fix(relay): preserve reasoning effort without implicit remapping

relay/channel/deepseek/adaptor.go

@@ -15,6 +15,7 @@ import (
 	"github.com/QuantumNous/new-api/relay/constant"
 	"github.com/QuantumNous/new-api/relaykit/dto"
 	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/setting/model_setting"
 	"github.com/QuantumNous/new-api/setting/reasoning"
 	"github.com/gin-gonic/gin"
 )
@@ -98,6 +99,9 @@ func applyDeepSeekV4OpenAIThinkingSuffix(info *relaycommon.RelayInfo, request *d
 	if info != nil && info.ChannelMeta != nil && info.UpstreamModelName != "" {
 		modelName = info.UpstreamModelName
 	}
+	if model_setting.ShouldPreserveThinkingSuffix(modelName) || info != nil && model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) {
+		return nil
+	}
 	baseModel, thinkingType, effort, ok := reasoning.ParseDeepSeekV4ThinkingSuffix(modelName)
 	if !ok {
 		return nil
@@ -125,6 +129,9 @@ func applyDeepSeekV4ClaudeThinkingSuffix(info *relaycommon.RelayInfo, request *d
 	if info != nil && info.ChannelMeta != nil && info.UpstreamModelName != "" {
 		modelName = info.UpstreamModelName
 	}
+	if model_setting.ShouldPreserveThinkingSuffix(modelName) || info != nil && model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) {
+		return nil
+	}
 	baseModel, thinkingType, effort, ok := reasoning.ParseDeepSeekV4ThinkingSuffix(modelName)
 	if !ok {
 		return nil
@@ -170,6 +177,9 @@ func applyDeepSeekV4ResponsesThinkingSuffix(info *relaycommon.RelayInfo, request
 	if info != nil && info.ChannelMeta != nil && info.UpstreamModelName != "" {
 		modelName = info.UpstreamModelName
 	}
+	if model_setting.ShouldPreserveThinkingSuffix(modelName) || info != nil && model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) {
+		return
+	}
 	baseModel, thinkingType, effort, ok := reasoning.ParseDeepSeekV4ThinkingSuffix(modelName)
 	if ok {
 		if thinkingType == "disabled" {

relay/channel/openai/adaptor.go

@@ -249,6 +249,18 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 	if info.ChannelType != constant.ChannelTypeOpenAI && info.ChannelType != constant.ChannelTypeAzure {
 		request.StreamOptions = nil
 	}
+	// Nested reasoning is an OpenRouter-compatible input dialect and needs
+	// projection even without a protocol conversion hop. Native top-level
+	// reasoning_effort stays untouched unless a modifier or conversion applies.
+	// OpenRouter retains its own dialect normalization below.
+	preserveSuffix := model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) || model_setting.ShouldPreserveThinkingSuffix(info.UpstreamModelName)
+	upstreamEffort, _ := reasoning.ParseOpenAIReasoningEffortFromModelSuffix(info.UpstreamModelName)
+	originEffort, _ := reasoning.ParseOpenAIReasoningEffortFromModelSuffix(info.OriginModelName)
+	renderReasoning := len(request.Reasoning) > 0 || len(info.RequestConversionChain) > 1 || request.ReasoningConversion != nil || info.ReasoningState() != nil ||
+		!preserveSuffix && (upstreamEffort != "" || originEffort != "")
+	if info.ChannelType != constant.ChannelTypeOpenRouter && !renderReasoning {
+		info.SetReasoningEffort(request.ReasoningEffort)
+	}
 	if info.ChannelType == constant.ChannelTypeOpenRouter {
 		initialIntent, err := kitreasoning.FromOpenAIChat(request)
 		if err != nil {
@@ -273,7 +285,6 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 			request.Usage = json.RawMessage(`{"include":true}`)
 		}
 		// 合并 effort 尾巴产生的意图
-		preserveSuffix := model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) || model_setting.ShouldPreserveThinkingSuffix(info.UpstreamModelName)
 		mergeEffortSuffix := func(modelName string) error {
 			rawEffort, _ := reasoning.ParseOpenAIReasoningEffortFromModelSuffix(modelName)
 			if rawEffort == "" {
@@ -374,8 +385,7 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 		}
 	}
 
-	if info.ChannelType != constant.ChannelTypeOpenRouter {
-		preserveSuffix := model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) || model_setting.ShouldPreserveThinkingSuffix(info.UpstreamModelName)
+	if info.ChannelType != constant.ChannelTypeOpenRouter && renderReasoning {
 		effort, baseModel := reasoning.ParseOpenAIReasoningEffortFromModelSuffix(info.UpstreamModelName)
 		if preserveSuffix {
 			effort = ""
@@ -412,7 +422,7 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 			info.UpstreamModelName = baseModel
 			request.Model = baseModel
 		}
-		if canonicalEffort := kitreasoning.OpenAIEffort(kitreasoning.EffectiveEffort(currentIntent)); canonicalEffort != "" {
+		if canonicalEffort := kitreasoning.EffectiveEffort(currentIntent); canonicalEffort != "" {
 			request.ReasoningEffort = string(canonicalEffort)
 			info.SetReasoningEffort(string(canonicalEffort))
 		}
@@ -666,6 +676,21 @@ func (a *Adaptor) ConvertOpenAIResponsesRequest(c *gin.Context, info *relaycommo
 	if preserveSuffix {
 		effort = ""
 	}
+	originEffort := ""
+	if info != nil && !preserveSuffix {
+		originEffort, _ = reasoning.ParseOpenAIReasoningEffortFromModelSuffix(info.OriginModelName)
+	}
+	crossProtocol := info != nil && len(info.RequestConversionChain) > 1
+	if (info == nil || info.ChannelType != constant.ChannelTypeOpenRouter) && !crossProtocol && effort == "" && originEffort == "" && request.ReasoningConversion == nil && info.ReasoningState() == nil {
+		if info != nil {
+			rawEffort := ""
+			if request.Reasoning != nil {
+				rawEffort = request.Reasoning.Effort
+			}
+			info.SetReasoningEffort(rawEffort)
+		}
+		return request, nil
+	}
 	currentIntent, err := kitreasoning.FromOpenAIResponses(&request)
 	if err != nil {
 		return nil, kitreasoning.AsClientError(err)
@@ -700,7 +725,7 @@ func (a *Adaptor) ConvertOpenAIResponsesRequest(c *gin.Context, info *relaycommo
 			info.UpstreamModelName = originModel
 		}
 	}
-	if canonicalEffort := kitreasoning.OpenAIEffort(kitreasoning.EffectiveEffort(currentIntent)); canonicalEffort != "" {
+	if canonicalEffort := kitreasoning.EffectiveEffort(currentIntent); canonicalEffort != "" {
 		if request.Reasoning == nil {
 			request.Reasoning = &dto.Reasoning{}
 		}

relay/channel/volcengine/adaptor.go

@@ -308,6 +308,7 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 	}
 
 	if !model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) &&
+		!model_setting.ShouldPreserveThinkingSuffix(info.UpstreamModelName) &&
 		strings.HasSuffix(info.UpstreamModelName, "-thinking") &&
 		strings.HasPrefix(info.UpstreamModelName, "deepseek") {
 		info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-thinking")

relay/channel/xai/adaptor.go

@@ -11,6 +11,7 @@ import (
 	relaycommon "github.com/QuantumNous/new-api/relay/common"
 	"github.com/QuantumNous/new-api/relaykit/dto"
 	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/setting/model_setting"
 
 	"github.com/QuantumNous/new-api/relay/constant"
 
@@ -78,10 +79,11 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 			request.MaxCompletionTokens = request.MaxTokens
 			request.MaxTokens = nil
 		}
-		if strings.HasSuffix(request.Model, "-high") {
+		preserveSuffix := model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) || model_setting.ShouldPreserveThinkingSuffix(request.Model)
+		if !preserveSuffix && strings.HasSuffix(request.Model, "-high") {
 			request.ReasoningEffort = "high"
 			request.Model = strings.TrimSuffix(request.Model, "-high")
-		} else if strings.HasSuffix(request.Model, "-low") {
+		} else if !preserveSuffix && strings.HasSuffix(request.Model, "-low") {
 			request.ReasoningEffort = "low"
 			request.Model = strings.TrimSuffix(request.Model, "-low")
 		}

relay/common/relay_info.go

@@ -496,9 +496,10 @@ func reasoningEffortFromRequest(request dto.Request) string {
 		}
 	case *dto.GeminiChatRequest:
 		if req != nil && req.GenerationConfig.ThinkingConfig != nil {
-			intent, err := kitreasoning.FromGemini(req)
-			if err == nil {
-				effort = string(kitreasoning.EffectiveEffort(intent))
+			config := req.GenerationConfig.ThinkingConfig
+			effort = config.ThinkingLevel
+			if effort == "" && config.ThinkingBudget != nil {
+				effort = string(kitreasoning.EffortFromBudget(*config.ThinkingBudget))
 			}
 		}
 	}

relay/helper/model_modifier.go

@@ -211,7 +211,7 @@ func applyModelControls(req dto.Request, parsed parsedModelModifiers) error {
 				delete(reasoningConfig, "max_tokens")
 				request.ReasoningEffort = ""
 				if parsed.intent.Effort != "" {
-					request.ReasoningEffort = string(reasoning.OpenAIEffort(parsed.intent.Effort))
+					request.ReasoningEffort = string(parsed.intent.Effort)
 				}
 			}
 			if len(reasoningConfig) == 0 {
@@ -237,7 +237,7 @@ func applyModelControls(req dto.Request, parsed parsedModelModifiers) error {
 				if request.Reasoning == nil {
 					request.Reasoning = &dto.Reasoning{}
 				}
-				request.Reasoning.Effort = string(reasoning.OpenAIEffort(parsed.intent.Effort))
+				request.Reasoning.Effort = string(parsed.intent.Effort)
 			} else if request.Reasoning != nil && parsed.intent.BudgetTokens == nil {
 				request.Reasoning.Effort = ""
 			}

relay/helper/reasoning_suffix.go

@@ -106,6 +106,9 @@ func ApplyReasoningModelSuffix(c *gin.Context, info *relaycommon.RelayInfo, outb
 	if info.Request != nil {
 		info.Request.SetModelName(info.UpstreamModelName)
 	}
+	if selected.hasThinking {
+		info.SetReasoningEffort(string(reasoning.EffectiveEffort(selected.intent)))
+	}
 	for i := range diagnostics {
 		diagnostics[i].From = info.RelayFormat
 	}

relaykit/relayconvert/internal/shared/claude/reasoning.go

@@ -35,14 +35,19 @@ func ApplyReasoning(ctx context.Context, req *dto.ClaudeRequest, info convmeta.M
 	// accounting metadata, but do not run the capability renderer or rewrite
 	// provider-native controls.
 	if !crossProtocol && source.IsEmpty() && suffix.IsEmpty() {
-		native, err := reasoning.FromClaude(req)
-		if err != nil {
-			return err
-		}
 		if info != nil {
-			if effort := reasoning.EffectiveEffort(native); effort != "" {
-				info.SetReasoningEffort(string(effort))
+			effort := req.GetEfforts()
+			if effort == "" && req.Thinking != nil {
+				switch {
+				case req.Thinking.Type == "disabled":
+					effort = string(reasoning.EffortNone)
+				case req.Thinking.BudgetTokens != nil:
+					effort = string(reasoning.EffortFromBudget(*req.Thinking.BudgetTokens))
+				case req.Thinking.Type == "enabled" || req.Thinking.Type == "adaptive":
+					effort = string(reasoning.EffortHigh)
+				}
 			}
+			info.SetReasoningEffort(effort)
 		}
 		return nil
 	}

relaykit/relayconvert/internal/shared/gemini/request.go

@@ -106,14 +106,15 @@ func ApplyThinkingConfig(geminiRequest *dto.GeminiChatRequest, info convmeta.Met
 	// modifier, read portable effort metadata without running the capability
 	// renderer or rewriting provider-native controls.
 	if !crossProtocol && suffix.IsEmpty() {
-		native, err := reasoning.FromGemini(geminiRequest)
-		if err != nil {
-			return err
-		}
 		if info != nil {
-			if effort := reasoning.EffectiveEffort(native); effort != "" {
-				info.SetReasoningEffort(string(effort))
+			effort := ""
+			if config := geminiRequest.GenerationConfig.ThinkingConfig; config != nil {
+				effort = config.ThinkingLevel
+				if effort == "" && config.ThinkingBudget != nil {
+					effort = string(reasoning.EffortFromBudget(*config.ThinkingBudget))
+				}
 			}
+			info.SetReasoningEffort(effort)
 		}
 		return nil
 	}

relaykit/relayconvert/reasoning/intent.go

@@ -363,8 +363,8 @@ func FromOpenAIChat(req *dto.GeneralOpenAIRequest) (Intent, error) {
 		BudgetSource:    SourcePivot,
 	}
 	if req.ReasoningEffort != "" {
-		projectedEffort := OpenAIEffort(EffectiveEffort(pivot))
-		if Effort(req.ReasoningEffort) == projectedEffort {
+		pivotEffort := EffectiveEffort(pivot)
+		if Effort(req.ReasoningEffort) == pivotEffort {
 			intent.Effort = ""
 			intent.Mode = ModeUnset
 		}
@@ -384,7 +384,7 @@ func ApplyToOpenAIChat(req *dto.GeneralOpenAIRequest, intent Intent) error {
 		return err
 	}
 
-	if effort := OpenAIEffort(EffectiveEffort(intent)); effort != "" {
+	if effort := EffectiveEffort(intent); effort != "" {
 		req.ReasoningEffort = string(effort)
 	}
 
@@ -412,7 +412,7 @@ func ApplyToOpenAIResponses(req *dto.OpenAIResponsesRequest, intent Intent) erro
 		return err
 	}
 
-	if effort := OpenAIEffort(EffectiveEffort(intent)); effort != "" {
+	if effort := EffectiveEffort(intent); effort != "" {
 		summary := "detailed"
 		if effort == EffortNone || (intent.IncludeThoughts != nil && !*intent.IncludeThoughts) {
 			summary = ""
@@ -436,16 +436,6 @@ func ApplyToOpenAIResponses(req *dto.OpenAIResponsesRequest, intent Intent) erro
 	return nil
 }
 
-// OpenAIEffort maps the canonical cross-provider vocabulary to the public
-// OpenAI reasoning_effort vocabulary. Claude/OpenRouter "max" has no direct
-// OpenAI equivalent and is represented by xhigh at that wire boundary.
-func OpenAIEffort(effort Effort) Effort {
-	if effort == EffortMax {
-		return EffortXHigh
-	}
-	return effort
-}
-
 func FromOpenAIResponses(req *dto.OpenAIResponsesRequest) (Intent, error) {
 	if req == nil {
 		return Intent{}, nil
@@ -481,8 +471,8 @@ func FromOpenAIResponses(req *dto.OpenAIResponsesRequest) (Intent, error) {
 		BudgetSource:    SourcePivot,
 	}
 	if req.Reasoning != nil && req.Reasoning.Effort != "" {
-		projectedEffort := OpenAIEffort(EffectiveEffort(pivot))
-		if Effort(req.Reasoning.Effort) == projectedEffort {
+		pivotEffort := EffectiveEffort(pivot)
+		if Effort(req.Reasoning.Effort) == pivotEffort {
 			intent.Effort = ""
 			intent.Mode = ModeUnset
 		}

relaykit/relayconvert/reasoning/intent_test.go

@@ -3,6 +3,7 @@ package reasoning
 import (
 	"testing"
 
+	"github.com/QuantumNous/new-api/relaykit/dto"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -123,3 +124,49 @@ func TestIntentStateRoundTrip(t *testing.T) {
 func boolPtr(v bool) *bool {
 	return &v
 }
+
+func TestOpenAIPivotRetainsExactStrengthAndBudget(t *testing.T) {
+	budget, include := 16384, false
+	for _, effort := range []Effort{EffortMax, EffortXHigh} {
+		t.Run(string(effort), func(t *testing.T) {
+			intent := Intent{Mode: ModeEnabled, Effort: effort, BudgetTokens: &budget, IncludeThoughts: &include}
+			chat := &dto.GeneralOpenAIRequest{}
+			require.NoError(t, ApplyToOpenAIChat(chat, intent))
+			assert.Equal(t, string(effort), chat.ReasoningEffort)
+			restored, err := FromOpenAIChat(chat)
+			require.NoError(t, err)
+			assert.Equal(t, effort, restored.Effort)
+			require.NotNil(t, restored.BudgetTokens)
+			assert.Equal(t, budget, *restored.BudgetTokens)
+			require.NotNil(t, restored.IncludeThoughts)
+			assert.False(t, *restored.IncludeThoughts)
+
+			responses := &dto.OpenAIResponsesRequest{}
+			require.NoError(t, ApplyToOpenAIResponses(responses, restored))
+			require.NotNil(t, responses.Reasoning)
+			assert.Equal(t, string(effort), responses.Reasoning.Effort)
+			restored, err = FromOpenAIResponses(responses)
+			require.NoError(t, err)
+			assert.Equal(t, effort, restored.Effort)
+			require.NotNil(t, restored.BudgetTokens)
+			assert.Equal(t, budget, *restored.BudgetTokens)
+			require.NotNil(t, restored.IncludeThoughts)
+			assert.False(t, *restored.IncludeThoughts)
+		})
+	}
+}
+
+func TestOpenAIPivotDoesNotTreatMaxAndXHighAsEquivalent(t *testing.T) {
+	intent := Intent{Mode: ModeEnabled, Effort: EffortMax}
+	chat := &dto.GeneralOpenAIRequest{}
+	require.NoError(t, ApplyToOpenAIChat(chat, intent))
+	chat.ReasoningEffort = "xhigh"
+	_, err := FromOpenAIChat(chat)
+	require.ErrorIs(t, err, ErrEffortConflict)
+
+	responses := &dto.OpenAIResponsesRequest{}
+	require.NoError(t, ApplyToOpenAIResponses(responses, intent))
+	responses.Reasoning.Effort = "xhigh"
+	_, err = FromOpenAIResponses(responses)
+	require.ErrorIs(t, err, ErrEffortConflict)
+}