[bk-agent-01]

Refs: no ticket (requested in T3 session "Update Stage With Upstream T3" by tushar).

Summary

This PR merges upstream pingdotgg/t3code main cefb544af3 (tag v0.1.7-nightly.20261103.4119) into stage. Upstream has 1,838 new commits since the last sync (05b697f6a2, 2026-10-21).

• 220 files conflicted (about 520 hunks). Opus subagents resolved them by area, then 12 CI rounds fixed what broke without a conflict.
• Every fork feature is kept. Fork code was replaced only where upstream now ships the same behaviour (table below).
• Upstream migrations 068–078 register at fork ids 1060–1070 (stage and bkt3 databases are both at 1059). 063 (DropUnusedV1Projections) stays unregistered.
• The fork marker check passes. CI is green: typecheck, lint, all 8 server shards, web, package tests, markers and the stage validate build.

Upstream replaced fork code

Area	Fork code	Upstream now	Why upstream is kept
Queue	held-queue adapter continuation (QueueHeldContinuation.expbkt3.ts, deleted)	wakeSkipsHold (pingdotgg#17885)	Same behaviour, stricter Stop check, own tests
Queue	nextStartableQueuedRun, steer-during-stop skip, canStartQueuedRun SQL	identical upstream code	Same behaviour
Questions	messageAnswerText multi-select join	isMessageQuestionAnswer validation	Validates before resolving; blank entries now rejected
Connect	connectDiscoveryCache.expbkt3.ts (deleted)	built-in stale-while-revalidate cache (pingdotgg#14519)	Same, no extra layer
Presence	background-policy lease redaction (deleted)	background-policy RPCs removed upstream	Nothing left to redact
SQLite	fork reader pool	upstream reader pool	Same core; fork busy weighting and configurable size kept on top
Composer	isConnecting send-block exceptions; plan-review dock scroll aliasing	isConnecting and the collapsing composer removed upstream	Offline durable sends are never blocked now
Traits picker	planModeEnabled → planModeAvailable prop rename	upstream exported TraitsSelectionInput	Cosmetic; planModeAvailable (default on) still drives it
Mobile	thread deep-link reset, overlay list, terminal remount, Shiki time limit, card toolbar	workspaceDetailLinkAction, registerPresentedRoutes, terminal writer, native tokenizer, new composer	Superset, or the problem no longer exists
Desktop	branded linuxWmClass, catalog clear lock	removed upstream (setDesktopName sets the class)	Brand still applied
MCP results	JsonToolResult workaround	effect 4.1.2 accepts any structuredContent	Fork helper kept, now harmless

Decisions

1. Kept the fork's client.connected analytics event (TEC-1631) although upstream removed its own (fix(server): analytics stop counting reconnects, retries, and dev servers pingdotgg/t3code#24601).
2. Restored GET /api/auth/pairing-links and GET /api/auth/clients (deleted upstream as unreached); fork web uses them for member self-service.
3. Restored V1 ProviderSession / providerRuntime schemas and the change-request presentation helpers; fork code still uses them.
4. Added the fork thread-access gate to new upstream thread RPCs (forceStopThread, getNativeChildWork, getThreadFamily, getWorktreeThreadIds, getOlderSettledThreads, searchThreads) and to scratch-thread promotion (every moved thread).
5. The fork shell never trims older settled threads (omitOlderSettled forced off), and filtered users get no cross-user counts; the fork sidebar has no settled pager.
6. Linked MCP peers always get the full tool list; the fork's short per-caller list had broken upstream's new peer links. Access checks still apply per call.
7. cliReleaseChannelOf reads -stage-nightly, -production-nightly and .upstream.g<sha> as nightly, so BK desktop keeps offering updates.
8. Fork settings pages moved under upstream's settings modal; the phase-grouped sidebar stays while Settings is open.
9. Re-exported GitHubRest and GitHubRequestPolicy from @t3tools/source-control-github for per-user GitHub profiles (no public replacement upstream).
10. Fork-owned files that predate upstream's new lint and Effect style rules are on marked ratchet lists (scripts/lib/forkLintStyleDebt.expbkt3.ts + vite.config.ts, tsconfig.base.json) instead of mass rewrites. Upstream-owned files are never on them.
11. The context-budget test measures upstream toolkits only, against a fork budget file (agentContextBudget.expbkt3.json); upstream's budget file is unchanged.
12. apps/web pins typescript: catalog:; otherwise the fork's Plate packages pulled real TypeScript and a second vite-plus, which broke every @effect/vitest web test.
13. Restored three fork features an earlier merge (a3f26f98de) had lost: source-control profile on clone/publish/lookup, the "convert origin to HTTPS" toast, the machine badge on the draft screen.
14. Fixed fork regressions found by CI: archive navigation after the archive guard, the update pill's React import, t3-merge-base read for new worktrees, fork control MCP tool services (now provided per request).

Acceptance check

Item	Proof
No conflict markers; every fork marker accounted for	check-fork-markers.ts passes; per-file marker drops traced to worker reports
Typecheck, lint, tests green	CI on the head (all required jobs green)
Cross-model review clean	Codex gpt-6.1-sol, 4 rounds; all findings fixed or deferred (below)
Migrations 1060–1070 apply on the stage DB	stage deploy, effect_sql_migrations max = 1070 (after merge)
stagebkt3 serves the merge	~/.t3/stage-dev/deployed-sha = merge commit, service active (after merge)

Follow-ups (not in this PR)

• P2 (pre-existing): tracked clones (projectCloneStart) ignore the user's source-control profile.
• Pre-existing: the thread-cache eviction sweep lost its call site in a3f26f98de.
• Pre-existing: startWithHandoffs rebuilds provider text from userText, which may drop fork thread-comment and plan-policy text on handoff turns.
• The stage deploy timer's GitHub token (~/.config/t3-deploy/gh-token) returns 401 since 2026-10-31; stage did not auto-install feat(claude-accounts): pace to reset and credits run-rate in the account picker #364.

How I tested

No local build or test (dev-server-1 policy). CI on the PR head is the gate. The stage deploy is the runtime proof.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for 361b354.

This comment will update automatically after the next completed run.

[bk-agent-01]

Cross-model review record (reviewer: Codex gpt-6.1-sol, author: Claude Opus 5.5)

Round	Head	Findings	Outcome
1	d4a29badba	P1 scratch-thread promotion ungated; P2 older-settled counts leak; P2 short pages after post-LIMIT filter; P2 fork sidebar loses older settled sessions	Fixed in 7db6e01aae (promotion authorizes every moved thread; fork shell never trims settled threads; filtered users get no cross-user counts). Short pages: accepted residual, no fork client pages.
2	7db6e01aae	P2 tracked clones (projectCloneStart) ignore the source-control profile	Pre-existing on stage before this merge; deferred as a follow-up.
3	7c9832f0cf	P2 fork snooze countdown only updates at wake time	Fixed in 361b354094.
4	361b354094	none	Clean.

[bk-agent-01]

Merged into stage as c08d81302f, pinned to reviewed head 361b354094 (CI green on newest runs; cross-model review clean, record above). Requested by tushar in the T3 session. Installing on stagebkt3 next.