Repository navigation
[Windows / Codex Desktop + VS Code] Intermittent stream disconnects and reconnect failures since Oct 10 #52919
Description
Activity
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
- app bug #52731
- [macOS] Repeated mid-task stream disconnections in a long gpt-6.1-sol conversation #52814
- Codex desktop stream disconnects repeatedly on macOS #52294
Powered by Codex Action
- changed the title
[-][Windows Codex Desktop 26.1007.2314.0] Intermittent streaming/reconnect failures on Oct 10; VS Code extension works[/-][+][Windows / Codex Desktop + VS Code] Intermittent stream disconnects and reconnect failures since Oct 10[/+]on Oct 10, 2026 Further observations from the affected user:
- Restarting Windows twice did not resolve the failures.
- Ordinary prompts and remote compaction intermittently fail in Codex.
- Both the Windows desktop app and the VS Code Codex extension have experienced disconnections.
- Browser ChatGPT and other applications remain usable.
The cause remains undetermined. Please investigate the failure events and correlate the request records through OpenAI Support.
Additional sanitized observation on 2026-10-10: the same remote-compaction error text occurred while continuing an existing long-running cloud task. This is a symptom-level correlation with this report, not confirmation of a shared root cause or of a Windows-specific fault.
Observed sequence (all times UTC):
- From approximately 17:27, multiple continuation attempts failed with "stream disconnected before completion".
- A small read-only inspection/archive-planning stage succeeded at approximately 17:32.
- A small assembly stage succeeded at approximately 17:45.
- Subsequent small continuation/archive attempts at approximately 17:47 and 17:48 failed before producing the requested output, with:
Error running remote compact task: stream disconnected before completion: An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error persists. Please include the request ID ... in your message.
Request IDs shown with the two later remote-compaction failures:
- 4b982e3e-5382-4287-bdc0-622e658b0699
- ef3f7c46-f1d6-43b8-83b3-d65c0438b2e4
Identifiers from the earlier stream-disconnect failures in this window:
- 4162594a-89d3-4cfe-8a03-69a0680f74d1
- 418451b2-98cc-4248-91b3-8ab72888f120
- e35f83eb-685d-4f88-a701-01e4652b9240
Observed reproduction sequence: continue an existing long-running task; retry with smaller stages, including a read-only check and archive preparation; observe intermittent success followed by another failure at remote compaction before the requested output. This is not yet a deterministic minimal reproduction. The two successful stages did not establish durable recovery.
Expected: the task can resume reliably, or a supported recovery path preserves completed work. Please correlate the identifiers and timestamps above and advise on safe recovery. The failure's cause is unknown. No source code, task content, images, raw conversations, credentials, account details, or internal filesystem paths are included.
I'm getting the same error, seems to be triggered by compaction of threads that have been working with images in long running (2h+) threads. This is using Codex desktop for WIndows Version 26.1007.21434. I've never had problems with these kinds of tasks before. Chatgpt Pro subscription.
"Error running remote compact task: stream disconnected before completion: An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error persists. Please include the request ID 33e8af0b-a076-45c3-8506-bda4b16b6105 in your message."
Additional sanitized Windows desktop diagnostics — October 11, 2026
I observed repeated failures on Codex Windows Desktop 26.1007.2314.0 (installed package metadata), using gpt-6.1-sol / ultra, with HTTP Responses transport and model WebSockets disabled.
The recurring error was:
stream disconnected before completion: An error occurred while processing your request.Read-only diagnostics found the following timeline. All timestamps below are UTC on 2026-10-11.
Time Observation Request ID 06:05:25 Responses endpoint returned HTTP 507 6eed0a36-9796-4ce5-8641-7d181e9901c106:05:27 Subsequent request returned HTTP 200 260d667a-8de9-44b3-9a52-4b0b4b7a783806:08:25 Turn failed after retries reached 5/5 0211763e-3293-48f7-b800-016bff7c2ed006:10:08 Another turn failed after 5/5 retries 6ae064cd-8078-453e-995b-4d00639a8c5e06:12:49 A third turn failed after 5/5 retries 00125a14-da0e-4624-87b4-010d35085951Requests during the retry sequences returned HTTP 200 before the response stream failed. The 507 response body was not retained in the local logs, so its cause and relationship to the later stream failures remain unknown.
Reopening the conversation and submitting another turn did not prevent recurrence: the new turn began at 06:26:53, followed by the same processing error and a 1/5 retry at 06:27:16, request ID
6c8107b1-c524-463c-9e08-6e80c7654742. A later authoritative status snapshot confirmed that this turn completed successfully at 06:29:30. This retry recovered; it was not another terminal failure. One successful completion does not establish that the intermittent problem is resolved.The local backend process remained running. Recorded input was approximately 69,227 tokens against a 258,400-token context window; no explicit context-limit or quota-exhaustion error was found. A local system proxy is present, and its logs are insufficient to exclude a network-path contribution.
Please correlate these request IDs with backend records. The root cause is not established. No prompts, project contents, credentials, full transcripts, or databases are attached.
Additional sanitized diagnostics from October 11, 2026. Adding evidence to this existing report because the visible symptom matches; a shared root cause is not established.
Environment observed locally:
- Windows 11 x64, build 26200.
- Installed desktop package: 26.1007.2314.0. Desktop technical logs report client version 26.1007.21434; the About dialog was not inspected.
- Running local Codex executable: codex-cli 0.162.0-alpha.17.2.
- The affected conversation used gpt-6.1-sol and gpt-6-astra with max reasoning. Later failures used HTTP Responses transport.
A read-only inspection of one current local project conversation found three confirmed failed turns. These are task_complete records with a non-null error, not counts of retry notifications. All times below are UTC, October 11:
Time Observation Request ID 03:15:58.215 Turn completed with a stream-processing error b7a64d37-90c1-4077-8267-71e3fbed7151 05:47:26 Sampling retry reached 5/5 1f816422-c28f-4103-9c20-020b5ad38649 05:47:32.839 Turn completed with a stream-processing error 29a61f87-67af-4a41-95f7-e48d9b9c5a05 07:16:11 Sampling retry reached 5/5 fe12d448-4184-475b-a3e6-fe83e5cd2aa3 07:16:20.167 Turn completed with a stream-processing error 8008e5f5-9ec8-43d5-8ced-c7a017d8b9c9 The terminal error text was:
stream disconnected before completion: An error occurred while processing your request.Additional observations:
- During the 07:15–07:16 retry sequence, six Responses HTTP requests logged HTTP 200 before subsequent stream failures. The final HTTP 200 was at 07:16:19, followed by the terminal failure at 07:16:20. Successful HTTP headers do not establish successful completion of the response.
- Retained sampling logs for the affected conversation contain 28 processing-error retry warnings plus one separate request-send connection warning. They are not 29 distinct failed turns. The three terminal errors above did not identify remote compaction as the failed stage.
- Separately, today's desktop technical logs contain 13 app_server_connection.websocket_error and 13 reconnect_failed records, beginning at 03:58:21, with ECONNRESET and "Client network socket disconnected before secure TLS connection was established". These concern the desktop connection layer; I have not correlated them causally with the later HTTP model-stream errors.
- A loopback proxy is configured in the diagnostic shell/user environment, and its port was listening when checked. A contemporaneous unauthenticated HTTPS HEAD probe completed TLS verification but returned HTTP 403. This does not validate authenticated streaming or exclude an intermittent proxy/network contribution.
- Configuration includes a 512000-token context-window override and a 400000-token compaction threshold. The last recorded request input was 303649 tokens, against the client-reported effective window of 486400. Current cached model metadata lists a 272000 default and 872000 maximum; the override is below that cached maximum. These client-side values do not establish backend acceptance or the failure's cause.
Observed reproduction: continue an existing long-running local conversation, observe intermittent stream failures, allow retries to reach 5/5, then receive a terminal error. A deterministic minimal reproduction is not available.
Please correlate the terminal request IDs and timestamps with backend records, and advise on diagnostics or a fix that distinguishes response-processing failures from the separate desktop TLS/reconnection failures. Raw transcripts, prompts, project contents, local paths, account identifiers, credentials, and database files are not attached. The inspection did not change settings or restart the running work.
Codex app version
26.1007.2314.0, taken from the installed Windows package metadata (not independently confirmed in the app's About dialog).
Subscription
ChatGPT Pro (user-reported).
Platform
Windows 11 25H2 (64-bit), Codex Windows desktop app.
Description
On 2026-10-10, the Windows Codex desktop app began repeatedly disconnecting and reconnecting. It had been working normally beforehand. The user signed in again after an unexpected disconnection; subsequently, ordinary prompts and longer tasks would intermittently succeed or fail after repeated
Reconnecting…attempts. Remote context compaction also fails. This makes ongoing work unreliable.Updated cross-client observation: Switching from the Windows desktop app to the VS Code Codex extension initially appeared to restore normal operation, but the VS Code extension subsequently disconnected as well (user report). The initial success in VS Code was therefore temporary and must not be treated as evidence of a desktop-only fault. The problem may affect a component shared by both clients. The root cause has not been established.
Locally collected evidence (2026-10-10)
Read-only analysis of the user's local Codex technical logs, covering the affected date, identified:
app_server_connection.websocket_erroreventsapp_server_connection.reconnect_failedeventschatgpt_pubsub_connection_failedeventschatgpt_pubsub_reconnect_exhaustedeventsThe desktop connection counts represent logged technical events, not distinct failed prompts. The session counts represent observed terminal failures in the inspected files, not a guaranteed complete incident total. These statistics alone do not establish which component is responsible.
Representative sanitized failed-task events (UTC; session/account identifiers and request IDs omitted):
UI errors included:
Reproduction / comparison
Reconnecting…attempts and/or fails a task withstream disconnected before completion.The failures are intermittent; this is not a claim that every request reproduces them. The VS Code observations are user-reported and not controlled transport-level experiments. Web ChatGPT chat remaining accessible does not rule out intermittent failures in Codex requests.
Expected
Desktop should complete requests reliably, or safely recover and resume interrupted tasks without requiring users to switch clients.
Request to maintainers
Additional diagnostic details can be provided privately to OpenAI Support if needed. Raw session files are not attached.
Possibly related: #52731, #52749, #37992. Symptom overlap does not establish a shared root cause.