Codex Desktop 26.616.71553 on Windows: Unsupported content type, queue/task mixing, disappearing sessions, broken model picker, unstable sidebar
What version of the Codex App are you using (From “About Codex” dialog)?
Codex Desktop App: Version 26.616.71553 Released: Jun 23, 2026
What subscription do you have?
ChatGPT Pro / highest Pro plan
What platform is your computer?
Windows Microsoft Windows NT 10.0.26200.0 x64 Device: ASUS Zenbook 14 OLED UX3405 CPU: Intel Core Ultra 9 185H RAM: 32 GB
What issue are you seeing?
What version of the Codex App are you using?
Codex Desktop App: Version 26.616.71553
Released: Jun 23, 2026
Previous build note:
The most stable recent build for me was the one released on Saturday. After that, the newer builds started introducing serious regressions. It would be very helpful to have a Beta / Stable channel or a way to roll back to a known-good build, because current updates are repeatedly breaking production work.
What subscription do you have?
ChatGPT Pro / highest Pro plan.
What platform is your computer?
Windows
Microsoft Windows NT 10.0.26200.0 x64
Device: ASUS Zenbook 14 OLED UX3405
CPU: Intel Core Ultra 9 185H
RAM: 32 GB
The computer and Windows environment have not changed compared with the previous reports.
Which model were you using?
Mostly GPT-5.5 / GPT-5.5 XHigh / High / Fast depending on the available selector.
Important: in some old chats/sessions, the model selector is now unavailable, so I cannot always select or verify the intended model.
What issue are you seeing?
Codex Desktop 26.616.71553 introduced several serious regressions on Windows.
This is a production-blocking issue for a paid Pro user. I use Codex with long queues, many tasks, and project sessions. These regressions make the app unreliable and unsafe for real work.
1. Frequent {"detail":"Unsupported content type"} failures
After updating to Codex Desktop 26.616.71553, I started frequently seeing:
{"detail":"Unsupported content type"}
This did not happen before.
There are two patterns:
- Sometimes I paste or submit a command/task and it immediately fails with
{"detail":"Unsupported content type"}. - Sometimes the task starts normally, works for around 10 minutes, and then stops with
{"detail":"Unsupported content type"}.
This is especially dangerous with long queues. If there are 20–50 queued tasks, Codex may fail a large portion of them with Unsupported content type, then continue to the next queued item. A user may not notice that 30–50% of the queue silently failed.
2. Queue/task context mixing
There is now a serious queue regression.
Sometimes the queue starts the next queued task, but Codex continues working on the previous task or applies the previous task context to the next queued item.
This is very harmful for real work because the task shown as active and the task being executed may not match.
Expected:
When the queue starts task N+1, Codex should execute task N+1 only.
Actual:
Codex sometimes appears to start the next queued task but continues the previous one or mixes context between queued tasks.
This can corrupt work, waste usage limits, and make the queue unreliable.
3. Left sidebar flickering for 1–3 minutes on startup
After launching Codex Desktop, the left sidebar/menu now flickers for 1–3 minutes.
The sidebar repeatedly appears and disappears.
A very short flicker used to happen before for maybe a millisecond, but now this is a visible and prolonged UI artifact. It looks like the app is repeatedly mounting/unmounting or rehydrating project/session state.
This makes the app feel unstable immediately after startup.
4. Chats/sessions created after the new build disappear automatically
The latest patch introduced automatic disappearing of chats/sessions inside projects.
This affects sessions created today after the new build was installed.
Interestingly, sessions created earlier, for example around Saturday Jun 20, appear to be less affected or not affected in the same way.
This suggests a regression in the current build’s session/project persistence, indexing, or hydration logic.
5. User has only around 5 seconds to pin a chat before it disappears
There is now a very painful workaround:
- Restart Codex Desktop.
- Immediately open the chat.
- You have around 5 seconds to pin it.
- If you do not pin it in time, the chat disappears again.
- Then you need to restart the app or even reboot the computer and try again.
This is not usable behavior.
Expected:
Chats/sessions should remain visible and accessible in their projects.
Actual:
Some newly created chats/sessions disappear unless they are pinned immediately during a short startup window.
6. Model behavior / quality became inconsistent and significantly worse
Model behavior has also become unstable.
Earlier in the month, GPT-5.5 / XHigh in Codex was much more useful and reliable.
Recently, the behavior became inconsistent:
- some requests still look reasonable;
- some requests behave like a much weaker model;
- the model often misses instructions;
- the model sometimes produces shallow or incorrect work;
- quality feels far below what GPT-5.5/XHigh provided earlier in the month.
This is especially frustrating because I am a loyal Pro user and rely on Codex for production work.
I understand that model behavior can be separate from the desktop app bugs, but from the user side the result is the same: Codex became unreliable.
What steps can reproduce the bug?
Reproduction path A: Unsupported content type
- Open Codex Desktop 26.616.71553 on Windows.
- Open an existing project.
- Add multiple tasks to the queue.
- Submit normal text commands/tasks.
- Observe that some tasks immediately fail with:
{"detail":"Unsupported content type"}
- In other cases, let a task run for around 10 minutes.
- Observe that it stops with the same error:
{"detail":"Unsupported content type"}
- The queue then continues to the next task, making the failure easy to miss.
Reproduction path B: queue/task context mixing
- Open a project with a task queue.
- Add several queued tasks.
- Let the queue auto-continue.
- Observe that when the next task starts, Codex may still act on the previous task or previous context.
- The visible active queue item and the actual work being performed can diverge.
Reproduction path C: sidebar flicker / project hydration instability
- Launch Codex Desktop 26.616.71553.
- Watch the left sidebar/project menu.
- For 1–3 minutes the sidebar repeatedly appears and disappears.
- This did not happen at this duration/severity in earlier builds.
Reproduction path D: disappearing new chats/sessions
- Open Codex Desktop 26.616.71553.
- Create a new chat/session inside a project.
- Close or restart Codex.
- Reopen Codex.
- The newly created chat/session may disappear from the project.
- If you restart and quickly open/pin the chat within around 5 seconds, it may remain accessible.
- If you miss that short window, it disappears again.
What is the expected behavior?
- Normal text commands should not fail with
{"detail":"Unsupported content type"}. - Long-running tasks should not stop after several minutes with
Unsupported content type. - Queue execution should be strictly aligned with the visible queued task.
- Queued tasks should not inherit or continue the previous task’s context unless explicitly requested.
- The left sidebar should not flicker for 1–3 minutes after startup.
- Newly created chats/sessions should not disappear from projects.
- Users should not need to restart Codex and pin a chat within 5 seconds to keep it visible.
- The model selector should remain available in old and new chats.
- GPT-5.5 / XHigh behavior should be consistent and should not randomly degrade to much weaker behavior.
What is the actual behavior?
- Many tasks fail with
{"detail":"Unsupported content type"}. - Some tasks fail immediately.
- Some tasks run for about 10 minutes and then fail with the same error.
- Queue continuation can hide these failures.
- Queue tasks can get mixed with previous task context.
- Sidebar/project menu flickers for 1–3 minutes on startup.
- New sessions/chats created under the current build can disappear.
- The user has a tiny window of around 5 seconds after restart to open and pin a chat before it disappears.
- Model selector is unavailable in some old chats.
- Model quality and instruction following became inconsistent and significantly worse.
Impact
This is production-blocking.
I am a ChatGPT Pro user on the highest plan and have been unable to use Codex reliably for serious work.
The current app behavior can silently lose work, corrupt queues, waste usage limits, and make long task queues unsafe.
This is especially damaging for users who rely on Codex as a working environment rather than a demo tool.
Regression information
This appears to be a regression introduced after the more stable Saturday build.
The Saturday build was significantly more stable for me. It would be reasonable to provide a Beta / Stable channel, or an official rollback option, so users are not forced into builds that break their work.
Request
Please escalate this to the Codex Desktop Windows team.
Please investigate:
Unsupported content typeerrors in normal queued text tasks.- Queue/task context mixing.
- Sidebar flickering / project hydration instability.
- New chat/session disappearance after restart.
- The 5-second window where a chat must be pinned before disappearing.
- Model selector unavailable in old chats.
- Inconsistent GPT-5.5 / XHigh behavior in Codex.
- Whether this is related to local project/session index, model catalog, app-server state, or queue persistence.
- Whether a rollback / beta / stable channel can be provided.
Please also provide guidance on:
- which logs are safe to share;
- whether logs may include prompts, code, screenshots, project data, secrets, or private information;
- how to export or recover disappearing sessions;
- whether there is a safe local reindex/rebuild procedure for Codex Desktop project/session state.
Additional information
This follows previous instability in Codex Desktop on Windows, including severe memory growth and inaccessible project/session history in earlier builds.
The machine and Windows environment did not change. The regressions appeared after Codex Desktop updates.
I can provide screenshots, exact timestamps, logs, and screen recordings if needed.
Additional high-severity update: {"detail":"Unsupported content type"} kills running tasks and makes queues unsafe
I need to emphasize that {"detail":"Unsupported content type"} is not a minor error message.
It is a severe queue-breaking bug.
Today, at least 20 tasks were affected by {"detail":"Unsupported content type"}.
The failures happened in multiple stages:
- At task start
- The task starts and immediately fails with
{"detail":"Unsupported content type"}.
- In the middle of a task
- The task works for several minutes, then suddenly stops with
{"detail":"Unsupported content type"}.
- At the final stage of a task
- The task appears to be close to completion, then fails with
{"detail":"Unsupported content type"}and the useful work is lost or left unfinished.
This is especially damaging because Codex currently has no reliable “stop queue” / “pause queue” / “cancel all queued tasks” control.
So when this error starts happening inside a long queue, Codex continues to the next queued task automatically.
If a user has 20–50 queued tasks, like I often do, the app can silently burn through the queue while many tasks fail with {"detail":"Unsupported content type"}.
This creates several serious problems:
- running work is killed;
- partially completed tasks are lost or left in an unknown state;
- queued tasks continue after failure;
- the user may not notice the failures until many tasks are already damaged;
- usage limits are wasted;
- work output becomes unreliable;
- it is not safe to leave Codex running unattended;
- a long queue becomes dangerous instead of useful.
This is a production-blocking issue for anyone using Codex queues seriously.
Expected behavior:
{"detail":"Unsupported content type"}should not happen for normal text tasks.- If this error happens, Codex should stop the queue automatically or ask the user what to do.
- The user should have a clear “Pause queue” / “Stop queue” / “Cancel remaining queued tasks” control.
- The failing task should be marked clearly as failed.
- Codex should not automatically continue running the next queued tasks after repeated unsupported-content failures.
- Codex should preserve the failed task state and make it easy to retry.
Actual behavior:
{"detail":"Unsupported content type"}can happen at the start, middle, or final stage of a task.- It kills or interrupts running tasks.
- Codex continues with the queue.
- There is no reliable stop-queue control.
- Many queued tasks can be affected before the user notices.
Impact today:
- At least 20 tasks were affected by
{"detail":"Unsupported content type"}. - The failures happened both at task start and during active execution.
- Some tasks failed after substantial work had already been done.
- This made the queue unsafe to use.
Please prioritize this as a high-severity queue/runtime bug, not just a cosmetic error message.
<img width="1359" height="603" alt="Image" src="https://github.com/user-attachments/assets/136acf5f-4661-4bfc-b603-edd10acb7648" />
What steps can reproduce the bug?
Update after Codex Desktop 26.616.71553, Released Jun 23, 2026:
The new build introduced several serious regressions on Windows:
- Frequent
{"detail":"Unsupported content type"}failures.
- Sometimes immediately after submitting a task.
- Sometimes after ~10 minutes of work.
- With long queues, many failed tasks can be missed because Codex continues to the next queued item.
- Queue/task context mixing.
- The queue starts the next task, but Codex continues working on the previous task or previous context.
- Left sidebar/project menu flickers for 1–3 minutes after startup.
- Previously this was only a very short artifact; now it is prolonged and disruptive.
- New chats/sessions created after this build disappear from projects.
- Sessions from the earlier Saturday build appear less affected.
- The only workaround is to restart Codex and pin the chat within ~5 seconds before it disappears.
- If I miss the short window, I need to restart again.
- Model selector is unavailable in some old chats/sessions.
- GPT-5.5 / XHigh behavior became much less consistent and significantly worse than earlier this month.
This is production-blocking for a paid ChatGPT Pro user.
Please escalate to the Codex Desktop Windows team and investigate session/project persistence, queue persistence, local index/state hydration, model catalog, and the Unsupported content type failures.
What is the expected behavior?
_No response_
Additional information
_No response_
7 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Update after Codex Desktop App Version 26.616.81150 • Released Jun 23, 2026
This update is even worse than the previous one.
All previous problems are still present, but this build introduced two new severe regressions:
After opening Codex Desktop, old sessions can take around 10–15 minutes to load.
This makes the app almost unusable for real work. A project cannot be opened and used normally anymore because Codex spends a very long time trying to load session history.
Expected behavior:
Actual behavior:
This is a new bug I have not seen before.
After this update, most sessions are cut off visually. In many sessions, around 90% of the conversation is missing or inaccessible in the UI, and I can only view around 10% of the session.
This is extremely serious because it makes existing work impossible to review or continue.
Expected behavior:
Actual behavior:
* Around 90% of the session appears to be cut off or inaccessible.
This is in addition to the previous problems already reported:
{"detail":"Unsupported content type"}failures;Impact:
This build makes Codex Desktop unsafe for production work.
The app now has:
Please escalate this to the Codex Desktop Windows team.
This looks like a serious regression in session hydration, local state/index loading, project history rendering, and queue/runtime handling.
Please provide:
I want to report a recurring Codex Desktop workflow issue.
after Codex Desktop App Update - Version 26.616.81150 • Released Jun 23, 2026
In a local structured project, routine single intake tasks used to take about 3-7 minutes. Recently, after automatic context compaction/reconnection, the agent often spends 20-40 minutes on ordinary cases before even creating the case artifact.
Example behavior:
Expected behavior:
Observed impact:
Feature request / fix:
Please add stronger compaction continuity for Codex Desktop:
and this is what usual does happen when you try submit report via CODEX (new features appears 3-5 updates ago, never has happened before that) memory consumption is around 50% only
An error has occurred
Codex crashed with the following error:
(code=3221226505, signal=null).
Most recent error: 5388746967 bytes failed
note: run with
RUST_BACKTRACE=1environment variable to display a backtraceAnd after 5 crashes in a row, a lot of queued tasks launched one after another, each time app was reloaded. They all got mixed together, and now I don’t understand what to do because Codex started mixing the data.
reporting issue cause crash in 80% cases
<img width="1373" height="1000" alt="Image" src="https://github.com/user-attachments/assets/221fa3c1-378a-4f8d-8f62-bfda39f1c385" />
Version 26.616.81150 • Released Jun 23, 2026
I want to report a recurring Codex Desktop issue with automatic context compaction causing repeated rediscovery loops and very slow task completion.
Context:
I am working in a large local structured repository with documented fast-path rules, validators, and task-specific workflows. Some routine tasks should normally finish in a few minutes.
Problem:
After automatic context compaction, Codex often repeats discovery work that was already completed before compaction.
Example:
I gave Codex a tightly scoped maintenance task:
Observed behavior:
This also happens in routine data-entry/update tasks:
Expected behavior:
After context compaction, Codex should preserve the current task state:
If something was already confirmed before compaction, Codex should not repeat that discovery unless:
Impact:
Suggested improvements:
“I already checked X; continuing to Y”
instead of checking X again.
This is not just normal slowness. It appears to be a compaction/recovery loop that causes Codex to repeat the same bounded inspection several times.
Codex app bug report: project chats disappeared twice within 3 days.
Version: 26.616.81150, released Jun 23, 2026.
Platform: Windows Codex desktop app.
Issue:
Two separate project chats/sessions disappeared from the Codex project sidebar while I was working in the app. The latest chat was created on Jun 25, 2026 and disappeared during active use, without closing, restarting, or leaving Codex.
Additional context:
This happened during a 3-day period when Codex versions had already changed multiple times. The latest occurrence happened on the latest available build at that time: 26.616.81150.
Expected:
Project chats should persist in the sidebar, remain recoverable, or appear in archived/filtered thread views. An active chat should not disappear without explicit user action.
Actual:
The project showed no chats even though the chat had existed and was being used moments earlier.
Impact:
This caused loss of visible session continuity and interrupted active work. Since it happened twice in a short period, it looks like a thread persistence, sidebar indexing, or local state issue rather than a one-time UI refresh.
Please investigate project chat persistence/sidebar indexing in the Windows Codex app and advise how to recover the missing sessions or which logs/session transcript files I should provide.
latest update for26th of june win desktop
-------------
An error has occurred
Codex crashed with the following error:
(code=3221225786, signal=null).
Most recent error: {"timestamp":"2026-06-26T07:54:08.755542Z","level":"WARN","fields":{"message":"configured non-curated plugin no longer exists in discovered marketplaces during cache refresh","plugin":"browser","marketplace":"openai-bundled"},"target":"codex_core_plugins::loader"}
{"timestamp":"2026-06-26T07:54:08.755580Z","level":"WARN","fields":{"message":"configured non-curated plugin no longer exists in discovered marketplaces during cache refresh","plugin":"documents","marketplace":"openai-primary-runtime"},"target":"codex_core_plugins::loader"}
{"timestamp":"2026-06-26T07:54:08.755599Z","level":"WARN","fields":{"message":"configured non-curated plugin no longer exists in discovered marketplaces during cache refresh","plugin":"pdf","marketplace":"openai-primary-runtime"},"target":"codex_core_plugins::loader"}
{"timestamp":"2026-06-26T07:54:08.755607Z","level":"WARN","fields":{"message":"configured non-curated plugin no longer exists in discovered marketplaces during cache refresh","plugin":"presentations","marketplace":"openai-primary-runtime"},"target":"codex_core_plugins::loader"}
{"timestamp":"2026-06-26T07:54:08.755615Z","level":"WARN","fields":{"message":"configured non-curated plugin no longer exists in discovered marketplaces during cache refresh","plugin":"spreadsheets","marketplace":"openai-primary-runtime"},"target":"codex_core_plugins::loader"}
{"timestamp":"2026-06-26T07:54:08.755622Z","level":"WARN","fields":{"message":"configured non-curated plugin no longer exists in discovered marketplaces during cache refresh","plugin":"template-creator","marketplace":"openai-primary-runtime"},"target":"codex_core_plugins::loader"}
Some things to try:
Check your config.toml for invalid settings
Check your settings to disable running in WSL if you are seeing compatibility issues
Try updating Codex or installing a different version
Click reload to restart Codex, or visit our documentation for additional help.