real dev environment verification, github pr 34977 main: 6a29d58d147522f85112c0c7149f2f9032cd955f pr: d35394f01165ca625b150413f4a9344b2d915684 linux; /usr/bin/google-chrome; repo Playwright; Codex gpt-6.1-sol (low) through local proxy setup: separate github/main worktree; vp i on each checkout. one dev server at a time, isolated .t3 homes, no seeding. project /home/user/release-proof and threads created through web UI. all turns used real Codex. scenario: after baseline turn completes, suspend only task-owned real codex child with SIGSTOP. call authenticated orchestration.forceStopThread over the running server's WebSocket API. while shutdown is pending, submit another message through the web UI. SIGSTOP deliberately widens teardown to exercise the race; no source patch, adapter substitute, fixture, or mocked protocol. main (times relative to force-stop request): 0ms force_stop_requested 59ms processes: 33394:S 33438:T 35555:S 681ms next_message_submitted 1112ms processes: 33394:S 33438:T 35555:S 36126:S 2113ms processes: 35555:S 36126:S 2132ms force_stop_response {"_tag":"Exit","requestId":0,"exit":{"_tag":"Success","value":{"outcome":"released","sharedWithThreadIds":[]}}} 7224ms replacement_turn_completed pr (times relative to force-stop request): 0ms force_stop_requested 57ms processes: 39993:S 40007:T 877ms next_message_submitted 2161ms force_stop_response {"_tag":"Exit","requestId":0,"exit":{"_tag":"Success","value":{"outcome":"released","sharedWithThreadIds":[]}}} 2302ms processes: 39993:S 40007:Z 2320ms processes: 39993:Z 40007:Z 2334ms processes: 39993:Z 40007:Z 2343ms processes: (old process gone; no replacement yet) 2858ms processes: 41661:S 7079ms replacement_turn_completed main: replacement wrapper 36126 appeared while old wrapper 33394 and stopped binary 33438 were still alive. server logged ProviderAdapterThreadInUseError / 'thread 01a1ee70-f01e-7550-ad87-e75fadea57f5 already has an active writer'. retry recovered and RELEASE_REOPEN_OK eventually rendered. pr: old wrapper 39993 and binary 40007 exited before replacement 41661 appeared. no active-writer error in dev server log. RELEASE_REOPEN_OK rendered. neighboring real-app checks: 10861ms clicked_restart_agent_session 17188ms restart_resumed_conversation 20754ms sibling_provider_turn_running 20865ms force_stop_while_sibling_working {"_tag":"Exit","requestId":0,"exit":{"_tag":"Success","value":{"outcome":"shared","sharedWithThreadIds":["31508d64-7ca8-4fcf-b238-3373dea4387b"]}}} 48146ms sibling_completed 58045ms detached_original_resumed restart agent session: clicked command palette action; next turn correctly recalled RELEASE_BASELINE_OK and RELEASE_REOPEN_OK. shared safety: sibling ran real terminal command 'timeout 20 sleep 12'; force-stop on original returned shared with sibling id, sibling completed SIBLING_OK; original resumed and completed RELEASE_SHARED_OK. final WebSocket projections: original runs 1-4 and all four provider turns completed; sibling run 1 and provider turn completed. blacksmith: vp i exit 0; vp test run apps/server/src/orchestration-v2/ProviderSessionManager.test.ts exit 0, 82/82 passed. existing head reports 36/36 successful checks via cafe pr status. no source changes, commits, or pushes. not verified: native desktop/mobile; macOS/Windows; remote/relay/tunnel; other providers; live 30-second timeout, caller cancellation, workspace-stat gap, credential races and retired-runtime fork races. session-manager tests cover these internal races, but they were not all induced through a real provider. repeated-close joining was checked by existing service tests, not separately through the client.