Skip to content

[macOS] Repeated mid-task stream disconnections in a long gpt-6.1-sol conversation #52814

Open
Open
[macOS] Repeated mid-task stream disconnections in a long gpt-6.1-sol conversation#52814
Labels
appIssues related to the Codex desktop appbugSomething isn't workingconnectivityIssues involving networking or endpoint connectivity problems (disconnections)

Description

@SMbey0nd
SMbey0nd
opened on Oct 10, 2026

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

26.1007.21159 (20052), read from the currently running desktop app bundle

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Darwin 25.5.0 arm64 arm (uname -mprs); macOS 26.5.2 (25F84)

What issue are you seeing?

An ongoing desktop task intermittently fails with the following error. Six failed turns occurred in the same conversation on October 10, 2026, between 14:43 and 15:06 (Asia/Shanghai, UTC+8). Some intervening turns completed, but repeated continuation attempts did not reliably recover the 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 271e7723-0c17-445b-ac90-c883a5869391 in your message.

The referenced turn failed at 15:05:08 after 358.346 seconds. The same error occurred with reasoning effort set to xhigh, low, and high on gpt-6.1-sol.

The original workflow concerned extracting document links and packaging the workflow as a reusable skill. Internal document URLs, names, and workspace details are redacted. A redacted English paraphrase of the skill request is: "Package the completed document-link extraction workflow as a reusable skill." Exact continuation prompts included 重试一次 (retry once), 继续任务 (continue the task), and 继续 (continue).

What steps can reproduce the bug?

Feedback ID: 01a124a1-d3d4-7a00-8266-dc7065da00b8
Affected conversation/session ID: 01a1246b-7753-7921-a34b-bd64f6ed307b

Observed reproduction sequence (intermittent; no isolated deterministic reproducer yet):

  1. Use the macOS desktop app with gpt-6.1-sol for an extended task involving tools and document processing.
  2. Continue in the existing conversation and ask it to package the workflow as a reusable skill.
  3. After a failed turn, send a short continuation prompt such as 继续.
  4. Observe progress on some turns, followed by another failed turn with the error above.
  5. Lowering reasoning effort to low did not prevent subsequent failures.

Before the later failures, the last recorded successful model request had 189,914 input tokens; the transcript reported a model context window of 258,400 tokens. This is a recorded successful request, not the exact input size of the failed request.

What is the expected behavior?

Codex should complete the turn or recover from a transient stream interruption and allow the task to continue. If the conversation cannot be processed, it should provide an actionable error and a supported recovery path while preserving completed work.

Additional information

Failure records from the affected local transcript (all times UTC+8, October 10, 2026):

Time Turn duration Request ID
14:43:19 627.7 s 6da720cd-a768-4454-b4b0-974ebfb505aa
14:44:54 65.6 s bcf0e158-0cb7-48d5-ba7b-29bfdf6b2294
14:45:49 32.7 s 0fcc88a6-915c-4d15-814b-3305df74a152
14:52:31 113.9 s 2fabe02b-cd0e-4a4c-8e0e-e0de97336eb5
15:05:08 358.3 s 271e7723-0c17-445b-ac90-c883a5869391
15:06:03 49.3 s eece003e-0602-4b64-a3aa-e007929ec2bb

All six errors were recorded as codex_error_info: "other". These transcript entries did not include HTTP status codes or an explicit context-length error, so they do not establish the root cause. Network changes, other models, account switching, and a fresh conversation have not been tested for this report.

The affected session metadata recorded core version 0.162.0-alpha.17.2.

I searched existing issues. Related reports include #17202 (closed), #19440, and #45925. They share the generic error symptom, but this report does not establish the model-specific or account-specific causes discussed there. Please consolidate if this is the same underlying issue.

Activity

  1. github-actions
    added
    bugSomething isn't working
    appIssues related to the Codex desktop app
    connectivityIssues involving networking or endpoint connectivity problems (disconnections)
    on Oct 10, 2026

  2. github-actions commented on Oct 10, 2026

    @github-actions
    github-actionsbot
    on Oct 10, 2026 – with GitHub Actions
    Contributor

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

    Powered by Codex Action

  3. Zelaron commented on Oct 10, 2026

    @Zelaron
    Zelaron
    on Oct 10, 2026

    Same error here from the Codex CLI (codex-cli 0.162.0, app-server 0.162.1) on Linux, ChatGPT Pro, gpt-6.1-sol at max effort.

    Uploaded thread: 01a0f42d-7501-7320-8b0a-88d91f8603f7

    In case it helps narrow this down, the failures seem tied to particular conversations rather than to load or the account. I run five long sessions on the same account through one app-server daemon. Between 16:29 and 17:39 UTC today, two of them sent 140 requests without a single failure, while the other three failed 60 to 75 percent of their requests. The uploaded thread failed 28 of 34 requests between 16:39 and 16:56 UTC while no other session was sending anything. Which conversations are affected has changed during the day, and remote compaction failed the same way in one session this morning. Each failure arrives 1 to 3 seconds after the server returns HTTP 200, with no 429 or rate-limit headers, and weekly usage is at 9 percent.

    Recent failed request IDs from the uploaded thread (UTC):

    17:36:50  1aebb2cb-4775-4f9a-9d5c-40e3e35f34d0
17:36:59  5615069f-217d-492c-92d8-45a0bab8aa04
17:37:04  83b5196a-0438-4cba-952b-795c79fc872e
17:47:09  2f09c70a-1c3b-442f-b419-1298f28729a0
17:47:24  04e0e919-ea60-4aa4-8f25-291b3719a9e8

  4. djgmd10 commented on Oct 12, 2026

    @djgmd10
    djgmd10
    on Oct 12, 2026

    Same issue here on the macOS ChatGPT desktop app (Codex), on the same build as the original report.

    Environment

    • ChatGPT for macOS 26.1007.21159 (20052) (bundle ID com.openai.codex), bundled codex-cli 0.162.0-alpha.17.2
    • ChatGPT Pro 20×
    • macOS 15.8.1 (24H32), Apple Silicon (arm64)
    • Location: South Korea
    • Observed: October 12, 2026 (KST, UTC+9)

    What happened
    Turns repeatedly fail after exhausting all reconnect attempts (Reconnecting 5/5) with:

    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 <id> in your message.

    This is not limited to long or heavy turns. A one-line prompt asking Codex to open a login window in the in-app browser failed twice in a row at Reconnecting 5/5. On the third attempt it got as far as opening the browser, then failed again mid-turn (Reconnecting 4/5, then 2/5).

    Request IDs

    • e5c6fba1-0414-4213-ae6f-e436215ccd4b
    • faa1fa06-2341-4946-86ed-0f697a846089
    • 182ba901-b8db-47bf-b33a-747b81a50e98

  5. baliiiiiiiii commented on Oct 12, 2026

    @baliiiiiiiii
    baliiiiiiiii
    on Oct 12, 2026 – with ChatGPT Codex Connector

    Additional report from a macOS Codex desktop user, based on a read-only inspection of the local transcript and request logs.

    The symptom closely matches this issue, especially the report that some conversations fail repeatedly while others continue normally.

    Environment

    • ChatGPT authentication, official Codex endpoint.
    • Session metadata: bundled CLI 0.162.0-alpha.17.2.
    • Affected conversation: gpt-6.1-sol, reasoning effort max.
    • An extended task with tool calls; no isolated deterministic reproducer.

    Three terminal failures on October 13, 2026 (Asia/Shanghai, UTC+8)

    Local time (UTC+8) UTC time on October 12 Request ID
    00:13:40 16:13:40 68a2b705-3b36-4586-830e-6993d6c8e1ac
    00:19:06 16:19:06 df600481-5c79-4a9f-96b8-1288e8e9f609
    00:31:37 16:31:37 d9409559-4039-4b3e-b336-ba6d18b1d17c

    All three turns ended with codex_error_info: "other" and:

    stream disconnected before completion: An error occurred while processing your request.

    The client repeatedly displayed reconnect attempts up to 5/5. Some retry sequences recovered and tool execution continued; other sequences exhausted retries and terminated the turn, requiring manual continuation.

    Request-log evidence

    • For the final failure, the request was POST https://chatgpt.com/backend-api/codex/responses, using HTTP streaming.
    • The HTTP response was 200 OK, with x-oai-request-id: d9409559-4039-4b3e-b336-ba6d18b1d17c.
    • A subsequent Turn error carried the same request ID and processing-error message. HTTP 200 therefore did not mean the model response completed.
    • The retained logs for 00:00:00–00:31:40 UTC+8 contain 49 sampling-retry entries for this conversation. Three other conversations have 2, 2, and 3 entries respectively; only the affected conversation has the three terminal turn errors in that window. These are retained-log counts, not a complete account-wide failure rate.

    Context and recovery

    • Before the three failed endings, the most recent successful requests had 212,364 / 220,124 / 246,021 input tokens. These are not measurements of the failed requests themselves.
    • The corresponding context-accounting estimates were 258,236 / 268,406 / 305,328, below the 360,000 automatic-compaction threshold and 380,000 effective context window. Both limit-reached flags were false.
    • Six earlier compactions had completion records. There is no direct evidence here of context exhaustion or a failed compaction causing these errors.
    • After manual continuation, tool calls succeeded and turns completed normally at 00:38:48 and 00:40:46 UTC+8. The conversation was not permanently stuck.

    The raw server error-event body/internal error code was not retained, so the specific backend cause remains unknown. No network changes, model changes, configuration resets, or fresh-conversation comparison were performed for this diagnosis.

    Please correlate the three request IDs with server-side logs to investigate the repeated processing failures and why they concentrated in this conversation. Project content, conversation text, credentials, and full raw logs are intentionally omitted.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    appIssues related to the Codex desktop appbugSomething isn't workingconnectivityIssues involving networking or endpoint connectivity problems (disconnections)

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions