[Windows][ChatGPT Work] Reconnecting 1/5–5/5, Error 10053, 16+ GB Network Usage, Severe Stalls — Reinstall Failed Copy #52271

What version of the Codex App are you using (From “About Codex” dialog)?

26.1002.7124.0

What subscription do you have?

Pro100

What platform is your computer?

Windows 10 22H2 (Build 19045) — x64

What issue are you seeing?

Summary

CRITICAL: ChatGPT Work / Codex Windows Desktop has become extremely unreliable since October 7, 2026, with repeated reconnections, Windows socket error 10053, severe thinking stalls, and excessive internet-data consumption.

The application worked smoothly before October 7. I was regularly using it for professional software development, image generation, and long-running Work conversations without these problems.

Since October 7, the application has become almost unusable.

Environment

• Windows 10 22H2, Build 19045, x64

• Codex Desktop: 26.1002.7124.0

• Codex CLI: 0.161.0

• Models affected: GPT-6 Astra Medium and High

• Main connection: iPhone 5G hotspot

• Windows proxy: Disabled

• WinHTTP: Direct connection

Problem 1 — Reconnecting 1/5 through 5/5

The application repeatedly displays:

Reconnexion en cours 1/5

Reconnexion en cours 2/5

Reconnexion en cours 3/5

Reconnexion en cours 4/5

Reconnexion en cours 5/5

The exact error recorded is:

stream disconnected before completion:
IO error: Une connexion établie a été abandonnée par un logiciel de votre ordinateur hôte. (os error 10053)

Tasks sometimes remain stuck for more than 10 minutes. Some eventually finish after multiple reconnect attempts, while others fail to complete.

This occurs in both existing and completely new Work conversations.

Problem 2 — Stuck on Thinking

The application frequently remains at:

Réflexion en cours

It can stay there for 5–10+ minutes without showing any new progress.

Sometimes the generated image or tool output is already visible, but the conversation remains stuck in the thinking state and subsequently reconnects.

This behavior occurs with GPT-6 Astra Medium and High.

Problem 3 — CRITICAL excessive internet consumption

The Windows desktop application is associated with unusually high internet usage.

My iPhone's mobile-data consumption increased dramatically, with approximately 30 GB reported across 48 hours of usage. I cannot attribute that entire amount to Codex.

However, Windows' own data-usage statistics for the iPhone Wi-Fi connection showed:

Total attributed to ChatGPT/Codex entries: approximately 16.18 GB over Windows' displayed 30-day period.

I also performed a 60-second Wi-Fi adapter measurement while the application was open:

• Total downloaded: 20.66 MB

• Total uploaded: 6.75 MB

• Combined: 27.41 MB

• Peak download: 2,477 KB/s

• Peak upload: 554 KB/s

These measurements include all PC traffic, not exclusively Codex.

Reproducible hotspot symptom: While ChatGPT/Codex Desktop is operating, my iPhone's internet can become extremely slow. TikTok videos sometimes stop loading on the phone itself. Completely closing the desktop application restores normal responsiveness.

This has also occurred when no new conversation request was actively being submitted.

Please investigate whether historical images, conversation data, synchronization activity, or repeated failed requests are being transferred unnecessarily.

Problem 4 — Confirmed diagnostic errors

I examined %USERPROFILE%\.codex\logs_2.sqlite with read-only SQLite queries.

One 24-hour diagnostic window showed:

• 10,413 total log records

• 14 ERROR entries

• 431 WARN entries

• 29 codex_core::responses_retry warnings

• 27 mentions of socket error 10053

• 44 disconnect mentions

• 75 timeout mentions

• 625 WebSocket mentions

A later 48-hour query recorded:

• Approximately 17,654 log entries

• 48 mentions of socket error 10053

• 7 mentions of HTTP fallback

• 11 localhost mentions

• 4 is_connect=true mentions

• 0 mcp_request_timeout mentions

• 0 unknown conversation mentions

These are diagnostic log-text matches, not necessarily unique failures.

Several consecutive retry warnings occurred during the same task at approximately 00:31–00:34 on October 8.

Problem 5 — Complete reinstall did not solve anything

Troubleshooting already performed:

• Restarted the application

• Repaired the Windows app

• Tested new conversations

• Tested existing conversations

• Tested different reasoning levels

• Inspected Windows network usage

• Checked proxy settings

• Investigated SQLite logs

• Tested an HTTPS/SSE custom-provider configuration

• Uninstalled the application completely

• Reinstalled it from Microsoft Store

The problem remained after reinstalling.

Problem 6 — Codex CLI works, Desktop Work fails

I installed Codex CLI 0.161.0 and tested a substantial GPT-6 Astra High reasoning request.

The request used approximately 9,725 tokens and completed successfully in 1–3 minutes without visible reconnect errors.

This was performed on the same computer and 5G hotspot.

The CLI test was text-only and therefore not identical to Desktop Work image generation, but it demonstrates that model access can work normally while the desktop experience fails.

Important conclusion

The issue is not explained solely by old or image-heavy conversations because completely new Work conversations can fail too.

The problem started suddenly around October 7, after the desktop application had previously worked reliably.

I am requesting an engineering investigation into Windows desktop response streaming, error 10053, WebSocket handling, excessive network transfers, and the persistent failure after reinstalling.

What steps can reproduce the bug?

Reproduction steps

1. Launch the official ChatGPT/Codex Windows desktop application, version 26.1002.7124.0.
2. Open ChatGPT Work.
3. Create a completely new conversation.
4. Select GPT-6 Astra High or Medium.
5. Submit a request such as:
   Create a professional e-commerce banner featuring a television, speakers, and a camera.
6. Wait for the response.
7. The application may remain at Réflexion en cours for several minutes.
8. It may then display Reconnexion en cours 1/5 through 5/5, sometimes with Windows socket error 10053.
9. Repeat in an existing Work conversation. Similar failures occur, sometimes after images or tools have already produced output.
10. Monitor the Windows network activity. During affected sessions, traffic can become elevated and the iPhone hotspot can become difficult to use.

Observed behavior

• Failure occurs intermittently but repeatedly.
• Tasks can take over ten minutes.
• Some tasks eventually resume after multiple retries.
• Both fresh and existing Work conversations are affected.
• Reinstalling the application does not resolve the problem.

Example diagnostic evidence

Multiple codex_core::responses_retry warnings were logged during the same turn on October 8 at:

• 00:31:57
• 00:32:33
• 00:33:10
• 00:33:47

All showed Windows socket error 10053.

Session and context information

• The issue affects multiple sessions, not one specific conversation.
• It reproduces in new conversations without accumulated history.
• Exact affected Work session IDs can be provided privately if required.
• Context-window and token-limit usage were not available for the affected Work turns.

I can provide sanitized logs and screenshots through a secure channel.

What is the expected behavior?

ChatGPT Work and Codex Desktop should operate reliably on Windows as they did before October 7, 2026.

Expected behavior:

• Work tasks should start and complete without repeated Reconnecting 1/5–5/5 failures.
• Responses should stream normally without prolonged silent thinking stalls.
• The desktop should recover gracefully from temporary network interruptions.
• Generated images and completed tool results should be followed by a proper final response.
• Existing and brand-new conversations should both function reliably.
• The application should not transfer unnecessary historical images, attachments, or conversation state.
• Background traffic should not cause excessive mobile-data consumption or make the connected iPhone hotspot unusable.
• Application updates and reinstalls should preserve access to existing conversations and generated files.

The expected outcome is a stable Windows desktop application that can complete substantial Work tasks without repeatedly disconnecting or generating abnormal network usage.

Additional information

Additional technical findings

Windows connectivity configuration

• WinHTTP direct connection
• Windows system proxy disabled
• No HTTP_PROXY, HTTPS_PROXY, or ALL_PROXY environment override detected
• No global custom model-provider override found

Basic ping tests to 1.1.1.1 showed 0% packet loss in two 20-packet samples.

These tests did not identify a constant packet-loss problem.

Custom provider experiment

During troubleshooting, I tested supports_websockets = false using a custom provider called openai_http.

It did not provide a lasting solution.

Some existing sessions subsequently required that provider definition to remain available so they could reopen.

The currently defined provider points directly to:

https://chatgpt.com/backend-api/codex

with WebSockets enabled.

No unavailable localhost proxy was found in the active configuration, so this does not appear identical to the confirmed local-proxy root cause in issue #45099.

Related reports

• [Desktop/Sidebar] Large-context sidebar fork repeatedly aborts WebSocket send before HTTP fallback #41667 — Windows socket error 10053
• ChatGPT / Codex Windows desktop client consumes excessive network traffic (~31 GB in 58 hours) #40513 — Excessive Windows desktop network consumption
• Codex keep reconnecting #28295 — Repeated Codex reconnection failures
• [Windows][26.908.40834] ChatGPT Work stuck in “Reconnecting… waiting for network” despite successful thread/start and turn/start #45099 — Windows Work reconnection loop, ultimately attributed to a custom local proxy

These reports may have different root causes.

Business impact

I rely on ChatGPT Work for professional development and visual design.

The problem is significantly delaying my work, disrupting long-running tasks, and consuming costly mobile data.

I have already escalated this case through OpenAI Support and can supply further sanitized diagnostics.

Please investigate whether this is a Windows desktop regression affecting WebSocket streaming, response retries, renderer synchronization, background network transfers, or image-history retransmission.

I would appreciate a confirmed fixed application version, an official supported workaround, or a specific diagnostic procedure that can identify and resolve the underlying problem.

Thank you.