fix(openai): keep answers in content when thinking is disabled

qwen3's chat template wraps <think> in {% if enable_thinking %}, but the
prefill detectors scan the raw template without evaluating the condition.
With streaming and no tools, the synthetic prefill made the Go-side extractor
classify the whole answer as reasoning whenever the client sent
metadata.enable_thinking=false.

Skip prefill detection when the request disables thinking. Real
<think>...</think> blocks that a model emits anyway are still extracted from
the content itself.

Fixes #12939

Assisted-by: ZCode:deepseek-flash
Signed-off-by: xianshujun <xianshupang@gmail.com>

core/config/model_config.go

@@ -1330,6 +1330,62 @@ func (c *ModelConfig) ApplyReasoningEffort(requestEffort string) {
 	}
 }
 
+// ReasoningDisabled reports whether the effective chat-template toggle asks the
+// model not to think.
+//
+// The backend resolves enable_thinking with a fixed precedence: the client's
+// request metadata overrides the operator's reasoning config (also set by
+// reasoning_effort=none through ApplyReasoningEffort), which overrides the model
+// YAML chat_template_kwargs. Callers that inspect the extracted reasoning have
+// to apply the same precedence and value coercion, or they disagree with the
+// template: a request can re-enable thinking over reasoning.disable=true, and a
+// model can disable it through chat_template_kwargs alone.
+func (c *ModelConfig) ReasoningDisabled() bool {
+	if c == nil {
+		return false
+	}
+	// Build the metadata the backend forwards, then let the same resolver that
+	// gRPCPredictOpts uses apply the layering and the string-to-bool coercion.
+	meta := map[string]string{}
+	if c.ReasoningConfig.DisableReasoning != nil {
+		if *c.ReasoningConfig.DisableReasoning {
+			meta["enable_thinking"] = "false"
+		} else {
+			meta["enable_thinking"] = "true"
+		}
+	}
+	if raw, ok := c.RequestMetadata["enable_thinking"]; ok {
+		meta["enable_thinking"] = raw
+	}
+	value, ok := c.ResolveChatTemplateKwargs(meta)["enable_thinking"]
+	if !ok {
+		return false
+	}
+	return !chatTemplateTruthy(value)
+}
+
+// chatTemplateTruthy reports how a jinja template reads a chat_template_kwargs
+// value: bools and numbers follow Go truthiness, and any non-empty string is
+// true. That last rule is why coerceChatTemplateKwarg turns metadata
+// "true"/"false" into real booleans before they reach the template.
+func chatTemplateTruthy(value any) bool {
+	switch v := value.(type) {
+	case nil:
+		return false
+	case bool:
+		return v
+	case string:
+		return v != ""
+	case int:
+		return v != 0
+	case int64:
+		return v != 0
+	case float64:
+		return v != 0
+	}
+	return true
+}
+
 // coerceChatTemplateKwarg coerces a request-metadata string value for use as a
 // jinja chat_template_kwarg. "true"/"false" become real booleans (so a jinja
 // `{% if preserve_thinking %}` reads false correctly, since any non-empty string

core/config/thinking_disabled_test.go

@@ -0,0 +1,76 @@
+package config_test
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+)
+
+// ReasoningDisabled decides whether the streaming reasoning extractor may
+// prefill a synthetic thinking-start token. It has to resolve enable_thinking
+// exactly like the backend: request metadata over the operator's reasoning
+// config, over the model's chat_template_kwargs.
+var _ = Describe("ModelConfig.ReasoningDisabled", func() {
+	It("is false when nothing disables reasoning", func() {
+		Expect((&config.ModelConfig{}).ReasoningDisabled()).To(BeFalse())
+	})
+
+	It("is true when the operator disabled reasoning", func() {
+		disabled := true
+		cfg := &config.ModelConfig{}
+		cfg.ReasoningConfig.DisableReasoning = &disabled
+		Expect(cfg.ReasoningDisabled()).To(BeTrue())
+	})
+
+	It("is false when the operator explicitly enabled reasoning", func() {
+		enabled := false
+		cfg := &config.ModelConfig{}
+		cfg.ReasoningConfig.DisableReasoning = &enabled
+		Expect(cfg.ReasoningDisabled()).To(BeFalse())
+	})
+
+	It("honors a client enable_thinking=false", func() {
+		cfg := &config.ModelConfig{RequestMetadata: map[string]string{"enable_thinking": "false"}}
+		Expect(cfg.ReasoningDisabled()).To(BeTrue())
+	})
+
+	It("stays enabled for enable_thinking=true", func() {
+		cfg := &config.ModelConfig{RequestMetadata: map[string]string{"enable_thinking": "true"}}
+		Expect(cfg.ReasoningDisabled()).To(BeFalse())
+	})
+
+	It("stays enabled for an unparsable metadata value", func() {
+		cfg := &config.ModelConfig{RequestMetadata: map[string]string{"enable_thinking": "sometimes"}}
+		Expect(cfg.ReasoningDisabled()).To(BeFalse())
+	})
+
+	It("lets request metadata re-enable thinking over the operator's disable", func() {
+		disabled := true
+		cfg := &config.ModelConfig{
+			RequestMetadata: map[string]string{"enable_thinking": "true"},
+		}
+		cfg.ReasoningConfig.DisableReasoning = &disabled
+		Expect(cfg.ReasoningDisabled()).To(BeFalse())
+	})
+
+	It("honors the model chat_template_kwargs enable_thinking=false", func() {
+		cfg := &config.ModelConfig{
+			ChatTemplateKwargs: map[string]any{"enable_thinking": false},
+		}
+		Expect(cfg.ReasoningDisabled()).To(BeTrue())
+	})
+
+	It("lets request metadata override the model chat_template_kwargs", func() {
+		cfg := &config.ModelConfig{
+			ChatTemplateKwargs: map[string]any{"enable_thinking": false},
+			RequestMetadata:    map[string]string{"enable_thinking": "true"},
+		}
+		Expect(cfg.ReasoningDisabled()).To(BeFalse())
+	})
+
+	It("tolerates a nil receiver", func() {
+		var cfg *config.ModelConfig
+		Expect(cfg.ReasoningDisabled()).To(BeFalse())
+	})
+})

core/http/endpoints/openai/chat.go

@@ -931,10 +931,7 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 				} else {
 					template = predInput
 				}
-				thinkingStartToken := reason.DetectThinkingStartToken(template, &config.ReasoningConfig)
-				if config.TemplateConfig.UseTokenizerTemplate {
-					thinkingStartToken = reason.DetectThinkingStartTokenInTemplate(template, &config.ReasoningConfig)
-				}
+				thinkingStartToken := thinkingStartTokenFor(config, template)
 
 				xlog.Debug("Thinking start token", "thinkingStartToken", thinkingStartToken, "template", template)
 

core/http/endpoints/openai/chat_stream_thinking_disabled_test.go

@@ -0,0 +1,104 @@
+package openai
+
+import (
+	"context"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/backend"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/pkg/model"
+)
+
+// Regression test for issue #12939. qwen3's chat template guards <think> with
+// {% if enable_thinking %}, but detection scans the raw template without
+// evaluating the condition. With streaming and no tools, prefilling that token
+// made the extractor classify the whole answer as reasoning whenever the client
+// sent metadata.enable_thinking=false.
+var _ = Describe("processStream with thinking disabled (issue #12939)", func() {
+	// The marker is only rendered when the template condition is true; the raw
+	// string still contains it, which is what detection sees.
+	const conditionalThinkTemplate = "{% if enable_thinking %}<think>{% endif %}"
+
+	var origInference modelInferenceFunc
+
+	BeforeEach(func() { origInference = backend.ModelInferenceFunc })
+	AfterEach(func() { backend.ModelInferenceFunc = origInference })
+
+	streamAnswer := func(answer string) {
+		backend.ModelInferenceFunc = func(
+			ctx context.Context, s string, messages schema.Messages,
+			images, videos, audios []string,
+			loader *model.ModelLoader, c *config.ModelConfig, cl *config.ModelConfigLoader,
+			o *config.ApplicationConfig,
+			tokenCallback func(string, backend.TokenUsage) bool,
+			tools, toolChoice string,
+			logprobs, topLogprobs *int,
+			logitBias map[string]float64,
+			metadata map[string]string,
+		) (func() (backend.LLMResponse, error), error) {
+			return func() (backend.LLMResponse, error) {
+				usage := backend.TokenUsage{Prompt: 1, Completion: 1}
+				tokenCallback(answer, usage)
+				return backend.LLMResponse{Response: answer, Usage: usage}, nil
+			}, nil
+		}
+	}
+
+	run := func(answer string, metadata map[string]string) []schema.OpenAIResponse {
+		streamAnswer(answer)
+		ctx, cancel := context.WithCancel(context.Background())
+		defer cancel()
+		req := &schema.OpenAIRequest{Cancel: cancel, Context: ctx}
+		req.Model = "qwen3-4b"
+		cfg := &config.ModelConfig{RequestMetadata: metadata}
+		responses := make(chan schema.OpenAIResponse, 16)
+		_, err := processStream(conditionalThinkTemplate, req, cfg, nil, config.NewApplicationConfig(), nil, responses, "req-1", 0)
+		Expect(err).ToNot(HaveOccurred())
+		return drainChannel(responses)
+	}
+
+	deltas := func(chunks []schema.OpenAIResponse) (reasoning, content string) {
+		for _, chunk := range chunks {
+			if len(chunk.Choices) == 0 || chunk.Choices[0].Delta == nil {
+				continue
+			}
+			if chunk.Choices[0].Delta.Reasoning != nil {
+				reasoning += *chunk.Choices[0].Delta.Reasoning
+			}
+			// Message.Content is an any: the streaming workers set it to a
+			// *string, while a plain string is also valid.
+			switch c := chunk.Choices[0].Delta.Content.(type) {
+			case *string:
+				if c != nil {
+					content += *c
+				}
+			case string:
+				content += c
+			}
+		}
+		return reasoning, content
+	}
+
+	It("keeps the answer in content when enable_thinking=false", func() {
+		reasoning, content := deltas(run("Hello!", map[string]string{"enable_thinking": "false"}))
+		Expect(content).To(Equal("Hello!"))
+		Expect(reasoning).To(BeEmpty())
+	})
+
+	It("reproduces the misclassification without the request lever (control)", func() {
+		reasoning, content := deltas(run("Hello!", nil))
+		Expect(reasoning).To(Equal("Hello!"),
+			"without enable_thinking=false the synthetic prefill classifies the answer as reasoning; "+
+				"this is the behaviour the fix guards against")
+		Expect(content).To(BeEmpty())
+	})
+
+	It("still extracts real thinking tags when the model ignores the toggle", func() {
+		reasoning, content := deltas(run("<think>plan</think>Hello!", map[string]string{"enable_thinking": "false"}))
+		Expect(reasoning).To(Equal("plan"))
+		Expect(content).To(Equal("Hello!"))
+	})
+})

core/http/endpoints/openai/chat_stream_workers.go

@@ -112,6 +112,26 @@ func chooseDeferredReasoning(
 	return extractor.Reasoning()
 }
 
+// thinkingStartTokenFor returns the thinking-start token to use for this
+// request, or "" when the request disables thinking.
+//
+// The detectors scan the raw prompt/template without evaluating jinja
+// conditionals, so qwen3's <think> is found even when the surrounding
+// {% if enable_thinking %} guard will not render it. Prefilling it in that case
+// makes the extractor classify the whole answer as reasoning (issue #12939),
+// so detection is skipped when the request asks for no thinking. Real
+// <think>…</think> blocks that a model emits anyway are still extracted from
+// the content itself.
+func thinkingStartTokenFor(cfg *config.ModelConfig, template string) string {
+	if cfg.ReasoningDisabled() {
+		return ""
+	}
+	if cfg.TemplateConfig.UseTokenizerTemplate {
+		return reason.DetectThinkingStartTokenInTemplate(template, &cfg.ReasoningConfig)
+	}
+	return reason.DetectThinkingStartToken(template, &cfg.ReasoningConfig)
+}
+
 // processStream is the streaming worker for chat completions with no
 // tool/function calling involved. It pushes SSE-shaped chunks onto
 // `responses` and returns the authoritative cumulative TokenUsage from
@@ -149,10 +169,7 @@ func processStream(
 	} else {
 		template = s
 	}
-	thinkingStartToken := reason.DetectThinkingStartToken(template, &cfg.ReasoningConfig)
-	if cfg.TemplateConfig.UseTokenizerTemplate {
-		thinkingStartToken = reason.DetectThinkingStartTokenInTemplate(template, &cfg.ReasoningConfig)
-	}
+	thinkingStartToken := thinkingStartTokenFor(cfg, template)
 	extractor := reason.NewReasoningExtractor(thinkingStartToken, cfg.ReasoningConfig)
 
 	// preferAutoparser is sticky: once the C++ autoparser has ever classified
@@ -250,10 +267,7 @@ func processStreamWithTools(
 	} else {
 		template = prompt
 	}
-	thinkingStartToken := reason.DetectThinkingStartToken(template, &cfg.ReasoningConfig)
-	if cfg.TemplateConfig.UseTokenizerTemplate {
-		thinkingStartToken = reason.DetectThinkingStartTokenInTemplate(template, &cfg.ReasoningConfig)
-	}
+	thinkingStartToken := thinkingStartTokenFor(cfg, template)
 	extractor := reason.NewReasoningExtractor(thinkingStartToken, cfg.ReasoningConfig)
 
 	result := ""