Update chat template: tool-call author in history, end_sampling after assistant turns, reasoning-effort validation (minimal=0.1, raise on invalid), raise on unknown roles

chat_template.jinja

@@ -1,4 +1,4 @@
-{#- Keep Python's floating type spelling when minja compacts 1.0 to 1. -#}
+{#- Keep Python's floating type spelling when the jinja engine compacts 1.0 to 1. -#}
 {%- macro json_scalar(value) -%}
   {%- set serialized = value | tojson(ensure_ascii=false, separators=(',', ':')) -%}
   {%- if value is float and '.' not in serialized and 'e' not in (serialized | lower) -%}
@@ -47,7 +47,7 @@
 {#-
   Decode a JSON string under Transformers, then re-encode it like json.dumps.
   llama.cpp converts string arguments to objects after capability detection;
-  the fallback keeps direct-minja parsing safe without runtime-only filters.
+  the fallback keeps engine-only parsing safe without runtime-only filters.
 -#}
 {%- macro json_string(source, state) -%}
   {%- if lipsum is defined -%}
@@ -298,23 +298,26 @@
 
 {#-
   Match Python f"{float(effort):.2f}" followed by trailing-zero removal.
-  Minja has no round filter; the bit table handles binary64 midpoint equality.
+  llama.cpp's jinja engine has no round filter; the bit table handles binary64 midpoint equality.
 -#}
 {%- macro reasoning_effort_text(effort) -%}
   {%- set eff = effort -%}
   {%- if eff is string -%}
     {%- set e = eff | trim | lower -%}
     {%- if e == 'none' -%}{%- set eff = 0.0 -%}
+    {%- elif e == 'minimal' -%}{%- set eff = 0.1 -%}
     {%- elif e == 'low' -%}{%- set eff = 0.2 -%}
     {%- elif e == 'medium' -%}{%- set eff = 0.7 -%}
     {%- elif e == 'high' -%}{%- set eff = 0.9 -%}
     {%- elif e == 'xhigh' -%}{%- set eff = 0.99 -%}
     {%- elif e == 'max' -%}{%- set eff = 0.99 -%}
+    {%- else -%}{%- set eff = e | float(-1.0) -%}
     {%- endif -%}
   {%- endif -%}
   {%- set value = eff | float -%}
-  {%- if value < 0 -%}{%- set value = 0.0 -%}{%- endif -%}
-  {%- if value > 0.99 -%}{%- set value = 0.99 -%}{%- endif -%}
+  {%- if value < 0 or value > 0.99 -%}
+    {{- raise_exception('Invalid reasoning_effort: ' + (effort | string) + '; expected none/minimal/low/medium/high/xhigh/max or a number in [0.0, 0.99]') -}}
+  {%- endif -%}
   {%- if value == value | int and value >= 0 and value <= 1 -%}
     {{- value | int -}}
   {%- else -%}
@@ -432,6 +435,7 @@
     {%- endif -%}
     {{- emit_message('tool', 'text', tool_content, resolved.name | string) -}}
   {%- elif role == 'user' or role == 'assistant' or role == 'system' or role == 'developer' -%}
+    {%- set turn_out -%}
     {%- if role == 'assistant' and message.get('reasoning_content') is string and message.get('reasoning_content') and
           ((preserve_thinking is not defined) or preserve_thinking is not false or message_index > thinking_state.last_user_index) -%}
       {{- emit_message('assistant', 'thinking', message.get('reasoning_content')) -}}
@@ -484,13 +488,19 @@
           {%- set raw_arguments = function.get('arguments') or {} -%}
           {%- set arguments_json = canonical_arguments(raw_arguments) -%}
           {%- set invocation = '{"name":' + canonical_json(function.get('name')) + ',"args":' + arguments_json + '}' -%}
-          {{- emit_message('assistant', 'invoke_tool_json', invocation) -}}
+          {{- emit_message('assistant', 'invoke_tool_json', invocation, function.get('name')) -}}
         {%- endif -%}
       {%- endfor -%}
     {%- endif -%}
-    {%- if role == 'assistant' and not (loop.last and not (add_generation_prompt is defined and add_generation_prompt)) -%}
+    {%- endset -%}
+    {{- turn_out -}}
+    {#- Close each historical model turn, but never emit a bare terminator for an
+        assistant message that rendered no blocks. -#}
+    {%- if role == 'assistant' and turn_out -%}
       {{- '<|content_model_end_sampling|>' -}}
     {%- endif -%}
+  {%- else -%}
+    {{- raise_exception('Unknown message role: ' + (role | string)) -}}
   {%- endif -%}
 {%- endfor -%}