Studio: Fix chat template disappearing after browser refresh (#5209)

* fix: preserve chat template on refresh

* chore: simplify chat template status lookup

studio/backend/core/inference/llama_cpp.py

@@ -1701,6 +1701,18 @@ class LlamaCppBackend:
             if chat_template_override:
                 import tempfile
 
+                self._chat_template = chat_template_override
+                flags = detect_reasoning_flags(
+                    self._chat_template,
+                    self._model_identifier,
+                    log_source = "GGUF chat template override",
+                )
+                self._supports_reasoning = flags["supports_reasoning"]
+                self._reasoning_style = flags["reasoning_style"]
+                self._reasoning_always_on = flags["reasoning_always_on"]
+                self._supports_preserve_thinking = flags["supports_preserve_thinking"]
+                self._supports_tools = flags["supports_tools"]
+
                 self._chat_template_file = tempfile.NamedTemporaryFile(
                     mode = "w",
                     suffix = ".jinja",

studio/backend/models/inference.py

@@ -283,6 +283,10 @@ class InferenceStatusResponse(BaseModel):
     supports_tools: bool = Field(
         False, description = "Whether the active model supports tool calling"
     )
+    chat_template: Optional[str] = Field(
+        None,
+        description = "Jinja2 chat template string for the active model",
+    )
     context_length: Optional[int] = Field(
         None, description = "Context length of the active model"
     )

studio/backend/routes/inference.py

@@ -985,6 +985,7 @@ async def get_status(
                 reasoning_always_on = llama_backend.reasoning_always_on,
                 supports_preserve_thinking = llama_backend.supports_preserve_thinking,
                 supports_tools = llama_backend.supports_tools,
+                chat_template = llama_backend.chat_template,
                 context_length = llama_backend.context_length,
                 max_context_length = llama_backend.max_context_length,
                 native_context_length = llama_backend.native_context_length,
@@ -998,12 +999,19 @@ async def get_status(
         is_audio = False
         audio_type = None
         has_audio_input = False
+        model_info = {}
         if backend.active_model_name:
             model_info = backend.models.get(backend.active_model_name, {})
             is_vision = model_info.get("is_vision", False)
             is_audio = model_info.get("is_audio", False)
             audio_type = model_info.get("audio_type")
             has_audio_input = model_info.get("has_audio_input", False)
+        chat_template_info = model_info.get("chat_template_info", {})
+        chat_template = (
+            chat_template_info.get("template")
+            if isinstance(chat_template_info, dict)
+            else None
+        )
 
         # Non-GGUF: only gpt-oss Harmony is wired through the transformers
         # generation path. Other template-level reasoning / tool kwargs
@@ -1041,6 +1049,7 @@ async def get_status(
             reasoning_always_on = False,
             supports_preserve_thinking = False,
             supports_tools = False,
+            chat_template = chat_template,
         )
 
     except Exception as e:

studio/backend/tests/test_native_context_length.py

@@ -320,11 +320,23 @@ class TestPydanticModels:
         """Field exists in InferenceStatusResponse.model_fields."""
         assert "native_context_length" in InferenceStatusResponse.model_fields
 
+    def test_status_response_has_chat_template_field(self):
+        """Status includes chat_template so the UI can rehydrate after refresh."""
+        assert "chat_template" in InferenceStatusResponse.model_fields
+
     def test_status_response_defaults_none(self):
         """Omitting native_context_length defaults to None."""
         resp = InferenceStatusResponse()
         assert resp.native_context_length is None
 
+    def test_status_response_chat_template_roundtrip(self):
+        """chat_template serializes and validates as part of status."""
+        resp = InferenceStatusResponse(chat_template = "{{ messages }}")
+        roundtripped = InferenceStatusResponse.model_validate_json(
+            resp.model_dump_json()
+        )
+        assert roundtripped.chat_template == "{{ messages }}"
+
     def test_roundtrip_preserves_value(self):
         """model_validate_json(model_dump_json()) round-trips."""
         resp = LoadResponse(

studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts

@@ -266,6 +266,10 @@ export function useChatModelRuntime() {
           ? (statusRes.native_context_length ?? null)
           : null;
         const currentSpecType = normalizeSpeculativeType(statusRes.speculative_type);
+        const nextDefaultChatTemplate =
+          statusRes.chat_template === undefined
+            ? useChatRuntimeStore.getState().defaultChatTemplate
+            : statusRes.chat_template;
         useChatRuntimeStore.setState({
           supportsReasoning,
           reasoningAlwaysOn,
@@ -282,6 +286,7 @@ export function useChatModelRuntime() {
           ggufNativeContextLength,
           modelRequiresTrustRemoteCode:
             statusRes.requires_trust_remote_code ?? false,
+          defaultChatTemplate: nextDefaultChatTemplate,
           speculativeType: currentSpecType,
           loadedSpeculativeType: currentSpecType,
         });

studio/frontend/src/features/chat/types/api.ts

@@ -129,6 +129,7 @@ export interface InferenceStatusResponse {
   reasoning_always_on?: boolean;
   supports_preserve_thinking?: boolean;
   supports_tools?: boolean;
+  chat_template?: string | null;
   context_length?: number | null;
   max_context_length?: number | null;
   native_context_length?: number | null;