Repository navigation
[Regression] Remote CLI and /side can no longer coordinate Windows Desktop threads; updates break task continuity #41580
Description
Summary
A cross-host workflow that previously worked has regressed:
Primary Codex thread on a remote Linux host
-> creates/opens or messages a normal user-owned thread in Windows Desktop
-> the Windows thread uses the local browser, Chrome extension, plugins, or Computer Use
-> the result is coordinated back to the primary thread
There are now three user-visible failures in that workflow:
- A Codex thread running on the remote Linux CLI/SSH-backed host can no longer create, open, hand off to, or send work to a normal thread in the local Windows Desktop app.
- A temporary
/sidechat can no longer reliably interact with the main thread or coordinate the creation/messaging of additional visible Desktop threads. - When a Windows Desktop update is installed, the app does not reliably relaunch. It must be opened manually, and active work on either side of the Desktop/remote boundary remains paused instead of recovering and continuing.
These symptoms may have independent underlying causes, but together they break one established orchestration and continuity workflow. This report intentionally groups them by user impact and does not claim a single root cause.
By “Desktop thread,” I mean a normal user-owned task that is visible and navigable in the Desktop sidebar. A host-only CLI session, an internal subagent, or a synthesized thread URI is not an equivalent replacement.
Environment
- Local host: Windows Codex Desktop
- Remote execution host: Linux
- Remote Codex CLI observed:
0.150.1 - Same authenticated account/workspace across the authorized surfaces
- Exact Desktop build is not available from the remote thread; it can be supplied privately if required
Reproduction
A. Remote thread -> local Windows Desktop thread
- Sign in to Codex Desktop on Windows and enable the authorized local capabilities needed for browser/Computer Use work.
- Connect/use a remote Linux host and start the primary Codex thread there.
- From that primary thread, ask Codex to create or message a normal user-owned thread on the Windows Desktop host.
- Observe that the effectful thread-management path is unavailable or cannot produce a real, navigable Windows Desktop thread.
- Attempts to substitute a remote CLI session or internal subagent do not provide the local Windows browser/plugin environment.
B. /side coordination
- Open a main Desktop thread.
- Start
/side. - Ask the side chat to inspect/coordinate with the main task, or create/message another visible Desktop task.
- Observe that the side chat cannot complete the coordination path that previously worked.
C. Update/restart continuity
- Keep one or more long-running tasks active in Windows Desktop and/or on a connected remote host.
- Install an available Desktop update.
- Observe that the Desktop app does not reliably reopen by itself.
- Open it manually.
- Observe that previously active work remains paused and does not automatically recover and continue.
Actual behavior
- The remote main task cannot route browser/Computer Use work to the already-authorized Windows Desktop environment.
/sideis reduced to an isolated temporary conversation when the required inter-task capabilities are absent.- Updating adds an unattended interruption: the app may remain closed until manually launched, and in-progress work stays paused afterward.
- The user must manually recreate coordination, find affected tasks, and resume them one by one.
- Context may still exist, but execution continuity is lost.
Expected behavior
Cross-host coordination
The authorized task-management surface should support an explicit, real handoff:
- remote primary thread -> normal Windows Desktop thread;
- Desktop thread -> remote primary thread;
- create/send/handoff operations should create or target user-owned tasks that are actually visible in Desktop;
- the destination task should use the destination host's browser, plugins, credentials, permissions, and Computer Use configuration;
- capability guidance and the callable tool catalog should agree.
/side
As documented, /side should remain a temporary side chat that does not interrupt the main chat, while still retaining the supported ability to inspect or coordinate the main task and authorized user-owned tasks.
Updates and active work
A Desktop update should have a transactional recovery flow:
- checkpoint active thread/goal/turn ownership;
- install the update;
- relaunch the Desktop app automatically, or show an explicit relaunch/progress state;
- reconnect to authorized remote hosts;
- recover each active task exactly once;
- automatically continue tasks that are safe to resume and do not require approval.
If a task cannot be resumed safely, it should remain in a durable “needs approval” or “resume required” state with a clear reason and one-click recovery. The app should not duplicate a partially completed action or silently mark interrupted work as complete.
Product context
The current documentation describes:
- Remote connections as using the connected host's projects, chats, plugins, browser setup, and Computer Use, including continuing the same work from another device:
https://learn.chatgpt.com/docs/remote-connections /sideas a temporary side chat that does not interrupt the main chat:
https://learn.chatgpt.com/docs/reference/slash-commands- long-running Goal work as supporting pause/resume:
https://learn.chatgpt.com/docs/long-running-work - the Desktop app as normally checking for and installing updates automatically:
https://learn.chatgpt.com/docs/enterprise/manage-app-updates
The documentation does not appear to state that post-update relaunch and task continuation are currently guaranteed; that portion is a reliability/UX request based on the observed interruption.
Related reports
These appear related but do not individually cover the complete Windows Desktop + remote Linux + /side + update-continuity workflow:
- [Desktop Remote SSH] inter-task tools stopped before update; 0.148 lacks codex_app MCP replacement #40865 — effectful inter-task tools missing on Desktop Remote SSH after the dynamic-tools migration
- [macOS][regression] Desktop cannot resume Remote Control / CLI thread:
already has an active writerafter latest update #37403 — Desktop cannot resume a Remote Control/CLI thread because an active writer remains - Windows: Codex auto-update causes first launch to run in background with no window (MainWindowHandle=0) #41523 / [Windows][26.825.5331.0] After Store auto-update, app starts headless for ~12 min — update-policy gate + cua_node runtime re-extraction block window creation; main process event loop synchronously blocked #41539 — Windows first launch after auto-update can remain headless
- Active goal auto-resume after an unclean restart leaves interrupted lifecycle state and diverges from persisted goal status after context-full retries (autogen selfreport from codex) #32973 — active Goal recovery after an unclean restart can diverge from persisted lifecycle state
Privacy and diagnostics
Account identifiers, organization/company names, hostnames, IP addresses, local usernames and paths, repository names, thread/turn IDs, browser profile names, signed-in sites, screenshots, and conversation contents are intentionally omitted.
Sanitized version information, timestamps, capability inventories, and relevant logs can be supplied through a private support channel if maintainers identify the exact diagnostics needed.
Activity
github-actions commentedon Aug 29, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
HaggisSupper commentedon Sep 5, 2026
Additional Windows Desktop reproduction observed on 2026-09-04:
turn_abortedbefore the side-conversation boundary appeared.Impact:
Expected behavior:
resume requiredstate rather than silently stopping.This appears to reinforce the
/sidenon-interruption and durable task-continuity portions of this issue. Local project names, paths, task contents, and identifiers are intentionally omitted.COOLak commentedon Oct 1, 2026
Adding a related Windows workflow observed on October 1, 2026: parallel cloud workers handled game-localization work, but visual QA needed a native local thread with the Windows app's Computer Use runtime. A dot could start cloud-backed tasks on the connected Windows executor and use shell/files/processes, yet could not create or directly coordinate the required native local thread.
After I manually created that native thread and supplied its desktop deep link, dot-side read failed with
unsupported placement format version 1, and send failed withCloudThreadNotFoundError: The cloud thread was not found.I added the detailed reproduction to #49729, which covers dot-created project tasks and known-ID follow-up access. The missing Computer Use runtime in delegated tasks also overlaps #49458; no shared root cause is established.The working workaround was a user-arranged file bridge: workers write handoff/status/results in the shared project folder and signal readiness to the native thread. That let the work continue, but first-class remote/cloud ↔ native-local thread coordination would avoid manually assembling this bridge.
Please make the cross-host create/read/status/message/steer route available to dots as well as remote Codex threads, preserve the selected project and destination capabilities, and retain authorization and permission boundaries. If a thread cannot be addressed through a particular route, show the supported route or the specific compatibility limitation.
This supports the cross-host coordination part of this issue. I have not reproduced the separate /side or update/restart symptoms in this test. Exact Desktop build was not collected; private thread IDs, local paths, screenshots, and files are omitted.