chat_template: replace no-user-query raise_exception with tolerant fallback

The multi_step_tool guard hard-raises when a conversation contains no plain user message (shape: [system, assistant, tool, ...]). Agent frameworks produce exactly this shape after history compression, and runtimes that render the template during automatic tool-parser generation (vLLM >= 0.24, llama.cpp, LM Studio, Ollama) fail the entire request with a 400 before generating a token — see vllm-project/vllm#36432 and Qwen/Qwen3.6-27B discussion #16 for the same failure on sibling models.

ns.last_query_index only gates reasoning-content preservation, so the safe fallback is -1 (treat every assistant turn as after the last query, i.e. preserve reasoning). Verified locally on vLLM 0.24: the failing shape renders and generates correctly with this template, and normal chat / multi-step tool conversations are byte-identical.

chat_template.jinja

@@ -76,7 +76,12 @@
     {%- endif %}
 {%- endfor %}
 {%- if ns.multi_step_tool %}
-    {{- raise_exception('No user query found in messages.') }}
+    {#- Tolerate conversations with no plain user turn (e.g. [system, assistant, tool, ...],
+        common in agent frameworks after history compression). The hard raise breaks every
+        runtime that renders the template at tool-parser-generation time (vLLM >=0.24,
+        llama.cpp, LM Studio, Ollama): the request 400s before a single token is generated.
+        -1 == "no user query" == preserve reasoning content on all assistant turns. #}
+    {%- set ns.last_query_index = -1 %}
 {%- endif %}
 {%- for message in messages %}
     {%- set content = render_content(message.content, true)|trim %}