AI: openai compatible provider can't not use tool

[Sherlock-Holo]

Summary

SUMMARY_SENTENCE_HERE

Description

Steps to trigger the problem:

1. use openai compatible provider (self-hosted qwen model)
2. as some questions about the project, which will let the LLM list dir
3. wait response

Expected Behavior:

zed AI can work

Actual Behavior:

zed logs report

2025-09-18T18:44:46+08:00 ERROR [open_ai] Failed to parse OpenAI response into ResponseStreamResult: `data did not match any variant of untagged enum ResponseStreamResult`
Response: `{"id":"8b98f5df7c9d49c5d519c01554ab5b30","model":"qwen3-coder-480b-a35b-instruct","object":"chat.completion.chunk","created":1758192286,"choices":[{"index":0,"delta":{"role":"assistant","content":"","tool_calls":[{"id":"call_63bfce97a05146819b8af1c1","type":"function","function":{"name":"list_directory","arguments":""}}]}}],"usage":{"prompt_tokens":0,"completion_tokens":0,"total_tokens":0,"prompt_tokens_details":{"cache_read_tokens":0,"cache_creation_tokens":0}}}`
2025-09-18T18:44:46+08:00 ERROR [agent2::thread] Turn execution failed: data did not match any variant of untagged enum ResponseStreamResult
2025-09-18T18:44:46+08:00 ERROR [agent2::agent] Error in model response stream: data did not match any variant of untagged enum ResponseStreamResult

Model Provider Details

• Provider: openai compatible provider (self-hosted qwen model)
• Model Name: qwen3-coder-480b-a35b-instruct
• Mode: Agent Panel
• Other Details (MCPs, other settings, etc): default rule is always resoned on 中文

Zed Version and System Specs

Zed: v0.204.1 (Zed) OS: Linux Wayland arch unknown Memory: 62.4 GiB Architecture: x86_64 GPU: Intel(R) Graphics (RPL-S) || Intel open-source Mesa driver || Mesa 25.2.3-arch1.1

[hiddenblue]

same problem for the model api from modelscope(qwen opensource platform)

[Messius58]

I'm also having trouble with Qwen3 coders using OpenAI compatibility, but not systematically.
Tool calls can work for hours and then crash suddenly, requiring a restart.
I have the impression:

• that unauthorized characters are present in the chat/context and are causing the entire process to crash.
• that this is due to the use of the "```" characters, which may be misused.

I'm trying to find the source and possibly correct it, as it's a shame to lose the entire thread.
The workaround is to avoid calling agents and to copy and paste.
I use Qwen3 coders locally via llama.cpp. perhaps :

• writing a provider specific to llama.cpp could also help ?
• writing code specific to Qwen3 ?

I could possibly look into implementing a solution if someone points me in the right direction to code it properly.