b4cc801 → bfd730292a	W1 diff8 full native surfaces 955b4cc801 → bfd730292a	12 pages. INCOMPLETE: portrait chat/plain/composer delivered. Actual orientation commands returned landscape but all PNGs stayed1170×2532; wide844×390 is NOT proved. Long-suffix/wrap and expanded standalone-composer proof incomplete.
W2	W2 before full native surfaces 955b4cc801	W2 after full native surfaces bfd730292a	W2 diff0 full native surfaces 955b4cc801 → bfd730292a	W2 diff8 full native surfaces 955b4cc801 → bfd730292a	8 pages. INCOMPLETE: actual OS accessibility-medium readbacks and chat/NewTask/standalone-composer initial shots captured in both appearances; surrounds scale while NewTask editor stays16pt. Long-size/suffix cases offscreen and native editor pagination at that OS size unverified. font22 remains a separate app-size supplement.
T1	T1 before full native surfaces 955b4cc801	T1 after full native surfaces bfd730292a	T1 diff0 full native surfaces 955b4cc801 → bfd730292a	T1 diff8 full native surfaces 955b4cc801 → bfd730292a	2 pages. COVERED per independent09:23 audit for stated combinations: Three actual thread paths visible in light/dark, including loaded GitHub favicon on AFTER. Baseline generic/raw presentation differs through inherited work; no same-main lower-layer-only pixel control.
T2	T2 before full native surfaces 955b4cc801	T2 after full native surfaces bfd730292a	T2 diff0 full native surfaces 955b4cc801 → bfd730292a	T2 diff8 full native surfaces 955b4cc801 → bfd730292a	8 pages. COVERED per independent09:23 audit for stated combinations: Emoji, explicit one/two-glyph and lucide fallback three-path states captured in light/dark. Source preservation only; no same-main lower-layer-only image comparison.
T3	T3 before full native surfaces 955b4cc801	T3 after full native surfaces bfd730292a	T3 diff0 full native surfaces 955b4cc801 → bfd730292a	T3 diff8 full native surfaces 955b4cc801 → bfd730292a	8 pages. COVERED per independent09:23 audit for stated combinations: Normal, punctuation-only, digit-containing and recognized-placeholder automatic monograms captured in all three thread paths, light/dark. No missing-placeholder-fixture claim applies.
T4	T4 before full native surfaces 955b4cc801	T4 after full native surfaces bfd730292a	T4 diff0 full native surfaces 955b4cc801 → bfd730292a	T4 diff8 full native surfaces 955b4cc801 → bfd730292a	10 pages. INCOMPLETE: missing/blank/scratch/pending/failed initial states captured. Shell-found/project-missing branch setup separately unseeded; pinned pending/failed do not establish release/completion.
T5	T5 before full native surfaces 955b4cc801	T5 after full native surfaces bfd730292a	T5 diff0 full native surfaces 955b4cc801 → bfd730292a	T5 diff8 full native surfaces 955b4cc801 → bfd730292a	4 pages. COVERED per independent09:23 audit for stated combinations: Original and renamed cross-environment snapshots captured sequentially; renamed satellite icon and both raw Remote renamed labels visible, context label intentionally Remote context. Snapshot replacement is deterministic fixture state, not a live server rename event.
G1	G1 before full native surfaces 955b4cc801	G1 after full native surfaces bfd730292a	G1 diff0 full native surfaces 955b4cc801 → bfd730292a	G1 diff8 full native surfaces 955b4cc801 → bfd730292a	4 pages. CAPTURED: All18 palette colors delivered across top/bottom pagination in light/dark. No same-main lower-layer-only image control; inherited baseline differences cannot prove contrast-only pixel equality.
G2	G2 before full native surfaces 955b4cc801	G2 after full native surfaces bfd730292a	G2 diff0 full native surfaces 955b4cc801 → bfd730292a	G2 diff8 full native surfaces 955b4cc801 → bfd730292a	14 pages. INCOMPLETE: recorded actual absent→placeholder→invalid-data→loaded→new-revision→blank-title native sequence, with cache readback false for first three and true for loaded/revision. Pending remote URL is not deterministically gated; already-loaded new revision does not expose fallback→new-image or establish no stale image at the intermediate instant.
G3	G3 before full native surfaces 955b4cc801	G3 after full native surfaces bfd730292a	G3 diff0 full native surfaces 955b4cc801 → bfd730292a	G3 diff8 full native surfaces 955b4cc801 → bfd730292a	38 pages. INCOMPLETE: screen/sidebar production pending/card/card-selected/slim0.4/slim-selected rows delivered across pagination; Settings48 and regular NewTask20 list delivered. Actual grouped NewTask24 class remains unseeded; intrinsic grid24 is not that caller composition. Complete wrapper-equivalence mapping remains unverified.
K1	K1 before full native surfaces 955b4cc801	K1 after full native surfaces bfd730292a	K1 diff0 full native surfaces 955b4cc801 → bfd730292a	K1 diff8 full native surfaces 955b4cc801 → bfd730292a	19 pages. INCOMPLETE: real palette queries for loaded favicon, emoji, one/two-glyph, lucide and automatic projects delivered in both appearances; first project action is selected and matching thread row unselected. Per-project glyph selected→unselected movement is not established.
K2	K2 before full native surfaces 955b4cc801	K2 after full native surfaces bfd730292a	K2 diff0 full native surfaces 955b4cc801 → bfd730292a	K2 diff8 full native surfaces 955b4cc801 → bfd730292a	3 pages. INCOMPLETE: separate exact v4 actual search fixture excerpt shows BOTH user (You) and assistant (Agent) native excerpts in LIGHT/DARK, with leave-normal and subsequent-palette isolation screenshots/action logs. v3 never enters excerpt. v3 query > and empty filtered results delivered; index-bottom LIGHT is paired, DARK index-bottom is BEFORE-only and unverified as a pair. Long title/detail, no-detail and indices8/9 complete visibility/selection mapping remain unverified.
X1	X1 before full native surfaces 955b4cc801	X1 after full native surfaces bfd730292a	X1 diff0 full native surfaces 955b4cc801 → bfd730292a	X1 diff8 full native surfaces 955b4cc801 → bfd730292a	3 pages. INCOMPLETE: actual unchanged-draft LIGHT→DARK→LIGHT recorded with paired sequence screenshots. Focused standalone composer remains occluded. Controlled no-document-assets→settled completion unseeded; pinned pending does not establish completion.
X2	X2 before full native surfaces 955b4cc801	X2 after full native surfaces bfd730292a	X2 diff0 full native surfaces 955b4cc801 → bfd730292a	X2 diff8 full native surfaces 955b4cc801 → bfd730292a	8 pages. INCOMPLETE: actual terminal destination and Close Terminal return delivered both appearances; native long-press menu and Copy context action delivered. Clipboard readback failed (Failed to read iOS simulator clipboard), so copied content is unverified. Mention/file/thread/element/preview/review destination+return and native selection remain unverified; thread composer focus is occluded.
Z1	Not captured	Not captured	Not captured	Not captured	0 pages. UNVERIFIED: Android native capture outside iOS authorization.

What else changed

Intended: light-surface chip labels/details and tintable symbols become darker when surrounding text is dark.
Dark accents remain unchanged on inspected branches. Native iris dark chips remain at the auditor-reported 2.42–2.80:1 legacy contrast; no all-theme readability or 4.5:1 guarantee is made.
Inherited from lower layers: stateful thread/PR presentation, inline project imagery, automatic monograms and palette project icons. Main-before differences in those images cannot establish contrast-only pixel preservation.
Fixture navigation shell: preliminary E4 retry retained Project overview versus New thread. Final E4 retries relaunch per scene; this history residue is unrelated to chip code. Other supplemental navigator titles may retain prior route state and are disclosed rather than attributed to contrast.
Focus/scroll diagnostics: old NewTask focused-visible pages include a failed first focus on AFTER C2-1 light and possible differing navigator/scroll state. They are partial diagnostics, not aligned full-atlas comparison. Interrupted aligned retries contribute only matching delivered subsets. Untested or unreviewed visible differences remain unverified.
OS chrome: charging-indicator rendering differs between some simulators despite identical normalization. Cursor/keyboard/selection and transient loading differences must be inventoried per final state; no claim of zero unrelated pixels is made.
Fixture-only: an offline alternate entry and static registry seeds, linked build profile/native arm64 hook, source snapshot adapter, and supplemental scene controls. Production auth/permissions were not changed. Final v3 composer/Home scenes mount. Focused thread-composer keyboard occlusion is diagnostic, not full-atlas readability proof.

Cleanup: both owned EAS simulator sessions stopped successfully by09:16 UTC; exact IDs/status in resource-cleanup.json and stop logs in replay archive. Both private simulator dotenv files reset mode0600. All four EAS builds FINISHED; no dev server started, source edit, PR write or push by capturer.

pixel-statistics.json build-manifest.json 22585-final-page-manifest.json [fixture-7d1650ea51.tar.gz](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ebdad1f823d950b2/fixture-[7d1650ea51](https://git.cafe/pingdotgg/t3code/commit/7d1650ea51f81a4000937f5c76d31221f1c89ac2).tar.gz) 22585-supplemental-overlay.tar 22585-supplemental-overlay-v4.tar 22585-supplemental-overlay-files.sha256 22585-supplemental-overlay-v4-files.sha256 22585-supplemental-REPLAY.md 22585-evidence-matrix-v2.md 22585-final-fixture-audit.md 22585-audit-recapture.png 22585-audit-recapture.json 22585-audit-pixel-delta.json capture-inventory.json capture-action-log.json resource-cleanup.json pair-accounting.json evidence-replay-scripts.tar.gz

Independent AFTER U1 recapture: zero product-region pixel delta;1438 full-image status/battery pixels differed. This is one-state reproduction, not whole-matrix acceptance.

X1-unchanged-draft-before.mp4 X1-unchanged-draft-after.mp4 G2-light-before.mp4 G2-light-after.mp4 G2-dark-before.mp4 G2-dark-after.mp4

Separate v4 K2 native query and isolation proof

State set	Before	After	Diff0	Diff8
Exact v4 builds · t3-code LIGHT/DARK · actual query and leave/subsequent state	K2 v4 before	K2 v4 after	K2 v4 diff0	K2 v4 diff8
k2-v4-build-manifest.json
k2-v4-pixel-statistics.json

Separate v4 K2 excerpt uses supplement cf24e99e7716078dcb49e2e2939cfdfc9bc0e95e97ce849dd222e1ca5445041d, BEFORE build b2371778-7494-4477-bccc-c4de8e37dfad, AFTER build d4415978-2a0a-46ca-8509-56b3014632c3. v3 supplement 1b163126bbaa361d8448498e87b795a508b7413321c0ae2a4552016ac13faf51 supplies all other rows. Both share base gzip 909f1369d4e742411630dc94b6ac6349d83085b205708274987059244dfdd1da. EAS gitCommitHash is null; source archives/manifests supply provenance.

Unverified

Android, including actual Material You rendering (expected on this VM; not an iOS blocker).
Required iOS fixture coverage remains incomplete. The matrix identifies every partial row; offscreen content, labels, passing tests and source equivalence are not substituted for captures. Final proof-audit verdict: INSUFFICIENT C1 P1 P2 P3 F1 F2 N1 E2 C2 W1 W2 T4 G2 G3 K1 K2 X1 X2. Exact missing requirements are listed in the independent audit.
Existing-thread focused composer: the standalone fixture is covered by the keyboard. Focused New Task composer captures are separately labelled and do not prove this path.
File/image size combinations and per-size long labels; supplied-empty-list/custom-accent fixtures; controlled delayed-asset completion and deterministic native pending/revision intermediates are not fully seeded or captured.
Landscape attempts remain portrait despite controller success; they are excluded as landscape proof. OS accessibility, scrolling and destination/copy interaction coverage remain partial under the final row audit. Copy-menu taps do not establish clipboard payload correctness: native clipboard readback failed.

Verification log

Source head: bfd730292a; merge-base/main: 955b4cc801. Main was integrated by merge commit. The shipped diff contains 17 mobile files; linked fixture tooling is excluded.

Tests: contrast suite 2 → 2 in one file, rewritten to fail independently when either light-mode correction or dark-surface preservation is reverted. Final source scoped run: 5 files, 37 passed, 0 failed. Mobile typecheck and targeted lint pass; existing warnings remain. Linked v4 fixture suite: 7 files, 39 passed, 0 failed, including a scoped-pin enter → hold → leave → later-update regression that fails with the previous permanent pin. Linked fixture typecheck, lint and formatting pass.

Reviews: seven fresh rounds with two reviewers each (GPT-6.1-Sol and Claude Opus 5.5); round 7 both NEEDS_JULIUS, with no production correctness findings. Rounds 1–6 found no production contrast defect. Fixture findings included missing providers, wrong streaming role, scene fallthrough, dirty-source snapshot provenance, scratch configuration, invalid shell defaults, and a search pin surviving cleanup. Those were corrected in the linked fixtures. The fixture scope objection was resolved by removing the harness from the shipped source. Round 6 compared a base-archive digest with a supplemental-archive digest; the replay identifies these distinct files and records both named hashes. Both round-7 reviewers verified the separately named hashes and accepted the v4 cleanup fix. Native proof remains independently audited; colour taste requires Julius.

Inherited lower-layer performance concern: repeated malformed thread-link prefixes can make the unchanged lower-layer matcher quadratic. This was not introduced by the contrast change and was preserved with the lower layer.

Native build recovery: materialized baseline dependencies inside its archive to avoid fingerprint path mismatch; used a fixture-only arm64 simulator hook because GhosttyKit lacks an x86_64 simulator slice. Normal release exports with fixture flag unset and flag 0 produced identical bundles without fixture markers. Obsolete cloud jobs were canceled or failed; all four final paired cloud builds finished, 405 paired captures were delivered, and the final audit is insufficient. Both simulator sessions are stopped; all owned commands and delegated work finished; private dotenv files were reset to mode 0600.

Final-head CI at 08:57 UTC: all 30 checks successful, 0 pending, 0 failing on bfd730292a, read through cafe pr status. No fork-run approval required. Parent holds the sole PR watch until final handoff.

Independent same-artifact U1 reset/recapture: product region y>=160 has exactly 0 changed pixels across 2,775,240 pixels; full surface has 1,438 changed pixels confined to battery/charging glyph at x1046–1102,y55–93. OS chrome differences are disclosed rather than described as anti-aliasing.

Dark-mode limit: preservation is not a 4.5:1 readability claim. The auditor measured unchanged iris dark user-bubble chips at about 2.42–2.80:1 (bubble #4b3d72); this inherited behavior is shown honestly and is outside the light-mode fix.

Final independent proof audit: INSUFFICIENT C1 P1 P2 P3 F1 F2 N1 E2 C2 W1 W2 T4 G2 G3 K1 K2 X1 X2. It accepts 10/29 iOS rows, with 18 partial and M1 a permitted Android-only limit. The final index contains 397 v3 plus 8 v4 exact pairs, each with stats, Diff0 and Diff8; 78 before-only diagnostics are excluded. K2 native user/agent excerpts and cleanup isolation are accepted, while its complete title/detail/selection mapping remains partial.

Inherited lower-layer commits, preserved unchanged:

0ba490263c — fix(mobile): thread links and PR URLs in chat render as chips
48f7d52c1b — fix(mobile): PR chips ignore pull requests unlinked from a stack
d3533d3024 — fix(mobile): PR chip state compares sync times as instants
ba04146da7 — fix(mobile): PR chips never hide an agent's link text
ad7e4fb78c — fix(mobile): chat chips cover thread links only
2ddce21aea — feat(mobile): thread chips show the thread's project icon
635b08dba1 — feat(mobile): thread chips show monogram and emoji project icons
cc8a549c38 — fix(mobile): each chip icon tile is captured once
8b0678c092 — fix(mobile): a chip icon capture that throws settles instead of crashing
4980e95581 — fix(mobile): projects without a favicon show a monogram, like desktop
fd101156b0 — fix(mobile): command palette project rows show project icons

Inherited lower-layer files (the enriched-link hook also contains this PR’s contrast hunk):

Takeover implemented with Claude Opus 5.5 and GPT-6.1-Sol through T3 Code.

Checks

All checks passed

CodeRabbitGitHub
Review paused
Successful
Native fingerprint diffGitHub
Successful
CheckGitHub
Successful
Native fingerprint diffGitHub
Successful
Mobile Native Static AnalysisGitHub
Skipped
TestGitHub
Successful
Build and TypecheckGitHub
Successful
Release SmokeGitHub
Successful
RustGitHub
Successful
Test WebGitHub
Successful
Label PR 22585GitHub
Successful
Sync PR size label definitionsGitHub
Skipped
Label PR sizeGitHub
Successful
Macroscope - Effect Service ConventionsGitHub
Skipped
Macroscope - Approvability CheckGitHub
Neutral
Macroscope - Effect Service ConventionsGitHub
Skipped
Deploy web previewGitHub
Skipped
Test Server 8GitHub
Successful
EAS PreviewGitHub
Skipped
Test Server 2GitHub
Successful
Test Server 4GitHub
Successful
Test Server 7GitHub
Successful
Test Server 1GitHub
Successful
Test Server 3GitHub
Successful
Test Server 5GitHub
Successful
Test Server 6GitHub
Successful
ChangesGitHub
Successful
LintGitHub
Successful
Macroscope - Correctness CheckGitHub
Successful
Native fingerprint diffGitHub
Cancelled
Collect PR targetsGitHub
Successful
Prepare PR size configGitHub
Successful

MmacroscopeappBot2d ago

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — This range contains substantial mobile runtime changes beyond contrast tuning, including thread/project chip rendering, project icon resolution, native rasterization, and state-driven asset lookup. It also adds a configuration-level static-analysis escape-hatch allowlist entry, so the scope and review-policy impact require human review.

^{You can add or adjust custom eligibility rules. Learn more.}

CcoderabbitaiBot2d ago

Review in Change Stack →

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Last reviewed commit: 11c8e9d1d178850ce90b3408676b71ee18d5782e
Latest commit, not reviewed: 598d97f9296df66240967552b28bba68a5afa90a

Use the following commands to manage reviews:

@coderabbitai resume to resume automatic reviews.
@coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

▶️ Resume reviews
🔍 Trigger review

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.config.ts
Review profile: CHILL
Plan: Advanced
Run ID: 37f2d9e2-32a6-48d8-9b27-1d0a0bacd9b2

📥 Commits

Reviewing files that changed from the base of the PR and between 954a2efc62c54cd73f4939577eeb30d64a5a15be and 11c8e9d1d178850ce90b3408676b71ee18d5782e.

📒 Files selected for processing: 6

apps/mobile/src/features/threads/ThreadFeed.tsx
apps/mobile/src/lib/enrichedLinkPresentation.test.ts
apps/mobile/src/lib/enrichedLinkPresentation.ts
apps/mobile/src/lib/markdownDocument.ts
apps/mobile/src/native/SelectableMarkdownText.types.ts
apps/mobile/src/native/useEnrichedLinkVariants.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

📝 Walkthrough

Mobile project icons now use explicit glyphs, valid favicons, or generated monograms. Thread-feed links support thread and GitHub pull-request chips with resolved icons and pull-request states. Chip accent selection now considers text color.

Changes: Mobile chip presentation

Layer / File(s)	Summary
Project icon resolution and rendering `apps/mobile/src/lib/projectIcon.ts`, `apps/mobile/src/lib/projectIcon.test.ts`, `apps/mobile/src/components/ProjectFavicon.tsx`, `apps/mobile/src/features/keyboard/CommandPalette.tsx`, `apps/mobile/src/features/keyboard/commandPaletteItems.ts`, `apps/mobile/src/native/ProjectChipIconImages.tsx`, `apps/mobile/src/App.tsx`	Project icons resolve to explicit glyphs, valid favicons, or generated monograms. The command palette and `ProjectFavicon` render project icons. The native rasterizer captures glyph tiles as PNG data URLs.
Enriched link and chip state `apps/mobile/src/lib/enrichedLinkPresentation.ts`, `apps/mobile/src/lib/enrichedLinkPresentation.test.ts`, `apps/mobile/src/lib/markdownDocument.ts`, `apps/mobile/src/state/entities.ts`	Pull-request chip states use the newest eligible snapshot. Thread chip targets come from markdown links and composer records, then resolve through entity state to project icons.
Thread-feed customization and contrast `apps/mobile/src/features/threads/ThreadFeed.tsx`, `apps/mobile/src/native/useEnrichedLinkVariants.ts`, `apps/mobile/src/native/SelectableMarkdownText.types.ts`, `apps/mobile/src/lib/composerChipPresentation.ts`, `apps/mobile/src/lib/composerChipPresentation.test.ts`, `apps/mobile/src/lib/mobileTheme.ts`, `apps/mobile/src/native/T3ComposerEditor.ios.tsx`, `apps/mobile/src/native/T3ComposerEditor.native.tsx`	The thread feed applies resolved icons and pull-request states to markdown link customizations. Chip accents use theme-derived text colors. Tests check light-theme contrast and dark-surface accent preservation.

Priority: ➖ Normal

Estimated code review effort: 3 (Moderate) | ~30 minutes

Change: Feature

Sequence Diagram(s)

mermaid

1	sequenceDiagram
2	participant ThreadFeed
3	participant Entities
4	participant LinkPresentation
5	participant LinkCustomization
6	participant MarkdownRenderer
7	ThreadFeed->>Entities: Request pull-request states and thread chip icons
8	Entities->>LinkPresentation: Derive states and thread targets
9	LinkPresentation-->>Entities: Return states and targets
10	Entities-->>ThreadFeed: Return states and resolved icons
11	ThreadFeed->>LinkCustomization: Build enriched link customizations
12	LinkCustomization->>LinkPresentation: Create chat and context presentations
13	LinkCustomization-->>ThreadFeed: Return customized links with chip icons
14	ThreadFeed->>MarkdownRenderer: Render markdown with customized links

Possibly related PRs

pingdotgg/t3code#12534: Changes mobile theme palettes and theme variables used as chip surface and text-color inputs in this PR.

Suggested reviewers: maria-rcks, juliusmarminge

Merge Risk: ⚪ Minimal · up to 11c8e

The inspected chip-rendering paths retain their fallback and link behavior. No merge-blocking issue was established.

Security architecture risk: 🔵 Low · up to 11c8e

Link enrichment changes presentation without changing navigation permissions. Displaying linked threads can now request project icons automatically, using existing environment-scoped asset handling. No introduced authorization bypass was established; remaining uncertainty concerns the reused asset-service controls and native rendering behavior.

Retained concerns No architecture-level concerns identified.

Security review details

Security Blast Radius

inferred — The new automatic resource-access surface is project-icon retrieval for known linked threads in configured environments. Markdown selects a thread reference, but workspace and favicon paths come from the resolved stored project, not directly from link text.

Trust Boundaries and Controls

observed — Link decoration does not replace navigation identity checks. Thread navigation continues through the existing linked-thread resolver; context navigation uses a matching stored record. The resolver's unique cross-environment fallback predates this PR, although icon lookup newly reuses it before navigation.
observed — Favicon lookup requests a project-favicon resource through environment-specific asset URL creation and resolves it against that environment's prepared connection. This preserves the existing asset-service boundary; the inspected client code does not itself establish the complete server-side authorization policy.

Resilience and Maintainability Implications

observed — The reused favicon cache scopes entries by environment and project resource. After a download, cancellation and generation/revision checks prevent cleared work from being persisted. Cache clearing serializes removal and pending writes, limiting stale-data recovery across environment cleanup.

Pre-merge checks: ^Passed 4

Check name	Status	Explanation
Title check	Passed	The title clearly identifies the mobile chat and composer chip readability fix in light mode, which matches the primary changes.
Linked Issues check	Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	Passed	Check skipped because no linked issues were found for this pull request.
Description check	Passed	The description is comprehensive and covers the problem, implementation, dependencies, verification, UI evidence, agent usage, and known verification gaps. It does not use the exact Change and Verific…

Full details: Description check

Explanation

The description is comprehensive and covers the problem, implementation, dependencies, verification, UI evidence, agent usage, and known verification gaps. It does not use the exact Change and Verification headings, and it does not clearly link an explicit maintainer approval comment, but the required information is mostly present.

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create a new PR

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

Added 13 commits2d ago

fix(mobile): thread links and PR URLs in chat render as chips0ba4902
fix(mobile): PR chips ignore pull requests unlinked from a stack48f7d52
fix(mobile): PR chip state compares sync times as instantsd3533d3
fix(mobile): PR chips never hide an agent's link textba04146
fix(mobile): chat chips cover thread links onlyad7e4fb
feat(mobile): thread chips show the thread's project icon2ddce21
feat(mobile): thread chips show monogram and emoji project icons635b08d
fix(mobile): a chip icon capture that throws settles instead of crashing8b0678c
fix(mobile): each chip icon tile is captured oncecc8a549
fix(mobile): projects without a favicon show a monogram, like desktop4980e95
fix(mobile): command palette project rows show project iconsfd10115
fix(mobile): composer chips use the light-surface accents too09de6c7
fix(mobile): chat chips stay readable in light modecd01fe7

JJulius Marminge2d ago

Taking over this one

Added 14 commits1h ago

Merge remote-tracking branch 'origin/main' into takeover/pr-22585f5e15e6
test(mobile): takeover tests for light-mode chip contrast5a9c6a7
refactor(mobile): derive light-mode chip accents with themeColorWithContrast22527a3
test(mobile): add offline native chip evidence fixtures23f2cf4
Merge remote-tracking branch 'origin/main' into takeover/pr-2258502b0a6d
test(mobile): cover native chip caller fixturesdfdde2f
test(mobile): stream large source snapshot archive379bc76
test(mobile): seed scratch and native favicon transitions6fc06a7
test(mobile): build chip evidence for arm64 simulators7d1650e
test(mobile): keep chip evidence tooling in linked replay artifactse579740
Merge main into mobile chip contrast takeoverbfd7302
fix(mobile): chips keep their vivid hue at readable contrast in every themeb85fcfe
Merge remote-tracking branch 'origin/main' into fix/mobile-chip-light-contrastbfcf418
fix(mobile): chips read at 5:1 instead of 4.5:1598d97f