feat: adopt the 2026-08-12 upstream batch (11 of 12 change sets)

[rynfar]

Problem

Twelve T3 Code commits landed past Pylon's review cursor (c196f422e..b73232bdd), covering sidebar chrome, the right panel, the composer, mobile thread actions, and the Connect CLI OAuth flow. Upstream is in polish mode — nine of the twelve are fix — so most of this is inherited-defect repair Pylon was still carrying.

What this does

Adopts eleven of the twelve change sets and skips one. Ten cherry-picked clean; O1 was ported by hand.

Three adaptations are worth review attention:

O1 (sidebar footer compaction) reverts a Pylon fix if taken verbatim. Upstream deletes useCanGoBack and makes the footer's Back button always navigate to /. Pylon had deliberately fixed that so Back returns to the thread the user actually came from. The port keeps Pylon's handler, keeps SidebarAccountDrainPill in the footer stack, and preserves PylonMark in the same file. Upstream's "Dismiss until next launch" goes away with the change — the pill is no longer a full-width banner, so there is nothing left to dismiss.

O7 (right panel empty state) shipped a real defect, fixed here rather than carried in. The new launcher claims bare B/T/F/D/P/A on a capture-phase window listener and treated an empty contenteditable as "not typing". Pylon's composer is a Lexical ContentEditable that is empty at rest, so with a thread open and the right panel empty, a message starting with any of those six letters lost its first keystroke to a surface opening instead — and stopPropagation meant nothing downstream could put it back. A focused text surface now always keeps its own keystrokes.

O9 (Connect CLI OAuth) is inert in Pylon, taken for drift reduction. It fixes a genuine upstream bug (parameters dropped across Clerk's sign-in redirect) but does nothing until Pylon owns a Clerk application — E21 was skipped precisely so a fresh clone does not point at T3 infrastructure.

O4 is skipped: it vouches a T3 contributor in .github/VOUCHED.td, which has no Pylon meaning. Same class as F24 and G4.

Full per-change-set reasoning is in .agents/upstream-review.md, which also advances the cursor to b73232bdd.

Verification

• Typecheck clean across web, mobile, shared, and contracts. Server reports 0 errors — only effect diagnostic suggestions in files this batch does not touch.
• Tests: web 2324 (244 files), mobile 657 (105 files), shared connectAuth 7, server publicConfig 12. All passing.
• vp lint clean over all 33 changed TypeScript files; vp fmt --check clean over all 36 changed files.
• O11's bundled contrast test asserts ≥4.5:1 across every built-in palette and passes against Pylon's F10-tuned palettes, not just upstream's.

Not verified in a real client. No browser or mobile pass ran. O1's compacted footer and O7's launcher plus its guard fix are the two that changed interaction rather than only layout, so they are where a UI pass would be worth the most. Happy to run one before merge if you want it.

Model: Claude Opus 5 via Claude Code.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}

[rynfar]

Real-client verification

Ran both passes. Web against a VACUUM INTO copy of the real database (9 threads); iOS on a purpose-booted iPhone 17 Pro simulator. Migrations 37–42 applied cleanly over that real database on first boot.

Six change sets confirmed against live behavior:

• O1 — footer is one 48px row (flex-direction: row) with three 32px icon buttons, replacing three stacked labeled rows. The Pylon-first adaptation is confirmed: from a thread, Usage → Back returned to /cec0464d…/6531739c…, the exact thread, not /. Upstream's version would have landed on root. On a footer page the row collapses to a single Back.
• O2 — at the 208px sidebar minimum, .sidebar-brand computes display: flex and renders 58px wide. The retired 13.5rem (216px) gate would have hidden it there.
• O5 — picker glyphs at x=294 vs prompt text at x=293 (1px); the button box still extends 10px left for the hit target, as intended.
• O7 — launcher renders with its six Kbd badges, and availability gating is right: data-surface-launcher-keys="TFDA", so Browser and Pull request stay visible-but-disabled without claiming their letters. Both guard paths confirmed — empty composer focused, t typed into the composer and nothing opened; launcher focused, t opened the Terminal. That is the reported defect and its fix, exercised.
• O12 — double-click cleared the persisted width (208 → null) and reset the live sidebar 208px → 256px.
• O3 — full round trip on iOS. Long-press showed the native menu as Un-settle / Regenerate title / Delete; tapping it regenerated through the provider: "Fork T3 Code With Pylon Branding" → "Build and Update Pylon Desktop Fork", title_regeneration_started_at back to null, new title rendering in the list.

Still unproven, with reasons:

• O6 cannot be verified on this host at all. It is gated to Platform.OS === "android", so on iOS includeOrderedLists is false and the changed path is inert — an iOS pass gives it zero coverage. There is no Android SDK or adb here. It rests on its unit tests.
• O11 and O8 are structurally unreachable locally — window.Clerk is undefined (Pylon has no Clerk app, see E21), and hosted-static onboarding does not exist in local mode. Both need a hosted deployment.
• O10 was not reached. Only one thread carries file-edit activities, its earlier turns sit behind E10 pagination, and it is heavy enough to wedge the automation bridge. Rests on its class-level unit test.
• O1's update pill was not exercised — SidebarUpdatePill returns null outside Electron, so the icon button, checking spinner, and release-notes tooltip need a desktop pass.

No before/after images attached: the measurements above are the evidence, and the browser tooling here returns screenshots to the agent rather than to a file I can upload. Say the word if you want images and I will capture them properly.

Two environment gotchas are now recorded in the ledger: the mobile dev bundle id is Pylon-owned (com.rynfar.pylon.dev, not the documented com.t3tools.t3code.dev), and ios/Pods goes stale whenever vp i changes a pnpm patch hash — needing pod install, which on this host also needs a UTF-8 LANG to survive CocoaPods 1.17.0 on Ruby 4.0.6.

[github-actions]

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for 7506bdb.

This comment will update automatically after the next completed run.