[andrewcai8]

Brings the fork up to upstream 291bcc392c (196 commits since 337c5858c3, upstream CI green on that sha). One new migration: upstream 080 ScheduledTaskThreadMode.

Notable upstream changes

• Server perf: each live event is projected once for all subscribers (perf(server): each live event is projected once for all subscribers pingdotgg/t3code#30860), batched projection recovery reads (perf(server): batch projection recovery reads into one query pingdotgg/t3code#31842), batched shutdown recovery commits (perf(server): batch shutdown recovery commits pingdotgg/t3code#31854), queued provider writes no longer wait (perf(server): batch queued provider writes without waiting pingdotgg/t3code#31656), the headless browser closes after five idle minutes (perf(server): the headless browser closes after five idle minutes (517 MiB → 1.6 MiB median) pingdotgg/t3code#30306).
• Client perf: a big server config on a slow link no longer drops the connection in a loop (fix(client): a big server config on a slow link no longer drops the connection in a loop pingdotgg/t3code#23758), PRs paint from data already held (perf(pr): pull requests paint from data already held, on web, desktop and mobile pingdotgg/t3code#27798), faster mobile sync and reconnects on bad wifi (perf(mobile): faster thread sync, connects, and reconnects on bad wifi pingdotgg/t3code#28487).
• Scheduled tasks can run in one thread with a fresh session each run (feat: scheduled tasks can run in one thread with a fresh session each run pingdotgg/t3code#30326) and link to their thread (feat(scheduled-tasks): tasks link to their thread pingdotgg/t3code#31764).
• Dictation on a chosen environment (feat(web): opt into environment dictation on this device pingdotgg/t3code#30623–feat(server): transcribe audio on a chosen environment pingdotgg/t3code#30625, feat(mobile): opt in to dictation on a paired environment pingdotgg/t3code#31059), secret-provider references for secret requests and more provider credentials (feat(settings): Simlock, Antigravity and OpenCode credentials can come from a secret provider pingdotgg/t3code#31652, feat(secrets): answer a secret request with a secret provider reference pingdotgg/t3code#31670), GitLab review features, PR comment insertion, /side on mobile.

Conflict resolutions (26 files)

• Scheduled tasks (contracts, server, web, mobile): the contract keeps the fork's target and upstream's threadMode. A cloud-target task can only use new_thread: its chat lives on the cloud machine, so this server has no thread to keep between runs. The server refuses cloud + continue/fresh_session (or a bound thread) with the existing "must start a new chat" message, and the web and mobile editors hide those combinations (scheduledTaskDestinationRules on web, cloudTargetAllowed/threadModeForTarget on mobile, each with a focused test).
• ProjectionStore: upstream's single batched recovery read (perf(server): batch projection recovery reads into one query pingdotgg/t3code#31842) wraps the fork's index-hinted live-row subqueries (recoveryTurnItems, INDEXED BY).
• ProviderInstanceRegistryHydration: upstream's initialSettings watcher is passed from the fork's layerWithDrivers.
• ws.ts / bin.ts: upstream's withOutputWindow rename and __speech-worker command, alongside the fork's Home/Fleet services and fork-exec.
• client-runtime shell: upstream's authoritativeOmission key replaces the per-session ref; the fork's followHostChat takes it.
• Usage-limit recovery (web banner, mobile card): upstream's new action layout with the fork's "Switch account" action re-added.
• Mobile: upstream moved the workspace sidebar to layout/workspace/; the fork's useNewThreadNavigation and cloud machine status stay wired, and the server-controls screen keeps the fork's Approvals section.
• package.json/knip.jsonc: upstream's split knip passes plus the fork's case:check and ignored deps; lockfile regenerated from upstream's.

Merge-caused fixes outside conflicts: ForkMigrations.test.ts pins upstream 080; ws.test.ts builds upstream's new push test from CoreWsRpcGroup; mobile Stack.tsx and two import paths follow upstream's moves; shell-sync.cloud.test.ts fixture gets environment.serverVersion; two Storage settings descriptions shortened for upstream's 200-character rule.

agentContextBudget.json rises (MCP tool list 54109 → 54730 bytes, from upstream's new tools), so this needs agent-context-increase.

Verification

• vp run typecheck: contracts, client-runtime, server, web, mobile, shared, desktop all clean.
• Tests: server scheduled tasks, projection store/recovery, provider hydration, ForkMigrations, AgentContextBudget, ws, OrchestratorMcpService (291 pass); client-runtime full suite (1999 pass); mobile unit + react suites (3387 pass); targeted web logic tests.
• vp lint on all 690 fork-changed TS files, vp run knip:check.

Not verified locally

• apps/web and apps/desktop full suites and browser tests (left to CI; they overload the Mac).
• SQLite query plans for the batched recovery read on a large database.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	21.5 KiB	21.5 KiB	0 B (0.0%)	24.4 KiB	✅
Codex	Live turn messages	1	1	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.9 KiB	21.9 KiB	0 B (0.0%)	24.4 KiB	✅
Claude	Live turn messages	1	1	0 (0.0%)	8	✅

Baseline: e911faa · PR result: 995541e · 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: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

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