Responses and compaction fail in Windows version 26.1007.2314.0, blocking work in long-running conversations #53203

Activity

1. github-actions

   added

   bugSomething isn't workingSomething isn't working

   appIssues related to the Codex desktop appIssues related to the Codex desktop app

   windows-osIssues related to Codex on Windows systemsIssues related to Codex on Windows systems

   connectivityIssues involving networking or endpoint connectivity problems (disconnections)Issues involving networking or endpoint connectivity problems (disconnections)

   contextIssues related to context management (including compaction)Issues related to context management (including compaction)

   on Oct 11, 2026
2. github-actions

   changed the title [-]Windows版26.1007.2314.0で応答と圧縮が失敗し、長期会話の仕事が止まる[/-] [+]Responses and compaction fail in Windows version 26.1007.2314.0, blocking work in long-running conversations[/+] on Oct 11, 2026
3. github-actions commented on Oct 11, 2026

   

   github-actionsbot

   on Oct 11, 2026 – with GitHub Actions

   Contributor

   More actions

   English translation:

   Conclusion: After the update, the previous screen crashes stopped on my PC. However, normal responses and conversation compaction now fail, and I cannot continue working in long-running conversations.

   • In multiple conversations, normal responses end with stream disconnected before completion.
   • In the long-running conversation I have used as my main workspace, Error running remote compact task: stream disconnected before completion occurred consecutively at 12:38 and 12:39 on October 11, 2026 (UTC). One response completed afterward, but I have not been able to confirm that compaction succeeded.
   • In another long-running conversation, Codex ran out of room in the model's context window appears repeatedly, and manual optimization also failed. Despite this, the screen displays “100% remaining (0 used out of 828K)”.
   • I have built up the history of my work and instructions in these conversations. Simply moving to a new conversation would mean losing that context.

   I ask OpenAI to do the following.

   • Investigate why normal responses and compaction fail, and provide a way to resume while preserving the original conversations and their history.
   • Provide the exact Windows version number containing the fix and its release timing.
   • I sent the failure times and request IDs to a private support chat. Please cross-reference these and investigate.

   The previous screen issues are documented in my report #48938. There is also a report of response and compaction failures in the same Windows version in #52919.
4. github-actions commented on Oct 11, 2026

   

   github-actionsbot

   on Oct 11, 2026 – with GitHub Actions

   Contributor

   More actions

   Potential duplicates detected. Please review them and close your issue if it is a duplicate.

   • [Windows / Codex Desktop + VS Code] Intermittent stream disconnects and reconnect failures since Oct 10 #52919
   • [Windows 10 / Codex Work] Remote compaction repeatedly fails across multiple threads (decode error and SSE timeout) since Oct 3 #53111

   Powered by Codex Action
5. huanghaitck commented on Oct 12, 2026

   

   huanghaitck

   on Oct 12, 2026

   More actions

   Additional Windows report on the same package/core version, observed October 10-12, 2026 (timestamps UTC+8).

   • Windows package OpenAI.Codex 26.1007.2314.0, bundled core 0.162.0-alpha.17.2; production tasks use gpt-6.1-sol, medium reasoning.
   • Existing and newly created local chats repeatedly fail with stream disconnected before completion: An error occurred while processing your request, retrying to 5/5. Multiple chats also receive the explicit high-demand message. Remote compaction failed in earlier observations.
   • Several processing failures were correlated to HTTP 200 response headers followed by a processing error carrying the same request ID. This is not proof of a completed stream or of a global outage; precise backend error codes need server-side investigation.
   • Minimal read-only, ephemeral CLI requests succeeded with Sol and Luna. A Sol minimal request using the existing user configuration also succeeded in about 12 seconds. These controls used low reasoning, unlike the medium production tasks, and do not replay the affected histories.
   • Local gateway/domestic connectivity and short endpoint probes were generally healthy. Three-minute persistent HTTPS/public WebSocket controls also completed. Proxy/transport involvement in actual authenticated model streams is not fully excluded; separate TLS/reset errors have occurred. Restarts, new chats and another proxy did not eliminate the upstream processing/high-demand messages.
   • A possible local contributor was found: explicit 1M context/900K compaction overrides while a recently fetched local model catalog reported 272K. These overrides were removed on October12 to use model defaults; recovery of long tasks has not yet been established. This discrepancy is not a proven cause of the high-demand errors.
   • After that change, the latest 15-minute sample still contained 20 processing-error and 13 high-demand retry records. Counts are retries, not independent requests or a failure rate.

   Request IDs and further diagnostics have been sent through the private official feedback channel and by email to OpenAI Support. Please investigate whether this shares a root cause with the response/compaction failures here, and provide a supported recovery preserving existing work and a fix/version when available. Private research content, email addresses, raw headers and transcripts are omitted from this public comment.
6. synraytech-lgtm commented on Oct 12, 2026

   

   synraytech-lgtm

   on Oct 12, 2026 – with ChatGPT Codex Connector

   More actions

   Additional independent Windows Codex Desktop report (October 12, 2026, UTC+8). This is an observation from a separate user; the app build number was not recorded.

   1. Existing thread repeatedly fails

   • An existing Codex Desktop conversation, used for several hours, started failing and could not be continued during the afternoon.
   • Error: stream disconnected before completion: An error occurred while processing your request.
   • The user reports this has blocked work in the original thread for several hours.
   • No compaction failure was directly observed in this case, so the cause should not be assumed to be compaction.

   2. Fresh conversation works, but model selection initializes incorrectly (possibly a separate issue)

   • Before sending the first message in a newly created chat, selecting any other available model immediately reverts the picker to GPT-5.6 Luna.
   • Send one message with the default model; the model then responds.
   • After that first message, choosing another model succeeds, and a simple test returns OK.
   • This aligns with the first-turn model-picker problem documented in [Linux Desktop 26.930.61225] New-chat model selection reverts to GPT-5.5; switching after start works #51301; it does not establish a shared root cause with the stream failures.

   3. Limits at the time of failure

   • 5-hour quota: 90% remaining.
   • Weekly quota: 60% remaining.
   • Thus exhaustion of the displayed usage quotas does not appear to explain the failure.

   Request to maintainers: Please investigate why an established thread can repeatedly fail while new threads can complete model inference; advise how to recover the original thread without losing its conversation history. Please also triage the new-chat pre-first-message model-picker reversion, if it is not already tracked by #51301.

   A request ID is available to the user for private diagnostic correlation, but is intentionally not posted publicly. No session IDs, local paths, credentials, or project contents are included.
7. jzhi65275-max commented on Oct 13, 2026

   

   jzhi65275-max

   on Oct 13, 2026 – with ChatGPT Codex Connector

   More actions

   Additional independent Windows Codex Desktop report (2026-10-11 to 2026-10-13, UTC+8).

   • Installed package: OpenAI.Codex 26.1007.2314.0, which the app reported as the latest available version at the time of testing.
   • Short, fresh conversations can respond normally.
   • Long-running sessions repeatedly fail when remote context compaction is triggered: Error running remote compact task: stream disconnected before completion. Automatic reconnect attempts can reach 5/5 without recovering the thread.
   • After preserving research state and starting an entirely new thread (not a fork), normal operation resumed temporarily, but compaction/stream disconnects occurred again later. This is not confined to one old session.
   • Other sustained tasks have intermittently failed with stream disconnected before completion: error sending request or An error occurred while processing your request; one was observed after Playwright tool activity. These observations do not establish that Playwright caused the failure or that every failure was a compaction request.
   • A read-only check of config.toml found no explicit overrides for remote_compaction_v2, model_auto_compact_token_limit, or model_context_window. No configuration changes were made.
   • A Clash proxy was in Global mode. A controlled alternative-network comparison has not been completed, so a network/transport contribution is not ruled out.
   • Temporary workaround: save project state and create a fresh independent conversation before compaction. This preserves work but does not resolve the underlying issue.

   Could maintainers investigate the remote compaction/streaming failure, provide a recovery path for affected threads that preserves conversation history, and identify a fix/build once available?

   Request IDs and any sensitive diagnostics should be shared via private OpenAI Support rather than this public issue. No private paths, account details, workspace content, or session logs are included here.
8. lanjunluck commented on Oct 13, 2026

   

   lanjunluck

   on Oct 13, 2026 – with ChatGPT Codex Connector

   More actions

   Additional independently verified Windows Codex Desktop report — 2026-10-13, Asia/Shanghai (UTC+08:00). Adding to this existing report because the symptoms overlap; a shared root cause is not established.

   Environment

   • Desktop version 26.1007.21434, build 13901, production channel. Installed-channel update check reported up_to_date.
   • Model gpt-6.1-sol, medium reasoning, default built-in provider.
   • User config has no explicit model_context_window or model_auto_compact_token_limit override.
   • Usage was allowed; no rate-limit or spend-control block was reported.

   Confirmed failures and request IDs

   Local time, 2026-10-13 (UTC+8)	Failure stage	Request ID
   13:16:35	Automatic remote history compaction, MidTurn	cdcc6f8a-8f7d-4829-8e84-08f747aabe5e
   13:19:15	Automatic remote history compaction, PreTurn	77f75b6b-8815-485f-98a3-733d00f0dbdf
   13:39:06	Recovery turn failed during remote compaction; time is terminal turn completion	ee71bf55-722f-44e1-83f2-5aa2046fd97c
   13:40:43	Normal response sampling failed after initial text and a completed local read; time is terminal turn completion	5935279e-2e00-4995-82e3-605acb8e9da2

   The first two requests logged HTTP 200 from the Responses endpoint, followed by the server processing-error message. HTTP headers do not establish successful stream completion.

   Error: stream disconnected before completion: An error occurred while processing your request. Compaction failures added Error running remote compact task.

   The compaction conversation's last retained input count was 365177 against a client-reported context window of 475000. The other affected conversation was around 276000–285000 input tokens. These counts are client accounting, not independent verification of the failed backend payload. A minimal deterministic reproducer is not available.

   Recovery observations and caveats

   • Preserving saved work and starting focused fresh conversations allowed work to resume. One short replacement conversation completed a full read-only turn in 88.9 seconds without an error.
   • A longer replacement conversation also logged five retries at 14:04:50–14:05:11 with websocket closed by server before response.completed, then continued working. These were recovered retries, not five terminal failed turns.
   • Short public HTTPS probes through the local proxy succeeded. One earlier transport/body-decoding error was logged at 13:03:47. Intermittent proxy or remote transport contributions remain unexcluded.
   • Local checks showed about 28 GiB free physical memory and 458 GiB free disk space.
   • No caches, histories or databases were deleted; no app/network/transport settings were changed or active tasks restarted.

   Please correlate the four request IDs with backend records and distinguish request processing, history-compaction/payload handling, capacity, and transport failures. Please provide a supported recovery preserving the original history and completed tool operations, or the exact fixed version when available. Fresh conversations are a workaround, not an established permanent fix.

   This feedback intentionally contains no conversation text, business materials, screenshots, files, raw logs, account identifiers, local paths, authentication headers or credentials. Request identifiers, diagnostic timestamps and sanitized technical facts are provided with the user's explicit authorization.
9. jibe-dev-ship-it commented on Oct 14, 2026

   

   jibe-dev-ship-it

   on Oct 14, 2026 – with ChatGPT Codex Connector

   Author

   More actions

   追記：GPT-6.1 Solで止まった会話が、GPT-5.6 Terraへの切替で復帰。他のモデルでは不可。

   -GPT-6.1 Solの通常応答が stream disconnected before completion で失敗した。

   • 同じ会話・同じ短い依頼文で、Solは失敗し、Terraは応答を完了した。
   • 他のモデルに切り替えても応答不能。Terraのみが成功する。

   OpenAIに、GPT-6.1 Solで通常応答が切断される原因、影響範囲、修正予定の説明を求める。元の会話と作業履歴を保ったまま使える修正が必要だ。