sync(upstream): merge upstream/main 48aa875c0 into fork/dev

[omegent-app]

Adopts upstream through 48aa875c0 — the ten commits that landed after #368, headlined by
pingdotgg#5566 (timeline live-follow), pingdotgg#5561 (reconnect loops during server stalls) and
pingdotgg#5558/pingdotgg#5551 (plans fold into chat, Build/Plan toggle removed).

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
48aa875c0	feat(web): remove Build/Plan toggle from the composer (pingdotgg#5551)
a8cd2ad2e	fix(web): plans stop hijacking the UI, fold into chat instead (pingdotgg#5558)
4a07c1ca9	docs: ship production T3 Connect public config in .env.example (pingdotgg#5573)
cf5c9948c	fix(web): keep agent panel rows stable (pingdotgg#5569)
2288d416a	fix(web): stop the "requests are slow" warning firing on every provider update (pingdotgg#5570)
ddfe45c66	test(server): catch client transfer regressions in CI (pingdotgg#5350)
9547cf246	fix(server): one disconnecting client no longer blocks every reconnect (pingdotgg#5572)
1c7d059f5	fix: scrolling up during a running thread no longer snaps back to the bottom (pingdotgg#5566)
6fa457607	fix(server): settle stopped Claude subagents (pingdotgg#5568)
ae7b27de8	fix: prevent reconnect loops during server stalls (pingdotgg#5561)

Behavioural decision you should know about

The fork's "keep the open WebSocket lease when a foreground liveness probe fails" is replaced by
upstream's probe model. Both address the same symptom — reconnect churn while the server is
stalled. The fork swallowed the probe failure; upstream pingdotgg#5561 instead splits the wake reasons
(application-active-probe vs application-active-reconnect), adds tolerance windows, and lets the
first post-probe attempt skip a backoff rung.

Keeping both was not an option: the auto-merge produced exactly that, and the fork's swallow made
upstream's wakeProbeFailed machinery unreachable dead code. I took upstream's model and kept the
fork's additions around it (connection diagnostics log, labelled ping-timeout errors,
followNetworkStatus resume handling). Say the word and I'll restore the lease-keeping behaviour
on top of upstream's model instead.

Resolutions

13 conflicted files. What actually needed judgement:

Where	Situation	Resolution
supervisor.ts	fork's probe swallow welded onto upstream's new wakeProbeFailed path — the swallow made it dead	probe failure propagates; log kept as tapCause (see above)
session.test.ts	upstream's patched pinger now tolerates two missed pong windows, so the fork's ping-timeout test fired too early and upstream's new test expected a bare transport close	fork's labelled timeout reporting kept; both tests re-timed to the new window
ChatView.tsx	fork's steering-queue send branch vs upstream's rebuilt anchoring; two follow-state variables coexisting after auto-merge (maintainTimelineAtEnd + timelineLiveFollowEnabled)	queue-chip branch kept, live-edge branch rebased onto upstream's follow state; the fork's superseded variable removed
ThreadFeed.tsx	same collision on mobile: upstream's follow latch vs the fork's isAtEnd, with the fork's stale handlers surviving as duplicates	upstream's latch adopted; the fork's unread-activity dot rebuilt on it rather than dropped
plan surfaces	upstream deleted PlanSidebar.tsx and folded plans into chat; the fork carried sidebar auto-open, dismissal and a toggle	sidebar wiring removed with the surface; the fork's settle-independent follow-up composer gate and findSidebarProposedPlan kept (the latter was silently dropped by the auto-merge while ChatView still called it)
MessagesTimeline.logic.ts	two row-union members merged into one malformed member; resolveOlderHistoryAutoLoad left over from the superseded pagination candidate, referenced only by its own tests	union split; dead helper and its tests removed
Omegent deep link	fork's #message-{id} scroll effect sat inside a block upstream rewrote	kept, retargeted at upstream's follow state
grokSegments.test.ts	upstream's new ThreadPlanProgressService missing from a fork test's layer	layer provided

Adversarial review

Both reviewers ran over the git show --cc resolution surface with the merged worktree available.

grok-4.5 — no high or medium findings; races/ordering, React effects and test-assertion integrity all reported clean. Four low findings, all fixed:

Finding	Disposition
supervisor test kept the fork's keep-lease title on upstream's reconnect body — a later "fix" could reintroduce keep-lease to match the name	Fixed. Renamed to quickly replaces a stalled mobile session after a probe timeout, plus a releaseCount === 1 assertion matching the desktop-side test
ThreadFeed: isAtEndRef never read or written after init; userNavigationInProgressRef write-only — leftovers of the retired fork latch	Fixed. Both removed
ThreadFeed: contentInset still destructured from the scroll event, unused since upstream's latch uses getState()	Fixed. Removed from the destructure
SidebarV2: upstream's role="status" sr-only "Regenerating title" absent from both title rows	Fixed, though not caused by this merge — the v2 rows are a fork rewrite of upstream's row, so this never came across with the surface. Derived locally from thread.titleRegeneration and restored in both rows

gpt-5.6-sol — two findings, both fixed:

Finding	Disposition
MED — findSidebarProposedPlan lost its only production consumer: it was restored during the merge because ChatView called it, but the plan-sidebar resolution removed that call	Fixed. Helper and its tests removed; activeProposedPlan uses findLatestProposedPlan directly
LOW — same misnamed supervisor test as above	Fixed as above

Both independently confirmed the intentional keep/drop surface is coherent: probe failures now reach wakeProbeFailed (tapCause, not swallowed), the live-follow rename is consistent through ChatView → MessagesTimeline → tests, steering-queue chips skip the live-edge re-arm, mobile unread only latches while follow is broken, and the transfer-budget test adds the omegent-t3 handshake rather than loosening the gate.

Verification

• Full recursive typecheck clean across 17 packages.
• 2,277 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.

Co-authored by @patroza

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