PR #27913 real dev verification, 2026-10-31, Linux Boat VM

Code under test
- baseline: github/main 15fe2832727893cbdd2214383f9f971474849695, separate /home/user/main worktree, vp i
- PR exact head: 1c67719e0c4fee83cd9115e884c24c4c82d8d446, perf/linked-task-parent-projection, vp i
- standalone linked peer: separate /home/user/peer worktree at baseline main, isolated .t3 state
- only one dev server at a time. main PID 45704 stopped before PR PID 66336 started.
- main server/web ports 16733/8693; PR 13773/5733; standalone peer 17774.

Real setup, no mocks/fixtures/demo data
- main project added through web UI's Add project > Local folder > /home/user/main.
- peer project added with the running app's `t3 project add /home/user/peer --base-dir .t3` CLI.
- enabled linked environments through main's Labs UI and peer's authenticated server.updateSettings WebSocket RPC.
- `t3 environment link http://127.0.0.1:17774 --base-dir .t3 --pairing-code [private] --access full-access` succeeded.
- Playwright from repository dependency, /usr/bin/google-chrome, real browser UI.
- composer selected Codex gpt-6.1-sol, low effort, full access.
- real parent and child provider turns used this VM's Codex proxy.
- remote child executed `timeout 45 bash -c 'sleep 20; git rev-parse --is-inside-work-tree; git status --porcelain'`: exit 0, output true, no changes.
- local child executed `timeout 20 git rev-parse --is-inside-work-tree`: exit 0, output true.
- provider replies: PEER_STATUS_OK and LOCAL_STATUS_OK.

Server proof from existing native trace spans, no source instrumentation
Tracing: T3CODE_TRACE_MIN_LEVEL=Trace, T3CODE_TRACE_FILE, T3CODE_TRACE_MAX_BYTES=104857600.
For each OrchestratorMcpService.readTask, traversed actual parentSpanId descendants and read db.query.text.
Counts below are SELECTs of payload_json from parent runs/context_transfers, not timing estimates.

MAIN
19:51:23.257 remote delegate_task initial running response, span 823dc65a43bb8717:
  4 SQL queries: thread, parent runs, selected subagent, parent context_transfers.
20:02:37.282 already-acknowledged remote task_status, span f9db7600e2b0e106:
  4 SQL queries; 2 unused parent history SELECTs.
20:02:37.637 repeated remote task_status, span f78848a9b42ec297:
  4 SQL queries; 2 unused parent history SELECTs.
20:02:37.698 completed local task_status, span 4f17e617f896e0aa:
  parent runs/context_transfers still read, plus child result reads.

PR, EXACT SAME PRESERVED TASKS AND DATA
20:06:26.565 remote task_status, span 0fe0f443fce0e5c8:
  2 SQL queries: canonical thread and selected subagent. ZERO runs/context_transfers SELECTs.
20:06:26.623 repeated remote task_status, span 56ce39a89774440b:
  2 SQL queries. ZERO runs/context_transfers SELECTs.
20:06:26.661 local task_status, span bc95d017c9af3bfe:
  parent runs/context_transfers still read, plus child result reads.
20:06:26.702 unknown task id:
  task_not_found, correct owning-thread error; no parent history SELECTs.

FULL RESPONSE PARITY
Node assert.deepEqual passed for the complete remote DTO, complete local DTO, and complete unknown-task error across main and PR.
- remote: completed, PEER_STATUS_OK, correct remoteChild, no pending question or approval.
- local: completed, LOCAL_STATUS_OK, same latestTerminalRunId and same non-null resultContextTransferId.
- error: OrchestratorMcpFailure / task_not_found, same message.

FRESH PR PROVIDER LIFECYCLE
20:07:08.184 remote delegate_task initial running, span 3b6214741c59b1be:
  2 SQL queries, ZERO parent history SELECTs.
20:07:40.789 first remote terminal task_status, span 65ea849e1eb959ee:
  completed / PEER_STATUS_OK, parent history read for terminal observer/acknowledgement, as intended.
20:07:40.809 repeated remote terminal task_status, span d9c0ebe3a6b219a6:
  completed / PEER_STATUS_OK, 2 SQL queries, ZERO parent history SELECTs.
20:07:40.832 local delegate_task returned running.
20:07:55.635 local task_status, span 0cd3536c13ac55cf:
  completed / LOCAL_STATUS_OK, non-null terminal run and result transfer, parent history retained.
Remote completion automatically woke the parent. Local completion automatically woke it too.

Existing focused checks
blacksmith-check run --id [task-owned worker] 'vp i && vp test run apps/server/src/mcp/OrchestratorMcpService.test.ts apps/server/src/peer/RemoteDelegation.test.ts'
Result: 2 test files passed, 63 tests passed (37 + 26). Worker stopped.
PR host status: 30 successful checks, 0 failing, 0 pending at checked head.
No source changes, commits or pushes were needed.

Limits
- Native desktop, iOS/Android, macOS/Windows, other providers, WAN/relay/tunnel were not exercised.
- This is two isolated real environments on loopback, not two physical machines.
- No 1,000-turn workload, whole-app latency, CPU or heap improvement is claimed.
- The provider's delegation contract forbids polling a running task. Running response proof is delegate_task's real readTask path; explicit task_status was exercised after completion, including first observation and already-acknowledged reads.
- Preserved main state includes migration 76, unknown to this older PR head; startup warned and skipped it. App/provider flows above succeeded. Fresh-install schema behavior was not separately checked.
- Initial cold Vite module compilation took several minutes before main UI became usable.