chore(sync): absorb upstream main through 024d49520

[NoahHendrickson]

Problem

custom was 821 commits behind upstream (last sync point 12391bd0d, 2026-09-08). Upstream's next commit after this range is pingdotgg#2829, the new orchestrator — a full server rewrite that landed today, deletes the hosts of two fork server customizations (CheckpointReactor.ts, server.test.ts) and adds ~40 migrations. That one is deliberately not in this PR; it gets its own sync.

What this does

Merges upstream/main through 024d49520 (v0.0.44–v0.0.45 cycles) into custom. 43 conflicted files resolved, 6 override shadows ported by hand (they never conflict, they drift).

Notable decisions, all fenced and recorded in .fork/customizations.yaml:

• CI — upstream split check into lint/typecheck/build; the fork's runner, lint-owned and shellcheck fences re-seated on the new jobs. Release gates follow the split macOS preview workflows (desktop-macos-preview-publish.yml is newly gated).
• Lint — upstream's new @shadcn/lint plugin (no-restyle, no-arbitrary-values, …) is error-level and fires ~280 times on fork components by design (restyling components/ui is the fork's purpose). A fenced vite.config.ts override turns those rules off for fork-owned files only; upstream code keeps the gate. React Compiler rule baselines in lint-owned.mjs re-measured after the vite-plus 1.0 bump.
• Composer — context records replace the per-kind context blocks; the design-mode extraction now runs ahead of upstream's resolver. The prompt editor is Tiptap; the fork's Lexical focus-skip hunk is covered natively (ProseMirror only writes the DOM selection when it owns focus) and the guard pins that condition instead. The overlay height is a ref upstream, so the transcript-mask inset is written straight onto the chat column rather than re-rendering ChatView.
• Sidebar — ported upstream's Working section (feat(web): beta Working section hides busy threads until they need you pingdotgg/t3code#13926) and the multi-PR badge control into the fork's card rows.
• Model menu — ported multi-model drafts, provider-cycling shortcuts and the fast/ultrafast speed icon into the fork's combined menu.
• Draft assignment — env mode resolves through upstream's shared resolveProjectSettings (environment setting now outranks t3.json; the fork previously had the reverse).
• Icon shim — 26 new lucide names upstream imports.
• Mobile — Expo 58; pnpm-patched deps pinned exactly for Release Smoke.
• Dropped nothing: fence census on every manifest file is accounted for (additions, plus three upstream-removed hosts: the WSL node-pty release job, select/combobox separators, the Lexical editor).

Verified

• web / desktop / server / mobile / client-runtime typecheck green
• fork guard suite 487/487; 1,118 focused web tests; 68 server tests for the fork's server customizations; desktop Clerk-skip and app-identity tests
• vp lint apps/web/src 0 errors; .fork/lint-owned.mjs no blocking warnings
• Known host-dependent failure only: the Windows cross-arch artifact test on Apple Silicon (pre-existing)

Not run in a browser. Worth a look after merge: provider switch with the combined model menu open via keyboard (Tiptap focus), and the transcript mask inset.

Follow-ups (not in this PR)

• Sync feat(orchestrator): introduce new orchestrator pingdotgg/t3code#2829 (new orchestrator) as its own PR once upstream's follow-up fixes settle.
• Fork sidebar trigger is not zoom-aware (upstream's --desktop-window-controls-inset emits 90px; the fork wants 80).
• The fork's liveness pill lacks upstream's new "View agents" link.

Claude Fable 5.1 via Claude Code (T3 Code), with three subagents for the sidebar, composer and override-shadow clusters.

🤖 Generated with Claude Code

[cursor]

Thermo-nuclear review of the fork-authored port (not the 821-commit upstream dump).

Do not approve. apps/web/src/overrides/components/chat/ModelPickerContent.tsx crosses 1k (898 → 1053). That shadow was allowed to keep the paged menu specifically because it stayed under the line. This sync pastes upstream's multi-model draft and provider-cycling into it and blows the budget.

The paged render belongs here. The increment that crossed does not: adjacentModelPickerProvider (L122), the selected-key set, the two cycle-eligibility memos, and cycleProvider. Move that machinery into a sibling under the override (pure helper + a small hook). Re-export the helper from this module so the shadow API stays compatible. The header already says four helpers are upstream-verbatim; the guard still pins three, and it never added selectedModels / onToggleModel to the API-compat list. Pin both when you extract.

Do not whole-file decompose this shadow. Do not pull ChatComposer / ChatView / yaml apart — those were already last-resort or already over 1k.

The rest of the absorb is structurally sound: Tiptap covers the Lexical focus-skip and the guard pins the condition; draft assignment uses resolveProjectSettings; the shadcn lint offs stay on fork-owned globs.

(Inline comments omitted: GitHub rejected file-level comments because this PR's diff exceeds the 300-file API cap.)

  

_{Sent by Cursor Automation: Thermo nuke 4.6}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.6 KiB	13.5 KiB	−104 B (−0.7%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.0 KiB	7.1 KiB	+23 B (+0.3%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.4 KiB	−127 B (−1.9%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.1 KiB	56.2 KiB	−866 B (−1.5%)	66.4 KiB	✅
Codex	Live turn messages	10	9	−1 (−10.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.5 KiB	−64 B (−0.5%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.0 KiB	7.1 KiB	+24 B (+0.3%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.5 KiB	6.4 KiB	−88 B (−1.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	57.8 KiB	57.0 KiB	−840 B (−1.4%)	66.4 KiB	✅
Claude	Live turn messages	9	9	0 (0.0%)	21	✅

Baseline: 6f9f29c · PR result: 81b776f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 113.9 KiB
• Claude decoded thread snapshot: 114.6 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[NoahHendrickson]

Reviewed commit 185d774b7d82bcbb392ca8543cf62b2d4d45c930. Changes needed. I focused on the fork integration: conflict resolutions, override ports, composer/context handling, sidebar integration, release gates, and the failing CI jobs. This is not an exhaustive audit of all 821 upstream commits.

1. [P2] Preserve design attachments when sending to multiple models — ChatView.tsx:7922–7927. In a project draft, attach a design change from the Browser panel, Shift-select a second model, and send. The single-model path includes forkDesignSend.text, but each multi-model target is built from messageTextForSend, which excludes the attached design request. Consequently every spawned agent misses those instructions; a design-only send instead receives the attachment bootstrap prompt. The multi-model branch also returns before forkDesignChanges.markSent, leaving the pill pending after successful sends. Include the captured design payload in each target and handle its completion bookkeeping for successful/failed targets.

2. [P2] Handle design blocks after legacy context serialization — MessagesTimeline.tsx:2011–2017. When connected to a server without inlineMessageContext, sending a design change together with a terminal excerpt or preview annotation now calls serializeLegacyContextMessage after adding the design block. That serializer appends the legacy context after the design block, so it is no longer trailing. Here extractTrailingDesignChanges returns no blocks; the following context resolution removes the terminal/annotation wrapper but leaves the full design request in the visible message, with no design chip or its verification summary. Support both the historical block order and this new serialization order. A focused reproduction using the real serializer, extractor, and context resolver fails with blocks: []; the inline-context control passes.

CI also needs these concrete repairs before merge:

• Test Web: TraitsPicker.test.ts:173 still expects showFastModeIcon: false; the migrated API returns speedIcon: null.
• Test: DesktopPreReadyPlatform.test.ts:98 and 130 still uses the upstream icon basename in both the copy mock and assertion. The implementation correctly writes com.t3tools.T3Code.Fork.desktop.png. Both Linux entry tests fail, including locally on macOS; these are not the documented Windows cross-architecture failure.
• Lint: SidebarThreadHeader.tsx has no consumers after the fork sidebar port. The CI knip:check step exits on this unused file before reaching lint. Account for this upstream component in the fork’s Knip configuration or reuse it where appropriate.

Validation: ran the focused TraitsPicker, DesktopPreReadyPlatform, ModelPickerContent, and composerModelTraits tests: 41 passed, 3 failed, matching CI. Ran a temporary two-case legacy design/context reproduction: 1 passed, 1 failed; removed it afterward. No browser verification performed. Workspace left unchanged.

Review by GPT-6 via Codex.

[NoahHendrickson]

Pushed d00baead8 addressing both reviews:

Codex review — all five items fixed:

1. Multi-model drafts now send forkDesignSend.text to every target and call markSent once any target started (keyed on the first started message id).
2. Legacy context serialization: new withDesignChangesTrailing helper serializes the prompt without the design run and re-appends it last, on both the single and multi-model paths. Test added covering the serializer order, the design-only send, and the no-blocks pass-through.
3. TraitsPicker test → speedIcon: null. 4. DesktopPreReadyPlatform tests → com.t3tools.T3Code.Fork.desktop.png (fenced). 5. knip → fenced ignore for SidebarThreadHeader.tsx under fork-sidebar-chrome, with knip.jsonc added to that entry's watch:.

Verified locally: vp run knip:check clean, web typecheck clean, the touched test files + design-mode and app-identity guards pass (190 tests), desktop pre-ready tests 6/6, lint-owned no blocking warnings.

Cursor review — not taking the ModelPickerContent.tsx split in this PR. There is no 1k-line budget recorded anywhere in .fork/ (README, AGENTS, notes, manifest), the file's growth is upstream's multi-model/provider-cycling machinery ported verbatim into the shadow, and the guard already pins the exported helpers. Extracting a sibling hook is reasonable cleanup but is its own change, not sync work. Noting it as a follow-up alongside the zoom-aware sidebar trigger.

[NoahHendrickson]

Ultra code review of the upstream sync (head 2b79e79). Nine finder agents ran over the fork-touched resolutions. I checked every finding below against the PR head; duplicates are merged and the rest were dropped (see the end).

Most important: design-mode pills on the new send paths (queue, multi-model), the dead Mod+Shift+E shortcut, assignDraftProject reading the wrong settings, and the composer placeholder CSS that the Tiptap swap broke. Items marked plausible were traced in source but not run.

Also: custom/DraftProjectPill.tsx isn't in this diff, but it drifted. Upstream now hides the per-environment Scratch project from its pickers (isScratchProject in DraftHeroHeadline and CommandPalette). The fork's pill lists every buildSidebarProjectPickerEntries row, so once Scratch has been opened it shows up as an ordinary assignable project.

Checked and dropped:

• The 400ms hero clock, the lint baseline re-measure and the SidebarThreadHeader knip ignore are deliberate, documented decisions.
• The unfenced TraitsPicker test predates this PR.
• The env-mode precedence flip matches upstream.
• The icon shim, model-picker ports, CI gating and app identity all came back clean.

GitHub only allows inline comments on the first 3,000 files of a PR, and this PR's vendored .repos/ push most fork files past that. So the findings are below with permalinks to the head commit, apart from the lint-baseline note, which is posted inline.

1. ChatView.tsx:7796

Queued follow-ups drop design-change pills. Upstream's new queue path (followUpBehavior: "queue" is the default) returns here, before forkDesignChanges.takeForSend at ~7870, and sendQueuedMessage never reads the pills.

• Prose plus pills during a running turn: the prose is queued without the <design_change_request> blocks. The pills stay pending and get attached to the next unrelated send.
• Pills only: hasSendableContent counts pills (~7607), so the send passes. It is queued with prompt: "", and sendQueuedMessage's deriveComposerSendState then sees nothing sendable and quietly finishes it. The design request never goes out.

Before this sync the fork had no queue, so this case went through the normal path that appends the pills. Either bypass the queue while pills are pending, or carry takeForSend's text and sent entries into the queued message and markSent when it actually dispatches.

---

2. ChatView.tsx:8250

Multi-model send + design pills: two problems with this markSent.

1. Partial failure loses the design request. startedCount > 0 clears the pills even when some targets failed. restoreFailedDraft puts back only messageTextForSend, which doesn't include the blocks. Example: A starts and B fails. Retrying B sends the bare prompt with no <design_change_request>. The comment above ("a failed target restores the draft text, not the pills, which the next send reads again") is wrong, because the pills are already gone. Fix: only markSent when failedSelections.length === 0, or re-add the sent entries on restore.
2. The resolution prompt can never show. The record is keyed on forkDesignChangeRef, which is the draft's thread. Every target runs on a fresh targetThreadId, so latestTurnAtom(draftRef) stays null. shouldOfferPreviewResolution is then always false, setVerifying never arms, and the chip keyed on forkFirstStartedMessageId never gets a verdict line. A later single send from the same draft would also wrongly "cover" this record. The record needs the started thread's ref, not the draft's.

Smaller, plausible: the composer is released (sendInFlightRef.current = false) before await starts, but the pills only clear after every start settles. A second send typed during worktree prep can pick up the same pills, and the design request goes out twice.

---

3. ChatView.tsx:6987

Mod+Shift+E (composer.effort) does nothing in the fork. openControl clicks button[data-composer-shortcut~="composer.effort"]. Upstream puts that attribute on TraitsPicker and on CompactComposerControlsMenu when it receives traitsMenuContent. The fork renders neither: effort lives in the combined model menu, and ChatComposer ~5295 no longer passes traitsMenuContent. The combined trigger in overrides/.../ProviderModelPicker.tsx never declares the attribute either. So the new default binding calls preventDefault and then finds no target, in both wide and compact layouts.

Fix: add data-composer-shortcut="composer.effort" to the combined model trigger when traits exist, or route composer.effort to toggleModelPicker.

---

4. useNewAgentDraft.ts:223

assignDraftProject only half-adopts upstream's project-scoped settings. useHandleNewThread ("New thread in X") does all of the following; the New agent pill does not:

• Settings source. Upstream resolves against environmentServerConfigs.get(projectRef.environmentId). This passes primaryServerSettings, so for a project on a non-primary environment, projectSettingsOverrides[project.id] is missing and the environment tier is the primary's. The doc comment on useAssignDraftProject ("the same reading upstream's handler makes") is no longer true.
• startFromOrigin (~259) reads raw settings.newWorktreesStartFromOrigin. It is now a project-scoped key, and upstream uses projectSettings.settings.newWorktreesStartFromOrigin.
• Default model (~265) still reads the aggregate project.defaultModelSelection. After the override fold, a model set or reset in project settings is ignored or overridden by a stale value. Upstream uses resolveProjectSettings(...).settings.defaultModelSelection.
• Permissions. useStartNewAgentDraft (~102) carries runtimeMode over from the viewed thread, and assignment never applies the new defaultRuntimeMode. Starting from a full-access thread keeps full access even when the project or environment default is approval-required. Upstream's handler now starts in the configured default.

One resolveProjectSettings(envSettings, project.id, project, projectFile) result, read from the project's own environment, would cover all four.

---

5. ComposerPromptEditorTiptap.tsx:1416

Fork composer placeholder CSS is dead after the Lexical → Tiptap swap. theme.custom.css ~698–735 targets [data-testid="composer-editor"] ~ div. That matched Lexical, where the placeholder was the editor's sibling. Here data-testid is on the ProseMirror node inside <EditorContent>'s wrapper div, and this placeholder is a sibling of that wrapper. So the rule never matches. The same goes for theme.custom.westworld.css ~197–204. On desktop the placeholder is back to upstream's text-placeholder/75 at upstream's size and leading, so the caret sits off the placeholder text, and Westworld's placeholder ink no longer applies. The guards (forkComposerShell.test.ts ~331–360, forkCoolDarkTheme.test.ts ~246–258) only check the CSS text, so they stay green.

Related, plausible: the fork moves the scrollport to [data-fork-composer-prompt]. The editor here now carries -m-1 … p-1 (line ~784), which can push its box 4px past the wrapper and bring back the phantom scrollbar that rule exists to prevent. The "12.5rem matches upstream's max-h-50" comment is also stale (upstream is max-h-52).

---

6. Sidebar.tsx:1289

Settled rows no longer mute their PR badge. Before the sync, settled slim rows drew the badge in text-secondary-label with settledPrHoverColorClass on hover. Upstream moved that muting onto its row surface (variantAction === "unsettle" && "[&:not(:hover):not(:focus-within)_*]:text-secondary-label/70"), but the resolution swapped that class list for threadRowSurfaceClassName. Neither version mutes it now. Expanding Settled shows full green or purple PR states on every row at rest, so the history tail looks as loud as live cards. No guard covers it.

---

7. ChatView.tsx:6394

Stale memo deps cause re-renders on panel toggles. The fork's composerLivenessPill body was kept, but the merge took upstream's dependency list. activeRightPanelSurface?.kind, addAgentsSurface and rightPanelOpen are never read in the body. So every right-panel open, close or surface switch rebuilds the pill. That invalidates composerContextStrip and re-renders the memoised composer, often mid-stream. Restoring the fork's five deps fixes it.

The same resolution also dropped upstream's "View agents" action from this pill without porting it. Decide whether that's intended; the spawn card still has its own link.

---

8. package.json:118

A patched dependency on a range. react-native-nitro-markdown is ^0.5.0, but pnpm-workspace.yaml patches @0.5.8. The same applies to node-pty (apps/server/package.json: ^1.2.0-beta.15, patch for @1.2.0-beta.15). Release Smoke deletes the lockfile and re-resolves. It passes today only because these are the newest matching versions. Once 0.5.9 or a newer node-pty beta ships, the patch goes unused and the required check turns red on every PR. Pin both exactly.

---

9. ws.ts:1432

Plausible: the bootstrap-branch fill-in misses upstream's new worktree fallback. preparingWorktree: bootstrap.prepareWorktree !== undefined (a few lines down) still assumes "a worktree will name its own branch". Since this sync, when hasCommit is false and requireWorktree isn't set, the thread runs in the project checkout instead (~1409). A client that sent branch: null then gets thread.create with branch: null on a real checkout, which is the gap server-local-thread-branch exists to close. Suggest preparingWorktree: bootstrap.prepareWorktree !== undefined && shouldPrepareWorktree.

---

10. theme.custom.css:2952

Plausible: "Show full message" lost its fork hover. The rule now covers only hover:bg-accent/20. "Show full message" is now a ghost-muted Button, which hovers with an opaque bg-accent: a stronger version of the cool smudge this customization exists to remove. The guard assertion for it was deleted rather than ported. .fork/customizations.yaml (~3799–3801) and the guard's header still describe two utilities. Consider keying a rule on the button inside .fork-timeline-cutoff and pinning it.

---

11. ChatView.tsx:3681

Plausible: mobile web lost the dock morph. The hero slide is forced on here (!prefersReducedMotion, 400ms), but runMobileComposerTransition (~7999) now takes active: panelAnimationsActive, which is off by default. On a mobile viewport with default settings, the view-transition morph is skipped, data-mobileComposerRouteTransition is never set, and the 400ms WAAPI translate runs instead. Before the sync the morph always ran on mobile and the slide stood down. The forced flag also ignores PanelAnimationSuppressionProvider.

---

12. ProviderModelPicker.tsx:355

Two small things in the trait suffix:

• The sr-only text below is hard-coded to "Fast mode on". With speedIcon === "ultrafast" the trigger shows UltrafastIcon, but a screen reader announces the wrong tier. Upstream's TraitsPicker now uses its speed label.
• Plausible: in a multi-model draft the trigger still appends the primary model's traits, and the tooltip names activeEntry. If the primary is shift-clicked out of the selection, the label shows effort and speed for a model nobody is sending to, and the effort slider edits that model's options.

---

13. UsagePage.tsx:110

Plausible: Escape can navigate away from the Usage popover. Upstream's new useEscapeToGoBack() attaches a window-level Escape → history.back() handler. It also runs in the fork's chrome="panel" mode, where this page lives in the sidebar popover rather than on a route. Base UI's dismiss stops the first Escape. But during the exit transition, or when a nested control consumes Escape without preventDefault, the keypress reaches the window and navigates away from the current thread. Skip the hook when chrome === "panel".

---

14. vite.config.ts:382

Convention: the shadcn/* rules are off for three whole upstream files. All five rules are disabled for Sidebar.tsx, SidebarChrome.tsx and MessagesTimeline.tsx. That also removes upstream's gate from all the upstream-authored code in those files, which contradicts the comment above ("stays on everywhere else so upstream-authored code keeps its gate"). It also lists MessagesTimeline, which isn't in FORK_ADOPTED_FILES, and omits AppSidebarLayout.tsx, which is. A narrower option is oxlint-disable-next-line shadcn/<rule> on the fenced lines; reportUnusedDisableDirectives: "error" keeps those honest.

---

15. AppSidebarLayout.tsx:66

Plausible: the traffic-light inset doesn't follow desktop zoom. Upstream's --desktop-window-controls-inset is 90 native points divided by the zoom factor. This sync also adds syncMacosWindowButtons, which keeps the native buttons at a fixed size. At Cmd− (zoom ~0.8), 80 CSS px ≈ 64pt, so the 28px sidebar toggle sits under the zoom button. Zoomed in, it drifts away to the right. Something like calc(80px * var(--desktop-window-controls-inset, 90px) / 90px) would keep the fork's figure and still scale with zoom.

---

16. compileCache.ts:24

Low: upstream-named cache directory. Upstream's new compile cache writes to <tmpdir | XDG cache>/t3code/compile-cache, the same directory an installed upstream T3 Code uses. The fork's app-identity work didn't rename it. On macOS it's outside $HOME, so launch-isolation-check.sh doesn't catch it. Functionally harmless because Node keys entries by path, but it breaks the fork-owned-paths rule.

---

17. MessagesTimeline.tsx:4566

Cleanup: this comment is inaccurate. Nothing renders AgentSpawnMemberRow anymore, and it never read this context anyway. That component and AGENT_MEMBER_STATUS_LABEL (~4579–4700) lost their only caller in this resolution, which leaves about 115 dead lines plus the two voids here. Either delete them or say plainly that they're kept so syncs merge cleanly.

🤖 Generated with Claude Code

[NoahHendrickson]

Pushed 427c3e5f9 addressing the ultra review at 2b79e79.

Fixed

1. Queued sends drop design pills. A send with pending pills now skips the queue and goes out with the <design_change_request> blocks appended, which is how it behaved before the queue existed.
2. Multi-model + pills. markSent now runs only when every target started. After a partial failure the pills stay, so the retry carries them. The sent record keys on the first started thread (markSent takes an optional sentThreadRef), not the draft. A unit test covers this.
3. Mod+Shift+E. The combined model trigger declares data-composer-shortcut="composer.effort" when the model has traits.
4. assignDraftProject. The pill resolves settings from the project's own environment through a single resolveProjectSettings(...). That result drives env mode, newWorktreesStartFromOrigin, the default model and defaultRuntimeMode. A test covers it.
5. Settled PR badge. Settled slim rows mute the badge at rest. Hover or focus on the row restores the color. Pinned in the card-rows guard.
6. Liveness pill memo is back to its five real deps. "View agents" is still listed as a follow-up in the PR body.
7. Patched deps. node-pty and react-native-nitro-markdown are pinned to their patched versions.
8. Bootstrap branch. preparingWorktree is now && shouldPrepareWorktree. New server test (confirmed it fails without the fix).
9. Mobile dock morph uses the fork's 400ms clock and yields only to reduced motion, matching the hero slide.
10. The sr-only label now says "Ultrafast mode on" for the ultrafast tier.
11. Usage popover. useEscapeToGoBack() only mounts on the routed page.
12. shadcn lint. The three upstream files are out of the override. The fork's own lines carry oxlint-disable-next-line, which reportUnusedDisableDirectives keeps honest.
13. Traffic-light inset is calc(var(--desktop-window-controls-inset, 90px) * 8 / 9): 80pt at every zoom level, and 80px when the preload hasn't run.
14. Compile cache moved to t3code-fork/compile-cache. The file is fenced, watched by fork-app-identity and pinned in its guard.
15. The comment now says plainly that AgentSpawnMemberRow and its label table are kept unedited so syncs merge cleanly.

• Scratch in the New agent pill. The pill filters out isScratchProject like upstream's pickers do. Pinned in the guard.
• Inline thread. The hooks now run above the early return, and the react(hooks) baseline entry is removed.

Handled in the parallel commit from the session that owns the composer CSS: #5 (Tiptap placeholder selectors and the 4px overhang) and #10 ("Show full message" hover, with its guard and manifest).

Not changed

• 2, smaller point (a second send during worktree prep picks up the same pills). It needs a send typed in the seconds before the starts settle. Fixing it means taking the pills out at commit time and restoring them on failure, which adds more machinery than the risk justifies here.
• 12, second bullet (primary model's traits in a multi-model draft). This matches upstream: its TraitsPicker also edits the primary selection when several models are selected.

Verified: web, server and desktop typecheck are clean. Fork guards pass 488/488, plus the touched unit tests (design store, new-agent draft, usage). The server branch fill-in tests pass 2/2. .fork/lint-owned.mjs reports no blocking warnings. vp lint on the touched files shows 0 errors, and knip:check is clean.

🤖 Generated with Claude Code

[NoahHendrickson]

Code review (/code-review max) at f9adf51

All 15 findings below were confirmed by an independent verifier at this head, and they are ordered most severe first. They are all in this review body because the PR changes 6,535 files and GitHub only indexes the first 3,000 for inline comments, so none of these lines can take one.

Several confirmed issues are left out because 427c3e5 already fixed them: queued messages dropping design pills, hooks after ChatView's early return, the liveness-pill memo deps, Mod+Shift+E (composer.effort) having no target, the Ultrafast screen-reader label, draft settings resolving against the primary environment, settled-row PR badge muting, the shadcn lint scope, and the node-pty/nitro-markdown pins. Roughly 20 lower-priority confirmed items were also cut: stale guards and manifest intents, upstream docs that describe upstream's UI, unfenced edits, and duplicated code.

---

1. apps/desktop/src/main.ts:142: SSH launches fail on packaged fork builds. This hands environment.appVersion to the SSH runner as its archive version. A packaged fork build (0.1.x, set by fork-release.yml's --build-version) therefore asks pingdotgg/t3code for a release that doesn't exist.

In archive mode the launch script runs "$RUNNER_FILE" --version under set -eu before any reuse check. The runner fetches github.com/pingdotgg/t3code/releases/download/v0.1.39/SHA256SUMS, gets a 404 (reproduced: curl exit 56), and the launch aborts. Before the sync, remotes with t3 on PATH or a server already running still worked; now they fail too. Dev builds report 0.0.45, which upstream does publish, so local testing never shows this. A fenced, pinned upstream base version would avoid it.

---

2. apps/web/src/theme.custom.css:736: every desktop composer gets phantom scrollbars. The fork's scroll box on [data-fork-composer-prompt] now overflows by 4px on both axes, because upstream's Tiptap editor uses -m-1 … p-1 inside a flow-root container (pingdotgg#13301). At ≥40rem with a one-line prompt, the editor's border box spans [-4, W+4] × [-4, 27] inside a W × 23 wrapper with overflow-y: auto (so overflow-x computes to auto). index.css styles ::-webkit-scrollbar, which makes scrollbars always-on in Chromium/Electron, so a 6px vertical thumb and a horizontal bar show under the prompt, and a trackpad can nudge the text 4px. The fork's own comments at :748-765 describe guarding against exactly this. Quick check: $('[data-fork-composer-prompt]').scrollHeight - clientHeight === 4.

---

3. apps/web/src/theme.custom.css:701 (also :727, :766, and theme.custom.westworld.css:198): the placeholder rules match nothing. They select [data-testid="composer-editor"] ~ div, a sibling layout only Lexical had. Under Tiptap the placeholder is a sibling of EditorContent's wrapper, so the colour, the 14/23 type, the overflow clip and Westworld's draft-placeholder ink are all gone (a jsdom render of the real editor matches 0 elements). The placeholder renders upstream's text-placeholder/75. On the Westworld draft hero it loses its white-70% ink over the portrait, which is the legibility bug the westworld rule exists to fix. At a non-14px prompt font size the placeholder and the caret disagree. The guards only regex the CSS text, so they stay green.

---

4. apps/web/src/components/chat/ChatComposer.tsx:1139: Empty chips in narrow strips. The merge tagged the fork's text-only Build/Plan label (and the runtime SelectValue at :1174) with data-composer-control-label. Upstream's new icon-only collapse step hides those labels.

Phone-width web client on a server thread, composer unfocused, crowded strip (PR chip, liveness pill, env chip). resolveRestingComposerControlsLayout returns iconOnlyCount: 1 for host widths between the icon-only and natural widths. [&_[data-composer-control-label]]:invisible absolute then leaves the Build/Plan toggle as an empty pill and the runtime chip as a bare chevron. Before the sync, the block moved into the ⋯ overflow menu with its labels intact. Dropping the attribute from both places restores that.

---

5. apps/web/src/components/ChatView.tsx:7784: Design-pill sends overtake queued messages. 427c3e5's gate skips the whole queue branch, including upstream's queueStillSending ordering guard, and ignores Mod+Enter's 'queue this one'.

A turn is running with followUpBehavior: "queue" (the default). The user sends A, which is queued until the next tool completes. They then attach a design change and send B. Because count(...) === 0 is false, B goes out immediately as a turn start on the running thread (a steer; on Grok this cancels the in-flight prompt) while A keeps waiting. The agent reads B before A, and B lands above A in the transcript. The same happens on an idle thread whose queue head is still preparing attachments. The real fix is probably a pills slot on QueuedComposerMessage; removing the gate brings back the dropped-request bug.

---

6. apps/web/src/custom/ComposerShell.css:26: the runtime-mode chip shows a chevron in every composer. The rule that hides its caret targets [data-slot="select-icon"], which upstream's rewritten ComposerSelectControl no longer emits; it renders a bare <SelectPrimitive.Icon>. That breaks the manifest intent "label only, no icon or caret".

---

7. apps/web/src/components/Sidebar.tsx:1353 (also DraftProjectPill.tsx:212 and :252): the fork's folder marks never render. Upstream pingdotgg#11572 made ProjectFavicon return a coloured monogram whenever the project has a title, before it looks at fallbackIcon. Any project without a favicon now shows a hashed-colour two-letter tile in Sidebar V2 cards, the draft project chip and its picker rows, and the tile also flashes while a real favicon loads. sidebarV2CardRows.test.ts only matches the source string, so it stays green.

---

8. AGENTS.md:97: Upstream's instruction leaks a credential in this fork. It says to keep the admin T3CODE_DEV_AUTH_TOKEN in "the main checkout's gitignored .env", but this fork force-tracks .env in a public repo.

A maintainer or agent follows this line (CLAUDE.md symlinks to it) or docs/operations/development.md:73. setup-worktree symlinks .env into every worktree and git shows the tracked file as modified, so the next git add -A or commit -a pushes an admin-scoped dev credential to NoahHendrickson/t3code. The t3ConnectOfficialConfig guard only fails after the push. This needs a fenced correction, e.g. point to a gitignored file that setup-worktree also links.

---

9. apps/web/src/components/ChatView.tsx:8260: Multi-model sends leave the design pills live after the composer is released. This lasts until every target has started, and indefinitely after a partial failure.

(1) The composer is released at 8242-8244 while await starts waits for checkout and setup scripts, which can take minutes. A new prompt sent in that window goes through takeForSend (a pure read), so the same <design_change_request> is appended again. (2) One target fails while the user has typed B: canRestoreDraft() is false and the toast says to send or clear B first. Sending B takes the old pills and markSent clears them. 'Restore prompt' then brings back text and attachments but no pills, so the failed model never gets the design request.

---

10. apps/web/src/components/BranchToolbarBranchSelector.tsx:792: A single-PR chip loses the fork's chip styling. It now renders through upstream's ThreadPullRequestBadgeControl, which emits an <a> with no data-slot for a single PR. None of the fork's :is(button, [data-slot="button"], [data-fork-context-chip]) context-chip rules match it.

A thread whose branch has exactly one linked PR, which is the common case. The chip renders with bare ComposerControl xs styling (transparent, 7px padding, --control-radius, upstream hover:bg-accent) next to the 24px glass-filled, 6px-radius chips; only order: 1 still applies. Stacks and multi-PR badges render as <button> and keep the styling, so the chips are inconsistent. Adding data-fork-context-chip to the render element would fix it.

---

11. apps/web/src/theme.custom.css:2958: long file names render in two tones. The changed-files rule targets span.truncate, which under upstream's new MiddleTruncate (pingdotgg#12805) is only the head of the name. For a 15+ character name such as MessagesTimeline.tsx, "MessagesTi" renders at full --foreground and "meline.tsx" at the wrapper's text-foreground/85, a visible seam in nearly every card until row hover.

---

12. apps/web/src/overrides/components/chat/ModelPickerContent.tsx:514: A Shift-toggle from search results still runs the single-pick tail. It clears the query and opens the provider page, so a multi-model selection can't be built from one search.

Multi-model draft, combined menu open. The user types '5.5' and Shift-clicks Opus 5.5. It is toggled in, then setSearchQuery('') and openPage(Claude models) run. A second Shift-click where GPT-5.5 was now hits whichever Claude row (or the back row) sits there. Upstream keeps the query on a toggle. Fix: if (!isSearching || additive) return;.

---

13. apps/web/src/components/ChatView.tsx:8710: Unsent design pills are orphaned when a draft is re-threaded. If the bootstrap thread is deleted (Cancel setup, or Work locally after thread creation), the draft gets a new id from newThreadId(), but its unsent pills stay keyed under the old id.

A new-worktree draft with a design pill is sent, and the user cancels setup after the thread is created. startPromise fails with disposition 'deleted' and turnStartSucceeded is false, so the pills aren't cleared. The draft is then re-threaded here; the background path does the same through restoreFailedBackgroundDraftThread around :8652. useDesignChangeTargetRef resolves the new id, so the composer shows no pills and the next or automatic send reads an empty key. The old entries stay orphaned in the store.

---

14. apps/web/src/components/ChatView.tsx:7190: Rewind restores raw design-request markup. Upstream's new rewind path uses recallableComposerPrompt(message.text), which only strips terminal, element and preview blocks, so rewinding a turn that carried a design request writes the raw <design_change_request> markdown into the composer.

The user rewinds 'Fix the header' plus a design pill. The composer gets Fix the header\n\n<design_change_request>…</design_change_request> with the caret after it. If they type and resend, the run is no longer trailing, extractTrailingDesignChanges finds nothing, and the transcript shows the raw request inline. On servers without inline message context, the legacy context blocks also survive (reproduced with the real module).

---

15. apps/web/src/components/ChatView.tsx:8266: 427c3e5's multi-model sent-preview fix doesn't take effect. The record is keyed on the first started thread but stored under the pill's runtimeTabId, which embeds the draft's thread id. No panel can ever match both.

Draft D has a preview tab and sends design pills to two models. markSent writes byTabId[runtimeTabId(D)] = { threadKey: key(S1) }. D's panel finds the record but compares it against key(D), so designSentPreviews.ts:248 returns false. S1's panel computes a runtimeTabId containing S1 and finds nothing. No 'did it land?' prompt appears, verification never arms, and the record stays until Discard or the tab closes (reproduced with the real modules).

🤖 Generated with Claude Code

[NoahHendrickson]

Pushed 2a0c5c5b0 addressing the /code-review max findings at f9adf51c8.

Fixed

1. SSH launches on packaged fork builds. The archive version is now serverPackageJson.version, the upstream release this fork is synced to, read from source and bundled at build time. It is no longer the fork's 0.1.x app version, which names no upstream release. Fenced under fork-app-identity, pinned in its guard, and noted in the intent.
2. Empty chips. The text-only Build/Plan label and the runtime SelectValue no longer carry data-composer-control-label, so upstream's icon-only collapse can't hide them.
3. Design pills vs. the queue. The bypass gate is gone. QueuedComposerMessage has a fenced forkDesignChanges slot. The pills leave the composer at queue time, sendQueuedMessage appends them as trailing blocks (and keeps them trailing on legacy-context servers), and records the send. Stop/remove hands them back to the composer. Ordering and Mod+Enter now behave as upstream intends. Covered by a new sendQueuedMessage test.
4. Runtime chip caret. The rule now targets the trigger child that holds [data-composer-control-chevron] (upstream's bare Select.Icon).
5. Folder marks. An explicit fallbackIcon now suppresses upstream's automatic title monogram; an icon the user picks in project settings still wins. Fenced in ProjectFavicon.tsx, watched and guarded under sidebar-v2-card-rows.
6. Dev credential. Fenced notes in AGENTS.md and docs/operations/development.md send T3CODE_DEV_AUTH_TOKEN to .env.local, and setup-worktree.ts now links that file too. Guarded under t3-connect-official-config. I also checked that no token appears in .env or its history.
7. Multi-model pills. They are detached when the composer clears, so a prompt sent while the threads start can't carry them again. Both the in-place restore and the "Restore prompt" toast path give them back with the failed draft.
8. Shift-toggle in search keeps the query.
9. Re-threaded drafts. Both re-thread paths move the unsent pills to the draft's new thread id with a new rekey.
10. Rewind. The trailing design-change run is stripped before recallableComposerPrompt, so the legacy context blocks come off too.
11. Sent-preview record for multi-model sends. You were right that it couldn't take effect. Multi-model sends no longer mint a record at all, since the draft's preview tab can't answer for the started threads, and I removed markSent's sentThreadRef parameter.

Already in the held commit eb2bab85a (pending a visual pass): #2 (scrollbar overhang), #3 (placeholder :has selectors), #10 (single-PR chip via [data-fork-pr-chip]).

Open: #11 (two-tone MiddleTruncate names, theme.custom.css ~2958). It sits next to that held commit's edit to the same rule block, so it should be fixed on top of it once it lands rather than now. The likely fix is to also ink the span.truncate + span.shrink-0 tail.

Verified: web and desktop typecheck pass, fork guards pass 491/491, plus the touched design-store, queue-sender, composer and favicon tests (693 in total). Lint on the touched files shows 0 errors, .fork/lint-owned.mjs reports no blocking warnings, and knip:check is clean.

🤖 Generated with Claude Code

[NoahHendrickson]

Reviewed the prior review bodies, the inline thread, and the follow-up fixes through 2a0c5c5b0199d115c07f5636dc1f73b9ce09da03. There is still work to do before calling the feedback fully addressed. This is a review of the reported fork-integration issues and their fixes, not an exhaustive audit of the upstream commit range.

Remaining reported issues

1. [P2] Land and verify the held composer CSS fixes. The PR head does not contain eb2bab85a. The desktop prompt still combines the fork's scroll container with Tiptap's -m-1 p-1 overhang, and the placeholder selectors still target a sibling of the contenteditable rather than a sibling of its EditorContent wrapper. The scrollbar and placeholder findings therefore remain in the code being merged, including the Westworld placeholder rule. The held changes need to reach this branch after their visual pass.

2. [P3] The other held styling fixes are also absent. A single PR renders as an anchor bearing data-fork-pr-chip, but the context-chip geometry selector still matches only buttons, button slots, and data-fork-context-chip. “Show full message” uses ghost-muted, while the transcript hover rule only targets hover:bg-accent/20. Both fixes are in the held CSS commit, not this PR head.

3. [P3] Finish the filename-color fix. The changed-files selector still colors only span.truncate. MiddleTruncate renders the tail in a separate span.shrink-0, inheriting the wrapper's text-foreground/85. Long filenames remain two-tone at rest. Apply the same ink to both halves without changing the intentionally muted header hint.

Omission in the latest queue fix

4. [P3] Show queued design attachments in the queued-message row. QueuedMessageTimelineRow counts images/files and terminal/preview/review contexts, but never reads queuedMessage.forkDesignChanges. Queue a design-only message during a running turn: the composer pill disappears correctly, but the queued row has no content or attachment summary—only “Queued” and its controls. With prose, the attached design request is likewise invisible until dispatch. Include a design-change summary/count so users can identify what is waiting and choose the correct row to send or restore. The sender itself now carries the payload correctly.

Explicit limitation, rather than a restored capability

The latest multi-model fix deliberately detaches the pills without creating a sent-preview record. This removes the invalid record, but multi-model sends still cannot offer the “did it land?” verification flow. Please record this as an accepted limitation/follow-up, or implement the missing flow, rather than treating that behavior as restored.

The other substantive fixes checked out on inspection, including payload delivery, queue ordering, failed-send restoration, draft rekeying, rewind cleanup, settings resolution, shortcuts, hooks ordering, SSH archive selection, dependency pins, and the original CI repairs. Deferring Cursor's model-menu extraction is reasonable: its asserted 1,000-line budget is not a repository rule. The liveness pill's “View agents” action remains an acknowledged follow-up.

Validation: 559 focused web tests passed using the web project's configuration, including the fork guard suite, design draft store, queued sender, new-agent draft, TraitsPicker, and model-picker tests. Current CI is green. No browser/visual verification performed; the remaining UI findings were checked against source and markup. No source changes made.

Review by GPT-6 via Codex.

[NoahHendrickson]

Re the review at 2a0c5c5b0:

4. Queued design attachments. Fixed in d3ef0f0e6. The queued row's summary counts design changes alongside attachments and context items ("1 design change"), so a design-only message no longer renders as an empty row.

• Multi-model "did it land?" This is recorded as an accepted limitation in the fork-design-mode manifest (watch note on ChatView.tsx), not as restored behavior. A multi-model send creates no sent-preview record, because preview tabs belong to the draft's thread and none of the started threads can answer for it.
  1–3. Composer CSS, PR chip and "Show full message" hover, and filename ink. These are ready as two local commits rebased onto d3ef0f0e6: the held Tiptap/PR-chip/hover commit and a new one for fix(fork): survive the Clerk bridge's registerSchemesAsPrivileged race #11 that inks MiddleTruncate's span.truncate + span.shrink-0 tail, pinned in forkChangedFilesCard.test.ts. Fork guards pass 492/492 on that tree. They'll be pushed after the visual pass.

🤖 Generated with Claude Code

[NoahHendrickson]

Pushed e2f7f20d6 and 81b776f30 after a visual pass in the dev app. They close the remaining items from the review at 2a0c5c5b0: #1 (Tiptap scroll overhang and placeholder selectors, including Westworld), #2 (single-PR chip and "Show full message" hover) and #3 (both halves of middle-truncated file names inked). Fork guards: 492/492 on this tree.

🤖 Generated with Claude Code