chore(v2): merge main into the V2 branch

[juliusmarminge]

Merges main (148e6de, 11 commits) into the V2 branch, so V2 stays current with main. A merge commit that resolves the conflicts and adapts two clean merges that were wrong for V2; nothing else.

A refactor PR is stacked on this one: it moves #14527 ("start a new project from just a name") into a service for the project folders T3 creates under its data dir, merged with ScratchWorkspace. Both merge together.

What comes in

• feat: start a new project from just a name #14527 new project from a name: contracts, RpcAuthorization, web, mobile, client-runtime, docs, and project/NewProject.ts (as-is; nothing calls it yet).
• feat(web): restart the agent session from cmd+k to load new skills and plugins #14542 restart the agent session from cmd+k: ProviderRegistry fresh workspace scans, AntigravityProvider rescans, the command palette action, docs.
• fix(server): Claude subagents with their own model no longer show the parent's model #14540 Claude subagent model: V1 adapter only, kept deleted (see below).
• CI and release perf: perf(ci): run PR checks side by side and balance server shards #14025, perf(release): start Windows builds without waiting for the Linux job #14027, perf(release): build and publish npm platform packages concurrently #14028, perf(release): build Vercel deployments early and go live after publish #14029, perf(release): trim Windows packaging setup #14037, fix(release): Windows CLI smoke test no longer fails on temp dir cleanup #14553.
• Web: fix(web): make composer suggestions usable with screen readers #10154 (composer suggestions a11y), fix(web): hotkey settings record plain keys and Tab #14548 (hotkey recorder).

Conflicts and how they were resolved

• provider/Layers/ClaudeAdapter(.test).ts, server.test.ts (V1, deleted on V2): kept deleted.
• ws.ts: V2's side everywhere. projects.createNew is a stub that fails with "Starting a new project from a name is not available yet.", and the config answer omits newProjectsRoot, so web and mobile hide the feature. The stacked PR implements both.
• contracts/rpc.ts: both import sets (V2's orchestrationV2, main's ProjectCreateNew*).
• ProviderRegistry.test.ts: V2's orchestrationAdapter plus feat(web): restart the agent session from cmd+k to load new skills and plugins #14542's invalidateCaches counter.
• knip.jsonc: V2's script entries plus update-test-shard-weights.ts.

Adapted after merging cleanly:

• ci.yml: git put V2's "Install C toolchain for process-tree fixtures" step into perf(ci): run PR checks side by side and balance server shards #14025's new Typecheck job. Moved it back to Test, which runs the ACP process-tree tests that compile with cc.
• CommandPalette.tsx: feat(web): restart the agent session from cmd+k to load new skills and plugins #14542 read thread.session, which V2's thread shell lacks. It now uses thread.runtime, like V2's other stopSession callers.

Follow-up: #14540 is only partly covered by V2

I fed #14540's test frames through ClaudeAdapterV2 in a throwaway probe (not committed).

• Covered: a subagent's snapshot model is pushed at once as subagent.updated (ClaudeAdapterV2.ts:5516-5535).
• Not covered: a task_started whose tool_use_id appears only in a subagent's assistant snapshot (main's toolu_nested) gets model: null, and its child thread takes the parent's model. Its parentNodeId is the root turn node, not the owning subagent. V2 records requested models only for Agent tool_use blocks (:5711) and has no owner map for snapshot-only tool calls.

The fix is being ported to ClaudeAdapterV2 as its own PR.

Sharding check (#14025)

• ci.yml now runs the server in 6 shards on 4-vCPU runners, ordered by WeightedShardSequencer.
• shardWeights.json lists main's files, 15 of which are V1 tests V2 deleted. The sequencer only looks up weights for files it is given, so stale names are ignored and V2-only files count as fast.
• I ran the sequencer over V2's 467 server test files for 6 shards: every file is in exactly one shard (53/84/84/81/82/83 files). transferBudgetV2.integration.test.ts lands in one shard, so the single transfer-report artifact still works. Re-recording the weights on V2 would balance the shards better; that is optional.
• Release workflows, scripts/build-npm-platform-packages.ts, smoke-cli-archive.ts and cli.ts are unchanged on V2 since the last sync, so they are main's versions verbatim, and the preview path (release_channel == 'preview') is untouched.

Verification

• vp i, then tsc --noEmit exits 0 for apps/server, apps/web, apps/mobile, packages/contracts, packages/client-runtime and packages/shared.
• Server tests, run in unshare -U --map-current-user -p -f --mount-proc:
  • ProviderRegistry.test.ts, AntigravityProvider.test.ts, NewProject.test.ts, RpcAuthorization.test.ts, weightedShardSequencer.test.ts and scripts/publishOrder.test.ts: 92 passed.
  • OrchestratorReplayFixtures.integration.test.ts, ClaudeReplayFixtures.integration.test.ts and ws.test.ts: 115 passed.
• Web: KeybindingsSettings.logic.test.ts, ComposerCommandMenu.test.tsx and composerMenuHighlight.test.ts: 34 passed. Shared: path.test.ts: 6 passed.
• vp lint on the conflicted and adapted files shows no errors. The CommandPalette.tsx warnings were already on V2.
• vp run knip:check passes.
• Not run: whole shards or repo-wide checks (CI runs them).

Model: Claude Opus 5.5 (Claude Code)

🤖 Generated with Claude Code

---

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.1 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	20.4 KiB	—	29.3 KiB	✅
Codex	Live turn messages	—	1	—	8	✅
Claude	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	20.8 KiB	—	29.3 KiB	✅
Claude	Live turn messages	—	2	—	8	✅

Baseline: unavailable · PR result: bb607b0 · 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: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a broad 53-file sync containing new project and session capabilities, provider behavior changes, npm publication changes, and production deployment workflow changes. It also modifies authorization code and adds static-analysis suppression directives, creating a review surface that requires human assessment.

^{You can add or adjust custom eligibility rules. Learn more.}