merge: sync upstream through 083fa4ab2 (15 commits)

[omegent-app]

Adopts upstream through 083fa4ab2 — the 15 commits after 9c7622dac. Fourteen conflicted
files, the largest batch this fork has taken.

Important

Merge with a merge commit, not squash. Squashing collapses the second parent and discards
upstream's commits as ancestors — the thing that keeps "commits behind upstream" honest.

Upstream commits entering the product

Commit
083fa4ab2	feat(web): use OKLCH for theme palettes (pingdotgg#6036)
35172010b	fix(web): use a clearer pull action icon (pingdotgg#6194)
57b105267	fix(web): thread error banner dismiss survives reconnect and rerenders (pingdotgg#6123)
1e355a2a3	fix(web): render dropdowns above toasts (pingdotgg#6165)
44621c345	feat(web): back buttons for the pull requests and usage pages (pingdotgg#6031)
220e573b1	fix(shared): detect Azure DevOps SSH remotes (pingdotgg#6187)
6befe42eb	fix(shared): normalize a bare Windows drive root (pingdotgg#6189)
b30a9bc41	feat(web): make environment artwork theme aware (pingdotgg#6183)
ac4780f45	fix(web): restore typography font sizes to defaults (pingdotgg#6172)
3da7f9c5c	fix(mobile): guard App Store release versions (pingdotgg#6177)
c842c6f5b	Add hourly past-24-hour usage view (pingdotgg#6170)
6676f9c83	fix(mobile): stabilize thread composer and interactions (pingdotgg#5986)
65b005f1e	feat(web): add Copy Thread ID to the thread context menus (pingdotgg#5574)
752acbf65	feat: shift+click to create a thread in the current project (pingdotgg#5994)
f5fce7416	fix(pull-requests): route self-hosted GitLab remotes (pingdotgg#6061)

The four that needed judgement

Copy Thread ID (pingdotgg#5574) — upstream shipped a feature the fork already had. The merge produced
two "Copy thread ID" entries in the thread action menu, and the fork's own menu test caught
it rather than it reaching a user. One entry survives, in upstream's position, with the state and
handlers collapsed onto upstream's naming — which is also better typed (ThreadId rather than
string, no String() coercion).

Thread error dismissal (pingdotgg#6123) — supersession, not duplication. The fork had its own dismissal
(a dismissedServerErrorsByThreadKey map feeding resolveServerThreadError). Upstream's
session-scoped masking fixes the same bug the fork's version had: a dismissal that did not survive
a reconnect. So upstream's is adopted, the fork's map and the now-dead arm of the resolver are
removed, and the fork's dismissal tests are retargeted at upstream's helpers rather than deleted
— upstream ships those helpers untested, so deleting would have quietly dropped the coverage.

Mobile composer (pingdotgg#5986) — upstream's overlay rewrite, with the fork's chips inside it. Upstream
restructures KeyboardStickyView, adds the scroll-to-end pill and keyboard-visibility gating; the
fork's overlay carried ComposerQueuedMessages (steering-queue chips). Upstream's structure is taken
whole and the chips moved inside it. Two follow-on repairs: the fork's copy of
use-thread-composer-state predates upstream's activeThreadBusy / selectedThreadQueueCount, so
both are computed there and threaded through — which also brings the fork upstream's honest "Queue"
vs "Send" label, where the fork had hard-coded "Send".

ThreadFeed keeps the fork's sticky empty→filled mount key over upstream's plain toggle:
remounting a feed that briefly empties mid-sync reads as the conversation reloading. The fork's
unread-activity latch is woven into upstream's rebuilt setEndFollow.

Smaller: app.config.ts takes upstream's 1.0.3 bump with the fork's MOBILE_RUNTIME_VERSION_OVERRIDE;
the mobile home screen keeps its single title, since upstream's native-stack title would render the
connection status twice against the fork's header; the release workflow stays fork-owned.

Checks I ran rather than assumed

• No test was silently nested by my re-bracing. Three conflict hunks were cut mid-function, so
  closing braces went back by hand — the exact shape that can bury a describe inside another and
  stop it running. Sidebar.logic.test.ts goes 129 fork + 3 upstream = 132 it( blocks, all
  executing; threadActivity.test.ts runs 22.
• Exactly one copy-thread-id entry remains, verified against the menu test.
• Typecheck clean across 18 packages, vp build passes.

Adversarial review

Both reviewers went through all seven resolutions. Neither found a HIGH in the conflict
resolutions themselves, but between them they found two real defects I introduced — both fixed.

Finding	Disposition
grok-4.5, HIGH — two scroll-to-end controls. pingdotgg#5986 removed the feed-local "Scroll to latest" chip and replaced it with the composer-overlay pill; the fork had built its own chip, and the merge kept both. Two chevrons in the same band, two controls for VoiceOver, and only one of them re-arming follow on press. Worse, the fork's chip positions against an estimated overlay height that excludes the queued-message chips, so with a queue it could sit under the overlay while the pill stayed correct	Fixed in e5b838f94. This is the failure class exactly — two implementations, both live, fighting. Upstream's pill wins since the rest of pingdotgg#5986's geometry is built around it. What the fork's chip had that upstream's lacks is the difference between "you scrolled up" and "you are missing something", so the feed now reports unread activity upward the same way it reports follow state, and the pill carries the dot and reads "New activity. Scroll to end". The chip's helpers went with it rather than lingering
gpt-5.6-sol, MEDIUM — the sticky mount latch tracked the bare thread id while the mount key it drives is environment-scoped. Two environments holding the same id would carry "already filled" across a switch, so the new feed skips its remount and computes the initial scroll against a zero composer inset — the positioning bug upstream's key exists to prevent	Fixed in 17adee9de. The fork's recurring shape: single-environment reasoning in a deliberately multi-environment client. Nothing typechecks differently and no test covers two environments sharing a thread id, so it would have shipped green
gpt-5.6-sol, LOW — the retargeted dismissal tests exercise upstream's helpers but not ChatView's integration; they would still pass if ChatView stopped consulting the session set	Acknowledged, not fixed. A component-level test is a bigger piece of work than belongs in a sync PR. It verified the wiring itself is correct at ChatView.tsx:1592 and :6745
grok-4.5, MEDIUM — board context menu ordering differs from the shared action menu builder	Pre-existing fork divergence, untouched by this merge

Confirmed sound by both: exactly one copy-thread-id entry in upstream's position, dispatched
through the typed handler from both call sites; resolveServerThreadError keeps local-over-server
precedence with no dismissal path left behind; the three useState declarations my removal regex ate
are restored correctly with all consumers intact; the queue chips sit inside the measured
composerOverlayRef so their height still feeds the list inset, with the fork's steering, editing
and start-new-thread plumbing whole; activeThreadBusy / queue count / the Queue-vs-Send label match
upstream; keeping the mobile route title out is right because HomeHeader already owns it on both
platforms; and the hand-restored braces close the intended blocks — no test is accidentally nested
or skipped.

Verification

• 2,763 tests pass; the only failure is the pre-existing CodexTextGeneration launch-args one,
  which reproduces on unmerged fork/dev.
• pnpm-lock.yaml regenerated rather than hand-merged.

Note

One full-suite run surfaced the pre-existing browserHistoryStore.test.ts flake (~1 run in 3).
It is not from this branch — I reproduced it earlier on clean fork/dev and at f067b34a1,
before any of this work. It passes in isolation.

Co-authored by @patroza

opened by Patrick Roza in chat thread Discord · Discord · T3