Ollama/Qwen3.5 returns empty final response when use_max_completion_tokens_param is enabled

[issb-gh]

To Reproduce

1. Configure integration_openai with an Ollama OpenAI-compatible endpoint: http://<ollama-host>:11434/v1

2. Use:

   • Nextcloud 33.0.7
   • integration_openai 4.5.2
   • context_agent 2.8.0
   • Ollama 0.32.7
   • Model qwen3.5-32k:latest

Note: qwen3.5-32k:latest is a local Ollama model tag based on qwen3.5:latest, created with PARAMETER num_ctx 32768. No other model changes were made.

3. Configure max_tokens=3000 and use_max_completion_tokens_param=1.

4. Run a multi-step Context Agent request, for example: How is the current weather in Dresden?

5. Observe that the coordinate lookup and weather lookup complete successfully, but the final assistant response is empty.

6. The resulting TaskProcessing task ends with STATUS_SUCCESSFUL and output: "".

7. Replay the same final /v1/chat/completions request directly against Ollama with the same message history and identical tool results.

8. With max_completion_tokens=3000, the response is:

   • finish_reason="stop"
   • content=""
   • tool_calls=null

9. Change only the token parameter to max_tokens=3000.

10. The same request now returns:

    • finish_reason="stop"
    • normal final assistant content
    • tool_calls=null

11. In Nextcloud, configure:

    • llm_extra_params={"reasoning_effort":"none"}
    • use_max_completion_tokens_param=0
    • max_tokens=3000

12. Repeat the same Context Agent request.

13. The complete flow now succeeds: coordinate tool → weather tool → final answer synthesis → visible answer in the Nextcloud UI.

[lukasdotcom]

Hi @issb-gh,

Do you mean that when this switch is enabled it doesn't work?

If that is the case that switch exists to allow support for providers that don't use the new endpoint.

[issb-gh]

Yes, exactly.

With the switch enabled, integration_openai sends max_completion_tokens, and with Ollama 0.32.7 + Qwen3.5 the final response becomes reproducible empty in our multi-step tool-calling case.

With the switch disabled, integration_openai sends the deprecated max_tokens parameter instead, and the exact same request returns the expected final answer.

So the switch itself is actually the workaround and seems to work as intended. The underlying compatibility issue is most likely on the Ollama/OpenAI-compatible API side.

I opened this issue mainly because the failure mode was difficult to diagnose: all tool calls succeeded and TaskProcessing reported success, but the final assistant content was empty.

I can also provide the isolated Ollama A/B reproducer if that is useful.

[lukasdotcom]

Yes, exactly.

With the switch enabled, integration_openai sends max_completion_tokens, and with Ollama 0.32.7 + Qwen3.5 the final response becomes reproducible empty in our multi-step tool-calling case.

With the switch disabled, integration_openai sends the deprecated max_tokens parameter instead, and the exact same request returns the expected final answer.

So the switch itself is actually the workaround and seems to work as intended. The underlying compatibility issue is most likely on the Ollama/OpenAI-compatible API side.

I opened this issue mainly because the failure mode was difficult to diagnose: all tool calls succeeded and TaskProcessing reported success, but the final assistant content was empty.

I can also provide the isolated Ollama A/B reproducer if that is useful.

Okay then I'll keep this issue open so that other can find this, but there isn't much that can be done in the source code. I think the default on a new install should be off already as long as you aren't using openai.

[issb-gh]

Hi Lukas,

Thanks, that makes sense. Keeping the issue open for discoverability is exactly what I hoped for.
I’ll leave the Nextcloud side as-is and investigate the remaining compatibility behavior on the Ollama side separately.

Thanks for the quick clarification.