Skip to content

integration_openai 5.0.0: multimodal-chatwithtools task hangs ~75s on streamed SSE responses #452

Closed

Description

@FurettoVolante

Which version of integration_openai are you using?

5.0.0

Which version of Nextcloud are you using?

35.0.2

Which browser are you using? In case you are using the phone App, specify the Android or iOS version and device please.

Brave Origin - Brave 1.97.57 (Build ufficiale) (x86_64) Chromium: 155.0.8059.80

Describe the Bug

Describe the bug

When integration_openai processes a core:text2text:multimodal-chatwithtools task with preferStreaming=true, and the upstream OpenAI-compatible endpoint closes the SSE connection after sending data: [DONE], the task takes ~75-78 seconds instead of <2s.

The delay is the same regardless of model, prompt, or upstream, and disappears entirely if $preferStreaming is forced to false inside MultimodalChatWithToolsProvider::process().

Expected Behavior

Expected behavior

Task should complete in about the time of the upstream response (<2s).

Actual behavior

Task consistently takes ~75-78s. Worker log shows:

Processing task 40 of type core:text2text:multimodal-chatwithtools
Finished processing task 40

(77 seconds apart)

To Reproduce

To Reproduce

  1. Nextcloud 35 + integration_openai 5.0.0 + assistant 4.1.0.
  2. Point integration_openai to any OpenAI-compatible local server that supports streaming and correctly sends data: [DONE] then closes the socket.
  3. From the assistant UI, submit any short prompt (e.g. "Rispondi solo OK").
  4. Watch oc_taskprocessing_tasks for the new row: started_at and ended_at will be ~77 seconds apart.
  5. In parallel, run curl -N against the same endpoint with the same payload: the stream finishes in <1s.

Root cause hypothesis

In StreamingService::parseStreamChatResponse():

while (!feof($body)) {
    $chunk = fread($body, self::CHUNK_SIZE);
    if ($chunk === false) { throw ...; }
    if ($chunk === '') { continue; }
    foreach ($this->parseSseChunk($chunk, $eventBuffer, $done, $usage, $choices) as $partial) {
        yield $partial;
    }
}

When the upstream closes the connection after sending data: [DONE], the PSR-7 stream returned by Guzzle in streamed mode does not make feof($body) return true right away; fread() remains blocked until a socket timeout fires. On my system the socket timeout is ~75s, matching net.ipv4.tcp_keepalive_intvl = 75.

The $done flag is set inside parseSseChunk when [DONE] is parsed, but the outer while (!feof($body)) does not check it, so the loop keeps reading.

Suggested fix

Break the outer loop when $done is true, and/or explicitly close the body stream once [DONE] is seen:

while (!$done && !feof($body)) {
    // ...
}

or, after the foreach:

if ($done) {
    fclose($body);
    break;
}

(I have not tested the exact fix; forcing $preferStreaming = false in the provider eliminates the symptom, which confirms the stream parsing is the culprit.)

Environment

  • Nextcloud 35.0.2
  • integration_openai 5.0.0
  • assistant 4.1.0
  • Raspberry Pi (ARM64), DietPi/Debian
  • Upstream: local OpenAI-compatible proxy on 127.0.0.1
  • request_timeout = 240, localai_text_generation_time = 0
  • PHP default_socket_timeout = 60, net.ipv4.tcp_keepalive_intvl = 75

Verification data

Test Time
curl direct to upstream, stream=false 0.29 s
curl direct to upstream, stream=true 0.77 s
Same task with $preferStreaming = false forced in provider 1 s
Same task with default $preferStreaming = true 77 s

Diagnosis performed by Claude (Anthropic) on behalf of the server administrator, who provided all environment access, executed all tests, and verified all reported data. The administrator delegated the writing and publication of this issue.

Activity

  1. edward-ly commented on Nov 2, 2026

    @edward-ly
    Contributor

    Hello! This may have been already fixed in #447. Please re-confirm if this bug is still happening on the master branch of the app.

    Also, since this is your first issue, please take a look at our AI contribution policy if you haven't already, especially regarding our rules around AI-assisted issue/PR descriptions.

  2. FurettoVolante commented on Nov 3, 2026

    @FurettoVolante
    Author

    2. Hello! This may have been already fixed in #447. Please re-confirm if this bug is still happening on the master branch of the app.
    Also, since this is your first issue, please take a look at our AI contribution policy if you haven't already, especially regarding our rules around AI-assisted issue/PR descriptions.

    You're right, I didn't see the PR #447, and thanks for pointing out the AI policy, I'll keep that in mind

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

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions