errors with openclaw tool calls.
I had to remove the exception for no user content as it broke tool calling in openclaw. I especially saw this after large queries that sat dormant when a /reset or a /new was sent to openclaw. since i removed the exception been working great.
diff -c froggeric-qwen3.6.jinja*
*** froggeric-qwen3.6.jinja 2026-04-29 20:34:44.063800543 -0400
--- froggeric-qwen3.6.jinja.orig 2026-04-29 20:29:14.919853638 -0400
*** 98,104 ****
{%- endif %}
{%- endfor %}
{%- if ns.multi_step_tool %}
! {{- '' }}
{%- endif %}
{%- for message in messages %}
{%- set content = render_content(message.content, true)|trim %}
--- 98,104 ----
{%- endif %}
{%- endfor %}
{%- if ns.multi_step_tool %}
! {{- raise_exception('No user query found in messages.') }}
{%- endif %}
{%- for message in messages %}
{%- set content = render_content(message.content, true)|trim %}
I am trying this and it works with qwen_xml qwen_code sucks balls, which is the right way to do?
i'm using .. Qwen/Qwen3.6-35B-A3B-FP8
I had to remove the exception for no user content as it broke tool calling in openclaw. I especially saw this after large queries that sat dormant when a /reset or a /new was sent to openclaw. since i removed the exception been working great.
Thanks for reporting. I had also noticed the bug, and had already fixed it in my locally revised variant of 27b heretic which I was still uploading. I have now applied the fix to all my qwen 3.6 releases, and updated the chat template.
Instead of replacing the raise_exception with an empty string, I used an explicit fallback:
{%- set ns.last_query_index = messages|length - 1 %}
This preserves the original default value for last_query_index, so the thinking display logic degrades gracefully. Assistant turns with reasoning content still render thinking tags when preserve_thinking is enabled, rather than losing them.
The fix is now in all Qwen 3.5 and 3.6 model repos (FernflowerAI 35B 8-bit/4-bit, Qwen3.6-27B 8-bit/4-bit, and both Heretic variants at 8-bit/6-bit/4-bit), as well as the standalone https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates repo (documented as bug #6).
ah thanks .. i think your update explains some of the sequencing and repeat messages I was seeing after a long tool execution. I'll implement your changes immediately. thanks!!
which reasoning parser should it be used with? qwen3_coder or qwen3_xml?