feat: subagent and tool presentation fixes (upstream G3)

[rynfar]

Ports upstream group G3: subagent and tool presentation. Claude skill calls show the skill name. Claude subagent threads show the model their agent file picks. Subagent descriptions are formatted as display names. Mobile shows full subagent details. OpenCode reasoning labels are no longer blank.

Sources

Upstream SHA	Title	Outcome
43bd667739	fix(web): Claude skill calls show the skill name (#14980)	Adapted. Conflict in ClaudeAdapterV2.ts: Pylon's tool-call projection has no input.presentation title. The title is now searchTitle ?? dynamicToolTitle(toolName, input) ?? null. computerUseToolTitle was renamed to dynamicToolTitle. No callers of the old name remain.
4f7760e6a0	fix(server): Claude subagents show the model their agent file picks (#14994)	Adopted. Clean cherry-pick. ProviderEventIngestor now depends on ThreadCommandExecutor. When a provider_native subagent reports a different model, it writes thread.model-selection-updated to the child thread under the thread lock. The layer wiring changed in runtimeLayer.ts, ProviderReplayHarness.ts, and the test layer in ProviderSessionManager.test.ts (wiring only).
151c241cd2	fix(client-runtime): format subagent descriptions as display names (#14999)	Adopted. Clean cherry-pick.
65731f986b	fix(mobile): show complete subagent details (#15189)	Adopted. Clean cherry-pick. Adds the new SubagentRow.tsx and the shared resolveSubagentMetadata / subagentDetailPreview helpers.
63f74e209d	fix(opencode): fill blank reasoning labels (#14978)	Adapted. Conflicts resolved in favor of Pylon in these places: the mobile useLiveDeviceState hook, the GlassBlurTargetContext composer wrapper with contextWindow/localOutboxCount, the web liveHandoffProjection and the supportsProviderSwitchingViaHandoff prop, and Pylon's model-change disabled-copy tests. Upstream's reportedModelSelection was added next to each of them. Pylon's mobile settings sheet has no providerInstanceId prop, so that prop was dropped from the memo deps. Adds the optional field OrchestrationV2ProviderThreadNativeMetadata.modelSelection. It is display-only and old clients ignore it.

No sources were skipped.

Pylon impact

• Prime: Prime child/subagent items come from LegacyRuntimeEventProjector with childThreadId: null. The new subagent model sync therefore never fires for Prime. formatSubagentDisplayTitle only rewrites Codex /root/... paths, so Prime labels display as before. The Prime and legacy projector tests pass.
• Providers: The variant label handling only affects OpenCode, the only provider that uses a variant option. Reported values are display-only and are never sent with a dispatch.
• Contracts: Adds one optional field. No migration.
• Scope: No changes to RunExecutionService, ProviderSessionManager, EventSink, or EffectOutbox. The only overlap is the test layer wiring in ProviderSessionManager.test.ts, which is needed because the ingestor's dependencies changed.

Review adaptations (96b335a7a9)

• Shared model labels: An OpenCode variant the user never chose and the provider never reported now shows "Default". Upstream showed "Unknown", which would have made existing drafts show "Unknown" instead of a level such as "Medium". "Unknown" now appears only when an explicit choice can't be resolved. The shared, client-runtime and web tests were updated.
• ProviderEventIngestor: The child-thread model sync is display-only. If it fails, it now logs Effect.logWarning and keeps the original stored events, instead of raising ProviderEventPublishError. A new regression test makes the thread.model-selection-updated write fail and checks that the stream continues. The test was confirmed to fail without the fix.
• Mobile: The subagent row Pressable gets back its accessibilityLabel (name and status).

Rebase onto #1002 / #997 / #998 / #1000 / #1001 (dbef955759)

• LegacyAdapterV2SessionManager.test.ts (added by feat(server): orchestrator v2 core fixes (upstream G1) #1002) now provides ThreadCommandExecutor to the ingestor layer, the same way runtimeLayer and ProviderReplayHarness do. Without it, the t3 typecheck failed with missingEffectContext.
• The subagent display-name test in threadExecution.test.ts now expects the waiting flag that fix: thread state and timeline fixes (upstream G4) #997 added.
• The rebase onto ProviderSessionManager.test.ts applied without conflicts.
• After the rebase: typecheck passed for contracts, shared, client-runtime, t3, web and mobile. The 13 server suites (including LegacyAdapterV2SessionManager, ProviderEventIngestor and ProviderSessionManager) passed 489 tests. The client suites passed: shared 41, client-runtime 52, web 156, mobile 7. Format and lint on the 39 changed files: clean.

Verification

• Typecheck passed: vp run -F @t3tools/contracts, @t3tools/shared, @t3tools/client-runtime, t3, @t3tools/web, @t3tools/mobile. Each one was confirmed to run tsc --noEmit.
• Shared tests: vp test run on model.test.ts and toolActivity.test.ts, 41 passed.
• Client-runtime tests: subagentDisplay.test.ts and threadExecution.test.ts, 50 passed.
• Web tests: TraitsPicker.test.ts and MessagesTimeline.logic.test.ts, 155 passed.
• Mobile tests: threadAgentsPresentation.test.ts, 7 passed.
• Server adapter and ingestor tests: ClaudeAdapterV2, OpenCode2AdapterV2, CodexAdapterV2, ProviderEventIngestor (includes the subagent model-sync test and its failure-path test), and ProviderSessionManager, all passed.
• After the review adaptations: model.test.ts 32, threadExecution.test.ts 39, TraitsPicker.test.ts 15, ProviderEventIngestor.test.ts 16, and threadAgentsPresentation.test.ts 7, all passed. Typechecks for shared, client-runtime, t3, web and mobile were re-run and passed.
• Server integration tests: OpenCode2OrchestratorV2.integration, ThreadFork.execution, SteeringCompletion.integration, LegacyAdapterV2Bridge, LegacyRuntimeEventProjector, PrimeAgentDaemonRuntimeEvents, and OrchestratorMcpToolkit.integration, 96 passed.
• vp fmt --check and vp lint on the 38 changed TS/TSX files: clean. Lint shows only warnings that were already in ChatView.tsx.

Not verified

• No browser, simulator, or device check was run, so there are no before/after screenshots. That covers the web tool row skill title, the subagent bar model label, the mobile SubagentRow / agents sheet, and the OpenCode traits picker labels.
• No live Claude agent-file or OpenCode session was tested. Coverage comes from the adapter and ingestor tests only.

Part of upstream cycle #996.

🤖 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	4.9 KiB	4.9 KiB	−40 B (−0.8%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.7 KiB	3.7 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.1 KiB	−40 B (−3.3%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.4 KiB	20.4 KiB	−41 B (−0.2%)	29.3 KiB	✅
Codex	Live turn messages	2	1	−1 (−50.0%)	8	✅
Claude	Total thread wire	4.9 KiB	4.9 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.7 KiB	3.7 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	20.8 KiB	20.8 KiB	0 B (0.0%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: cda6633 · PR result: dbef955 · 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.