当输出出现tool call的时候，对话被终止

[zt449569708]

Description

在模型对话输出的时候，如”Thinking: 您说得对！我需要重新检查。让我删除之前的所有修改文件，重新解压原始文档，然后仔细诊断问题：<tool_call>bash<arg_key>command</arg_key><arg_value> ...“，当出现“<tool_call>bash<arg_key>command....”的时候对话中断，不再继续输出。今天更新到1.1.23才出现的，之前使用都没出现。

Plugins

Oh My OpenCode

OpenCode version

1.1.23

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Function call auto-truncation #8012: Function call auto-truncation - Reports that conversation stops and cannot continue after a function call, with output appearing as XML
• Response terminates prematurely when using Gemini 3 via LiteLLM #6244: Response terminates prematurely when using Gemini 3 via LiteLLM - Reports OpenCode stops processing as soon as the model triggers a tool call

Both issues describe conversations being interrupted when tool calls appear in the output, similar to your report. Feel free to ignore if your specific case differs.

[bi-kai]

same issue, with version 1.1.27

mark & focus

[imjoey]

same issue, version 1.1.51 with GLM-4.7 model

[kade06]

same issue, version 1.2.15 with GLM-5 model。 end with </arg_value></tool_call>

[liuyibo123]

same issue ,version 1.2.15 with GLM-5 。</arg_value></tool_call> shows.

[SupetZYK]

same issue version 1.2.20 with GLM-5 </arg_value></tool_call> shows.

[wangeek]

same issue version 1.2.26 with GLM-4.7 </arg_value></tool_call> shows.

[MNCloudwerksTechnology]

Same issue version v1.2.27 and qwen3.5-35B-A3B

Stops directly after thinking block or after toolcalling. Happens in longer tasks more often. (+100k Context)

`
┃ Thinking: The test failure "should emit logo click event for dashboard FAILED" suggests an issue with how logoClick is emitting. Let me check the platform component's test to see if that's the failing test. I'll skip it for now and run all platform tests: src/app/pages/platform/platfo +18 -42
┃
┃ <tool_call>
┃ <function=bash>
┃ <parameter=command>
┃ npm test -- --browsers=ChromeHeadless --watch=false --include=src/app/pages/platform/platform.component.spec.ts 2>&1 | grep -E "(FAILED|SUCCESS|TOTAL|passed)" | tail -10
┃
┃ <parameter=description>
┃ Run platform tests
┃
┃ <parameter=timeout>
┃ 120000
┃
┃
┃ </tool_call>

 ▣  Build · Qwen3.5-35B-A3B · 18.7s  

`

[hashelq]

If you use qwen (or a fork of it, like omnicoder) -> ollama/ollama#14492

Interestingly, seems like switching to

RENDERER qwen3-coder
PARSER qwen3-coder

for qwen3.5:9b fixes the tool calling problems for a trade off of broken think tag parsing.