[flamboh]

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.

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.

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:

1. https://git.cafe/pingdotgg/t3code/pulls/22237: Thread links in chat render as chips
2. https://git.cafe/pingdotgg/t3code/pulls/22243: Thread chips show the thread's project icon
3. https://git.cafe/pingdotgg/t3code/pulls/22481: Projects without a favicon show a monogram, like desktop
4. 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.

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

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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	955b4cc801 → bfd730292a	6 pages. INCOMPLETE: user/composer initial review and PR states. Expanded thread-composer review detail remains unverified.
P3	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	955b4cc801 → bfd730292a	6 pages. INCOMPLETE: user/assistant and streaming initial scenes captured. Scene replacement does not prove actual incremental streaming→settled transition.
F1	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	955b4cc801 → bfd730292a	12 pages. INCOMPLETE: seven MIME/extension distinctions in user/assistant/composer initial pages; expanded composer full atlas remains occluded.
F2	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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	955b4cc801	bfd730292a	955b4cc801 → bfd730292a	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-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):

• apps/mobile/src/App.tsx
• apps/mobile/src/components/ProjectFavicon.tsx
• apps/mobile/src/features/keyboard/CommandPalette.tsx
• apps/mobile/src/features/keyboard/commandPaletteItems.ts
• apps/mobile/src/features/threads/ThreadFeed.tsx
• apps/mobile/src/lib/enrichedLinkPresentation.test.ts
• apps/mobile/src/lib/enrichedLinkPresentation.ts
• apps/mobile/src/lib/projectIcon.test.ts
• apps/mobile/src/lib/projectIcon.ts
• apps/mobile/src/native/ProjectChipIconImages.tsx
• apps/mobile/src/native/useEnrichedLinkVariants.ts
• apps/mobile/src/state/entities.ts

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

[macroscopeapp]

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.}

[coderabbitai]

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: 11c8e9d
• Latest commit, not reviewed: 598d97f

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 954a2ef and 11c8e9d.

📒 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)

sequenceDiagram
  participant ThreadFeed
  participant Entities
  participant LinkPresentation
  participant LinkCustomization
  participant MarkdownRenderer
  ThreadFeed->>Entities: Request pull-request states and thread chip icons
  Entities->>LinkPresentation: Derive states and thread targets
  LinkPresentation-->>Entities: Return states and targets
  Entities-->>ThreadFeed: Return states and resolved icons
  ThreadFeed->>LinkCustomization: Build enriched link customizations
  LinkCustomization->>LinkPresentation: Create chat and context presentations
  LinkCustomization-->>ThreadFeed: Return customized links with chip icons
  ThreadFeed->>MarkdownRenderer: Render markdown with customized links

Loading

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: 4

Check name	Status	Explanation
Title check		The title clearly identifies the mobile chat and composer chip readability fix in light mode, which matches the primary changes.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.
Description check		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.}

[juliusmarminge]

Taking over this one