chore(v2): merge main into the V2 branch (48 commits)

[juliusmarminge]

Merges origin/main at a21b42cec4, 48 commits past the merge base a107f8a079, into the V2 branch. It keeps main's fixes, carries main's performance work over to V2's replacement code, and keeps everything V2 added.

Merge, not rebase. The request was a rebase. This PR is one merge commit plus follow-up port commits. A rebase would rewrite all 715 V2 commits, break every open stacked PR on t3code/codex-turn-mapping, and need a force-push to a shared branch. Nothing here needs a real rebase.

Commits:

• e42ccbddfa merge with the conflict resolutions
• 433ee8b640 port #13698
• ace70caecc port #13673
• af6754b1f1 port #13704
• 437b334c0c port #13691
• 2b9af8fe71 port #13689
• fe76961c94 port #11594

Each port commit covers one main PR, so each can be reverted on its own.

Refresh (second batch). The branch now also includes:

• V2 head 2c388dae9a, merged in d07c035047 and b049c24b13. Both merges were conflict-free.
• A second merge of origin/main at 95030dc674, 18 commits after a21b42c, in 8119814a66.
• 4 port commits for that batch: 1b10d0874c (#10158), 11acae1f2f (#13765), df06b8b97b (#13368), d45989af57 (#13339).

Pushes were fast-forward only. #13742 (mobile PR icon, one file) landed on main after the second merge and is not included; it belongs in the next merge.

Decisions

First batch (main a107f8a..a21b42c)

Legend: A = ACCEPT, P = PORT, S = SUPERSEDED, D = NEEDS-DISCUSSION. Paths are relative to apps/server/src/ unless they start with apps/ or packages/.

PR	Title	Decision	Reason / V2 evidence	Port commit
#11477	keep nested task states out of parent bullets	A	Web-only markdown clipboard fix; merged cleanly.
#13575	ship a Linux .deb that updates itself	A	Release and desktop updater; not touched by V2.
#13501	cache compiled JavaScript between launches	A	compileCache.ts is new. The flushCompileCache call conflicted with V2's imports in serverRuntimeStartup.ts and is kept.
#13589	one t3.json setup action for every OS	A	Dev tooling.
#13590	new worktree threads no longer say "checkout"	A + P in merge	Logic merged cleanly. V2's thread details panel still passed the old effectiveEnvModeOverride prop. The merge wires envMode through ChatView → ThreadDetailsPanel → BranchToolbar and applies main's label and icon rules to V2's env-mode selector markup.	in e42ccbddfa
#13585	t3 app survives a second desktop quit	A	Desktop only.
#13599	price Claude fast mode at the fast rate	A	usage/*; not touched by V2.
#13611	update OpenAI logo	A	Icons.
#12810	render project icons in mobile chat list	A	Mobile; merged cleanly.
#10409	read cursor, opencode, antigravity usage history	A + conflict fixes	See "V2 changes altered" below: Cursor keychain opt-in, a file-name case collision, usage presentation, and the usageSummary refresh trigger.
#10584	retry failed statement preparations	A	packages/shared/nodeSqliteClient; applies to V2's SQLite too.
#13356	Android controls scale with text size	A	Mobile. One conflict in ThreadComposer.tsx: V2's iconUrl moved onto main's renderIcon(size) API.
#13450	focus returns to composer after citation note	A	Web.
#13492	honor standard OTLP env vars	A	CLI config and shared.
#13673	settling a thread closes its idle shells	P	TerminalManager.closeIdle and the setup-script cleanup merged cleanly (shared code). The settle trigger lived in V1 ProviderCommandReactor, which V2 deleted, and V2's settle only detached the provider session. ThreadSettlementServiceV2 now handles thread.settled: it re-reads the thread and calls closeIdle only if the thread is still settled.	ace70caecc
#13678	load Cursor keyring with createRequire	A	Came in with the renamed provider/cursorKeychainToken.ts (see below).
#13395	avoid rereading unchanged review-preview files	A	vcs/GitVcsDriverCore; shared by V2.
#13641	OTLP vars in desktop main process and WSL	A	Desktop.
#13700	restore compact provider instance badges	A	The Sidebar conflict is resolved to V2's SidebarProviderStack. V2's one-off .provider-stack-badge-text { font-size: 7px } is replaced with main's new text-5xs token (same 7px), and the one-off class is removed.
#13702	lighter mobile Home rows off-screen	A	Mobile.
#13731	price Cursor cache savings by base model	A	Usage.
#13705	render mobile Home rows further ahead	A	Mobile.
#13699	name each service after its application	A	Also removed the now-deleted otlpServiceName field from V2 testkits (CodexAdapterV2.testkit.ts, CursorAdapterV2.testkit.ts, testkit/ProviderReplayHarness.ts, cli/project.test.ts).
#13720	stop remapping every thread on each thread event	A (V1 file) / S for V2	Applied to orchestration/projector.ts, which still backs V1 OrchestrationEngine: V2's runtimeLayer.ts:86 still wires it for ProjectService commands. V2's own read side has no in-memory list of all threads. ProjectionStoreV2.apply writes one row per event, and shell streams re-read only changed threads (orchestration-v2/ProjectionStore.ts getThreadShell). The client still has a cheaper version of this cost (memory notes below).
#13718	remove unused items tracking from Claude adapter	S	V1 provider/Layers/ClaudeAdapter.ts is deleted in V2. ClaudeAdapterV2 never collected per-turn SDK messages: readThreadSnapshot fails with "does not implement snapshots" (orchestration-v2/Adapters/ClaudeAdapterV2.ts:6832), and there is no turns[].items state.
#13701	name the command on subprocess spans	A	processRunner; shared.
#13685	t3 triage points at log files that exist	A	CLI.
#13684	SQLite WAL file shrinks after large writes	A	persistence/Layers/Sqlite.ts; applies to the V2 database. Sqlite.test.ts passes.
#13698	summarize the server trace file from the CLI	P	V2 moved the command list out of bin.ts into binCli.ts, so traceCommand was left unregistered (knip flagged it). Registered in binCli.ts:78.	433ee8b640
#13704	PR sync reads only threads with linked PRs	P	V2's PullRequestSyncReactor called OrchestratorV2.getShellSnapshot() every minute. That call builds every thread's shell, archived ones included, then everything unlinked was dropped. Added ProjectionStoreV2.getThreadsWithPullRequests (orchestration-v2/ProjectionStore.ts:5046): one SELECT with json_array_length(payload_json,'$.pullRequests') > 0, plus a matching in-memory store implementation. Main's listThreadsWithPullRequests on V1 ProjectionSnapshotQuery was not taken: it queries projection_thread_pull_requests, which V2 never writes (links live in the V2 thread payload).	af6754b1f1
#13694	server heap snapshot on SIGUSR2	A	HeapSnapshot.layer is wired into V2's server.ts. Useful for the V2 memory work below.
#13688	shutdown no longer rewrites every stopped session row	S	V1 ProviderService.ts is deleted in V2. ProviderSessionManagerV2.shutdown (orchestration-v2/ProviderSessionManager.ts:1515) iterates only live in-memory sessions; stopped rows are never touched.
#13693	thread list snapshot without decoding twice	A (V1 file) / S for V2	Applied to V1 ProjectionSnapshotQuery (still used for projects and some reads). V2's ProjectionStore.getShellSnapshot decodes each payload once and builds plain objects (shellFromState); there is no whole-snapshot re-decode.
#13683	slow servers load the thread list once	A	client-runtime HTTP snapshot timeouts; the V2 client uses these loaders.
#13686	hidden terminal drawers don't keep full history	S	V2's drawer already reads only the shell: const serverThread = useThreadShell(threadRef) (apps/web/src/components/ChatView.tsx:942). Main's active-drawer useThread fallback exists for archived V1 threads without a shell. V2 needs only the worktree path and project, so V2's side is kept.
#13691	per-thread settlement and PR checks don't rebuild the list	P	Both V2 services read every thread for a one-thread request. getSettlementCandidates(threadId?) now filters in SQL (ProjectionStore.ts:4944), and a sweep with no candidates stops early. ThreadPullRequestServiceV2 reads getThreadEventSequence then getThreadShell for one thread (ThreadPullRequestService.ts:126), and a one-thread read only clears that thread's pending backfill (as on main).	437b334c0c
#13530	running threads open at the latest message	A	Mobile.
#13697	record event loop stalls in the trace	A	EventLoopMonitor.layer is wired into V2's ApplicationObservabilityLive.
#13689	stop re-running git for every project each minute	A + P	The 15-minute identity cache TTL and the publish-repository refresh in ws.ts merged cleanly. V2's ThreadPullRequestService resolved identities without refresh, so after a turn it could miss a newly added remote for up to 15 minutes. Discovery requests already marked refresh: true now pass it through to the resolver.	2b9af8fe71
#13714	hide the Cursor keychain prompt when Cursor isn't set up	A	Web, mobile, and client-runtime usage state.
#11594	chat width setting for wide screens	A + P	The settings row, contract field, root data-chat-width, and minimap gutter logic merged; the minimap tests are ported onto V2's test. V2 sizes the lane with --chat-content-max-width rather than main's --chat-max-width, so the setting did nothing until the port. Wide and Full now override V2's token. D: V2's comfortable width stays 46rem, not main's 48rem.	fe76961c94
#13651	accept opencode v2 serve ready line	A	provider/opencodeRuntime.ts is shared with OpenCodeAdapterV2.
#7079	stop treating agy CLI as Antigravity IDE	A	Editor contract; editors are not stored, so decoding is safe.
#13713	empty workspace draggable on desktop	A	Web.
#13669	editors don't vanish when discovery is slow	A	Shared.
#12537	exclude SSH ports from provider URLs	A	Shared.
#13409	Mod+B bolds on non-Latin layouts	A	Web.
#13695	prune expired replay-protection files	A	ReplayMarkers.layer is added to V2's RuntimeCoreDependenciesBaseLive (server.ts:525) because main's ReactorLayerLive chain doesn't exist in V2.

Second batch (main a21b42c..95030dc)

PR	Title	Decision	Reason / V2 evidence	Port commit
#13408	terminal links drop a trailing colon	A	Web; merged cleanly.
#13748	bump node-pty to 1.2.0-beta.15	A	The lockfile was regenerated with vp i from main's lockfile, keeping V2's deps (@cursor/sdk, @napi-rs/keyring).
#13344	focus ring on sidebar thread rows	A	Merged into V2's Sidebar.tsx.
#13310	mobile composer stays within folded screen	A	Mobile.
#13368	let OpenCode generate session titles	P	V1 ProviderCommandReactor is deleted. V2's OpenCodeAdapterV2 sent a placeholder title: "T3 Code <threadId>" on every session.create, which stops OpenCode from generating a title. It now sends no title. D: main also forwards a hand-renamed title; V2 has no title-provenance field and never sent the thread title, so that part is left out.	df06b8b97b
#13339	Antigravity inspects unsupported files by path	S (code) + P (docs)	V1 AntigravityAdapter is deleted. The shared buildAntigravityPrompt merged, but V2 has no production caller for it. V2's Antigravity runs through AcpAdapterV2, which embeds only images and passes every other file as a saved path (AttachmentPrompt.providerMessageTextWithAttachmentPaths). Reads are confined by AntigravityClientFiles (#13613). Main's doc text described V1's native PDF/text/audio limits, so providers-antigravity.md now describes V2.	d45989af57
#13795	link URLs with ports and single-label hosts	A	Mobile markdown patch.
#10158	keyboard navigation for Usage	A + P	Merged; usagePageOpen added to V2's keybindings doc list. Main left usage.open unranked in the settings sort, which made the comparator inconsistent. With V2's extra default bindings, main's own test failed. usage.open now ranks first, and the test checks both input orders.	1b10d0874c
#13756	stop writing empty spans	A (shared) / S for V2	The processRunner, shell, GitVcsDriverCore and PortScanner changes merged. Main's skip of the empty attachment-cleanup span is kept in V1 ProjectionPipeline. V2's ProjectionStoreV2.apply is a plain untraced Effect.gen, so it writes no per-event span.
#13763	Diagnostics no longer loads the whole trace ring	A	diagnostics/TraceDiagnostics and cli/trace are shared with V2.
#13759	sort without re-parsing dates per comparison	A	The shared sortSettledThreads (client-runtime) and precomputed project sort apply to V2. V2's duplicate sortSettledThreadsForSidebar is dropped for main's shared one (same resolver). Its tests now live in client-runtime/threadSort.test.ts, converted to V2's latestRun.
#13761	renderer trace proxy stops tracing itself	A	untracedRequestsLayer is added last in V2's makeRoutesLayer, after V2's MCP wiring.
#13765	background sweeps only read threads that can still settle	P / S for settlement	Settlement is superseded: getSettlementCandidates already filters settled, pinned, archived and busy threads in SQL (ProjectionStore.ts, #13691 port). PR discovery was not: its minutely sweep read every thread, active and archived. getShellSnapshot gains unsettledOnly (SQL and memory store), and discovery reads active, unsettled threads unless backfill is pending, as on main.	11acae1f2f
#13767	saving the thread list cache no longer freezes the UI	S / D	Main skips Schema encoding for thread shells because V1 shells are already in encoded form. V2 shells hold DateTime values, so that shortcut would write the wrong format. V2 encodes through its V2 cache schema, and mobile already yields during large saves (apps/mobile/src/connection/shell-cache-encoding.ts, #12117). Main's IndexedDB abort listener fix is kept. D: the V2 web encoder still does not yield (see memory notes).
#13774	idle wakeups from Connect relay and session reaper	A (relay, in merge) / S (reaper)	Relay: V2 keeps its publish worker and drain, and adds main's unlinked backoff (5s → 60s) and requestCatchUp, which cloud/http now calls. Main's three catch-up tests are rewritten against V2's relay harness. Reaper: V1 ProviderSessionReaper is deleted, and V2 has no polling reaper; idle release is one timer per live session (ProviderSessionManager.ts scheduleIdleReleaseInternal). The repository's new excludeStopped option merged; V2's only list() caller (AgentSessionImporter) needs stopped rows.	in 8119814a66
#13807	trailing underscores and tildes in URLs	A	Mobile patch.
#13764	queued messages send while their thread isn't open	S	V2 deleted the client-side queue (queuedMessageStore, f7a764c8ee). Queued messages are server runs, and Orchestrator.startNextQueuedRun starts the next queued run when a run ends, whether or not any client has the thread open. resumeQueuedRuns also recovers them after restart. Main's QueuedMessageSender, sendQueuedMessage and root mount are not taken.
#13812	background git fetches no longer fill the disk	A	vcs/GitVcsDriverCore is shared with V2.

Migrations

Second batch: main still has no new migration (latest 054), and its only contract change is keybindings.ts (new commands and usagePageOpen, all additive). Both upgrade paths are unchanged, and the migration and import tests still pass.

At merge time, and rechecked against origin/main at 2679d279ce: main has no new migration files. Its latest is still 054_ProjectionThreadsAutoSettleDisabledAt, and V2 adds 055_OrchestrationV2 and 056_RemoveRedundantProjectionIndexes, so numbers don't collide. No migration was added here.

1. V1 users (at 054) upgrading to V2. 055 and 056 run in order. Main changed no V1 projection table or column, so LegacyV1ThreadImporter and LegacyV1Cutover see the same schema they were written against. Main's Sqlite.ts change is a WAL checkpoint after large writes, which only helps the import. The 055/056 migration tests and the LegacyV1Cutover/Importer tests pass.
2. V2 users (at 056) upgrading. Nothing re-runs or reorders. The new getThreadsWithPullRequests and the getSettlementCandidates filter only read existing V2 columns (payload_json, deleted_at, updated_at) and need no index change.
3. Contracts decode. ClientSettings.chatWidth and ServerSettings.cursorKeychainUsageEnabled both have decoding defaults, so stored V2 settings still decode. usage.ts bumps USAGE_CONTRACT_VERSION to 6 with added provider literals and optional fields. Usage is computed per request, not stored. The editor.ts change edits a static list only.

V2 changes altered during conflict resolution

• bin.ts: kept V2's thin entry. Main's command list edit went into binCli.ts (port commit).
• server.ts: kept V2's layer graph and added ReplayMarkers.layer. The PR sync worker now also gets ProjectionStoreV2.layer (from the perf(server): pull request sync reads only threads with linked pull requests #13704 port).
• V1 ProjectionSnapshotQuery (still in V2): dropped the whole-snapshot decode (perf(server): build the thread list snapshot without decoding it twice #13693). Did not add listThreadsWithPullRequests.
• CursorDriver.ts / cursorUsageLimits.ts: kept V2's SDK-based driver and passed main's cursorKeychainUsageEnabled through. Main's provider/cursorCredentialStore.ts differs only by case from V2's provider/CursorCredentialStore.ts: tsc fails with TS1149, and it would break on macOS. It is renamed to provider/cursorKeychainToken.ts, with its test and imports updated. Main's Cursor usage tests moved from the V1-shaped CursorProvider.test.ts into V2's cursorUsageLimits.test.ts, with labels updated to main's "Overall / Cursor Models / Other Models".
• apps/web/.../usage/usageProviders.ts: kept V2's driverKind-based presentation instead of main's mark icons, and added cursor, opencode and antigravity in V2's shape.
• packages/client-runtime/.../server.ts: kept V2's searchAcpRegistry. The single usageSummary atom now uses main's usageScanSettingsAtom trigger.
• MessagesTimeline*, ChatComposer, ComposerSurface, ChatView: kept V2's chat-content-lane markup and dropped main's TimelineLoadEarlierHeader, which V2 already removed. The drawer keeps V2's shell-only read.
• SettingsPanels.tsx: both rows kept (V2's composer-context row and main's chat-width row).
• scripts/lib/cli-external-packages.test.ts: the expected list now has both @cursor/sdk (V2) and @napi-rs/keyring (main).
• Files V2 deleted and main modified: all kept deleted. The main-side changes are ported where they matter (above). These are the V1 ClaudeAdapter, ProviderService, ProviderCommandReactor, ThreadSettlementReactor, ThreadPullRequestReactor, and OrchestrationEngine/ProjectionSnapshotQuery/ProviderSessionReaper tests, plus bin.test.ts, server.test.ts, and integration/OrchestrationEngineHarness.
• V2 test fixtures that mocked V1 ProjectionSnapshotQuery kept V2's versions. The one auto-merged mock that gained listThreadsWithPullRequests (AgentSessionScanner.test.ts) had it removed.

Second batch

• AgentAwarenessRelay: the service has both drain (V2) and requestCatchUp (main). cloud/http.test stubs drain.
• server.ts: untracedRequestsLayer goes after V2's MCP route layer.
• Sidebar.logic.ts: V2's sortSettledThreadsForSidebar is removed in favor of main's shared sortSettledThreads.
• KeybindingsSettings.logic.ts: usage.open is ranked (port commit).
• OpenCodeAdapterV2: no placeholder session title (port commit).
• Docs: composer.md keeps V2's server-queue text; main's Antigravity video line merged and is true for V2. providers-antigravity.md is rewritten for V2's path-based attachments.
• Not taken from main: QueuedMessageSender, sendQueuedMessage, queuedMessageStore, encodeShellSnapshotForCache and its test, and the unsettledOnly plumbing on V1 ProjectionSnapshotQuery (V2 has no caller for it).
• Cursor rename still needed. Main still ships lowercase provider/cursorCredentialStore.ts, and batch 2 did not touch it, so V2 keeps it as cursorKeychainToken.ts.

V2 memory and CPU hotspots worth a follow-up (not fixed here)

1. V1 engine still resident. orchestration-v2/runtimeLayer.ts:86,305 wires OrchestrationLayerLive, whose OrchestrationEngine hydrates a full in-memory V1 read model at boot (orchestration/Layers/OrchestrationEngine.ts:344-345) and keeps it for the process lifetime. The only user is ProjectService's V1 commands. This is probably the largest fixed memory cost in V2.

2. Unbounded event buffers. EventSink.ts:200 (live PubSub) and :626 (one unbounded PubSub per event type, never removed). ProviderSessionManager.ts:1219 has one unbounded queue per session subscriber. A slow consumer, such as a stalled streamDomainEvents subscriber, buffers without limit.

3. Adapter maps that grow per item and are cleared only on reset.

   • ClaudeAdapterV2.ts:2788 itemOrdinals is copied (new Map(current)) on every new item and never pruned. That is O(items) memory and O(items²) copying over a long session.
   • The same pattern exists in CursorAdapterV2, AcpAdapterV2, and CodexAdapterV2.ts:1639 planDeltas, which keeps the full plan text per item.

4. Shell stream per streaming delta. Every V2 event, including turn-item deltas, becomes a shell event. Each busy thread then costs one heavy getThreadShell query per 50ms window per client. On the client, each thread.updated maps the whole thread array (client-runtime/state/shellReducer.ts:7) and rebuilds the thread index (threadShell.ts:71). This is the client-side counterpart of perf(server): stop remapping every thread on each thread event #13720.

5. Remaining full-shell reads on background paths. relay/AgentAwarenessRelay.ts:689, CheckpointRestoreSafety.ts:28 (reads the full shell, then checkpoint context for every other thread), ThreadManagementService listProjectThreads, and the minutely full sweeps in both settlement and PR discovery.

6. Unbounded, non-deduplicating worker queues. DrainableWorker does not merge repeated requests, so a burst of run.updated events queues one sweep per event even after perf(server): per-thread settlement and PR checks no longer rebuild the whole thread list #13691.

7. Small leaks.

   • assistantStreaming.ts:93: the per-message buffer is removed only on a final update, so interrupted streams leave it behind.
   • CheckpointService.ts:252: one semaphore per working directory, never removed.

8. Web shell cache encode (perf(clients): saving the thread list cache no longer freezes the UI #13767 counterpart). apps/web/src/connection/storage.ts encodes the whole V2 shell snapshot through Schema in one go on every save. Mobile yields every 32 rows; the web path doesn't, so a large thread list can still stall a web frame on save.

Please check these (riskiest calls)

1. Merge instead of rebase (see the top of this description).
2. fix(terminal): settling a thread closes its idle shells #13673 port location. Idle-shell cleanup lives in ThreadSettlementServiceV2 as a domain-event reaction, not as an effect in the V2 outbox. It is not replayed after a crash, which matches main's reactor. An outbox effect would need a new effect type and more machinery.
3. Cursor file rename. Main's cursorCredentialStore.ts is now cursorKeychainToken.ts on this branch. The next main merge will conflict on it again unless main renames it too.
4. perf(server): pull request sync reads only threads with linked pull requests #13704 port adds a projection-store method instead of reusing main's V1 query. Main's query reads a table V2 no longer writes.
5. Comfortable chat width stays at V2's 46rem instead of main's 48rem.
6. Terminal drawer keeps V2's shell-only read. Archived V2 threads get no drawer thread data, which V2 already accepted before this merge.

Second batch:
7. #13764 superseded. V2 drains queued runs on the server, so a thread nobody has open still sends. I confirmed this in code, not in a live run.
8. #13368 drops the placeholder title but does not forward hand-renamed titles.
9. #13767 not ported. The V2 web cache encode is untouched (memory note 8).
10. #13765 port. The PR discovery sweep now reads unsettled-only unless backfill is pending. A settled thread whose PR changes waits for backfill or a per-thread event, as on main.

Verification

Second batch, on head b049c24:

• tsc --noEmit: server, contracts, client-runtime, effect-acp, web, mobile all report 0 error TS and 0 warning TS.
• vp run knip:check: passes.
• Replay suite plus 055/056 migration and LegacyV1Cutover/Importer tests: 7 files, 105 tests, all pass.
• Port tests: I reverted each production change and confirmed its new test fails.
  • KeybindingsSettings.logic.test.ts (Usage order)
  • ProjectionSettlement.test.ts (unsettled-only SQL/memory parity)
  • ThreadPullRequestService.test.ts (periodic sweep is unsettled-only after backfill)
  • OpenCodeAdapterV2.test.ts (no session title)
• Merged areas: 16 server files (345 tests: relay, cloud http, projection pipeline, trace diagnostics, git, port scanner, Antigravity support, Cursor), 17 web files (588 tests), 2 client-runtime files (41 tests), all pass.

First batch:

• tsc --noEmit: server, web, mobile, client-runtime, contracts, effect-acp all report 0 error TS and 0 warning TS.
• vp run knip:check: passes.
• Migrations and import: 055_OrchestrationV2*.test.ts (3 files), 056_RemoveRedundantProjectionIndexes.test.ts, LegacyV1Cutover.integration.test.ts, LegacyV1ThreadImporter.test.ts, Sqlite.test.ts all pass.
• Replay: OrchestratorReplayFixtures.integration.test.ts, 93/93 pass.
• Ports:
  • ThreadSettlementService.test.ts (new settle-terminal and single-thread tests), ThreadPullRequestService.test.ts (new one-thread-read and refresh tests), PullRequestSyncReactor.test.ts (asserts no shell read), ProjectionSettlement.test.ts (new SQL/memory parity test for getThreadsWithPullRequests), cli/trace.test.ts all pass.
  • I reverted each new test's production change and confirmed the test fails, then restored it.
• Merged areas:
  • Server: cursor, usage, observability, terminal, VCS, CLI, auth, project, and runtime-startup test files, 34 files, all pass.
  • Web: 15 files (MessagesTimeline, BranchToolbar, ThreadDetailsPanel, settings, usage, keybindings, markdown-clipboard), 436 tests, all pass.
  • Other: client-runtime usage, contracts settings, and 8 shared test files all pass.
• scripts/build-desktop-artifact.test.ts has 15 failures, and they are pre-existing. The same 15 fail with V2 HEAD's versions of the file on this machine (asar unpack under the custom TMPDIR).
• Not run: repo-wide suites, the desktop build, and any browser, device, or live-provider run.

Model: Claude Opus 5.5 (Claude Code)

🤖 Generated with Claude Code

---

[macroscopeapp]

Tailwind ownership: this Cursor swatch uses a static inline color. Please put bg-[#8b8b8b] on the owning element instead.

Suggested change

	className="size-2 shrink-0 rounded-full"
	className="size-2 shrink-0 rounded-full bg-[#8b8b8b]"

^{Posted via Macroscope — UI Consistency}