ChatGPT Dots not showing or not working? Plan and region checks, how to get a missing Dot back on desktop, and why not to delete it (OpenAI Dots)
Never seen Dots? Check plan, region and rollout first: eligible Pro outside the EEA, UK and Switzerland, Business Premium, or Enterprise with an admin switch. Dot gone from the desktop sidebar? It usually still works on the web; open it through search or a notification, and set a custom model provider back to OpenAI. Silent Dot? Answer pending requests, then Pause and Resume, then Reboot, and don't delete it.

The short version: why Dots are missing, how to get a vanished Dot back, and what to do when it goes silent
In one line
If you have never seen Dots, it is almost always eligibility rather than a bug: your plan, your region, the gradual rollout, or an Enterprise switch, and reinstalling won't change any of them. If your Dot was working and then vanished or went quiet, it usually still exists. After the desktop update of 20 to 21 October, many people lost only the sidebar entry on Mac and Windows while their Dot kept working at chatgpt.com/dots/home, and the update of 23 to 24 October didn't bring it back for everyone. The most common cause found so far is the model_provider line in the Codex config, even when it still points at OpenAI. For a silent Dot, look for a decision or approval it is waiting on, then try Pause followed by Resume, then Reboot. Don't delete it to fix a problem: several people who deleted their Dot couldn't create a new one for days.
What to do, in order
- Check the OpenAI status page for a Dot incident. On Oct 7 it listed two Dot incidents in the same afternoon.
- Make sure the account and workspace you meant to use are active, then open chatgpt.com/dots in a desktop browser. If the Dot answers there, it is fine, and the problem is in one app.
- If there is no Dot anywhere, compare your plan with sections 1 to 4. Dots need an eligible Pro plan for adults outside the EEA, UK and Switzerland, Business Premium (which can take several days), or Enterprise with Use dots (Beta) switched on by an admin under Workspace settings, Permissions & roles, Workspace capabilities. Set up in the ChatGPT mobile app, the desktop app or a desktop browser, because mobile web isn't supported. Travelling into the EU, or an account first created in the UK, has also made Dots disappear.
- If only the desktop app lost it, open the Dot another way first: search for "dot" or your Dot's name in the desktop search and open the result under Chats, try the Cmd+K command menu on a Mac, click a notification from the Dot, or open
codex://dots. Then open~/.codex/config.toml. Unless the top-levelmodel_providerreads exactly"openai", the app hides the Dot, even when the provider is only another name for OpenAI, such as"custom"or"openai-http". Back up the file, change that one line to"openai", quit the app fully (on Windows, the tray process too) and relaunch it. If you use CC Switch, turn off its "Unified Codex session history" option first, or it writes"custom"back. Reinstalling didn't help in these reports. - If the Dot is missing from the phone app, or the steps above don't bring it back, update that app first. On Windows, a restart followed by a Repair of the ChatGPT app brought a desktop entry back. Keep using the web in the meantime.
- If the Dot stopped replying, open the task in Activity (in the Dot's profile) and answer any pending decision, approval, sign-in or connection request. If nothing changes, choose Pause and then Resume straight away, and after that Reboot (on desktop, right-click the Dot in the sidebar). Restarting the app is not the same thing.
- If errors only appear when the Dot works on your own computer, note the desktop app version and the exact error text, then read section 9 before you reconnect or reinstall anything.
- Don't delete the Dot to start over, and don't delete a "New chat" entry in Recents that shows your Dot's avatar, because for one person that deleted the whole Dot. If none of this works, contact support with what section 13 lists.
Where it gets stuck, and what fixed it
- Gone from the desktop sidebar after the 26.1016 update, still there on the web and the phone. From 20 October, people on Windows and Mac lost the Your dot entry, and on some machines the dictation microphone with it, while the Dot kept replying in the browser (#55795, #55783). The 26.1021.31345 build of 23 to 24 October didn't fix it for several of them. Two causes have been found. The first is the provider line. The app shows the Dot only when
model_provideris exactly"openai", so a custom provider hides it, and so does an alias that still signs in with ChatGPT and has no custom address. On 24 October, changing"custom"to"openai"and quitting the app, tray process included, brought the entry back on Windows for two people, with no reset, sign-out or reinstall (recovery steps). CC Switch's "Unified Codex session history" option rewrites the line to"custom"every time it applies a setting, and its maintainer suggests turning the option off (CC Switch). The second is a start-up interrupted by a network change: the entry stayed hidden until the app was fully quit and reopened on a stable connection. Some people with neither cause were still waiting on 24 October, and for one of them searching for "dot" found nothing, so keepcodex://dots, notifications and the web as fallbacks. - Gone from one app for other reasons. On Windows, a restart and then a Repair of the app helped. On a Mac showing "Couldn't load your dot", a faster connection (a startup download was being cut off after 10 seconds) or, for one person, a macOS update helped. A task missing from Recents came back once it was opened from the Dot chat. The Android disappearance on Business Premium seats had no fix by Oct 14, but the desktop app and the phone's browser still reached the Dot.
- Chats the Dot created are missing from the desktop lists, or open blank. On 21 October, sessions a Dot had created showed in Codex on the web but not in the desktop lists or the Cmd+K search. Choosing one from the Dot's own session list and clicking Go to Chat opened it and put it in Recent, but only until the next day (OpenAI Community). On 26.1021.31345, one Mac user could still select old Dot sessions in the sidebar, but they opened blank, while the web and Slack worked (#55783). Ask the Dot for the links to the chats and tasks it created, and open them on the web when the desktop app shows nothing.
- Gone because of where you are. Personal Pro excludes the EEA, UK and Switzerland. One user's Dot disappeared when they travelled to the EU, even with a VPN, and a US Pro user's Dot kept working in Italy for three days before every app showed "Dots aren't available for this account" (#55996). Reinstalling doesn't help; see EU & UK availability.
- Messages marked Read, the computer shows Ready, and no reply ever comes. This hit several Pro accounts on Oct 17, on the web, the Mac app and the iPhone at once, while ordinary ChatGPT chats kept working. OpenAI Support replied in the thread that they were aware and engineering was working on a fix. For one person, pausing the Dot and resuming it straight away brought replies back; for another, neither Pause and Resume nor Reboot helped. A third traced it to a pending decision that couldn't be cleared, and a brand-new Dot on the same account went just as quiet, so deleting didn't help either (OpenAI Community, follow-up).
- The Dot won't open anywhere, with HTTP 409 "root thread is unavailable". A Business user lost access on macOS, web and mobile from 20 October, and the support case was closed as a duplicate of an unrelated call problem (OpenAI Community). Reply to support that it is a separate problem, ask for the Dot, its memories and cloud files to be kept, and don't delete it.
- Answers turn shallow and approvals are asked for again. One heavy user saw this on 20 October and found the Dot back to normal on 22 October (OpenAI Community). Before rebuilding your setup, give it a day and test again with a familiar task.
- Files can't be read or saved. From 20 October, Dots couldn't attach or save generated files ("hosted apps tools/list request failed with HTTP status 403"), and some couldn't read an image uploaded into the conversation ("scope_violation"). Saving to Google Drive worked as a stopgap, and uploads worked again for one user on 22 October (saving files, reading images).
- Slow, or "still working" when nothing is running. Status lines such as Last active can freeze while the Dot works normally. Send a short message and see whether a reply arrives, open the task in Activity to see whether it shows a blocker or a pending approval, and ask the Dot directly whether anything new has come in. On a Mac, a local task that was waiting for approval showed only as "in progress" in the main Dot chat.
- Deleted to start over, and now there is no new Dot. The page shows "Failed to load your dot" or "Couldn't create your dot", or creation fails with HTTP 409 and
cloud_aeon_recovery_requiredwhile the Dot list is empty. Two people got a new Dot by creating it on the web instead of in the app. Others still couldn't after five days on every client, and on Oct 19 support told one of them it knew about the problem but had no clear plan yet (OpenAI Community, #55120). Section 11 has the details. - "Your dot is on a break". It is an abuse-prevention pause, and people have hit it with plan usage left. The breaks people timed lasted from about 1 hour 40 minutes to about 13.5 hours, and Reboot didn't end one early. On long jobs, have the Dot save progress somewhere you can see; the full answer is on "Your dot is on a break".
- The cloud computer says "Unavailable", or files are gone. Reboot often didn't bring it back, and in several reports anything not pushed to Git or saved elsewhere was lost. Keep finished work in Google Drive, a GitHub repo or the chat, and if the shell still responds, copy important files out before you try a reboot.
- Errors on your own Windows PC. Behind a local proxy, the
respect_system_proxysetting brought the PC back online for four people. With "setup refresh had errors", commands ran again for several people after the 26.1007 desktop update, and until then, approving each command to run outside the sandbox got around it. Browser control failing with "HRESULT 0x80070003" had no confirmed fix by Oct 10, though the shell kept working. - "user cancelled MCP tool call" when you cancelled nothing. Check the destination first, because the action sometimes did happen. If it didn't, one explicit retry usually worked. Scheduled runs don't retry on their own, so check their results.
- Calls ring but never connect. Update the app, Reboot the Dot, or type to it while it rings. OpenAI rolled out call fixes on 8 October, and some calls still failed after that. If a connected call returns
voice_parent_context_pending, type the request into the chat instead. On Android, leaving the Dot's chat ends the call, so stay in that chat until you are done (OpenAI Community). More call reports are on the voice calls page. - The Windows desktop app says "Voice chat unavailable" and the Dot's call button is gone, while calls work on the phone and the web (#56846). The check that hides the Dot also covers dictation and voice in the app (analysis), so look at the
model_providerline first, and call from the phone or the web in the meantime.
What you can fix yourself, and what only OpenAI can
You can usually fix:
- the wrong account or workspace being active;
- a missing desktop sidebar entry, by opening the Dot through search, a notification or
codex://dots, and by settingmodel_providerto exactly"openai", with CC Switch's session history option turned off if you use it; - a chat the Dot created that the desktop app doesn't list, by opening it from the Dot's session list with Go to Chat, or from a link the Dot sends you;
- an outdated or broken app, with an update, a Repair, or on a Mac a macOS update;
- a decision, approval or sign-in the Dot is waiting on, by answering it in the task;
- a stuck Dot, with Pause and Resume or with Reboot;
- a Windows PC that stays Offline behind a proxy, with one setting;
- a tool call that failed for no reason, by checking and then retrying once.
Only OpenAI, or your admin, can fix:
- plan and region eligibility, rollout timing, and the Enterprise switch (your admin); the full list is in the availability guide;
- an "on a break" pause;
- a desktop entry that stays missing with
model_providerset to"openai"and a clean start on a stable connection; - a Dot that stays silent after Pause, Resume and Reboot, or one that returns HTTP 409 on every device;
- an account that can't create a new Dot after a delete;
- files lost from the cloud computer, or file saves that fail with a 403;
- old computer registrations piling up ("Maximum number of local executor registrations reached").
For these, wait, or contact support with the details in section 13.
One message to copy
If the Dot still answers somewhere, on the web, on your phone or through a notification, send it this before you change anything:
Something isn't working with you on [device or app]. Before you change anything, check and give me one short list: any decision, approval or sign-in you are waiting on from me, whether you can still reach your cloud computer and my connected computers, any task in Activity that is stuck, any app or tool connection that has dropped, whether you can still read and save files, the links to the chats and tasks you created in the last few days, and the exact text of the last error you got, with its time in UTC. Don't retry, reboot, delete or reconnect anything yourself.
1. Confirm account, plan, and workspace
Open the account menu and verify the active email and workspace. Compare with the availability guide: launch access covers eligible Pro 100 / Pro 200 / Pro 500 (outside excluded regions), Business Premium, and Enterprise when enabled. Switching from a personal workspace to an organization can change what you see.
2. Separate “Pro in Europe” from “still rolling out”
Personal Pro launch access excludes the EEA, UK, and Switzerland. The full plan matrix and FAQ are on ChatGPT Dots EU & UK availability. If that is your situation, the missing entry matches official Access, and reinstalling won't change it. One user, rbondi, commented on Oct 7 that their existing Dot disappeared from the web and macOS app when they travelled to the EU, even with a VPN, while it kept sending them notifications. jts-nyc, a US Pro user, arrived in Italy on Oct 18 and kept using their Dot, Jdot, for three days. On Oct 21 it vanished from the iPhone and desktop apps, and chatgpt.com/dots on the phone showed "Dots unavailable. Dots aren't available for this account." It came back on the phone briefly and then disappeared again (#55996). So a trip can cut access some days after you arrive, not only on the first day.
If you are on Business Premium (including in Europe) or Pro outside those regions, Dots still roll out gradually. Meet dots · Access says you may not see Dots immediately even when eligible, and eligible Business Premium accounts can take several days after announcement. An eligible plan is not the same as the feature being live on your account today.
3. Create it in the ChatGPT app or a desktop browser, not mobile web
You can create your Dot in the ChatGPT mobile app, the ChatGPT desktop app, or https://chatgpt.com/dots in a desktop browser; after setup the same Dot follows you across devices. If you set up on your phone, update the ChatGPT app to the latest version first. Mobile web is not supported, so opening chatgpt.com/dots in a phone browser won't get you through setup (Get started with your dot). Connecting your own computer still happens in the desktop app, whichever device you created the Dot on.
4. Enterprise needs an admin to turn Dots on
In an Enterprise workspace (incl. Edu / Healthcare), ask the administrator to enable Use dots (Beta) under Workspace settings → Permissions & roles → Workspace capabilities. It is off by default (Manage dots in ChatGPT workspaces). Confirm you are in the org workspace before troubleshooting a personal account.
5. Your Dot disappeared from the sidebar, or the page won't load
These users already had a working Dot, and then the entry was gone or wouldn't open.
-
Gone from the Mac and Windows sidebar after the 26.1016 update, still working on the web. From Oct 20, people who updated the desktop app to 26.1016.51333 found the Your dot entry missing from the sidebar, while chatgpt.com/dots/home still showed the Dot, its history and a working message box. On Windows, wanbo-ai found it missing in both the ChatGPT and Codex views, reinstalled the app and tried the Ctrl + . shortcut without success; rolling back to the earlier version brought the entry back, and on the current version, searching for "dot" in the desktop app and opening the result under Chats opened the existing Dot, a tip from phoenixt10 (#55795). On 23 Oct the entry was still missing for wanbo-ai on 26.1016.62342. On a Mac, Stahl-G saw the same on 26.1016.51333 (#55783), and others in that thread opened the Dot from a desktop notification, by searching for its name, or with the
codex://dotslink.- The custom provider cause. bmpolonsky noticed that the Dot entry and the dictation microphone disappeared together whenever a custom
model_providerwas selected in the Codexconfig.toml, and came back when that one line was commented out. ARTISTA666 confirmed it: even a custom provider that still signs in with ChatGPT makes the app treat the setup as not OpenAI's own, and hides the Dot. The fix that worked was to quit the app completely, back up~/.codex/config.toml, changemodel_providerto"openai"(the custom provider block can stay), and relaunch. singularitti confirmed it on 26.1016.62342. One Mac user found a provider setting there that they didn't remember adding (#56392), so check the file even if you never edited it. The provider's name doesn't matter. sprocsw had a provider keyedcustomwith the name OpenAI and no custom address, and Yunyou20xiya had one keyedopenai-httpand still had no entry on 26.1021.31345, because the app's code checks for the literal"openai"(sprocsw, Yunyou20xiya). Yunyou20xiya found the Dot through the Cmd+K command menu on the Mac. On 24 October, ccuiab, on Windows 26.1021.31345, backed up the file, changed onlymodel_provider = "custom"to"openai", quit ChatGPT including its tray process and reopened it, and the entry came back with no Dot reset, sign-out, cache clearing or reinstall. spdw666 followed the same steps and it worked (#55795, confirmation). ccuiab'scustomline came from CC Switch. With its "Unified Codex session history" option on, CC Switch writes"custom"for the official sign-in each time it applies a setting, so a manual fix gets overwritten. The maintainer's advice for desktop users is to turn the option off, and the desktop app still lists sessions from every provider (CC Switch). - The start-up cause. Later the same day ARTISTA666's entry disappeared again with the provider already set to
"openai", and this time the microphone was still there. The app had started while the network was changing, and the Dot never loaded during that session. Bringing the window to the front didn't help; quitting fully and reopening on a stable connection did. - Not everyone had either cause. YouAI-Liu had never used a custom provider and still had no entry on 23 Oct, and DRKTCoke could reach the Dot only by clicking its notifications (#55845). nguyen-ammonite, on Windows 26.1016.62342 with no top-level
model_providerin the base config, couldn't find the Dot by searching for "dot" either (comment). Until a fix arrives, open the Dot from search, a notification,codex://dotsor the web.
- The custom provider cause. bmpolonsky noticed that the Dot entry and the dictation microphone disappeared together whenever a custom
-
Deleting a chat in Recents deleted the whole Dot. On Oct 23, WaleyChAn saw an entry in the Recents list titled "New chat" with the Dot's avatar. It opened the same conversation as the Dot itself, so they deleted it as a stray chat. The deletion gave no warning about the Dot, but the conversation and the Dot both disappeared, and the next start began a fresh Dot from the naming step (#56348). If a Recents entry shows your Dot's avatar, leave it alone.
-
Gone from the Windows sidebar while it kept working. On Oct 6, M4rkzzz's Dot vanished from the Windows desktop app, and several cloud chats showed a greyed-out input box. The Dot itself kept running and posting messages to those chats, signed with an old task title instead of its name (#51213). Another Windows user, gaoshuang811227-GHB, still had the Dot on their phone but not on desktop after updating. A Windows restart followed by a Repair of the ChatGPT app brought the desktop entry back, though their own computer still showed as offline.
-
"Your dot could not be loaded" on the web. BryanYue saw that message with a Retry button on the web on Oct 7, the same morning calls to the Dot were failing. A dot-specific restart fixed the calls; restarting the app alone hadn't (#51533).
-
Gone on Windows and Ubuntu after an update, still there on the web. On Oct 9, CRV5230 updated the desktop app on both computers (Windows build 26.1002.52244). The Dots entry disappeared from both sidebars, and so did the tasks the Dot had created in a project. Switching between the Codex and ChatGPT views didn't bring them back. On the web the Dot still answered, and it could still pull up the results of three earlier tasks (#52457).
-
Gone after a run of loading errors. Before wyxxxxxxx19-collab's Dot vanished from the Windows sidebar on Oct 10 (desktop 26.1007.21434), the app kept showing "Local Work executor registration failed (HTTP 403)", "Couldn't load your dot" and "We couldn't check access to messages", with short recoveries in between. At the same time the Dot was active at chatgpt.com/dots/home, with its computer shown as ready. They mention a local proxy running in TUN mode as part of the setup, without claiming it caused the problem (#52774).
-
A task missing from Recents. A cloud task that NaturalMatchaKing's Dot created on Windows wasn't in Recents at first, showed up hours later, and disappeared again after an app restart, although the task itself still existed. Opening it from the Dot chat put it back in Recents (#52733).
-
Dot-created sessions missing from the desktop app, or blank. On 21 October, cryptocean could see dozens of sessions created by the Dot in Codex on the web, but not in the desktop lists or the Cmd+K search, even in Codex mode. After a long wait, selecting the Dot loaded its session list on the right. Picking a session, opening its preview and clicking Go to Chat opened it in full and added it to Recent, but by the next day it was gone again, and there was no option to move these sessions into a project (OpenAI Community). On 24 October, amirtaherkhani, on a Mac with 26.1021.31345, had no Dot entry and could select old Dot-created sessions in the sidebar, but they opened blank, while the Dot answered on the web and in Slack (#55783).
-
Gone after a mid-October update: UK-origin account, Dot created in Sri Lanka. On 14 Oct, Jack (@j4ppleby) wrote that the latest update had removed his Dot. He created it in Sri Lanka and was still there, but his ChatGPT account was originally created in the UK. A follow-up asked whether gating now uses ChatGPT account setup location rather than where the Dot was created. No fix was reported in the thread. This sits next to the Pro-in-Europe travel report in §2 and the EU & UK availability page. Reinstalling does not restore a region-gated Dot. X post (follow-up)
-
Gone from the Android sidebar on a Business Premium seat. david.varkonyi, in Hungary, had a Dot that worked on PC and Android for several days. Around Oct 8 it disappeared from the Android sidebar, together with its chat and its notifications, while it kept answering normally on PC. Signing out and back in, restarting and reinstalling the app, updating ChatGPT and Android, and clearing the app cache didn't bring it back (community report). On Oct 13, serhatoezdemir reported the same: the Dot had been in the Android app, then disappeared completely after signing out and back in, while it still worked on the PC and in the browser on the same phone. By Oct 14 the thread had six posts and nobody had reported a fix or an answer from support. On Oct 11, Dimitris Papadopoulos (@dimitrispap) reported the same on X for an EU Business Premium seat: the Dot had disappeared from the Android sidebar, and reinstalling the app or clearing the cache did not help. Until there is a fix, the desktop app still reaches the Dot, and, as in serhatoezdemir's case, the phone's browser may too.
-
Wrong Dot name on Android: header and sidebar show "tibro". On 1 Oct, dot_ai_assistant reported Android ChatGPT 1.2026.265 (41) showing "tibro" in the conversation header and "Your tibro" in the sidebar, while the saved assistant name was still "dot" and iPhone plus web showed "Dot" correctly. The user had not renamed it. On 14 Oct, OpenAI Support asked them to restart the Android app, then update or reinstall if it persisted, confirm the name still looked right on desktop, and send a screenshot if Android stayed wrong. By 15 Oct the thread still had only those two posts.
-
Gone from the Mac app, or from the iPhone sidebar, after an update (Oct 10–13). curiousgangsta, on a Pro 200 plan, has posted since Oct 11 that the Dot is missing from the sidebar of the macOS desktop app, while it is still there in the browser and in the iOS app on the same account and workspace (X, Oct 11). On Oct 13 it was still missing from the Mac app, and curiousgangsta put it down to an update (X, Oct 13). On Oct 12 joeker wrote that their Dot had "just completely disappeared", and that it didn't matter much because they had given up on ChatGPT and Codex (X). Matt Zatorski (@matzatorski) said on Oct 10 that the whole feature vanished right after the Dot was configured (X). On Oct 13 Matt Zatorski posted two iPhone screenshots: in the earlier one the Dot, tobby, sits at the top of the ChatGPT sidebar; in the later one the list starts at Space and the Dot entry is gone. The post asks OpenAI to bring it back. None of the three had reported a fix by the evening of Oct 13 (X).
-
Push notification opens a new chat instead of the Dot. On 14 Oct, Jeremy Waller (@JeremyWaller) got a ChatGPT push when his Dot replied, but opening the notification landed him in a new chat. He saw no way to open the Dot in the app. X
-
Dot-created threads on the website, missing in the desktop and mobile apps. Eliot Gevers (@EliotGevers) on 14 Oct listed eight issues; the first was that threads his Dot created show up on the website but not in the desktop or mobile app, and approvals given to the Dot often don't reach the tasks (he has to approve again in each thread). Vish (@vishp_93) on 13 Oct reported the same gap from another angle: when the Dot makes a new chat it is a cloud chat, and it does not appear in the Codex sidebar or the mobile app unless he asks the Dot for the link. EliotGevers · vishp_93
-
Gone from the mobile app only; still on browser and desktop. Todd Wahl (@ToddWahl) on 11 Oct wrote that Dots only shows up in the browser and desktop ChatGPT for him; his Dot disappeared from the mobile app after a couple of days, with a support ticket still open and no resolution. X
-
Hard to find on the desktop app; web still works. Dale Soon (@DaleSoon94566) on 13 Oct often can't find Dots on desktop and has to use the web version instead, and a lot of the time Dots also says it can't connect to the local computer. X
-
Cloud task created, but no entry in the iOS Cloud section. dahuyouya on 10 Oct asked the Dot to create a cloud task; the Dot confirmed creation and returned progress and results, but there was no entry or link to open it in the iOS app, including in the Cloud section. X
-
Gone from the Codex Mac app after it was already in daily use. On Oct 10, Alan E. (@storiesofalan) wrote that their Dot disappeared from the Codex app for macOS (version 26.1007.21159) after they had already been using it frequently. X
-
Suddenly disappeared after the Dot was working on its own. On 14 Oct, Marcus (@AgentDreamer_) said their Dot suddenly disappeared after they had finally got it working autonomously; two days later they were still waiting on support, and "a lot of good work is gone." X
-
Gone from the Mac, Windows and iPhone apps at once, while travelling. On 15 Oct, Casey Williams (@Packet7hrower) wrote that he couldn't see his Dot in the macOS app, the Windows app or the iOS app. He was out of the country, and the Dot was holding all of his travel plans and documents. His post names the three apps, not chatgpt.com/dots/home in a browser, which is where other people in this section still found their Dot; if you are travelling, also keep the region reports in §2 in mind. X
If the Dot works on one device, it still exists. In the reports above, chatgpt.com/dots/home kept working while the desktop entry was gone. Check the other device's account and workspace first, then update or repair that app before touching the Dot itself.
On a Mac: "Couldn't load your dot" or "Couldn't load workspace settings"
Between October 2 and 10, several Mac users found the desktop app wouldn't load their Dot, or wouldn't let them start any new chat, while the same account worked on chatgpt.com and on their phone, Dot included. Six people's reports across three GitHub issues point to two different causes, and only one has a reported fix.
An Apple device check that fails on that Mac. bakira99 opened #52342 on Oct 9: after updating to desktop 26.1002.52244, every new chat showed "Failed to load workspace settings. Try again to continue." Their app logs showed that Apple's DeviceCheck token couldn't be generated on that Mac. georgesamuelson's logs go back to September 25 and show the check never succeeded once on that MacBook Pro. Before the October 7 app update the failure didn't stop new chats; from that update on, every request that needs the check fails. georgesamuelson later added screenshots from October 2 showing "Couldn't load your dot", "We couldn't load your tools" and "Couldn't load color customization" on the same Mac. Ocean326 traced it further, into Apple's own device-certificate service, and also kept seeing "Couldn't check Lockdown mode". potassium-cyber saw the same on desktop 26.1007.21159.
Across these reports, quitting and relaunching, restarting the Mac, signing out and in, reinstalling the app, a newer app build, and turning off VPNs or changing proxy settings didn't help. One thing did, for one person: potassium-cyber updated macOS from 26.6.2 to 26.7.1 (build 25G241) and restarted, and the warning was gone and new chats worked again. A macOS update isn't a sure fix: Ocean326 went from 26.3 to 27.0.1 and still had the error, and georgesamuelson, already on 27.0.1, had no newer update offered.
To see whether this is what's happening on your Mac, open Terminal and run the line below. If it prints file names, your app logs contain the same DeviceCheck failure.
grep -rl "DeviceCheck token generation" ~/Library/Logs/com.openai.codex
A startup download that can't finish in 10 seconds. On Oct 8, shaanguan's Dot on macOS 26.4 kept vanishing from the sidebar or showing "Unable to load your dot", and the app's language kept switching between Chinese and English (#52010). The network log in the report showed the app fetching a settings file of about 8.3 MB at startup and cancelling each attempt at about 10 seconds, after receiving at most about 1 MB. Every retry started again from zero with the same limit. So if your Dot comes and goes only on the Mac, it's worth trying the app on a faster connection before reinstalling anything. kffz's "cannot load your dot" report from the same day (#52070) has no logs and was flagged as a possible duplicate of this one.
In both cases the Dot itself is fine: keep using it at chatgpt.com/dots/home or on your phone while the Mac app is stuck.
Gone with no error at all. On Oct 13, Faris (@FarisWayne) simply noticed that the Dot agent he had set up was gone. If yours vanished the same way, check chatgpt.com/dots/home and your phone before you create a new one; the delete-and-recreate route in section 11 has its own problems.
6. It stopped replying, or the reply never arrived
-
Unreachable everywhere with HTTP 409. From Oct 20, marcodigital's Dot in a Business workspace wouldn't open on macOS, the web or mobile, on any network. The desktop logs showed HTTP 409 "Your dot's root thread is unavailable" and "thread access denied". OpenAI's incident report of Oct 20 included Dots, but access didn't come back afterwards, and the new support case was closed automatically as a duplicate of an older one about voice calls (OpenAI Community). If this happens to you, reply in the case that it is a separate problem, ask support to keep the Dot, its conversations, memories and cloud files, and don't delete it.
-
Shallow answers for two days. On Oct 20, wjddusghd44, who uses their Dot, Eve, for engineering and automation work, found its answers had become shallow: it paraphrased their latest sentence instead of using the earlier context, dropped decisions already agreed, and asked again for permission it had already been given. On Oct 22 they wrote that Eve was back, following the context and handling corrections well again (OpenAI Community). A sudden drop like this has passed on its own, so test again a day later with a familiar task before you rebuild your rules.
-
Read, Ready, and silent on every device. On Oct 17, AlexN's Dot marked every message Read on the Mac app, the web and an iPhone but never answered. The status panel said the computer was Ready, no limit or break notice appeared, and Chrome's inspector showed a 429 rate-limit error. Pause and Resume and Reboot didn't help. OpenAI Support replied the same afternoon that they were aware of the issue and engineering was working on a fix. TMan76's Dot also looked as if it was typing and never replied; TMan76 paused it and resumed it straight away, and it answered again, though they couldn't say whether that or a service recovery did it. swortslandon's Dot went quiet after a task raised a pending decision that couldn't be undone. Stopping the task and deleting the thread didn't clear it, and after they deleted the Dot, losing its saved memories, a brand-new one was silent in the same way (OpenAI Community, swortslandon's write-up).
-
Replies that take 5–10 minutes for the simplest questions. On 14 Oct, RusBus09 (@russellclarke9) asked whether anyone else was seeing extreme degradation on OpenAI Dots: their Dot was thinking for 5–10 minutes before replying to the simplest questions. X
-
Dot conversations that do not continue when you switch devices. On 14 Oct, Valley Edge Digital (@ValleyEdgeDigit) wrote that Dot conversations were not continuing, and that projects in general were not continuing on any device they moved to, while they were still deciding whether to keep the Codex/ChatGPT subscription. X
-
Twenty-seven setup failures in two working sessions. On 13 Oct, Adam (@itsactuallyadam) asked their Dot to list every OpenAI bug and failure from setting up workflows; it came back with twenty-seven from two sessions. Adam wrote that the Dot did not know how Dots works, and that they had to explain features that then failed or had no documentation. They shared 19 of the 27 publicly and offered to dig into the rest with the OpenAI Devs / Codex team. X
-
A reply that silently went missing. On Oct 6 at 23:50 UTC, RowboTony (Pro 100, macOS) asked their Dot, Pete, a question and got nothing back: no thinking indicator, no error, no retry. Four minutes later RowboTony asked for a status update, and Pete answered that its earlier send had returned "Message delivery could not be confirmed. Do not retry this send." A second user, kairytwo2, saw the same message twice on Android the next day (#51523).
-
Messages show as sent and read, but no reply text arrives. On 2 Oct, DotLover (Pro $200, macOS 27.0.1, desktop app 26.928.40906) wrote that messages to the Dot were marked sent and read, a loading indicator or empty reply bubble appeared, and no reply text ever arrived. The Dot's cloud computer stayed reachable. They asked for either a real reply or a clear failure message with which limit applied and a retry time.
-
Dot can list Codex Cloud tasks but create/continue returns UNKNOWN. On 15 Oct, TDove reported that across multiple Dots, listing saved Codex Cloud environments and reading finished results still worked, while creating or continuing a task intermittently failed with
AppServerBackendRequestError/error_code: UNKNOWNbefore any new execution appeared. Typing the same instruction directly into the Cloud task UI succeeded (verified file download with matching SHA256 around 15:58 UTC). Request IDs from failed attempts includea0779ca2-bb01-43b1-b478-76c229418345(~10:18–10:19 UTC) anda7143718-2c6c-4d54-ae2b-ca6e3f98c1d9(~15:44 UTC). Related GitHub reports from robjshore, b0ggs and hx123321 describe the same delegated-vs-direct gap. -
Follow-ups to an existing Codex Cloud task also returned UNKNOWN (same day pattern). On 15 Oct, CrossS reported that assistant-initiated continuation messages to a saved GitHub-backed cloud task (coordinated as Atlas on ChatGPT Web / macOS) repeatedly failed with
AppServerBackendRequestError/ UNKNOWN around 10:07–10:13 UTC; a fresh-task create around 10:21 UTC also failed, while reading tasks and listing environments still worked. Service recovered the same afternoon: the original task completed a new turn around 15:41 UTC and a continuation was accepted around 15:42 UTC. CrossS points to the same delegated-vs-direct UNKNOWN pattern as TDove and earlier GitHub Dot reports. -
An outbound prompt appeared that the owner did not send. On 14 Oct, Peter Petrash (@petekp) found an outbound prompt to his Dot in the ChatGPT app that he did not send, apparently sent while he was asleep. The Dot has full access to his filesystem; he wondered whether users' prompts were being misdirected in the backend. X
-
An email watch that stayed quiet for two hours. anotb had an event-based Gmail watch switched on. An email arrived at 21:17 UTC on Oct 7, no alert came, and at 23:25 they asked about it. An on-demand mailbox lookup then found the email (#51900).
-
Notifications for messages you already read. dean1012 got notifications on a second device 30 to 60 minutes after reading the same messages on the first one, so old messages looked like new updates (#51574).
-
A safety pause that wouldn't lift. 2018wzh's Dot showed the "Your dot paused out of precaution" banner. Clicking Resume your dot failed with "Something went wrong" on both the Windows app and the website, while the sidebar listed the Dot as active (OpenAI Community).
-
"Selected model is at capacity" in the middle of work. On Oct 9, GGBondBlueWhale's Dot Lulu stopped several times in one coding conversation in the macOS app with "Selected model is at capacity. Please try a different model." The retry countdown ran out and the same error came back (#52341). jhs1121 got it with different model choices on Oct 7 and 8, while ordinary ChatGPT chats on the web kept working.
-
A task still "running" nine days later. NemuKei's task "Check osaka report sources" started around Oct 1. Stop returned "Could not stop this task" on Oct 4, the deliverable was finished on Oct 6, and on Oct 10 the Mac app still showed the task running, with a Stop button. Their other device didn't list it at all (#52777).
-
"Last active" frozen while chat works. On several Windows machines, MonoCo3217's Dot kept showing a Last active time of Oct 7, 19:44 UTC and a stale activity indicator, while messages went out and replies came back normally (#52767). On 15 Oct, kline reported the same split in the OpenAI Community: on 13 Oct the Windows app showed "Last active: 2 days ago" for a Dot that the web showed as active 3 minutes earlier, and app updates since then hadn't fixed it. A status line that doesn't move isn't proof the Dot has stopped; send it a message and see whether a reply arrives.
-
It said the work was running, but nothing was. On Oct 1, Nigel999 (macOS app) asked their Dot to restore an existing local application and report concrete progress. Over a long exchange the Dot kept saying the task was running or being handled, then admitted that no task was running and the app had not been restored; the menu steps it gave didn't match Nigel999's screen either. OpenAI Support replied by asking for the task's status under the Dot's profile, Activity: whether it showed a blocker, a failed step or a pending approval, and whether the Dot kept reporting progress after Activity showed the task had stopped (community thread). Kamimoto saw the same thing with Codex: no progress for 20 minutes until he asked "status?", and only then did the Dot find that Codex had failed (Codex page). If "still working" goes on for a while, open the task in Activity and check what it actually shows.
-
"In progress" that was really waiting for you. On Sep 30, linerking's Dot started a local task on a Mac that needed approval. The main Dot chat showed it only as "in progress", with no approval notice; the approval sat inside the desktop task, and linerking found and approved it about 16 minutes later. In the same issue, Wulfgardr (macOS 26.1002.52244, Oct 7–8) got stuck input requests to show up again by having the coordinating Dot read its child tasks one by one (#49531, comment). If a local task seems stuck, ask the Dot to read that task itself and tell you whether it is waiting for an approval.
-
It can no longer reply in Slack. On Oct 13, Joseph Vitko (@JosephVitko) found his Dot could no longer reply in Slack and called it a blocker. The same day Lethabo Lee Ntshudisane (@DistinctLee) wrote that Dots "still lose access to tools they already had" and logged feedback on it. When a connector that worked yesterday goes quiet, open the Dot's tools or plugin settings and check the connection is still there before you rewrite the job; for Slack setups, see Slack assistant.
-
It forgets which plugins to use, and drops tasks. On Oct 13, rununmanned (@rununmanned), who hands agents a wide mix of jobs across several companies, wrote that their Dot keeps forgetting the right plugins to use, says it can't do things a plain ChatGPT thread could do, and drops tasks and forgets about them completely. They still rated ChatGPT's voice models the best they have used. If this sounds familiar, name the plugin in the request itself, and ask the Dot at the end of the day to list what it still has open, so a dropped task shows up while you can still pick it up.
-
A connected app that keeps disconnecting. On 14 Oct, FX7 wrote in the OpenAI Community that one integration, their phone provider Quo, sometimes has to be reconnected five times in a single day. They run two Dots, and at night one of them sends a message that the connection dropped again; each time FX7 has to be at the computer and click a button to reconnect (community thread). If a job depends on an app that drops like this, ask the Dot to tell you the moment the connection goes, as FX7's does, so the job doesn't fail silently.
Before you assume it's your Dot, check the OpenAI status page. On Oct 7 it listed "Dot turn failures, delayed responses, and reduced proactivity" from 17:21 to 18:23 UTC, and "Dot in Codex may fail to create new threads" from 16:31 to 18:06 UTC. Both were marked resolved. If something time-sensitive should have arrived, ask the Dot directly ("did anything new come in from X since 9 p.m.?"). In anotb's case the lookup worked even though the watch had stayed quiet.
7. "Your dot is on a break"
This section is the short version. The full answer, with every stop message ("on a break", "resting", the preview limit and a delegated task's usage limit), the reports people sent and a step-by-step plan, is on "Your dot is on a break": what it means and what to do. Cost and allowance questions are answered in the FAQ.
The full message reads: "Your dot is on a break. Congrats you are in the very top users and have hit our abuse prevention limit. Check back in a bit!" Users have seen it while their normal plan allowance still had room. Talking with your Dot doesn't count toward ChatGPT usage limits and deeper work has its own allowance (Access), but this limit and when it lifts aren't described anywhere public. Here is how long it lasted for the people who timed it:
- About 13.5 hours. bshane54's Dot went on a break at 7:43 a.m. PDT on Oct 7 and answered again at 9:15 p.m., saying it could reply but hadn't checked that its background tools were back (OpenAI Community).
- More than 3 hours, with usage left. In the same thread, c2chq had asked their Dot at bedtime to build a Google Sheet of jobs they qualified for. The job never finished, and in the morning the Dot was on a break, although their $200 plan showed 47% usage left and about 59,000 credits.
- Under 2.5 hours, then about 4 hours the next day. GGA981 got a similar "Your Dot is resting" message on Oct 5 and again about 20 hours after it lifted on Oct 6. By then one task had been open for about 116 hours (OpenAI Community).
- About 1 hour 40 minutes. ivg-design's Dot stopped at about 6:50 p.m. EDT on Oct 6, kept showing the notice even right after they approved a pending task, and replied normally again by about 8:30 p.m. (#51540).
The message is still turning up: on 13 Oct, Mr Leefy (@SirLeefy) posted the exact wording above, without saying how long the break lasted.
A different message, "You've hit the preview limit. Try again at 4:10 AM", does give a time. rabbitdaisuke saw it after more than four hours of calls with their Dot, and in the same thread astroianu described their Dot going quiet from 1:48 a.m. to 9:48 a.m. two nights running.
OpenAI Support's reply in that thread couldn't give a cooldown length. It asked anyone still blocked to email support@openai.com with a screenshot of the message, the affected task, when it stopped, and their timezone. None of these users reported that Reboot ended a break early. If an overnight job matters, ask the Dot to save progress somewhere you can see as it goes, such as a Google Sheet or a file in the chat, so a break doesn't cost you the whole run.
8. The Dot's cloud computer says "Unavailable", or files are gone
This is different from a missing Dot: chat still works, but the Dot's own cloud computer, or the files on it, can't be reached. These come from Pro users in the public openai/codex issue tracker, plus one from Threads and one from X.
- "Unavailable", with no reason or retry button. On Oct 1, GGBondBlueWhale (macOS app) saw the cloud computer switch to "Unavailable" while chat kept working (#50115). The next day kmagpiedev reported "Offline / Computer unavailable" on Windows.
- Reboot didn't bring it back. aaronmorris-dev (Pro 100, macOS) found the computer Offline on Oct 4. Desktop control returned "failed to connect native pipe: Connection refused", and choosing Reboot several times didn't restore it (#50938).
- "Root executor preparation for your dot failed." werty4870-boop (Windows, Oct 3) first saw "Environment is not connected" and then this error. Reboot didn't help, the same failure showed on the web, and the cloud browser couldn't load google.com (#50545).
- Desktop gone after a process restart. JanSimeonMetzger (macOS, Oct 2) approved a script that restarted the cloud desktop's dock process. The desktop then showed "Computer unavailable" while chat, shell commands, and files kept working. The reporter warns against treating a dock restart as a safe fix, and used the remaining shell access to start a backup (#50407).
- A project folder missing the next morning. nervouna (Oct 3) had a 230-file project folder under
/workspace/sharedone evening and found it gone the next morning (#50577). - Most of the disk emptied mid-task. During active work on Oct 4, lufesiare's workers lost previously readable folders, and disk use dropped from about 29 GB to about 1.3 GB. Code already pushed to Git survived; unpublished work had to be rebuilt and tests rerun (#50985). Two more users said the same happened to them, one of them three times, roughly once a day, even with backups set up.
- Smaller versions of the same thing. After a week, Nicholas Wu hit files that wouldn't open, a prepared .md file that couldn't be found, and finished work that didn't pick up again in the original chat.
- "No usable execution environment is available." On Oct 9, elitink's cloud terminal and cloud browser stopped responding at about 09:46 UTC. By 10:02 new processes failed with "exec-server transport disconnected", and from 10:36 even a
pwdwas refused with that message. Chat and connected-app reads, including the Library, kept working, and the computer was still down at 10:53 UTC (#52531). - A Reboot that paused the Dot instead. COOLak got the same message after a file transfer timed out. Updating the Windows app to 26.1007.21434 changed nothing. On Oct 9, Reboot sometimes failed with "Couldn't request a reboot. Please try again." and sometimes showed "Reboot requested", but either way the computer stayed "Computer unavailable". After one Reboot the Dot was left paused until they chose Resume, and the computer was still offline after that. They sent in-app feedback and opened a support conversation (#52509).
- Recent files gone, older ones still there. On Oct 10, youcantbearme's cloud computer stalled during an offline Rust test. When commands worked again, recent worktrees, Git history and test results were missing while older project files remained, and disk use had dropped from about 30 GB to 3.3 GB. They hadn't touched Pause, Reboot or Delete (#52765).
- Two days of building, then the project files and the GitHub login were gone. On 15 Oct, ChAn (@WaleyCh) wrote that after nearly two days of building with Dots, the recent project files were nowhere to be found, and the
gh(GitHub CLI) login was gone too. ChAn said this had happened before and asked whether the workspace could be recovered. X - Files the Dot made but couldn't hand over. On Oct 20, Armywas's Dot generated Word and Excel files on the web but couldn't attach them to the chat, failing with "hosted apps tools/list request failed with HTTP status 403" before any upload started, while reading the Library still worked. Others saw the same on Android and when replacing saved files. NuLL_00 delivered files through Google Drive in the meantime, and on Oct 22 uploads through the normal route worked again for them (OpenAI Community).
- An image in the conversation that the Dot can't open. On Oct 22, fathdsci uploaded images straight into the Dot conversation on macOS (26.1016.61520), iOS (1.2026.279) and the web. The file showed up in the Library, but reading it failed with "HTTP 403: scope_violation" and "The requested file is outside the Library folders selected for this user turn." A ZIP and one screenshot opened, other new images didn't, and uploading again didn't help (OpenAI Community). There is no reliable fix yet, so if the Dot says it can't read an attachment, give support the time, the app version and the exact error.
The cloud computer "can keep its state between periods of use" (Connect computers and apps), but how long files are kept isn't stated. In lufesiare's case, what survived was the code already pushed to Git. Ask the Dot to keep finished work in Google Drive, a GitHub repo, or the chat as it goes. If chat and shell still respond while the desktop is down, have it copy anything important out before you try a reboot.
9. Errors when the Dot works on your own computer
- "The invoking thread is not attached to an active Aeon." On Oct 7, LucioLiu's Dot on macOS could no longer create tasks or list connected computers, though it could still read old task results (#51730). Within an hour about a dozen people reported the same. Restarting the app and reconnecting the computer didn't help. Rebooting the Dot did: classner, yolanmees, and Dean-AltAIus confirmed it on macOS, and ve6btc on Windows by clicking the Dot's name in the upper-left corner and choosing Reboot. For ve6btc, local commands still failed afterwards with "setup refresh had errors". The status page listed "Dot in Codex may fail to create new threads" for the same afternoon.
- One Mac connected, the other "offline". ashbhati's Dot could use a Mac mini but showed the MacBook Air in use at the time as offline (#51229).
- "Local Work executor registration failed (HTTP 409)". From Oct 2 to Oct 10, singkia's MacBook Air showed "Your computer / Connection unavailable", and the same error appeared on a second Mac and a Windows PC on the account. Sending the request again returned "Maximum number of local executor registrations reached". The account had 32 old registrations (25 named Mac-mini.local, 7 named Windows), none of them attached. Disconnecting one reported success, but it stayed on the list (#52709). The report found no setting that clears them, so this one is for OpenAI support.
- "Browser is not available" on a Mac. On Oct 9, chrislsavage could see the in-app browser but the app couldn't use it, and the Chrome extension reported "native transport disconnected", even though the Mac showed as connected (#52619). When the Dot tried to start a diagnostic task on that Mac, it got "DesktopTaskWorkspaceUnavailableError: The desktop could not prepare a task workspace." Reopening the browser, clearing the cache and re-granting Accessibility and screen-sharing permissions didn't help. shaynesherman found that telling ChatGPT to use a longer timeout let the browser connect in the end, after 93 seconds.
- It can't change a scheduled task it created itself. On Oct 13, Brandon (@BrandonVanB) posted that his Dot, Odin, couldn't change the very scheduled task it had created, and that the ChatGPT app itself rejected it. In his screenshot, Odin explains that the desktop tool blocks it from controlling the ChatGPT app itself, so it couldn't inspect the task editor on his Mac. It then asked him to open Scheduled Tasks, select the task "Keep Inboxes Useful", open its editor without saving, and send a cropped screenshot with any error. When a Dot needs to fix one of its own scheduled tasks on your computer, expect to open the editor yourself and show it what you see.
- "Could not complete setup" when you connect a Mac. On 14 Oct, mac_connect_report tried to connect their Mac to an existing Dot from ChatGPT Desktop and got "Could not complete setup. Please try connecting again" every time. Chatting with the Dot in the cloud still worked, but the Mac connection never came up. Restarting, signing out and back in, updating the app and switching networks didn't help, and they had already sent diagnostics to support (community thread). The next day no_idea_coder replied with the same problem on macOS 26.7.1, an M3 Pro and app version 26.1007.71326. No workaround had been posted by 15 Oct.
- Tests that kept running on your computer with no sign of it. On 15 Oct, ruscopeland wrote that they had asked their Dot to run some tests and told it to stop the beep it played after each run, not knowing the beep was the only sign it was working. Five days later a message said their computer had disconnected; asked why that mattered, the Dot said it needed the computer to keep running the tests. It had been running them over and over in a terminal on their computer, which, unlike a Codex task, shows no spinning item in the conversation list. It used up a week of their usage in one day, and when the usage reset, it simply started again (community thread). When you give a Dot a repeating job on your own computer, say when it should stop, and look at Activity in the Dot's profile now and then.
"Offline" on your own Mac or PC means that computer is currently unavailable. It doesn't remove the connection (Connect computers and apps). Check that the ChatGPT app is open and signed in to the same account on that computer before reconnecting anything.
Your Windows PC stays "Offline" and Allow access is greyed out behind a proxy
Several Windows users who reach the internet through a local proxy such as Clash found that ChatGPT itself worked, but their Dot couldn't reach the PC. The PC stayed Offline in the Dot's computer list and the Allow access button was greyed out, even while they were using the app on that same PC. Updating and restarting the app didn't change it.
- weite76 (Windows 10, Sep 30). They tested the connection with the codex.exe that ships with the desktop app. Ordinary HTTPS checks passed, but the WebSocket connection timed out after about 15 seconds. With an HTTP proxy set for that one test process, it connected in about 2.8 seconds. Putting proxy variables in a
.codex\.envfile made the app hang at its startup logo, so they removed that file again. What worked: with the app closed, they rancodex features enable respect_system_proxy, which savesrespect_system_proxy = trueunder[features]in the Codexconfig.toml. After a normal relaunch the PC was no longer Offline, Allow access worked, and a real read-only test on the PC succeeded. They back upconfig.tomlfirst and warn against replacing the whole file or adding a second[features]section (#49626). - diosmentiras (Oct 2), springmirage (Oct 8) and lxhcxyjpl-dotcom (Oct 10) confirmed the same setting on their own PCs. diosmentiras uses the Windows system proxy with TUN mode off; after the change the PC connected, access could be authorized, and a test command printed
dot-connection-ok(comment). springmirage, on Clash Verge Rev, had found that Clash's TUN mode also brought the PC online, but it made Tailscale, RustDesk and at times ChatGPT itself unstable. With the setting added, the PC showed Connected with TUN off (comment). lxhcxyjpl-dotcom, on app version 26.1007.11041, could connect and authorize the PC, but hadn't yet confirmed whether an intermittent "Failed to resume chat / Codex app-server is not available" also went away (comment). - kaileysong (Windows 10, Oct 8) went a different way. The Dot starts a separate
codex.exe exec-serverprocess, and that process didn't follow the desktop app onto the Clash system proxy; the logs kept showing failed WebSocket connections. TUN mode was out because it affects other software, so they had a launch wrapper written that sets the Clash proxy for that exec-server only, which fixed the connection. In a follow-up they noted the Dot still hands the actual work to Codex tasks and called itself only a "foreman" (linux.do, Chinese). - huangxuanminggithub (Windows, Oct 3) had the opposite picture. The PC was online and delegated work ran, but every entry in the Dot's Activity view showed the same generic "Task created" label, and opening any of them gave "Couldn't load this chat". The desktop log showed the app's main WebSocket closing with code 1006 about 15 seconds into each attempt. They use a ClashR proxy; switching it to global mode didn't help, and they hadn't tried
respect_system_proxy. The report is still open (#50561).
What to try: if ChatGPT works, you use a system proxy, and your PC shows Offline with Allow access greyed out, quit the app, back up config.toml, add the setting with the command above, reopen the app and run a read-only test on the PC. If you'd rather not touch the Codex config, TUN mode is what brought springmirage's PC back, at a cost to other network tools. If the PC is connected but Activity tasks won't open, that is the separate problem in #50561, and nobody has posted a confirmed fix for it yet.
"setup refresh had errors" on Windows: what it looks like and what worked
From October 7 a wave of Windows users found their Dot could still see their PC but couldn't run anything on it. Their reports, spread over 12 issues on OpenAI's Codex GitHub, describe the same pattern.
What it looks like. The PC shows as connected, attached and authorized. The Dot can create a task on it, and can often list projects and read old results. Then the first command in that task, even Get-Location, fails before PowerShell starts:
exec-server rejected request (-32603): helper_unknown_error: setup refresh had errors
The same command works when you open an ordinary local Codex chat on the same PC. Almost every report makes that comparison (#51882, #51735, #49953, #52334). It hit x64 and ARM machines (#51911, a Snapdragon laptop) and showed up after updating to desktop 26.1002.51308 or 26.1002.52244 (package 26.1002.7124.0). freevern's timeline is typical: everything worked until the October 7 app update, and the next update didn't fix it.
Why the Dot fails when ordinary Codex doesn't. Several reporters found the same line in their sandbox log: Windows error 32 ("the file is being used by another process") on node_repl.exe in the app's cua_node runtime folder (#51851, #52116, Ken-Chen-Ori). alecwardd explained the split. Every fresh Dot task gets a new dated folder (Documents\Codex\<date>\task, then task-2) as its write location, and a new folder forces the sandbox to re-check its runtime files. That check trips over the file the app itself is holding open. Ordinary tasks in an existing project usually skip the check, so they pass. Resuming the same Dot task doesn't help, because every command repeats the check. The underlying bug is #51634. A fix (PR #51822) was merged on October 7. alecwardd found it in command-line builds 0.162.0-alpha.20 and 0.163.0-alpha.2, while desktop 26.1002.52244 still bundles 0.162.0-alpha.2. On October 9 the Windows app updated to 26.1007.21434 (package 26.1007.2314.0, bundled command-line build 0.162.0-alpha.17.2). After that update, tadeogutierrez found that a fresh Dot task ran Get-Location with normal permissions and the error didn't come back, though browser control then failed with the new error described below. Not every PC had the update yet: on October 10, rafimyu580-pixel was still on 26.1002.52244 with the update check saying it was current.
What people tried that didn't help. Rebooting Windows, restarting the app (including killing it in Task Manager), disconnecting and reconnecting the PC, revoking and re-granting access, signing out and in, reinstalling, Windows Repair, and creating fresh tasks. Rebooting the Dot fixed the separate "active Aeon" error above, but not this one.
What worked for individual reporters. None of these is an official fix, and each comes from one or two people:
-
Run each command with your approval, outside the sandbox. jreverett found that the identical
Get-Locationsucceeded when the task requested escalated permission, and alecwardd confirmed it in a fresh Dot task. Every command has to ask, because anything run the normal way still fails. -
Roll back the app. ve6btc installed package 26.930.7945.0 and the same Dot task ran
Get-Location, a directory listing andgit --version. The error came back after a later update. -
Change a local setting. fuler2026 recovered after turning on
features.prefer_mxcand restarting the local executor. huangkai-boy switchedsandboxunder[windows]in~/.codex/config.tomlfrom "elevated" to "unelevated". JoonSeongJeong installed the 0.162.0-alpha.20 command-line build with the standalone installer. Andyhsui confirmed the "unelevated" switch on another PC on Oct 9. -
Let the app update itself. motoradmir-ship-it's Windows virtual machine still failed after a full reboot on the evening of Oct 9. Right after Codex updated itself, the same read-only command ran with exit code 0, with no settings changed; the new version number wasn't recorded.
-
Update, then test the route the Dot actually uses. Creer, a civil-engineering design firm, added a postscript on 10 October to its write-up of this error. An older office PC turned out to be on the same 26.1002.52244 build. It was offered an update to 26.1007.21434, and afterwards a new Work could list file names; Creer hadn't run the same test before updating, so the update isn't proven to be what fixed it. The PC with the error said it was up to date on 26.1002.52244. Installing from OpenAI's Windows download page then stopped the app from starting at all, with "failed to initialize sqlite state runtime". A Windows restart, with no settings or app data deleted, brought it up on 26.1007.21434. The Dot then restarted the Work that had been failing, confirmed it was running on that PC and listed the work folder: 63 items, exit code 0. Japanese file names came out garbled until that step's output was set to UTF-8. Reading, editing and saving file contents were still untested, and Creer is keeping staff on the old version until a normal day's work runs end to end. "Up to date" on screen settled nothing; a small Dot → Work → PC check did.
While you wait, work that doesn't need your PC can run on the Dot's own cloud computer, and the Codex orchestration page shows how others split jobs between the two.
An ask to copy before work on a Windows PC.
Before you start work on my Windows PC, open one task and run only
Get-Location. If it fails, retry once at most, then stop. Tell me the exact error, the folder the task was given, and the desktop app version, and don't open new tasks to try again.
This avoids the Dot opening task after task. Each fresh task gets a new folder and fails the same way (alecwardd; Jowjm's Dot reported the same two errors after repeated attempts).
Browser control fails with "HRESULT 0x80070003" after the 26.1007 update
From October 9, Windows users whose Dot could run commands on their PC again found it couldn't use the browser or the screen. The first cua.getState() call fails before any tab list or page appears:
MXC launcher: cannot locate required Windows platform directory: HRESULT 0x80070003
mrnguyenlk-byte opened #52407 on Oct 9, right after a setting change had brought shell commands back on 26.1002.52244. evanziwen, Stage4000 and Jollendi91 then hit it on desktop 26.1007.21434 (package 26.1007.2314.0), with the Chrome extension and the in-app browser alike. Ordinary shell commands kept working on the same PCs.
What reporters found. BladexZN traced it to the SystemDrive environment variable: a new process started without it can't look up the ProgramData folder and gets exactly this error. On another PC, alleng666-cmd found that the same lookup had already been failing quietly on Oct 8 under the older 0.160.1 core, while browser control still worked. It only became fatal after the switch to 0.162.0.
What didn't help. Restarting the app, the browser and the PC, and a successful Windows Repair of the app (kaza0418-design). BladexZN's own allow_mxc = false workaround worked for a while and then failed again the same day. They withdrew it and asked others not to rely on it.
Still open on October 10. Later reports add more of the same. nothismann and yuseong2099-bit ran the official Microsoft Store Repair through WinGet. It completed, and the next cua.getState() still failed. nothismann also saw a newly created delegated task fail on its first try, so it isn't only older tasks. vgg7pjtbrt-a11y hit it on two separate physical Windows PCs. In #52858, kidsfunlearningsc-coder had Any App, Chrome, Edge and Excel turned on in the Computer use settings, and file access worked after a full Windows restart. The Dot's new local task still failed with the same error, and the tool said "Native computer APIs are disabled". Rebolvsk restarted both Chrome and the app and got the same error in a fresh runtime. By the afternoon of Oct 10, nobody in these threads had reported a fix from OpenAI.
A route that worked in one setup. Pi-Alba-Medi8 asked their Dot to start a new task the same way as a conversation that was still working. The Dot created it through the local-session route (host local instead of durable), and that task read the current Chrome page, took screenshots and carried on with the document work. Recreating tasks the old way hadn't helped. A second local task, though, used window-based control instead of the Chrome tabs and stopped because it couldn't confirm the page URL. They describe it as a working route for one task, not a fix.
Two asks to copy.
Before any browser work on my Windows PC, run one
cua.getState()check. If it fails with "HRESULT 0x80070003", don't retry it and don't open new tasks for it. Tell me, finish the parts that only need the shell, and list the browser steps that are waiting for me.
For these reporters the shell kept working while the browser didn't, so this keeps the shell-only part of a job moving and stops the Dot from looping on a step that fails the same way every time.
If a command on my PC fails with "setup refresh had errors", don't retry it the same way. Ask me to approve running that one command outside the sandbox, with a one-line reason, and ask again for each command after that.
This avoids silent retries of a route that can't work right now, and it keeps every unsandboxed command behind your yes (jreverett, alecwardd).
10. Calls ring but never connect
The longest thread is #50870: calls to the Dot ring and then show "Couldn't connect" or "Could not connect call", on iPhone, Android, Mac, and web, on Wi‑Fi and cellular, while normal ChatGPT Voice and texting the Dot keep working. Three things helped people there:
-
Restart the Dot, not the app. sffiltech-openclaw had failures on a MacBook Air, a Mac Studio, and an iPhone. Restarting the devices and pausing and resuming the Dot didn't help. Restarting the Dot's own computer from its computer controls did. BryanYue fixed the same thing with the dot-specific restart and closed the report (#51533).
-
Text it while it rings. FLi79Za, who opened the thread, found that typing to the Dot while the phone was ringing and asking it to answer worked several times.
-
When Reboot itself fails. KingIVthe1st got "Couldn't request a reboot. Please try again.", and mrkillbob saw no response at all from Reboot on Oct 8. Both added their in-app feedback reference to the thread.
-
Update the app first. In #49499, where calls failed within seconds on desktop, web and iOS, dmzu-cn said updating fixed it on Oct 2, and hmchen fixed it on Oct 7 by rebooting the Dot from its right-click menu in the desktop app.
-
OpenAI's 8 October fix, and reports after it. In the OpenAI Community thread on failed calls, OpenAI Support said on 9 October that fixes for Dot call connection problems had been rolled out on 8 October, and asked people to try again. Not everyone got through. Shismo still had calls timing out on 10 October while ordinary ChatGPT Voice and texting the Dot worked, and kenrend reported on 11 October that iPhone calls stayed on "Calling…" on both Wi‑Fi and 4G, even after rebooting the Dot. Earlier in the same thread, Saentzee, whose calls had failed on Windows desktop and web with "Voice transport hangup was not confirmed" in the logs, started and ended an ordinary Voice conversation in a new chat, then called the Dot again, and the call connected. Saentzee describes it as one successful attempt, not a confirmed fix. If calls still fail after updating and rebooting, note the date, device and network for support (see section 13 below).
The call connects, but what you ask for never becomes work
A different failure shows up once the call is up. Amplifeye (OpenAI Community, 6 October, Windows desktop app and Android) asked the Dot on a call to start a task, update ongoing work or just check a document, and every handoff came back with "Not sent. The spoken context could not be confirmed yet" and the code voice_parent_context_pending. Ending the call, restarting both apps and retrying didn't change it; the same request typed into the chat went through. A day later DevinOxman hit the same error, then paused and resumed the Dot. After that, requests from the call reached the task, but messages coming back from the task still failed, so on the Dot's own suggestion DevinOxman had those replies sent to the text chat instead. If you see this code, type the request into the chat and check there that the work started. The voice calls page has more reports of spoken requests that didn't turn into work.
Still unanswered or hit-and-miss after the mid-October fix
Connection fixes rolled out around 8–9 October did not clear every call problem. DistinctLee (10 Oct) still had a Dot that never answered on mobile on the latest iOS ChatGPT app, while chat worked. Shreyansh Parakh (13 Oct) separately called Dot calling "hit and miss" and asked for a Home Screen shortcut because ChatGPT → Dot felt laggy. If your call still fails after an app update and a Dot restart, keep the job in the text chat and check the voice calls page for more first-hand reports.
Calls that end when you leave the chat, and voice missing on Windows
- Leaving the chat ends the call on Android. On 24 October, JipyTech found that a call with the Dot in the Android app ended as soon as they left that chat to look at something else in ChatGPT (OpenAI Community). Stay in the Dot's chat until the call is over.
- Voice gone from the Windows desktop app. On 25 October, djyonicohen's Windows app (26.1021.31345) showed "Voice chat unavailable" and "Your account or workspace does not have access to voice chat" in the voice settings. The call button for the Dot and regular voice chat were both missing, and the Dot was gone from the sidebar, reachable only by searching for its name, Lucy, while calls still worked on the phone and the web with the same account (#56846). The check that hides the Dot when
model_providerisn't"openai"covers voice too, so look at that line first, and call from the phone or the web in the meantime.
More call cases, including the ones that worked, are on the voice calls page.
11. Reboot vs Delete
- Reboot restarts the Dot. On desktop, click your Dot's name (upper left) or right-click it in the sidebar, then choose Reboot. By Oct 9, COOLak could only find it in the sidebar right-click menu (Rename, Pause, Reboot, Delete) of the Windows app 26.1007; for them the conversation header menu and the web and Android apps no longer offered it. If the menu shows Resume instead of Pause after a Reboot, the Dot has been paused, so choose Resume. One user quotes its confirmation as saying conversations, memories, and scheduled tasks are kept, though work in progress may be interrupted. OpenAI's docs don't describe Reboot yet; a user has asked for that. In the reports above it fixed the "active Aeon" error and stuck calls for several people. Nobody reported it ending an "on a break" pause, and it didn't bring back aaronmorris-dev's or COOLak's offline cloud computer.
- Delete, which the Help Center still calls Reset, removes your Dot, including its conversations, saved memories, and scheduled tasks (Getting started with your dot). It can't be undone, and it doesn't undo changes already made in your apps or recall messages already sent to other people (Manage your dot). Afterwards you create a new Dot in the ChatGPT mobile app, the desktop app or a desktop browser. On Oct 8, txchiang reported that after deleting the Dot from the Windows sidebar, there was no apparent way to restore it (#51998).
Try Reboot first. Save anything you need before you consider deleting.
Deleted your Dot to start over, and now it won't load
Some people delete the Dot on purpose, hoping a fresh one will clear a problem. On Oct 7, ewerton.lourenco (Pro 500, Windows 11, Chrome) deleted theirs to set it up again. The deletion looked fine, but the "Your dot" page then showed only "Failed to load your dot". Try again did nothing, and there was no way to start setup again. liahua01170 had the same problem the next day, and garymjr saw "Couldn't load your dot" after deleting their Dot for the same reason. Jens (@JensLee1e8) had only one Dot on a Pro account. On Oct 11 Jens deleted it from the menu, so that its existing context wouldn't influence a new idea. Afterwards the Dot's name was gone, the entry just said "Your Dot", and the chat pane on the right couldn't load it. There was no Initialize or Rebuild option to start again (X article). Two days later Jens wrote that it still wouldn't launch at all (X).
How it ended. On Oct 8 at 17:34 UTC, about a day after deleting, ewerton.lourenco could create a new Dot and connect a computer. They hadn't changed anything, so it looks like it was fixed on OpenAI's side. On Oct 10, Fotios wrote that after deleting, they could recreate their Dot only from the web, and on Oct 12 DRE.BLOCK said the same: it kept failing in the app and worked on the web. In the same thread dzguy, whose calls had failed for over 24 hours, decided not to delete theirs after reading it.
It didn't end for everyone. On Oct 14 yardenisraelbenaviv still got "Couldn't create your dot" and "Something went wrong during setup" in Chrome on Windows, the Windows desktop app and the iOS app, so the web didn't help them, and their case had been escalated to a support specialist. On Oct 18 unsignedint wrote that their Dot had been dead for five days. Leon-Rivers deleted a Dot through Codex, and after that every attempt to create a new one failed in Chrome, Incognito, iOS and Codex: the create request returned HTTP 409 with cloud_aeon_recovery_required, asking them to recover or delete an existing Dot, while their Dot list was empty (#55120). On Oct 19 vnt passed on support's answer: OpenAI knew about the problem but had no clear plan to solve it yet.
The safer route. None of the fixes above needed a delete. Calls (section 10), an offline cloud computer (section 8) and a missing sidebar entry (section 5) each came back for someone with a narrower step. If you still want to start over, do it in a desktop browser, and expect that the new Dot may not load right away, or for several days.
Send this before you delete. It asks for everything that deleting removes.
I'm thinking about deleting you and starting over. Before I do, write one list I can save: every scheduled task with its time and exact wording, the custom rules I've given you, the apps and computers you're connected to, and what you remember about me that I'd have to tell a new Dot again. Don't delete or change anything.
12. "user cancelled MCP tool call" when you didn't cancel anything
Sometimes a Dot stops partway through a job and reports that a tool call was cancelled by you, with the exact words "user cancelled MCP tool call". You didn't press anything, and in the reports that mention it, no approval card appeared either. The Gmail and GitHub guides already cover this message for email sends and large uploads (odedd, cedar_circuit, tndd, konard). The reports below show it in other places: Slack, a private mail server, a spreadsheet, a work-management app, the Dot's own task tools, and one account where every tool call has returned the same message for more than two weeks.
Where it has shown up: Slack, a mail server, a spreadsheet and task tools
-
Slack, twice. gebezis began seeing the message on Oct 2, on both reads and writes in a connected app. In one case an update wasn't saved, and the Dot confirmed the old version was still there. An explicit "try again" finished the job (#29857). On Oct 8, a Slack message gebezis had authorized came back with only the cancellation text. A read of the destination thread showed no message, and a later authorized retry went through (#50769).
-
A scheduled check that only reads. Zachariah-Jack's Dot ran a daily review of a JobTread account. On Oct 3, 4 and 5, one read in each run (an activity feed, a schema lookup, a related-task read) returned the message, so none of the three runs completed. On Oct 5 Zachariah-Jack told the Dot they hadn't cancelled anything and asked it to carry on. The retry finished the cross-check about a minute later (#29857).
-
A send that did happen. EzraCerpac's Dot uses a privately hosted mail server. On Oct 6 two calls failed this way, and the server's logs showed no matching request at all. But on Oct 5, the server received a call, ran it and returned success in about 1.6 seconds, while the Dot still reported that the user had cancelled it. "Allow all actions" was already switched on for that app (#29857).
-
Gmail and a new task. On Oct 7 kairytwo2 saw the message on a Gmail send and on creating a cloud task. After checking that no sent email and no new task existed, authorized retries worked. Two identical package installs failed with a related message, "automatic approval review was cancelled", even after explicit authorization (#29857).
-
Half of a read-only check. On Oct 10 iamtornado authorized a read-only look at a spreadsheet. Listing the sheets worked; reading a range from the same file came back cancelled, with no error code or request ID. One ChatGPT Space page update was cancelled the same way while another one went through (#50769).
-
A task that was told to do nothing. On Oct 10 dean1012 (Windows app 26.1007.21434, Pro Lite) asked their Dot to create a test task in a saved cloud environment, with a prompt that only asked for an acknowledgment. The create call came back cancelled after about 17 seconds, with no task ID, and no new task showed up in the task lists. dean1012 approved a retry, which worked about 20 seconds later. Their earlier cases included two task messages on Oct 5 that never arrived and a GitHub merge on Oct 6 that left the pull request open (#52916, summary).
-
Still every call, more than two weeks later. On 14 Oct, Uzulor (@Uzulortft) wrote that for more than two weeks their Dot had stayed unusable: everything they asked returned "user cancelled MCP tool call". They asked not to reset the Dot, only to be able to use it again. X
What the message does and doesn't tell you
OpenAI hasn't explained the message in these threads. odedd pointed out that in the public Codex source, a failed or missing approval answer ends up with the same wording, which would fit dean1012 seeing no prompt at all.
- "User cancelled" doesn't mean you did anything. None of the seven had cancelled. dean1012 saw no prompt, and kairytwo2 couldn't say whether one had appeared.
- The action may or may not have happened. gebezis's Slack message and dean1012's task weren't there. EzraCerpac's mail call had actually run. So you can't tell from the message alone whether a retry will create a duplicate.
- It hits reads too. Zachariah-Jack's and iamtornado's failures were read-only lookups. It isn't only risky writes waiting for permission.
- One clear retry usually worked. In the four reports that describe retrying the cancelled action (gebezis, Zachariah-Jack, kairytwo2, dean1012), one explicit retry went through. kairytwo2's package installs, which failed with the related approval-review message, did not. For a scheduled run, though, nobody retries automatically: Zachariah-Jack's daily reviews simply stayed incomplete until someone noticed.
What to do when you see it
- Write down the time of the failed call (in UTC if you can) and what it was meant to do.
- Look at the destination yourself, or ask the Dot to read it back: the Slack thread, the Sent folder, the sheet, the task list.
- If nothing changed, ask for exactly one retry of that same action. If it did change, don't retry.
- If it happens inside a scheduled job, check that run's result instead of assuming the next run will fill the gap.
- If it keeps happening, send support the times and the destination check from steps 1 and 2.
An ask to copy:
If a tool call comes back as "user cancelled", don't treat it as my decision. First check whether the action actually happened in the destination and tell me what you found. Retry once only if it didn't happen, and tell me if the retry fails too.
13. What to send support
If the account meets the published plan and region rules and enough time has passed for rollout, collect the plan name, workspace type, region, app version, entry URL, and the on-screen message. For a Dot that stopped working, add the exact error text, the time in UTC, and what you already tried (answering pending requests in Activity, Pause and Resume, Reboot, app restart, another device). For a "user cancelled" error, also say whether the destination changed.
Related
Copy this task brief
Can't see Dots? Check plan, region, rollout and the Enterprise switch. Dot missing from the desktop app, silent, or unable to read and save files? Open it from search or the web, answer pending requests, try Pause and Resume or Reboot, and keep Delete for last.Download Markdown ↓