[GPT-6.1 Sol][Codex] on behalf of shivam.
Verification report for head 08167260879bcbd57ef2bbcd30c16dcc96219218, based on 05c0ed7da69ee6aceab8fad84e8a0dc0c474138a.
The two patches are unchanged by the rebase from 96bac431e27a82c79c57e99198a78ac46b16332a, confirmed with git range-diff. The intervening upstream commits change client presentation, folder references and Usage filtering, not the backend paths under test.
Final-head checks
vp run check:changed exited 0 on the exact published head, on physical corex after rebasing. Formatting, lint, both knip passes, GraphQL generation and all 35 package typechecks passed. The adjacent tests passed as shown below.
| 1 | Test Files 21 passed (21) |
| 2 | Tests 734 passed (734) |
Regression and review evidence
- The retained three-level Claude adapter case failed before the fix at L2's parent. Disabling the hook lookup then failed at L3, whose launch frame is absent from the live SDK stream.
- The extended mixed-chain integration case failed before the progress fix with
result_availableinstead ofwaiting_for_children. It now keeps the outer task working until its nested native child finishes. - Integration coverage includes concurrent caller ownership, retry keys, sibling isolation, credential replacement, first-call ingestion, status/cancel, root and nested Stop, questions, completion delivery, remote controls and restart placement. Tests await persisted events or explicit worker completion.
- Independent Codex reviews resolved findings R1 through R10. The last review on pre-rebase head
96bac431e27a82c79c57e99198a78ac46b16332afound no remaining blocker or should-level finding and passed 10 focused tests. The same patch is present here.
Live providers
Physical corex, isolated source server and test state, Claude Haiku 5.5 and Codex GPT-6 Luna through mcli. No live user database was opened for writing.
On 30b8423bc455cffb8b79d260f98586f85d5187e6, stored tasks and child-thread parents confirmed a three-level native Claude chain and native Claude/Codex to delegate_task. Each child belonged to its immediate caller. All provider CLI runs completed with exit 0.
On 96bac431e27a82c79c57e99198a78ac46b16332a, the live mixed chain was root to app X to native A to native B. A and X's own turn completed while B remained blocked on a fixture FIFO. Stored events confirmed the outer task remained running. After releasing B, B completed at event 1050 and X completed at event 1053. The CLI exited 0.
The actual nested Stop command on 30b8423bc455cffb8b79d260f98586f85d5187e6 reached Claude's root session and interrupted L2. Both that head and baseline 01a93e7b51a39810c02d87701fa6ab23a2d1bea4 subsequently changed the cancelled label to failed after a late provider tool-result error. That existing label defect is outside this fix.
Desktop recordings
The embedded recordings use the actual Electron desktop app on a physical Mac mini M6, macOS 27.0.1, Electron 44.10.0. vp run build:desktop passed 5/5. Client source was 30b8423bc455cffb8b79d260f98586f85d5187e6; final backend was 96bac431e27a82c79c57e99198a78ac46b16332a. Client files were identical between those commits. Baseline backend was 01a93e7b51a39810c02d87701fa6ab23a2d1bea4.
The recordings inspect the completed live-provider runs, with direct child navigation and immediate-parent headers. Pairing used direct backend SSH forwards. Desktop profile, keychain and T3 home were isolated. The job removed its app process, forwards, worktree and private state; the Linux test servers are also stopped.
Limits
No live cross-environment provider delegation, signed-in Connect run, production-scale query benchmark or separate mobile/Windows/Linux desktop pass. Remote controls have focused tests, but remote child backlinks retain existing root behavior. The actual Codex first-call ingestion race and workflow-member admission were tested in separate layers, not as a combined live run. A preexisting prose-link hydration issue can show “Thread no longer available”; direct Lineage navigation passed.