feat: support OpenRouter reasoning

common/conv/any.go

@@ -1,6 +1,9 @@
 package conv
 
 func AsString(v any) string {
-	str, _ := v.(string)
+	if str, ok := v.(string); ok {
 		return str
 	}
+
+	return ""
+}

relay/adaptor/openai/adaptor.go

@@ -17,6 +17,7 @@ import (
 	"github.com/songquanpeng/one-api/relay/adaptor/geminiv2"
 	"github.com/songquanpeng/one-api/relay/adaptor/minimax"
 	"github.com/songquanpeng/one-api/relay/adaptor/novita"
+	"github.com/songquanpeng/one-api/relay/adaptor/openrouter"
 	"github.com/songquanpeng/one-api/relay/channeltype"
 	"github.com/songquanpeng/one-api/relay/meta"
 	"github.com/songquanpeng/one-api/relay/model"
@@ -95,6 +96,21 @@ func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.G
 		return nil, errors.New("request is nil")
 	}
 
+	meta := meta.GetByContext(c)
+	switch meta.ChannelType {
+	case channeltype.OpenRouter:
+		includeReasoning := true
+		request.IncludeReasoning = &includeReasoning
+		if request.Provider == nil || request.Provider.Sort == "" {
+			if request.Provider == nil {
+				request.Provider = &openrouter.RequestProvider{}
+			}
+
+			request.Provider.Sort = "throughput"
+		}
+	default:
+	}
+
 	if request.Stream && !config.EnforceIncludeUsage {
 		logger.Warn(c.Request.Context(),
 			"please set ENFORCE_INCLUDE_USAGE=true to ensure accurate billing in stream mode")

relay/adaptor/openai/main.go

@@ -27,6 +27,7 @@ const (
 
 func StreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*model.ErrorWithStatusCode, string, *model.Usage) {
 	responseText := ""
+	reasoningText := ""
 	scanner := bufio.NewScanner(resp.Body)
 	scanner.Split(bufio.ScanLines)
 	var usage *model.Usage
@@ -62,6 +63,10 @@ func StreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*model.E
 			}
 			render.StringData(c, data)
 			for _, choice := range streamResponse.Choices {
+				if choice.Delta.Reasoning != nil {
+					reasoningText += *choice.Delta.Reasoning
+				}
+
 				responseText += conv.AsString(choice.Delta.Content)
 			}
 			if streamResponse.Usage != nil {
@@ -94,7 +99,7 @@ func StreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*model.E
 		return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), "", nil
 	}
 
-	return nil, responseText, usage
+	return nil, reasoningText + responseText, usage
 }
 
 // Handler handles the non-stream response from OpenAI API
@@ -150,20 +155,26 @@ func Handler(c *gin.Context, resp *http.Response, promptTokens int, modelName st
 			CompletionTokens: completionTokens,
 			TotalTokens:      promptTokens + completionTokens,
 		}
-	} else if textResponse.PromptTokensDetails.AudioTokens+textResponse.CompletionTokensDetails.AudioTokens > 0 {
+	} else if (textResponse.PromptTokensDetails != nil && textResponse.PromptTokensDetails.AudioTokens > 0) ||
+		(textResponse.CompletionTokensDetails != nil && textResponse.CompletionTokensDetails.AudioTokens > 0) {
 		// Convert the more expensive audio tokens to uniformly priced text tokens.
 		// Note that when there are no audio tokens in prompt and completion,
 		// OpenAI will return empty PromptTokensDetails and CompletionTokensDetails, which can be misleading.
+		if textResponse.PromptTokensDetails != nil {
 			textResponse.Usage.PromptTokens = textResponse.PromptTokensDetails.TextTokens +
 				int(math.Ceil(
 					float64(textResponse.PromptTokensDetails.AudioTokens)*
 						ratio.GetAudioPromptRatio(modelName),
 				))
+		}
+
+		if textResponse.CompletionTokensDetails != nil {
 			textResponse.Usage.CompletionTokens = textResponse.CompletionTokensDetails.TextTokens +
 				int(math.Ceil(
 					float64(textResponse.CompletionTokensDetails.AudioTokens)*
 						ratio.GetAudioPromptRatio(modelName)*ratio.GetAudioCompletionRatio(modelName),
 				))
+		}
 
 		textResponse.Usage.TotalTokens = textResponse.Usage.PromptTokens +
 			textResponse.Usage.CompletionTokens

relay/adaptor/openrouter/model.go

@@ -0,0 +1,22 @@
+package openrouter
+
+// RequestProvider customize how your requests are routed using the provider object
+// in the request body for Chat Completions and Completions.
+//
+// https://openrouter.ai/docs/features/provider-routing
+type RequestProvider struct {
+	// Order is list of provider names to try in order (e.g. ["Anthropic", "OpenAI"]). Default: empty
+	Order []string `json:"order,omitempty"`
+	// AllowFallbacks is whether to allow backup providers when the primary is unavailable. Default: true
+	AllowFallbacks bool `json:"allow_fallbacks,omitempty"`
+	// RequireParameters is only use providers that support all parameters in your request. Default: false
+	RequireParameters bool `json:"require_parameters,omitempty"`
+	// DataCollection is control whether to use providers that may store data ("allow" or "deny"). Default: "allow"
+	DataCollection string `json:"data_collection,omitempty" binding:"omitempty,oneof=allow deny"`
+	// Ignore is list of provider names to skip for this request. Default: empty
+	Ignore []string `json:"ignore,omitempty"`
+	// Quantizations is list of quantization levels to filter by (e.g. ["int4", "int8"]). Default: empty
+	Quantizations []string `json:"quantizations,omitempty"`
+	// Sort is sort providers by price or throughput (e.g. "price" or "throughput"). Default: empty
+	Sort string `json:"sort,omitempty" binding:"omitempty,oneof=price throughput latency"`
+}

relay/model/general.go

@@ -1,5 +1,7 @@
 package model
 
+import "github.com/songquanpeng/one-api/relay/adaptor/openrouter"
+
 type ResponseFormat struct {
 	Type       string      `json:"type,omitempty"`
 	JsonSchema *JSONSchema `json:"json_schema,omitempty"`
@@ -68,6 +70,9 @@ type GeneralOpenAIRequest struct {
 	// Others
 	Instruction string `json:"instruction,omitempty"`
 	NumCtx      int    `json:"num_ctx,omitempty"`
+	// Openrouter
+	Provider         *openrouter.RequestProvider `json:"provider,omitempty"`
+	IncludeReasoning *bool                       `json:"include_reasoning,omitempty"`
 }
 
 func (r GeneralOpenAIRequest) ParseInput() []string {

relay/model/message.go

@@ -24,6 +24,11 @@ type Message struct {
 	// Prefix Completion feature as the input for the CoT in the last assistant message.
 	// When using this feature, the prefix parameter must be set to true.
 	ReasoningContent *string `json:"reasoning_content,omitempty"`
+	// -------------------------------------
+	// Openrouter
+	// -------------------------------------
+	Reasoning *string `json:"reasoning,omitempty"`
+	Refusal   *bool   `json:"refusal,omitempty"`
 }
 
 type messageAudio struct {

relay/model/misc.go

@@ -5,9 +5,9 @@ type Usage struct {
 	CompletionTokens int `json:"completion_tokens"`
 	TotalTokens      int `json:"total_tokens"`
 	// PromptTokensDetails may be empty for some models
-	PromptTokensDetails usagePromptTokensDetails `gorm:"-" json:"prompt_tokens_details,omitempty"`
+	PromptTokensDetails *usagePromptTokensDetails `gorm:"-" json:"prompt_tokens_details,omitempty"`
 	// CompletionTokensDetails may be empty for some models
-	CompletionTokensDetails usageCompletionTokensDetails `gorm:"-" json:"completion_tokens_details,omitempty"`
+	CompletionTokensDetails *usageCompletionTokensDetails `gorm:"-" json:"completion_tokens_details,omitempty"`
 	ServiceTier             string                        `gorm:"-" json:"service_tier,omitempty"`
 	SystemFingerprint       string                        `gorm:"-" json:"system_fingerprint,omitempty"`
 }