fix(template): align Qwen reasoning parser state

- Make enable_thinking the authoritative reasoning-mode control used by the vLLM Qwen parser.\n- Remove message-level thinking toggles and prevent tools or repeated failures from silently changing parser state.\n- Add ultra-to-xhigh reasoning-effort compatibility while preserving the xhigh default.\n- Reject the unsafe legacy auto-disable option and update the README to use enable_thinking explicitly.\n- Preserve XML tools, OpenAI and Codex history, multimodal input, reasoning history, and agent error guidance.

README.md

@@ -280,7 +280,6 @@ response = client.chat.completions.create(
     extra_body={
         "chat_template_kwargs": {
             "enable_thinking": False,
-            "auto_disable_thinking_with_tools": True,
         }
     },
 )
@@ -317,9 +316,10 @@ Always use the current `chat_template.jinja`. It is intentionally separate from
 | `enable_thinking` | `true` | Enable or disable thinking output |
 | `preserve_thinking` | `true` | Preserve earlier thinking in conversation history |
 | `tool_call_format` | `xml` | Select `xml` or `json` tool-call output |
-| `auto_disable_thinking_with_tools` | `false` | Disable thinking automatically when tools are present |
 | `continue_final_message` | `false` | Continue an assistant message instead of opening a new turn |
 
+Use `enable_thinking=false` explicitly for non-thinking tool calls. The legacy `auto_disable_thinking_with_tools` option is rejected because changing the prompt mode without changing vLLM's Qwen reasoning-parser state would misclassify reasoning as normal content.
+
 ## Client integrations
 
 The model and template support developer/system instructions, reasoning history, OpenAI-style tools, parallel calls, tool results, and multimodal placeholders. The serving layer translates the native format for each client.

chat_template.jinja

@@ -1,8 +1,8 @@
-{%- set template_version = "qwen3.8-agentic-nvfp4-final" -%}
+{%- set template_version = "qwen3.8-agentic-nvfp4-v1-stable" -%}
 {%- set requested_tool_format = tool_call_format if tool_call_format is defined and tool_call_format is not none else "xml" -%}
 {%- set add_vision_id = add_vision_id if add_vision_id is defined else false -%}
-{%- set thinking_was_explicit = (enable_thinking is defined) or (thinking is defined) -%}
-{%- set enable_thinking = enable_thinking if enable_thinking is defined else (thinking if thinking is defined else true) -%}
+{%- set thinking_was_explicit = enable_thinking is defined -%}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else true -%}
 {%- if preserve_reasoning is defined and preserve_reasoning is not none -%}
   {%- set preserve_history_thinking = preserve_reasoning -%}
 {%- elif preserve_thinking is defined and preserve_thinking is not none -%}
@@ -10,7 +10,6 @@
 {%- else -%}
   {%- set preserve_history_thinking = true -%}
 {%- endif -%}
-{%- set auto_disable_thinking_with_tools = auto_disable_thinking_with_tools if auto_disable_thinking_with_tools is defined else false -%}
 {%- set max_tool_arg_chars = max_tool_arg_chars if max_tool_arg_chars is defined else 0 -%}
 {%- set max_tool_response_chars = max_tool_response_chars if max_tool_response_chars is defined else 0 -%}
 {%- set agent_error_hardening = agent_error_hardening if agent_error_hardening is defined else false -%}
@@ -18,8 +17,11 @@
 {%- set image_counter = namespace(value=0) -%}
 {%- set video_counter = namespace(value=0) -%}
 {%- set state = namespace(thinking=enable_thinking, previous_role="", consecutive_failures=0) -%}
-{%- if auto_disable_thinking_with_tools and has_tools -%}
-  {%- set state.thinking = false -%}
+{# vLLM's Qwen parser reads only chat_template_kwargs.enable_thinking. A template-only
+   auto-disable would desynchronize parser and prompt state, so reject the legacy knob
+   instead of silently producing incorrectly parsed reasoning. #}
+{%- if auto_disable_thinking_with_tools is defined and auto_disable_thinking_with_tools -%}
+  {{- raise_exception("auto_disable_thinking_with_tools is unsafe with the Qwen vLLM reasoning parser. Set chat_template_kwargs.enable_thinking=false explicitly instead.") -}}
 {%- endif -%}
 
 {%- macro render_json_arguments(value) -%}
@@ -107,7 +109,7 @@
 {%- set effort_raw = (reasoning_effort | string | lower) if reasoning_effort is defined and reasoning_effort is not none else "xhigh" -%}
 {%- set effort = "" -%}
 {%- if effort_raw in ("none", "off") -%}
-  {# Match current vLLM semantics: an explicit enable_thinking/thinking kwarg wins. #}
+  {# Match current vLLM semantics: an explicit enable_thinking kwarg wins. #}
   {%- if not thinking_was_explicit -%}
     {%- set state.thinking = false -%}
   {%- endif -%}
@@ -115,27 +117,16 @@
   {%- set effort = "low" -%}
 {%- elif effort_raw == "medium" -%}
   {%- set effort = "medium" -%}
-{%- elif effort_raw in ("high", "xhigh", "max", "ultracode", "extreme") -%}
+{%- elif effort_raw in ("high", "xhigh", "max", "ultra", "ultracode", "extreme") -%}
   {%- set effort = "xhigh" -%}
 {%- elif effort_raw -%}
   {# Unknown future effort names keep the model's normal thinking behavior and do not hard-fail rendering. #}
   {%- set effort = "" -%}
 {%- endif -%}
 
-{# Resolve explicit think on/off controls before building the system prefix. Only trusted
-   conversation roles are scanned; tool outputs cannot flip reasoning mode. A deliberate
-   mode change may invalidate the prefix cache for that turn, which is preferable to a
-   contradictory system prefix and generation scaffold. #}
-{%- for control_message in messages -%}
-  {%- if control_message.role == "system" or control_message.role == "developer" or control_message.role == "user" -%}
-    {%- set control_text = render_content(control_message.content, false, false) -%}
-    {%- if "<|think_off|>" in control_text -%}
-      {%- set state.thinking = false -%}
-    {%- elif "<|think_on|>" in control_text -%}
-      {%- set state.thinking = true -%}
-    {%- endif -%}
-  {%- endif -%}
-{%- endfor -%}
+{# `enable_thinking` is the single source of truth for generation mode. Do not
+   mutate reasoning state from magic strings embedded in conversation content: vLLM's
+   Qwen parser cannot see those mutations and would parse the output in the wrong state. #}
 
 {# Keep the generation grammar Qwen-native and stable by default. Current vLLM
    normalizes OpenAI wire-format argument strings into mappings before rendering.
@@ -165,12 +156,6 @@
 {%- set system_acc = namespace(text="") -%}
 {%- for message in messages[:head.count] -%}
   {%- set part = render_content(message.content, false, true) | trim -%}
-  {%- if "<|think_off|>" in part -%}
-    {%- set state.thinking = false -%}
-  {%- elif "<|think_on|>" in part -%}
-    {%- set state.thinking = true -%}
-  {%- endif -%}
-  {%- set part = part | replace("<|think_off|>", "") | replace("<|think_on|>", "") | trim -%}
   {%- if part -%}
     {%- if system_acc.text -%}
       {%- set system_acc.text = system_acc.text ~ "\n\n" ~ part -%}
@@ -250,14 +235,6 @@
   {%- set content = rendered_content if tool_like else (rendered_content | trim) -%}
 
   {%- set wrapped_tool_response = role == "user" and content[:15] == "<tool_response>" and content[-16:] == "</tool_response>" -%}
-  {%- if system_like or (role == "user" and not wrapped_tool_response) -%}
-    {%- if "<|think_off|>" in content -%}
-      {%- set state.thinking = false -%}
-    {%- elif "<|think_on|>" in content -%}
-      {%- set state.thinking = true -%}
-    {%- endif -%}
-    {%- set content = content | replace("<|think_off|>", "") | replace("<|think_on|>", "") | trim -%}
-  {%- endif -%}
 
   {%- if system_like -%}
     {{- "<|im_start|>system\n" ~ content ~ "<|im_end|>\n" -}}
@@ -429,7 +406,7 @@
 {%- set continuing_final = continue_final_message is defined and continue_final_message and conversation and conversation[-1].role == "assistant" -%}
 {%- if add_generation_prompt and not continuing_final -%}
   {{- "<|im_start|>assistant\n" -}}
-  {%- if state.thinking and (not agent_error_hardening or state.consecutive_failures < 2) -%}
+  {%- if state.thinking -%}
     {{- "<think>\n" -}}
   {%- else -%}
     {# Qwen's own templates use this closed empty think marker to force non-thinking generation. #}