refactor: extract protocol conversion layer into standalone relaykit module (#6369)

* test(relayconvert): add golden snapshot matrix and relaykit boundary guard

Phase 0 of the relaykit extraction plan: pin byte-level output of every
registered (from,to) request/response/stream conversion route, and
forbid kit-bound packages from growing host-only imports.

* wip(relayconvert): drop gin.Context from converter signatures; add convmeta draft

Phase 1 in progress: relayconvert now takes context.Context; host media
resolver adapts gin.Context back at the service boundary.

* refactor(relayconvert): decouple converters from RelayInfo, gin, and settings

Phase 1 of the relaykit extraction plan:
- converters now depend on convmeta.Meta (implemented by RelayInfo) instead
  of *relaycommon.RelayInfo; ClaudeConvertInfo and the format guesser move
  to convmeta with aliases left behind
- host settings reach converters via a convmeta.Options snapshot built in
  RelayInfo.ConvOptions; no more model_setting/reasoning global reads inside
  the conversion layer
- effort-suffix helpers move to service/relayconvert/reasoning (old package
  forwards); chat-to-responses upgrade policy moves to service (host routing
  logic, not conversion)
- golden conversion matrix unchanged

* test(relayconvert): tighten boundary — kit packages now free of gin/setting imports

* refactor(dto): drop gin and logger dependencies

Phase 2 (part 1): dto.Request.IsStream now takes *http.Request instead of
*gin.Context (Gemini's impl reads query/path off the std request); dto's
three logger calls become common.SysError. Boundary test allowlist is now
empty — kit-bound packages import no gin/setting/logger/model.

* refactor(kit): extract dependency-free kitutil; dto/types/relayconvert stop importing common

Phase 2 of the relaykit extraction plan:
- new service/relayconvert/kitutil holds the pure helpers the kit needs
  (JSON wrappers, pointer/string/uuid/timestamp utils, MaskSensitiveInfo,
  pluggable LogInfo/LogError hooks, Debug flag)
- dto, types, and all relayconvert packages now use kitutil; their only
  remaining internal deps are dto/types/constant
- common keeps every original symbol (MaskSensitiveInfo delegates to
  kitutil) so host code is untouched; main.go routes kit logging into
  common.SysLog/SysError and mirrors DebugEnabled
- golden conversion matrix unchanged

* refactor(kit): move EndpointType/FinishReason to types; OpenRouter dialect via Options

Kit packages (dto/types/relayconvert/reasonmap) no longer import constant:
- EndpointType and finish-reason values live in types; constant re-exports
- the OpenRouter special-case in claude->openai request conversion reads
  Options.OpenRouterDialect, set by the host from the channel type;
  InitChannelMeta invalidates the cached snapshot on channel switch

* refactor: extract relaykit submodule (dto/types/relayconvert/reasonmap)

Phase 3 of the relaykit extraction plan:
- new go module github.com/QuantumNous/new-api/relaykit containing dto
  (minus task family), types, relayconvert (with convmeta/kitutil/reasoning),
  and reasonmap; host consumes it via require + replace, go.work for dev
- task-family dto (task/suno/midjourney/video) stays in the host dto
  package; dual-consumer host files alias it as taskdto
- relaykit builds and tests standalone (GOWORK=off): no host imports,
  no gin, no DB, no settings
- golden conversion matrix unchanged

* build(docker): copy relaykit/go.mod before go mod download

The local-replace submodule's go.mod must exist inside the build context
for the main module graph to resolve.

* fix: address relaykit extraction regressions

* fix: address relaykit review regressions

* docs: document Meta nil receiver contract

* fix(relaykit): fail OpenAI→Claude conversion without max_tokens; reject negative default_max_tokens

The Claude Messages API requires max_tokens (omitting it is a 400
"Field required"), but with a nil Options.Claude.DefaultMaxTokens hook
the converters silently emitted a request the upstream is guaranteed to
reject. Both OpenAI Chat and Responses → Claude conversions now return
sharedclaude.ErrMissingMaxTokens when no path (client value, default
hook, thinking-adapter floor) supplied one. Unreachable in the host,
which always configures the hook.

Host side, claude.default_max_tokens now rejects negative values at the
option API before persisting — they would wrap into huge unsigned values
during conversion. Zero stays allowed: the current API treats
max_tokens: 0 as cache pre-warming.

* fix: make Gemini safety settings read path race-free

relaykit/relayconvert/internal/claude_messages/to_oai_chat_req.go

@@ -0,0 +1,219 @@
+package claudemessages
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+const (
+	webSearchMaxUsesLow    = 1
+	webSearchMaxUsesMedium = 5
+	webSearchMaxUsesHigh   = 10
+)
+
+type openRouterRequestReasoning struct {
+	Enabled   bool   `json:"enabled"`
+	Effort    string `json:"effort,omitempty"`
+	MaxTokens int    `json:"max_tokens,omitempty"`
+	Exclude   bool   `json:"exclude,omitempty"`
+}
+
+func ClaudeMessagesRequestToOpenAIChat(claudeRequest dto.ClaudeRequest, info convmeta.Meta) (*dto.GeneralOpenAIRequest, error) {
+	openAIRequest := dto.GeneralOpenAIRequest{
+		Model:       claudeRequest.Model,
+		Temperature: claudeRequest.Temperature,
+	}
+	if claudeRequest.MaxTokens != nil {
+		openAIRequest.MaxTokens = kitutil.GetPointer(*claudeRequest.MaxTokens)
+	}
+	if claudeRequest.TopP != nil {
+		openAIRequest.TopP = kitutil.GetPointer(*claudeRequest.TopP)
+	}
+	if claudeRequest.TopK != nil {
+		openAIRequest.TopK = kitutil.GetPointer(*claudeRequest.TopK)
+	}
+	if claudeRequest.Stream != nil {
+		openAIRequest.Stream = kitutil.GetPointer(*claudeRequest.Stream)
+	}
+
+	isOpenRouter := convmeta.OptionsOf(info).OpenRouterDialect
+	if isOpenRouter {
+		if effort := claudeRequest.GetEfforts(); effort != "" {
+			effortBytes, _ := kitutil.Marshal(effort)
+			openAIRequest.Verbosity = effortBytes
+		}
+		if claudeRequest.Thinking != nil {
+			var reasoningConfig openRouterRequestReasoning
+			if claudeRequest.Thinking.Type == "enabled" {
+				reasoningConfig = openRouterRequestReasoning{
+					Enabled:   true,
+					MaxTokens: claudeRequest.Thinking.GetBudgetTokens(),
+				}
+			} else if claudeRequest.Thinking.Type == "adaptive" {
+				reasoningConfig = openRouterRequestReasoning{
+					Enabled: true,
+				}
+			}
+			reasoningJSON, err := kitutil.Marshal(reasoningConfig)
+			if err != nil {
+				return nil, fmt.Errorf("failed to marshal reasoning: %w", err)
+			}
+			openAIRequest.Reasoning = reasoningJSON
+		}
+	} else if info != nil {
+		thinkingSuffix := "-thinking"
+		if strings.HasSuffix(info.GetOriginModelName(), thinkingSuffix) &&
+			!strings.HasSuffix(openAIRequest.Model, thinkingSuffix) {
+			openAIRequest.Model = openAIRequest.Model + thinkingSuffix
+		}
+	}
+
+	if len(claudeRequest.StopSequences) == 1 {
+		openAIRequest.Stop = claudeRequest.StopSequences[0]
+	} else if len(claudeRequest.StopSequences) > 1 {
+		openAIRequest.Stop = claudeRequest.StopSequences
+	}
+
+	tools, _ := kitutil.Any2Type[[]dto.Tool](claudeRequest.Tools)
+	openAITools := make([]dto.ToolCallRequest, 0)
+	for _, claudeTool := range tools {
+		openAITool := dto.ToolCallRequest{
+			Type: "function",
+			Function: dto.FunctionRequest{
+				Name:        claudeTool.Name,
+				Description: claudeTool.Description,
+				Parameters:  claudeTool.InputSchema,
+			},
+		}
+		openAITools = append(openAITools, openAITool)
+	}
+	openAIRequest.Tools = openAITools
+
+	openAIMessages := make([]dto.Message, 0)
+	if claudeRequest.System != nil {
+		if claudeRequest.IsStringSystem() && claudeRequest.GetStringSystem() != "" {
+			openAIMessage := dto.Message{
+				Role: "system",
+			}
+			openAIMessage.SetStringContent(claudeRequest.GetStringSystem())
+			openAIMessages = append(openAIMessages, openAIMessage)
+		} else {
+			systems := claudeRequest.ParseSystem()
+			if len(systems) > 0 {
+				openAIMessage := dto.Message{
+					Role: "system",
+				}
+				isOpenRouterClaude := isOpenRouter && strings.HasPrefix(convmeta.UpstreamModelName(info), "anthropic/claude")
+				if isOpenRouterClaude {
+					systemMediaMessages := make([]dto.MediaContent, 0, len(systems))
+					for _, system := range systems {
+						message := dto.MediaContent{
+							Type:         "text",
+							Text:         system.GetText(),
+							CacheControl: system.CacheControl,
+						}
+						systemMediaMessages = append(systemMediaMessages, message)
+					}
+					openAIMessage.SetMediaContent(systemMediaMessages)
+				} else {
+					systemStr := ""
+					for _, system := range systems {
+						if system.Text != nil {
+							systemStr += *system.Text
+						}
+					}
+					openAIMessage.SetStringContent(systemStr)
+				}
+				openAIMessages = append(openAIMessages, openAIMessage)
+			}
+		}
+	}
+
+	for _, claudeMessage := range claudeRequest.Messages {
+		openAIMessage := dto.Message{
+			Role: claudeMessage.Role,
+		}
+		if claudeMessage.IsStringContent() {
+			openAIMessage.SetStringContent(claudeMessage.GetStringContent())
+		} else {
+			content, err := claudeMessage.ParseContent()
+			if err != nil {
+				return nil, err
+			}
+			var toolCalls []dto.ToolCallRequest
+			mediaMessages := make([]dto.MediaContent, 0, len(content))
+
+			for _, mediaMsg := range content {
+				switch mediaMsg.Type {
+				case "text", "input_text":
+					message := dto.MediaContent{
+						Type:         "text",
+						Text:         mediaMsg.GetText(),
+						CacheControl: mediaMsg.CacheControl,
+					}
+					mediaMessages = append(mediaMessages, message)
+				case "image":
+					imageData := fmt.Sprintf("data:%s;base64,%s", mediaMsg.Source.MediaType, mediaMsg.Source.Data)
+					mediaMessage := dto.MediaContent{
+						Type:     "image_url",
+						ImageUrl: &dto.MessageImageUrl{Url: imageData},
+					}
+					mediaMessages = append(mediaMessages, mediaMessage)
+				case "tool_use":
+					toolCall := dto.ToolCallRequest{
+						ID:   mediaMsg.Id,
+						Type: "function",
+						Function: dto.FunctionRequest{
+							Name:      mediaMsg.Name,
+							Arguments: requestToJSONString(mediaMsg.Input),
+						},
+					}
+					toolCalls = append(toolCalls, toolCall)
+				case "tool_result":
+					toolName := mediaMsg.Name
+					if toolName == "" {
+						toolName = claudeRequest.SearchToolNameByToolCallId(mediaMsg.ToolUseId)
+					}
+					oaiToolMessage := dto.Message{
+						Role:       "tool",
+						Name:       &toolName,
+						ToolCallId: mediaMsg.ToolUseId,
+					}
+					if mediaMsg.IsStringContent() {
+						oaiToolMessage.SetStringContent(mediaMsg.GetStringContent())
+					} else {
+						mediaContents := mediaMsg.ParseMediaContent()
+						encodedJSON, _ := kitutil.Marshal(mediaContents)
+						oaiToolMessage.SetStringContent(string(encodedJSON))
+					}
+					openAIMessages = append(openAIMessages, oaiToolMessage)
+				}
+			}
+
+			if len(toolCalls) > 0 {
+				openAIMessage.SetToolCalls(toolCalls)
+			}
+			if len(mediaMessages) > 0 && len(toolCalls) == 0 {
+				openAIMessage.SetMediaContent(mediaMessages)
+			}
+		}
+		if len(openAIMessage.ParseContent()) > 0 || len(openAIMessage.ToolCalls) > 0 {
+			openAIMessages = append(openAIMessages, openAIMessage)
+		}
+	}
+
+	openAIRequest.Messages = openAIMessages
+	return &openAIRequest, nil
+}
+
+func requestToJSONString(v interface{}) string {
+	b, err := kitutil.Marshal(v)
+	if err != nil {
+		return "{}"
+	}
+	return string(b)
+}

relaykit/relayconvert/internal/claude_messages/to_oai_chat_resp.go

@@ -0,0 +1,400 @@
+package claudemessages
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/reasonmap"
+	sharedclaude "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/shared/claude"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+type ClaudeResponseInfo struct {
+	ResponseId   string
+	Created      int64
+	Model        string
+	ResponseText strings.Builder
+	Usage        *dto.Usage
+	Done         bool
+}
+
+func StopReasonClaudeToOpenAI(reason string) string {
+	return reasonmap.ClaudeStopReasonToOpenAIFinishReason(reason)
+}
+
+func StreamResponseClaude2OpenAI(claudeResponse *dto.ClaudeResponse) *dto.ChatCompletionsStreamResponse {
+	var response dto.ChatCompletionsStreamResponse
+	response.Object = "chat.completion.chunk"
+	response.Model = claudeResponse.Model
+	response.Choices = make([]dto.ChatCompletionsStreamResponseChoice, 0)
+	tools := make([]dto.ToolCallResponse, 0)
+	fcIdx := 0
+	if claudeResponse.Index != nil {
+		fcIdx = *claudeResponse.Index
+	}
+	var choice dto.ChatCompletionsStreamResponseChoice
+	if claudeResponse.Type == "message_start" {
+		if claudeResponse.Message != nil {
+			response.Id = claudeResponse.Message.Id
+			response.Model = claudeResponse.Message.Model
+		}
+		choice.Delta.SetContentString("")
+		choice.Delta.Role = "assistant"
+	} else if claudeResponse.Type == "content_block_start" {
+		if claudeResponse.ContentBlock != nil {
+			if claudeResponse.ContentBlock.Type == "text" && claudeResponse.ContentBlock.Text != nil {
+				choice.Delta.SetContentString(*claudeResponse.ContentBlock.Text)
+			}
+			if claudeResponse.ContentBlock.Type == "tool_use" {
+				tools = append(tools, dto.ToolCallResponse{
+					Index: kitutil.GetPointer(fcIdx),
+					ID:    claudeResponse.ContentBlock.Id,
+					Type:  "function",
+					Function: dto.FunctionResponse{
+						Name:      claudeResponse.ContentBlock.Name,
+						Arguments: "",
+					},
+				})
+			}
+		} else {
+			return nil
+		}
+	} else if claudeResponse.Type == "content_block_delta" {
+		if claudeResponse.Delta != nil {
+			choice.Delta.Content = claudeResponse.Delta.Text
+			switch claudeResponse.Delta.Type {
+			case "input_json_delta":
+				tools = append(tools, dto.ToolCallResponse{
+					Type:  "function",
+					Index: kitutil.GetPointer(fcIdx),
+					Function: dto.FunctionResponse{
+						Arguments: *claudeResponse.Delta.PartialJson,
+					},
+				})
+			case "signature_delta":
+				signatureContent := "\n"
+				choice.Delta.ReasoningContent = &signatureContent
+			case "thinking_delta":
+				choice.Delta.ReasoningContent = claudeResponse.Delta.Thinking
+			}
+		}
+	} else if claudeResponse.Type == "message_delta" {
+		if claudeResponse.Delta != nil && claudeResponse.Delta.StopReason != nil {
+			finishReason := StopReasonClaudeToOpenAI(*claudeResponse.Delta.StopReason)
+			if finishReason != "null" {
+				choice.FinishReason = &finishReason
+			}
+		}
+	} else if claudeResponse.Type == "message_stop" {
+		return nil
+	} else {
+		return nil
+	}
+	if len(tools) > 0 {
+		choice.Delta.Content = nil
+		choice.Delta.ToolCalls = tools
+	}
+	response.Choices = append(response.Choices, choice)
+
+	return &response
+}
+
+func ResponseClaude2OpenAI(claudeResponse *dto.ClaudeResponse) *dto.OpenAITextResponse {
+	choices := make([]dto.OpenAITextResponseChoice, 0)
+	fullTextResponse := dto.OpenAITextResponse{
+		Id:      fmt.Sprintf("chatcmpl-%s", kitutil.GetUUID()),
+		Object:  "chat.completion",
+		Created: kitutil.GetTimestamp(),
+	}
+	var responseText string
+	var responseThinking string
+	if len(claudeResponse.Content) > 0 {
+		responseText = claudeResponse.Content[0].GetText()
+		if claudeResponse.Content[0].Thinking != nil {
+			responseThinking = *claudeResponse.Content[0].Thinking
+		}
+	}
+	tools := make([]dto.ToolCallResponse, 0)
+	thinkingContent := ""
+
+	fullTextResponse.Id = claudeResponse.Id
+	for _, message := range claudeResponse.Content {
+		switch message.Type {
+		case "tool_use":
+			args, _ := kitutil.Marshal(message.Input)
+			tools = append(tools, dto.ToolCallResponse{
+				ID:   message.Id,
+				Type: "function",
+				Function: dto.FunctionResponse{
+					Name:      message.Name,
+					Arguments: string(args),
+				},
+			})
+		case "thinking":
+			if message.Thinking != nil {
+				thinkingContent = *message.Thinking
+			}
+		case "text":
+			responseText = message.GetText()
+		}
+	}
+	choice := dto.OpenAITextResponseChoice{
+		Index: 0,
+		Message: dto.Message{
+			Role: "assistant",
+		},
+		FinishReason: StopReasonClaudeToOpenAI(claudeResponse.StopReason),
+	}
+	choice.SetStringContent(responseText)
+	if len(responseThinking) > 0 {
+		choice.ReasoningContent = &responseThinking
+	}
+	if len(tools) > 0 {
+		choice.Message.SetToolCalls(tools)
+	}
+	if thinkingContent != "" {
+		choice.Message.ReasoningContent = &thinkingContent
+	}
+	fullTextResponse.Model = claudeResponse.Model
+	choices = append(choices, choice)
+	fullTextResponse.Choices = choices
+	return &fullTextResponse
+}
+
+func UsageFromClaudeAPIUsage(usage *dto.ClaudeUsage) *dto.Usage {
+	if usage == nil {
+		return &dto.Usage{}
+	}
+	semanticUsage := &dto.Usage{
+		PromptTokens:     usage.InputTokens,
+		CompletionTokens: usage.OutputTokens,
+		UsageSemantic:    "anthropic",
+		UsageSource:      "anthropic",
+		BillingUsage:     dto.CloneBillingUsage(usage.BillingUsage),
+	}
+	if semanticUsage.BillingUsage == nil {
+		semanticUsage.BillingUsage = dto.NewClaudeMessagesBillingUsage(usage)
+	}
+	semanticUsage.PromptTokensDetails.CachedTokens = usage.CacheReadInputTokens
+	semanticUsage.PromptTokensDetails.CachedCreationTokens = usage.CacheCreationInputTokens
+	semanticUsage.ClaudeCacheCreation5mTokens = usage.GetCacheCreation5mTokens()
+	semanticUsage.ClaudeCacheCreation1hTokens = usage.GetCacheCreation1hTokens()
+	return UsageFromClaudeUsage(semanticUsage)
+}
+
+func UsageFromClaudeUsage(usage *dto.Usage) *dto.Usage {
+	mapped := buildOpenAIStyleUsageFromClaudeUsage(usage)
+	return &mapped
+}
+
+func cacheCreationTokensForOpenAIUsage(usage *dto.Usage) int {
+	if usage == nil {
+		return 0
+	}
+	splitCacheCreationTokens := usage.ClaudeCacheCreation5mTokens + usage.ClaudeCacheCreation1hTokens
+	if splitCacheCreationTokens == 0 {
+		return usage.PromptTokensDetails.CachedCreationTokens
+	}
+	if usage.PromptTokensDetails.CachedCreationTokens > splitCacheCreationTokens {
+		return usage.PromptTokensDetails.CachedCreationTokens
+	}
+	return splitCacheCreationTokens
+}
+
+func buildOpenAIStyleUsageFromClaudeUsage(usage *dto.Usage) dto.Usage {
+	if usage == nil {
+		return dto.Usage{}
+	}
+	clone := *usage
+	clone.BillingUsage = dto.CloneBillingUsage(usage.BillingUsage)
+	clone.ClaudeCacheCreation5mTokens, clone.ClaudeCacheCreation1hTokens = sharedclaude.NormalizeCacheCreationSplit(
+		usage.PromptTokensDetails.CachedCreationTokens,
+		usage.ClaudeCacheCreation5mTokens,
+		usage.ClaudeCacheCreation1hTokens,
+	)
+	cacheCreationTokens := cacheCreationTokensForOpenAIUsage(usage)
+	// Expose the standard OpenAI cache-write field alongside the legacy
+	// cached_creation_tokens so OpenAI-format clients can bill cache writes.
+	clone.PromptTokensDetails.CacheWriteTokens = cacheCreationTokens
+	totalInputTokens := usage.PromptTokens + usage.PromptTokensDetails.CachedTokens + cacheCreationTokens
+	clone.PromptTokens = totalInputTokens
+	clone.InputTokens = totalInputTokens
+	clone.TotalTokens = totalInputTokens + usage.CompletionTokens
+	clone.UsageSemantic = "openai"
+	clone.UsageSource = "anthropic"
+	return clone
+}
+
+func BuildMessageDeltaPatchUsage(claudeResponse *dto.ClaudeResponse, claudeInfo *ClaudeResponseInfo) *dto.ClaudeUsage {
+	usage := &dto.ClaudeUsage{}
+	if claudeResponse != nil && claudeResponse.Usage != nil {
+		*usage = *claudeResponse.Usage
+	}
+
+	if claudeInfo == nil || claudeInfo.Usage == nil {
+		return usage
+	}
+
+	if usage.InputTokens == 0 && claudeInfo.Usage.PromptTokens > 0 {
+		usage.InputTokens = claudeInfo.Usage.PromptTokens
+	}
+	if usage.CacheReadInputTokens == 0 && claudeInfo.Usage.PromptTokensDetails.CachedTokens > 0 {
+		usage.CacheReadInputTokens = claudeInfo.Usage.PromptTokensDetails.CachedTokens
+	}
+	if usage.CacheCreationInputTokens == 0 && claudeInfo.Usage.PromptTokensDetails.CachedCreationTokens > 0 {
+		usage.CacheCreationInputTokens = claudeInfo.Usage.PromptTokensDetails.CachedCreationTokens
+	}
+	cacheCreation5m := 0
+	cacheCreation1h := 0
+	if usage.CacheCreation != nil {
+		cacheCreation5m = usage.CacheCreation.Ephemeral5mInputTokens
+		cacheCreation1h = usage.CacheCreation.Ephemeral1hInputTokens
+	} else {
+		cacheCreation5m = claudeInfo.Usage.ClaudeCacheCreation5mTokens
+		cacheCreation1h = claudeInfo.Usage.ClaudeCacheCreation1hTokens
+	}
+	cacheCreation5m, cacheCreation1h = sharedclaude.NormalizeCacheCreationSplit(
+		usage.CacheCreationInputTokens,
+		cacheCreation5m,
+		cacheCreation1h,
+	)
+	if usage.CacheCreation == nil && (cacheCreation5m > 0 || cacheCreation1h > 0) {
+		usage.CacheCreation = &dto.ClaudeCacheCreationUsage{}
+	}
+	if usage.CacheCreation != nil {
+		usage.CacheCreation.Ephemeral5mInputTokens = cacheCreation5m
+		usage.CacheCreation.Ephemeral1hInputTokens = cacheCreation1h
+	}
+	return usage
+}
+
+func claudeBillingUsageFromSemanticUsage(usage *dto.Usage) *dto.BillingUsage {
+	if usage == nil {
+		return nil
+	}
+	cacheCreation5m, cacheCreation1h := sharedclaude.NormalizeCacheCreationSplit(
+		usage.PromptTokensDetails.CachedCreationTokens,
+		usage.ClaudeCacheCreation5mTokens,
+		usage.ClaudeCacheCreation1hTokens,
+	)
+	claudeUsage := &dto.ClaudeUsage{
+		InputTokens:              usage.PromptTokens,
+		CacheCreationInputTokens: usage.PromptTokensDetails.CachedCreationTokens,
+		CacheReadInputTokens:     usage.PromptTokensDetails.CachedTokens,
+		OutputTokens:             usage.CompletionTokens,
+	}
+	if cacheCreation5m > 0 || cacheCreation1h > 0 {
+		claudeUsage.CacheCreation = &dto.ClaudeCacheCreationUsage{
+			Ephemeral5mInputTokens: cacheCreation5m,
+			Ephemeral1hInputTokens: cacheCreation1h,
+		}
+	}
+	return dto.NewClaudeMessagesBillingUsage(claudeUsage)
+}
+
+func PatchClaudeMessageDeltaUsageData(data string, usage *dto.ClaudeUsage) string {
+	if data == "" || usage == nil {
+		return data
+	}
+
+	data = setMessageDeltaUsageInt(data, "usage.input_tokens", usage.InputTokens)
+	data = setMessageDeltaUsageInt(data, "usage.cache_read_input_tokens", usage.CacheReadInputTokens)
+	data = setMessageDeltaUsageInt(data, "usage.cache_creation_input_tokens", usage.CacheCreationInputTokens)
+
+	if usage.CacheCreation != nil {
+		data = setMessageDeltaUsageInt(data, "usage.cache_creation.ephemeral_5m_input_tokens", usage.CacheCreation.Ephemeral5mInputTokens)
+		data = setMessageDeltaUsageInt(data, "usage.cache_creation.ephemeral_1h_input_tokens", usage.CacheCreation.Ephemeral1hInputTokens)
+	}
+
+	return data
+}
+
+func setMessageDeltaUsageInt(data string, path string, localValue int) string {
+	if localValue <= 0 {
+		return data
+	}
+
+	upstreamValue := gjson.Get(data, path)
+	if upstreamValue.Exists() && upstreamValue.Int() > 0 {
+		return data
+	}
+
+	patchedData, err := sjson.Set(data, path, localValue)
+	if err != nil {
+		return data
+	}
+	return patchedData
+}
+
+func FormatClaudeResponseInfo(claudeResponse *dto.ClaudeResponse, oaiResponse *dto.ChatCompletionsStreamResponse, claudeInfo *ClaudeResponseInfo) bool {
+	if claudeInfo == nil {
+		return false
+	}
+	if claudeInfo.Usage == nil {
+		claudeInfo.Usage = &dto.Usage{}
+	}
+	if claudeResponse.Type == "message_start" {
+		if claudeResponse.Message != nil {
+			claudeInfo.ResponseId = claudeResponse.Message.Id
+			claudeInfo.Model = claudeResponse.Message.Model
+		}
+
+		if claudeResponse.Message != nil && claudeResponse.Message.Usage != nil {
+			claudeInfo.Usage.PromptTokens = claudeResponse.Message.Usage.InputTokens
+			claudeInfo.Usage.UsageSemantic = "anthropic"
+			claudeInfo.Usage.PromptTokensDetails.CachedTokens = claudeResponse.Message.Usage.CacheReadInputTokens
+			claudeInfo.Usage.PromptTokensDetails.CachedCreationTokens = claudeResponse.Message.Usage.CacheCreationInputTokens
+			claudeInfo.Usage.ClaudeCacheCreation5mTokens = claudeResponse.Message.Usage.GetCacheCreation5mTokens()
+			claudeInfo.Usage.ClaudeCacheCreation1hTokens = claudeResponse.Message.Usage.GetCacheCreation1hTokens()
+			claudeInfo.Usage.CompletionTokens = claudeResponse.Message.Usage.OutputTokens
+			claudeInfo.Usage.BillingUsage = claudeBillingUsageFromSemanticUsage(claudeInfo.Usage)
+		}
+	} else if claudeResponse.Type == "content_block_delta" {
+		if claudeResponse.Delta != nil {
+			if claudeResponse.Delta.Text != nil {
+				claudeInfo.ResponseText.WriteString(*claudeResponse.Delta.Text)
+			}
+			if claudeResponse.Delta.Thinking != nil {
+				claudeInfo.ResponseText.WriteString(*claudeResponse.Delta.Thinking)
+			}
+		}
+	} else if claudeResponse.Type == "message_delta" {
+		if claudeResponse.Usage != nil {
+			claudeInfo.Usage.UsageSemantic = "anthropic"
+			if claudeResponse.Usage.InputTokens > 0 {
+				claudeInfo.Usage.PromptTokens = claudeResponse.Usage.InputTokens
+			}
+			if claudeResponse.Usage.CacheReadInputTokens > 0 {
+				claudeInfo.Usage.PromptTokensDetails.CachedTokens = claudeResponse.Usage.CacheReadInputTokens
+			}
+			if claudeResponse.Usage.CacheCreationInputTokens > 0 {
+				claudeInfo.Usage.PromptTokensDetails.CachedCreationTokens = claudeResponse.Usage.CacheCreationInputTokens
+			}
+			if cacheCreation5m := claudeResponse.Usage.GetCacheCreation5mTokens(); cacheCreation5m > 0 {
+				claudeInfo.Usage.ClaudeCacheCreation5mTokens = cacheCreation5m
+			}
+			if cacheCreation1h := claudeResponse.Usage.GetCacheCreation1hTokens(); cacheCreation1h > 0 {
+				claudeInfo.Usage.ClaudeCacheCreation1hTokens = cacheCreation1h
+			}
+			if claudeResponse.Usage.OutputTokens > 0 {
+				claudeInfo.Usage.CompletionTokens = claudeResponse.Usage.OutputTokens
+			}
+			claudeInfo.Usage.TotalTokens = claudeInfo.Usage.PromptTokens + claudeInfo.Usage.CompletionTokens
+			claudeInfo.Usage.BillingUsage = claudeBillingUsageFromSemanticUsage(claudeInfo.Usage)
+		}
+
+		claudeInfo.Done = true
+	} else if claudeResponse.Type == "content_block_start" {
+	} else {
+		return false
+	}
+	if oaiResponse != nil {
+		oaiResponse.Id = claudeInfo.ResponseId
+		oaiResponse.Created = claudeInfo.Created
+		oaiResponse.Model = claudeInfo.Model
+	}
+	return true
+}

relaykit/relayconvert/internal/gemini_chat/to_oai_chat_req.go

@@ -0,0 +1,174 @@
+package geminichat
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/internal/jsonutil"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+func GeminiGenerateContentRequestToOpenAIChat(geminiRequest *dto.GeminiChatRequest, info convmeta.Meta) (*dto.GeneralOpenAIRequest, error) {
+	modelName := ""
+	isStream := false
+	if info != nil {
+		isStream = info.GetIsStream()
+	}
+	modelName = convmeta.UpstreamModelName(info)
+	openaiRequest := &dto.GeneralOpenAIRequest{
+		Model:  modelName,
+		Stream: kitutil.GetPointer(isStream),
+	}
+
+	var messages []dto.Message
+	for _, content := range geminiRequest.Contents {
+		message := dto.Message{
+			Role: convertGeminiRoleToOpenAI(content.Role),
+		}
+
+		var mediaContents []dto.MediaContent
+		var toolCalls []dto.ToolCallRequest
+		for _, part := range content.Parts {
+			if part.Text != "" {
+				mediaContent := dto.MediaContent{
+					Type: "text",
+					Text: part.Text,
+				}
+				mediaContents = append(mediaContents, mediaContent)
+			} else if part.InlineData != nil {
+				mediaContent := dto.MediaContent{
+					Type: "image_url",
+					ImageUrl: &dto.MessageImageUrl{
+						Url:      fmt.Sprintf("data:%s;base64,%s", part.InlineData.MimeType, part.InlineData.Data),
+						Detail:   "auto",
+						MimeType: part.InlineData.MimeType,
+					},
+				}
+				mediaContents = append(mediaContents, mediaContent)
+			} else if part.FileData != nil {
+				mediaContent := dto.MediaContent{
+					Type: "image_url",
+					ImageUrl: &dto.MessageImageUrl{
+						Url:      part.FileData.FileUri,
+						Detail:   "auto",
+						MimeType: part.FileData.MimeType,
+					},
+				}
+				mediaContents = append(mediaContents, mediaContent)
+			} else if part.FunctionCall != nil {
+				toolCall := dto.ToolCallRequest{
+					ID:   fmt.Sprintf("call_%d", len(toolCalls)+1),
+					Type: "function",
+					Function: dto.FunctionRequest{
+						Name:      part.FunctionCall.FunctionName,
+						Arguments: jsonutil.ToJSONString(part.FunctionCall.Arguments),
+					},
+				}
+				toolCalls = append(toolCalls, toolCall)
+			} else if part.FunctionResponse != nil {
+				toolMessage := dto.Message{
+					Role:       "tool",
+					ToolCallId: fmt.Sprintf("call_%d", len(toolCalls)),
+				}
+				toolMessage.SetStringContent(jsonutil.ToJSONString(part.FunctionResponse.Response))
+				messages = append(messages, toolMessage)
+			}
+		}
+
+		if len(toolCalls) > 0 {
+			message.SetToolCalls(toolCalls)
+		} else if len(mediaContents) == 1 && mediaContents[0].Type == "text" {
+			message.Content = mediaContents[0].Text
+		} else if len(mediaContents) > 0 {
+			message.SetMediaContent(mediaContents)
+		}
+
+		if len(message.ParseContent()) > 0 || len(message.ToolCalls) > 0 {
+			messages = append(messages, message)
+		}
+	}
+
+	openaiRequest.Messages = messages
+
+	if geminiRequest.GenerationConfig.Temperature != nil {
+		openaiRequest.Temperature = geminiRequest.GenerationConfig.Temperature
+	}
+	if geminiRequest.GenerationConfig.TopP != nil && *geminiRequest.GenerationConfig.TopP > 0 {
+		openaiRequest.TopP = kitutil.GetPointer(*geminiRequest.GenerationConfig.TopP)
+	}
+	if geminiRequest.GenerationConfig.TopK != nil && *geminiRequest.GenerationConfig.TopK > 0 {
+		openaiRequest.TopK = kitutil.GetPointer(int(*geminiRequest.GenerationConfig.TopK))
+	}
+	if geminiRequest.GenerationConfig.MaxOutputTokens != nil && *geminiRequest.GenerationConfig.MaxOutputTokens > 0 {
+		openaiRequest.MaxTokens = kitutil.GetPointer(*geminiRequest.GenerationConfig.MaxOutputTokens)
+	}
+	if len(geminiRequest.GenerationConfig.StopSequences) > 0 {
+		openaiRequest.Stop = geminiRequest.GenerationConfig.StopSequences[:min(len(geminiRequest.GenerationConfig.StopSequences), 4)]
+	}
+	if geminiRequest.GenerationConfig.CandidateCount != nil && *geminiRequest.GenerationConfig.CandidateCount > 0 {
+		openaiRequest.N = kitutil.GetPointer(*geminiRequest.GenerationConfig.CandidateCount)
+	}
+
+	if len(geminiRequest.GetTools()) > 0 {
+		var tools []dto.ToolCallRequest
+		for _, tool := range geminiRequest.GetTools() {
+			if tool.FunctionDeclarations == nil {
+				continue
+			}
+			functionDeclarations, err := kitutil.Any2Type[[]dto.FunctionRequest](tool.FunctionDeclarations)
+			if err != nil {
+				kitutil.LogSystemError(fmt.Sprintf("failed to parse gemini function declarations: %v (type=%T)", err, tool.FunctionDeclarations))
+				continue
+			}
+			for _, function := range functionDeclarations {
+				openAITool := dto.ToolCallRequest{
+					Type: "function",
+					Function: dto.FunctionRequest{
+						Name:        function.Name,
+						Description: function.Description,
+						Parameters:  function.Parameters,
+					},
+				}
+				tools = append(tools, openAITool)
+			}
+		}
+		if len(tools) > 0 {
+			openaiRequest.Tools = tools
+		}
+	}
+
+	if geminiRequest.SystemInstructions != nil {
+		systemMessage := dto.Message{
+			Role:    "system",
+			Content: extractTextFromGeminiParts(geminiRequest.SystemInstructions.Parts),
+		}
+		openaiRequest.Messages = append([]dto.Message{systemMessage}, openaiRequest.Messages...)
+	}
+
+	return openaiRequest, nil
+}
+
+func convertGeminiRoleToOpenAI(geminiRole string) string {
+	switch geminiRole {
+	case "user":
+		return "user"
+	case "model":
+		return "assistant"
+	case "function":
+		return "function"
+	default:
+		return "user"
+	}
+}
+
+func extractTextFromGeminiParts(parts []dto.GeminiPart) string {
+	texts := make([]string, 0)
+	for _, part := range parts {
+		if part.Text != "" {
+			texts = append(texts, part.Text)
+		}
+	}
+	return strings.Join(texts, "\n")
+}

relaykit/relayconvert/internal/gemini_chat/to_oai_chat_resp.go

@@ -0,0 +1,298 @@
+package geminichat
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/QuantumNous/new-api/relaykit/types"
+)
+
+func UsageFromGeminiMetadata(metadata *dto.GeminiUsageMetadata, fallbackPromptTokens int) *dto.Usage {
+	if metadata == nil {
+		if fallbackPromptTokens <= 0 {
+			return nil
+		}
+		usage := &dto.Usage{PromptTokens: fallbackPromptTokens}
+		usage.PromptTokensDetails.TextTokens = fallbackPromptTokens
+		return usage
+	}
+
+	promptTokens := metadata.PromptTokenCount + metadata.ToolUsePromptTokenCount
+	if promptTokens <= 0 && fallbackPromptTokens > 0 {
+		promptTokens = fallbackPromptTokens
+	}
+
+	usage := &dto.Usage{
+		PromptTokens:     promptTokens,
+		CompletionTokens: metadata.CandidatesTokenCount + metadata.ThoughtsTokenCount,
+		TotalTokens:      metadata.TotalTokenCount,
+		BillingUsage:     dto.CloneBillingUsage(metadata.BillingUsage),
+	}
+	if usage.BillingUsage == nil {
+		usage.BillingUsage = dto.NewGeminiChatBillingUsage(metadata)
+	}
+	usage.CompletionTokenDetails.ReasoningTokens = metadata.ThoughtsTokenCount
+	usage.PromptTokensDetails.CachedTokens = metadata.CachedContentTokenCount
+
+	for _, detail := range metadata.PromptTokensDetails {
+		if detail.Modality == "AUDIO" {
+			usage.PromptTokensDetails.AudioTokens += detail.TokenCount
+		} else if detail.Modality == "IMAGE" {
+			usage.PromptTokensDetails.ImageTokens += detail.TokenCount
+		} else if detail.Modality == "TEXT" {
+			usage.PromptTokensDetails.TextTokens += detail.TokenCount
+		}
+	}
+	for _, detail := range metadata.ToolUsePromptTokensDetails {
+		if detail.Modality == "AUDIO" {
+			usage.PromptTokensDetails.AudioTokens += detail.TokenCount
+		} else if detail.Modality == "IMAGE" {
+			usage.PromptTokensDetails.ImageTokens += detail.TokenCount
+		} else if detail.Modality == "TEXT" {
+			usage.PromptTokensDetails.TextTokens += detail.TokenCount
+		}
+	}
+	for _, detail := range metadata.CandidatesTokensDetails {
+		switch detail.Modality {
+		case "IMAGE":
+			usage.CompletionTokenDetails.ImageTokens += detail.TokenCount
+		case "AUDIO":
+			usage.CompletionTokenDetails.AudioTokens += detail.TokenCount
+		case "TEXT":
+			usage.CompletionTokenDetails.TextTokens += detail.TokenCount
+		}
+	}
+
+	if usage.TotalTokens > 0 && usage.CompletionTokens <= 0 {
+		usage.CompletionTokens = usage.TotalTokens - usage.PromptTokens
+	}
+
+	if usage.PromptTokens > 0 && usage.PromptTokensDetails.TextTokens == 0 && usage.PromptTokensDetails.AudioTokens == 0 {
+		usage.PromptTokensDetails.TextTokens = usage.PromptTokens
+	}
+
+	return usage
+}
+
+func ResponseGeminiChat2OpenAI(id string, created int64, response *dto.GeminiChatResponse) *dto.OpenAITextResponse {
+	fullTextResponse := dto.OpenAITextResponse{
+		Id:      id,
+		Object:  "chat.completion",
+		Created: created,
+		Choices: make([]dto.OpenAITextResponseChoice, 0, len(response.Candidates)),
+	}
+	isToolCall := false
+	for _, candidate := range response.Candidates {
+		choice := dto.OpenAITextResponseChoice{
+			Index: int(candidate.Index),
+			Message: dto.Message{
+				Role:    "assistant",
+				Content: "",
+			},
+			FinishReason: types.FinishReasonStop,
+		}
+		if len(candidate.Content.Parts) > 0 {
+			var content strings.Builder
+			var inlineGrow int
+			for _, part := range candidate.Content.Parts {
+				if part.InlineData != nil {
+					inlineGrow += len(part.InlineData.MimeType) + len(part.InlineData.Data) + 32
+				}
+			}
+			if inlineGrow > 0 {
+				content.Grow(inlineGrow)
+			}
+			appended := 0
+			writeSep := func() {
+				if appended > 0 {
+					content.WriteByte('\n')
+				}
+				appended++
+			}
+			var toolCalls []dto.ToolCallResponse
+			for _, part := range candidate.Content.Parts {
+				if part.InlineData != nil {
+					if strings.HasPrefix(part.InlineData.MimeType, "image") {
+						writeSep()
+						content.WriteString("![image](data:")
+						content.WriteString(part.InlineData.MimeType)
+						content.WriteString(";base64,")
+						content.WriteString(part.InlineData.Data)
+						content.WriteByte(')')
+					} else {
+						writeSep()
+						content.WriteString("[media](data:")
+						content.WriteString(part.InlineData.MimeType)
+						content.WriteString(";base64,")
+						content.WriteString(part.InlineData.Data)
+						content.WriteByte(')')
+					}
+				} else if part.FunctionCall != nil {
+					choice.FinishReason = types.FinishReasonToolCalls
+					if call := geminiResponseToolCall(&part); call != nil {
+						toolCalls = append(toolCalls, *call)
+					}
+				} else if part.Thought {
+					choice.Message.ReasoningContent = &part.Text
+				} else {
+					if part.ExecutableCode != nil {
+						writeSep()
+						content.WriteString("```")
+						content.WriteString(part.ExecutableCode.Language)
+						content.WriteByte('\n')
+						content.WriteString(part.ExecutableCode.Code)
+						content.WriteString("\n```")
+					} else if part.CodeExecutionResult != nil {
+						writeSep()
+						content.WriteString("```output\n")
+						content.WriteString(part.CodeExecutionResult.Output)
+						content.WriteString("\n```")
+					} else if part.Text != "\n" {
+						writeSep()
+						content.WriteString(part.Text)
+					}
+				}
+			}
+			if len(toolCalls) > 0 {
+				choice.Message.SetToolCalls(toolCalls)
+				isToolCall = true
+			}
+			choice.Message.SetStringContent(content.String())
+		}
+		if candidate.FinishReason != nil {
+			switch *candidate.FinishReason {
+			case "STOP":
+				choice.FinishReason = types.FinishReasonStop
+			case "MAX_TOKENS":
+				choice.FinishReason = types.FinishReasonLength
+			case "SAFETY", "RECITATION", "BLOCKLIST", "PROHIBITED_CONTENT", "SPII", "OTHER":
+				choice.FinishReason = types.FinishReasonContentFilter
+			default:
+				choice.FinishReason = types.FinishReasonContentFilter
+			}
+		}
+		if isToolCall {
+			choice.FinishReason = types.FinishReasonToolCalls
+		}
+
+		fullTextResponse.Choices = append(fullTextResponse.Choices, choice)
+	}
+	return &fullTextResponse
+}
+
+func StreamResponseGeminiChat2OpenAI(geminiResponse *dto.GeminiChatResponse) (*dto.ChatCompletionsStreamResponse, bool) {
+	choices := make([]dto.ChatCompletionsStreamResponseChoice, 0, len(geminiResponse.Candidates))
+	isStop := false
+	for _, candidate := range geminiResponse.Candidates {
+		if candidate.FinishReason != nil && *candidate.FinishReason == "STOP" {
+			isStop = true
+			candidate.FinishReason = nil
+		}
+		choice := dto.ChatCompletionsStreamResponseChoice{
+			Index: int(candidate.Index),
+			Delta: dto.ChatCompletionsStreamResponseChoiceDelta{},
+		}
+		var content strings.Builder
+		var inlineGrow int
+		for _, part := range candidate.Content.Parts {
+			if part.InlineData != nil {
+				inlineGrow += len(part.InlineData.MimeType) + len(part.InlineData.Data) + 32
+			}
+		}
+		if inlineGrow > 0 {
+			content.Grow(inlineGrow)
+		}
+		appended := 0
+		writeSep := func() {
+			if appended > 0 {
+				content.WriteByte('\n')
+			}
+			appended++
+		}
+		isTools := false
+		isThought := false
+		if candidate.FinishReason != nil {
+			switch *candidate.FinishReason {
+			case "STOP":
+				choice.FinishReason = &types.FinishReasonStop
+			case "MAX_TOKENS":
+				choice.FinishReason = &types.FinishReasonLength
+			case "SAFETY", "RECITATION", "BLOCKLIST", "PROHIBITED_CONTENT", "SPII", "OTHER":
+				choice.FinishReason = &types.FinishReasonContentFilter
+			default:
+				choice.FinishReason = &types.FinishReasonContentFilter
+			}
+		}
+		for _, part := range candidate.Content.Parts {
+			if part.InlineData != nil {
+				if strings.HasPrefix(part.InlineData.MimeType, "image") {
+					writeSep()
+					content.WriteString("![image](data:")
+					content.WriteString(part.InlineData.MimeType)
+					content.WriteString(";base64,")
+					content.WriteString(part.InlineData.Data)
+					content.WriteByte(')')
+				}
+			} else if part.FunctionCall != nil {
+				isTools = true
+				if call := geminiResponseToolCall(&part); call != nil {
+					call.SetIndex(len(choice.Delta.ToolCalls))
+					choice.Delta.ToolCalls = append(choice.Delta.ToolCalls, *call)
+				}
+			} else if part.Thought {
+				isThought = true
+				writeSep()
+				content.WriteString(part.Text)
+			} else {
+				if part.ExecutableCode != nil {
+					writeSep()
+					content.WriteString("```")
+					content.WriteString(part.ExecutableCode.Language)
+					content.WriteByte('\n')
+					content.WriteString(part.ExecutableCode.Code)
+					content.WriteString("\n```\n")
+				} else if part.CodeExecutionResult != nil {
+					writeSep()
+					content.WriteString("```output\n")
+					content.WriteString(part.CodeExecutionResult.Output)
+					content.WriteString("\n```\n")
+				} else if part.Text != "\n" {
+					writeSep()
+					content.WriteString(part.Text)
+				}
+			}
+		}
+		if isThought {
+			choice.Delta.SetReasoningContent(content.String())
+		} else {
+			choice.Delta.SetContentString(content.String())
+		}
+		if isTools {
+			choice.FinishReason = &types.FinishReasonToolCalls
+		}
+		choices = append(choices, choice)
+	}
+
+	response := dto.ChatCompletionsStreamResponse{
+		Object:  "chat.completion.chunk",
+		Choices: choices,
+	}
+	return &response, isStop
+}
+
+func geminiResponseToolCall(item *dto.GeminiPart) *dto.ToolCallResponse {
+	argsBytes, err := kitutil.Marshal(item.FunctionCall.Arguments)
+	if err != nil {
+		return nil
+	}
+	return &dto.ToolCallResponse{
+		ID:   fmt.Sprintf("call_%s", kitutil.GetUUID()),
+		Type: "function",
+		Function: dto.FunctionResponse{
+			Arguments: string(argsBytes),
+			Name:      item.FunctionCall.FunctionName,
+		},
+	}
+}

relaykit/relayconvert/internal/jsonutil/stringify.go

@@ -0,0 +1,15 @@
+package jsonutil
+
+import (
+	"fmt"
+
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+func ToJSONString(v interface{}) string {
+	bytes, err := kitutil.Marshal(v)
+	if err != nil {
+		return fmt.Sprintf("%v", v)
+	}
+	return string(bytes)
+}

relaykit/relayconvert/internal/media/media.go

@@ -0,0 +1,46 @@
+package media
+
+import (
+	"errors"
+	"sync"
+
+	"context"
+	"github.com/QuantumNous/new-api/relaykit/types"
+)
+
+type MediaResolver struct {
+	GetBase64Data        func(c context.Context, source types.FileSource, reason ...string) (string, string, error)
+	DecodeBase64FileData func(base64String string) (string, string, error)
+}
+
+var (
+	mediaResolverMu sync.RWMutex
+	mediaResolver   MediaResolver
+)
+
+func SetMediaResolver(resolver MediaResolver) {
+	mediaResolverMu.Lock()
+	defer mediaResolverMu.Unlock()
+
+	mediaResolver = resolver
+}
+
+func ResolveBase64Data(c context.Context, source types.FileSource, reason ...string) (string, string, error) {
+	mediaResolverMu.RLock()
+	resolver := mediaResolver.GetBase64Data
+	mediaResolverMu.RUnlock()
+	if resolver == nil {
+		return "", "", errors.New("relayconvert media resolver is not configured")
+	}
+	return resolver(c, source, reason...)
+}
+
+func DecodeBase64FileData(base64String string) (string, string, error) {
+	mediaResolverMu.RLock()
+	resolver := mediaResolver.DecodeBase64FileData
+	mediaResolverMu.RUnlock()
+	if resolver == nil {
+		return "", "", errors.New("relayconvert media resolver is not configured")
+	}
+	return resolver(base64String)
+}

relaykit/relayconvert/internal/oai_chat/to_claude_messages_req.go

@@ -0,0 +1,409 @@
+package oaichat
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"context"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	relaymedia "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/media"
+	sharedclaude "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/shared/claude"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/reasoning"
+)
+
+const (
+	webSearchMaxUsesLow    = 1
+	webSearchMaxUsesMedium = 5
+	webSearchMaxUsesHigh   = 10
+)
+
+type openRouterRequestReasoning struct {
+	Enabled   bool   `json:"enabled"`
+	Effort    string `json:"effort,omitempty"`
+	MaxTokens int    `json:"max_tokens,omitempty"`
+	Exclude   bool   `json:"exclude,omitempty"`
+}
+
+func OpenAIChatRequestToClaudeMessages(c context.Context, info convmeta.Meta, textRequest dto.GeneralOpenAIRequest) (*dto.ClaudeRequest, error) {
+	opts := convmeta.OptionsOf(info)
+	claudeTools := make([]any, 0, len(textRequest.Tools))
+
+	for _, tool := range textRequest.Tools {
+		if params, ok := tool.Function.Parameters.(map[string]any); ok {
+			claudeTool := dto.Tool{
+				Name:        tool.Function.Name,
+				Description: tool.Function.Description,
+			}
+			claudeTool.InputSchema = make(map[string]interface{})
+			if params["type"] != nil {
+				claudeTool.InputSchema["type"] = params["type"].(string)
+			}
+			claudeTool.InputSchema["properties"] = params["properties"]
+			claudeTool.InputSchema["required"] = params["required"]
+			for key, value := range params {
+				if key == "type" || key == "properties" || key == "required" {
+					continue
+				}
+				claudeTool.InputSchema[key] = value
+			}
+			claudeTools = append(claudeTools, &claudeTool)
+		}
+	}
+
+	if textRequest.WebSearchOptions != nil {
+		webSearchTool := dto.ClaudeWebSearchTool{
+			Type: "web_search_20250305",
+			Name: "web_search",
+		}
+
+		if textRequest.WebSearchOptions.UserLocation != nil {
+			anthropicUserLocation := &dto.ClaudeWebSearchUserLocation{
+				Type: "approximate",
+			}
+
+			var userLocationMap map[string]interface{}
+			if err := kitutil.Unmarshal(textRequest.WebSearchOptions.UserLocation, &userLocationMap); err == nil {
+				if approximateData, ok := userLocationMap["approximate"].(map[string]interface{}); ok {
+					if timezone, ok := approximateData["timezone"].(string); ok && timezone != "" {
+						anthropicUserLocation.Timezone = timezone
+					}
+					if country, ok := approximateData["country"].(string); ok && country != "" {
+						anthropicUserLocation.Country = country
+					}
+					if region, ok := approximateData["region"].(string); ok && region != "" {
+						anthropicUserLocation.Region = region
+					}
+					if city, ok := approximateData["city"].(string); ok && city != "" {
+						anthropicUserLocation.City = city
+					}
+				}
+			}
+
+			webSearchTool.UserLocation = anthropicUserLocation
+		}
+
+		switch textRequest.WebSearchOptions.SearchContextSize {
+		case "low":
+			webSearchTool.MaxUses = webSearchMaxUsesLow
+		case "medium":
+			webSearchTool.MaxUses = webSearchMaxUsesMedium
+		case "high":
+			webSearchTool.MaxUses = webSearchMaxUsesHigh
+		}
+
+		claudeTools = append(claudeTools, &webSearchTool)
+	}
+
+	claudeRequest := dto.ClaudeRequest{
+		Model:         textRequest.Model,
+		StopSequences: nil,
+		Temperature:   textRequest.Temperature,
+		Tools:         claudeTools,
+	}
+	if maxTokens := textRequest.GetMaxTokens(); maxTokens > 0 {
+		claudeRequest.MaxTokens = kitutil.GetPointer(maxTokens)
+	}
+	if textRequest.TopP != nil {
+		claudeRequest.TopP = kitutil.GetPointer(*textRequest.TopP)
+	}
+	if textRequest.TopK != nil {
+		claudeRequest.TopK = kitutil.GetPointer(*textRequest.TopK)
+	}
+	if textRequest.IsStream(nil) {
+		claudeRequest.Stream = kitutil.GetPointer(true)
+	}
+
+	if textRequest.ToolChoice != nil || textRequest.ParallelTooCalls != nil {
+		claudeToolChoice := sharedclaude.MapOpenAIToolChoice(textRequest.ToolChoice, textRequest.ParallelTooCalls)
+		if claudeToolChoice != nil {
+			claudeRequest.ToolChoice = claudeToolChoice
+		}
+	}
+
+	if claudeRequest.MaxTokens == nil || *claudeRequest.MaxTokens == 0 {
+		if defaultMaxTokens, configured := opts.Claude.DefaultMaxTokensFor(textRequest.Model); configured {
+			value := uint(defaultMaxTokens)
+			claudeRequest.MaxTokens = &value
+		}
+	}
+
+	if baseModel, effortLevel, ok := reasoning.TrimEffortSuffix(textRequest.Model); ok && effortLevel != "" &&
+		(strings.HasPrefix(textRequest.Model, "claude-opus-4-6") ||
+			strings.HasPrefix(textRequest.Model, "claude-opus-4-7") ||
+			strings.HasPrefix(textRequest.Model, "claude-opus-4-8")) {
+		claudeRequest.Model = baseModel
+		claudeRequest.Thinking = &dto.Thinking{
+			Type: "adaptive",
+		}
+		claudeRequest.OutputConfig = json.RawMessage(fmt.Sprintf(`{"effort":"%s"}`, effortLevel))
+		if strings.HasPrefix(baseModel, "claude-opus-4-7") ||
+			strings.HasPrefix(baseModel, "claude-opus-4-8") {
+			claudeRequest.Thinking.Display = "summarized"
+			claudeRequest.Temperature = nil
+			claudeRequest.TopP = nil
+			claudeRequest.TopK = nil
+		} else {
+			claudeRequest.TopP = nil
+			claudeRequest.Temperature = kitutil.GetPointer[float64](1.0)
+		}
+	} else if opts.Claude.ThinkingAdapterEnabled &&
+		strings.HasSuffix(textRequest.Model, "-thinking") {
+
+		trimmedModel := strings.TrimSuffix(textRequest.Model, "-thinking")
+		if strings.HasPrefix(trimmedModel, "claude-opus-4-7") ||
+			strings.HasPrefix(trimmedModel, "claude-opus-4-8") {
+			claudeRequest.Thinking = &dto.Thinking{Type: "adaptive", Display: "summarized"}
+			claudeRequest.OutputConfig = json.RawMessage(`{"effort":"high"}`)
+			claudeRequest.Temperature = nil
+			claudeRequest.TopP = nil
+			claudeRequest.TopK = nil
+		} else {
+			if claudeRequest.MaxTokens == nil || *claudeRequest.MaxTokens < 1280 {
+				claudeRequest.MaxTokens = kitutil.GetPointer[uint](1280)
+			}
+
+			claudeRequest.Thinking = &dto.Thinking{
+				Type:         "enabled",
+				BudgetTokens: kitutil.GetPointer[int](int(float64(*claudeRequest.MaxTokens) * opts.Claude.ThinkingAdapterBudgetTokensPercentage)),
+			}
+			claudeRequest.TopP = nil
+			claudeRequest.Temperature = kitutil.GetPointer[float64](1.0)
+		}
+		if !opts.ShouldPreserveThinkingSuffix(textRequest.Model) {
+			claudeRequest.Model = trimmedModel
+		}
+	}
+
+	if textRequest.ReasoningEffort != "" {
+		switch textRequest.ReasoningEffort {
+		case "low":
+			claudeRequest.Thinking = &dto.Thinking{
+				Type:         "enabled",
+				BudgetTokens: kitutil.GetPointer[int](1280),
+			}
+		case "medium":
+			claudeRequest.Thinking = &dto.Thinking{
+				Type:         "enabled",
+				BudgetTokens: kitutil.GetPointer[int](2048),
+			}
+		case "high":
+			claudeRequest.Thinking = &dto.Thinking{
+				Type:         "enabled",
+				BudgetTokens: kitutil.GetPointer[int](4096),
+			}
+		}
+	}
+
+	if textRequest.Reasoning != nil {
+		var reasoningConfig openRouterRequestReasoning
+		if err := kitutil.Unmarshal(textRequest.Reasoning, &reasoningConfig); err != nil {
+			return nil, err
+		}
+
+		budgetTokens := reasoningConfig.MaxTokens
+		if budgetTokens > 0 {
+			claudeRequest.Thinking = &dto.Thinking{
+				Type:         "enabled",
+				BudgetTokens: &budgetTokens,
+			}
+		}
+	}
+
+	if textRequest.Stop != nil {
+		switch stop := textRequest.Stop.(type) {
+		case string:
+			claudeRequest.StopSequences = []string{stop}
+		case []interface{}:
+			stopSequences := make([]string, 0)
+			for _, item := range stop {
+				stopSequences = append(stopSequences, item.(string))
+			}
+			claudeRequest.StopSequences = stopSequences
+		}
+	}
+
+	formatMessages := make([]dto.Message, 0)
+	lastMessage := dto.Message{
+		Role: "tool",
+	}
+	for i, message := range textRequest.Messages {
+		if message.Role == "" {
+			textRequest.Messages[i].Role = "user"
+		}
+		fmtMessage := dto.Message{
+			Role:    message.Role,
+			Content: message.Content,
+		}
+		if message.Role == "tool" {
+			fmtMessage.ToolCallId = message.ToolCallId
+		}
+		if message.Role == "assistant" && message.ToolCalls != nil {
+			fmtMessage.ToolCalls = message.ToolCalls
+		}
+		if lastMessage.Role == message.Role && lastMessage.Role != "tool" {
+			if lastMessage.IsStringContent() && message.IsStringContent() {
+				fmtMessage.SetStringContent(strings.Trim(fmt.Sprintf("%s %s", lastMessage.StringContent(), message.StringContent()), "\""))
+				formatMessages = formatMessages[:len(formatMessages)-1]
+			}
+		}
+		if fmtMessage.Content == nil || (fmtMessage.IsStringContent() && fmtMessage.StringContent() == "") {
+			fmtMessage.SetStringContent("...")
+		}
+		formatMessages = append(formatMessages, fmtMessage)
+		lastMessage = fmtMessage
+	}
+
+	claudeMessages := make([]dto.ClaudeMessage, 0)
+	isFirstMessage := true
+	var systemMessages []dto.ClaudeMediaMessage
+
+	for _, message := range formatMessages {
+		if message.Role == "system" {
+			if message.IsStringContent() {
+				if text := message.StringContent(); text != "" {
+					systemMessages = append(systemMessages, dto.ClaudeMediaMessage{
+						Type: "text",
+						Text: kitutil.GetPointer[string](text),
+					})
+				}
+			} else {
+				for _, ctx := range message.ParseContent() {
+					if ctx.Type == "text" && ctx.Text != "" {
+						systemMessages = append(systemMessages, dto.ClaudeMediaMessage{
+							Type: "text",
+							Text: kitutil.GetPointer[string](ctx.Text),
+						})
+					}
+				}
+			}
+			continue
+		}
+
+		if isFirstMessage {
+			isFirstMessage = false
+			if message.Role != "user" {
+				claudeMessage := dto.ClaudeMessage{
+					Role: "user",
+					Content: []dto.ClaudeMediaMessage{
+						{
+							Type: "text",
+							Text: kitutil.GetPointer[string]("..."),
+						},
+					},
+				}
+				claudeMessages = append(claudeMessages, claudeMessage)
+			}
+		}
+
+		claudeMessage := dto.ClaudeMessage{
+			Role: message.Role,
+		}
+		if message.Role == "tool" {
+			if len(claudeMessages) > 0 && claudeMessages[len(claudeMessages)-1].Role == "user" {
+				lastClaudeMessage := claudeMessages[len(claudeMessages)-1]
+				if content, ok := lastClaudeMessage.Content.(string); ok {
+					lastClaudeMessage.Content = []dto.ClaudeMediaMessage{
+						{
+							Type: "text",
+							Text: kitutil.GetPointer[string](content),
+						},
+					}
+				}
+				lastClaudeMessage.Content = append(lastClaudeMessage.Content.([]dto.ClaudeMediaMessage), dto.ClaudeMediaMessage{
+					Type:      "tool_result",
+					ToolUseId: message.ToolCallId,
+					Content:   message.Content,
+				})
+				claudeMessages[len(claudeMessages)-1] = lastClaudeMessage
+				continue
+			}
+
+			claudeMessage.Role = "user"
+			claudeMessage.Content = []dto.ClaudeMediaMessage{
+				{
+					Type:      "tool_result",
+					ToolUseId: message.ToolCallId,
+					Content:   message.Content,
+				},
+			}
+		} else if message.IsStringContent() && message.ToolCalls == nil {
+			text := message.StringContent()
+			if text == "" {
+				text = "..."
+			}
+			claudeMessage.Content = text
+		} else {
+			claudeMediaMessages := make([]dto.ClaudeMediaMessage, 0)
+			for _, mediaMessage := range message.ParseContent() {
+				switch mediaMessage.Type {
+				case "text":
+					if mediaMessage.Text != "" {
+						claudeMediaMessages = append(claudeMediaMessages, dto.ClaudeMediaMessage{
+							Type: "text",
+							Text: kitutil.GetPointer[string](mediaMessage.Text),
+						})
+					}
+				default:
+					source := mediaMessage.ToFileSource()
+					if source == nil {
+						continue
+					}
+					base64Data, mimeType, err := relaymedia.ResolveBase64Data(c, source, "formatting image for Claude")
+					if err != nil {
+						return nil, fmt.Errorf("get file data failed: %s", err.Error())
+					}
+					claudeMediaMessage := dto.ClaudeMediaMessage{
+						Source: &dto.ClaudeMessageSource{
+							Type: "base64",
+						},
+					}
+					if strings.HasPrefix(mimeType, "application/pdf") {
+						claudeMediaMessage.Type = "document"
+					} else {
+						claudeMediaMessage.Type = "image"
+					}
+
+					claudeMediaMessage.Source.MediaType = mimeType
+					claudeMediaMessage.Source.Data = base64Data
+					claudeMediaMessages = append(claudeMediaMessages, claudeMediaMessage)
+					continue
+				}
+			}
+
+			if message.ToolCalls != nil {
+				for _, toolCall := range message.ParseToolCalls() {
+					inputObj := make(map[string]any)
+					if args := toolCall.Function.Arguments; args != "" {
+						if err := kitutil.Unmarshal([]byte(args), &inputObj); err != nil {
+							kitutil.LogInfo("tool call function arguments is not a map[string]any: " + fmt.Sprintf("%v", toolCall.Function.Arguments))
+						}
+					}
+					claudeMediaMessages = append(claudeMediaMessages, dto.ClaudeMediaMessage{
+						Type:  "tool_use",
+						Id:    toolCall.ID,
+						Name:  toolCall.Function.Name,
+						Input: inputObj,
+					})
+				}
+			}
+			claudeMessage.Content = claudeMediaMessages
+		}
+		claudeMessages = append(claudeMessages, claudeMessage)
+	}
+
+	if len(systemMessages) > 0 {
+		claudeRequest.System = systemMessages
+	}
+
+	claudeRequest.Prompt = ""
+	claudeRequest.Messages = claudeMessages
+	// Checked last so every injection path (default hook, thinking adapter
+	// floor) has had its chance to satisfy the required field.
+	if claudeRequest.MaxTokens == nil {
+		return nil, sharedclaude.ErrMissingMaxTokens
+	}
+	return &claudeRequest, nil
+}

relaykit/relayconvert/internal/oai_chat/to_claude_messages_resp.go

@@ -0,0 +1,475 @@
+package oaichat
+
+import (
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/reasonmap"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/samber/lo"
+)
+
+func generateStopBlock(index int) *dto.ClaudeResponse {
+	return &dto.ClaudeResponse{
+		Type:  "content_block_stop",
+		Index: kitutil.GetPointer[int](index),
+	}
+}
+
+func buildClaudeUsageFromOpenAIUsage(oaiUsage *dto.Usage) *dto.ClaudeUsage {
+	if oaiUsage == nil {
+		return nil
+	}
+	if billingUsage := dto.CloneBillingUsage(oaiUsage.BillingUsage); billingUsage != nil && billingUsage.ClaudeUsage != nil {
+		if billingUsage.Source == dto.BillingUsageSourceClaudeMessages || billingUsage.Semantic == dto.BillingUsageSemanticAnthropic {
+			return billingUsage.ClaudeUsage
+		}
+	}
+	billingUsage := dto.NewOpenAIChatBillingUsage(oaiUsage)
+	if existingBillingUsage := dto.CloneBillingUsage(oaiUsage.BillingUsage); existingBillingUsage != nil && existingBillingUsage.OpenAIUsage != nil {
+		if existingBillingUsage.Source == dto.BillingUsageSourceOAIChat ||
+			existingBillingUsage.Source == dto.BillingUsageSourceOAIResponses ||
+			existingBillingUsage.Semantic == dto.BillingUsageSemanticOpenAI {
+			billingUsage = existingBillingUsage
+		}
+	}
+	cacheCreation5m, cacheCreation1h := NormalizeCacheCreationSplit(
+		oaiUsage.PromptTokensDetails.CachedCreationTokens,
+		oaiUsage.ClaudeCacheCreation5mTokens,
+		oaiUsage.ClaudeCacheCreation1hTokens,
+	)
+	cacheCreationTokens := oaiUsage.PromptTokensDetails.CacheCreationTokensTotal()
+	inputTokens := oaiUsage.PromptTokens
+	if oaiUsage.PromptTokensDetails.CacheWriteTokens > 0 {
+		// OpenAI native cache-write usage counts cached and cache-write tokens
+		// inside prompt_tokens, while Claude semantics reports input_tokens
+		// excluding both. Both counts are unadjusted prefixes and may overlap,
+		// so clamp a negative remainder at zero.
+		inputTokens = oaiUsage.PromptTokens - oaiUsage.PromptTokensDetails.CachedTokens - cacheCreationTokens
+		if inputTokens < 0 {
+			inputTokens = 0
+		}
+	}
+	usage := &dto.ClaudeUsage{
+		InputTokens:              inputTokens,
+		OutputTokens:             oaiUsage.CompletionTokens,
+		CacheCreationInputTokens: cacheCreationTokens,
+		CacheReadInputTokens:     oaiUsage.PromptTokensDetails.CachedTokens,
+		BillingUsage:             billingUsage,
+	}
+	if cacheCreation5m > 0 || cacheCreation1h > 0 {
+		usage.CacheCreation = &dto.ClaudeCacheCreationUsage{
+			Ephemeral5mInputTokens: cacheCreation5m,
+			Ephemeral1hInputTokens: cacheCreation1h,
+		}
+	}
+	return usage
+}
+
+func NormalizeCacheCreationSplit(totalTokens int, tokens5m int, tokens1h int) (int, int) {
+	remainder := lo.Max([]int{totalTokens - tokens5m - tokens1h, 0})
+	return tokens5m + remainder, tokens1h
+}
+
+func StreamResponseOpenAI2Claude(openAIResponse *dto.ChatCompletionsStreamResponse, info convmeta.Meta) []*dto.ClaudeResponse {
+	if info == nil {
+		info = &convmeta.Values{}
+	}
+	state := info.EnsureClaudeConvertInfo()
+	if state.Done {
+		return nil
+	}
+
+	var claudeResponses []*dto.ClaudeResponse
+	// stopOpenBlocks emits the required content_block_stop event(s) for the currently open block(s)
+	// according to Anthropic's SSE streaming state machine:
+	// content_block_start -> content_block_delta* -> content_block_stop (per index).
+	//
+	// For text/thinking, there is at most one open block at state.Index.
+	// For tools, OpenAI tool_calls can stream multiple parallel tool_use blocks (indexed from 0),
+	// so we may have multiple open blocks and must stop each one explicitly.
+	stopOpenBlocks := func() {
+		switch state.LastMessagesType {
+		case convmeta.LastMessageTypeText, convmeta.LastMessageTypeThinking:
+			claudeResponses = append(claudeResponses, generateStopBlock(state.Index))
+		case convmeta.LastMessageTypeTools:
+			base := state.ToolCallBaseIndex
+			for offset := 0; offset <= state.ToolCallMaxIndexOffset; offset++ {
+				claudeResponses = append(claudeResponses, generateStopBlock(base+offset))
+			}
+		}
+	}
+	// stopOpenBlocksAndAdvance closes the currently open block(s) and advances the content block index
+	// to the next available slot for subsequent content_block_start events.
+	//
+	// This prevents invalid streams where a content_block_delta (e.g. thinking_delta) is emitted for an
+	// index whose active content_block type is different (the typical cause of "Mismatched content block type").
+	stopOpenBlocksAndAdvance := func() {
+		if state.LastMessagesType == convmeta.LastMessageTypeNone {
+			return
+		}
+		stopOpenBlocks()
+		switch state.LastMessagesType {
+		case convmeta.LastMessageTypeTools:
+			state.Index = state.ToolCallBaseIndex + state.ToolCallMaxIndexOffset + 1
+			state.ToolCallBaseIndex = 0
+			state.ToolCallMaxIndexOffset = 0
+		default:
+			state.Index++
+		}
+		state.LastMessagesType = convmeta.LastMessageTypeNone
+	}
+	if info.GetSendResponseCount() == 1 {
+		msg := &dto.ClaudeMediaMessage{
+			Id:    openAIResponse.Id,
+			Model: openAIResponse.Model,
+			Type:  "message",
+			Role:  "assistant",
+			Usage: &dto.ClaudeUsage{
+				InputTokens:  info.GetEstimatePromptTokens(),
+				OutputTokens: 0,
+			},
+		}
+		msg.SetContent(make([]any, 0))
+		claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+			Type:    "message_start",
+			Message: msg,
+		})
+		//claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+		//	Type: "ping",
+		//})
+		if openAIResponse.IsToolCall() {
+			state.LastMessagesType = convmeta.LastMessageTypeTools
+			state.ToolCallBaseIndex = 0
+			state.ToolCallMaxIndexOffset = 0
+			var toolCall dto.ToolCallResponse
+			if len(openAIResponse.Choices) > 0 && len(openAIResponse.Choices[0].Delta.ToolCalls) > 0 {
+				toolCall = openAIResponse.Choices[0].Delta.ToolCalls[0]
+			} else {
+				first := openAIResponse.GetFirstToolCall()
+				if first != nil {
+					toolCall = *first
+				} else {
+					toolCall = dto.ToolCallResponse{}
+				}
+			}
+			resp := &dto.ClaudeResponse{
+				Type: "content_block_start",
+				ContentBlock: &dto.ClaudeMediaMessage{
+					Id:    toolCall.ID,
+					Type:  "tool_use",
+					Name:  toolCall.Function.Name,
+					Input: map[string]interface{}{},
+				},
+			}
+			resp.SetIndex(0)
+			claudeResponses = append(claudeResponses, resp)
+			// 首块包含工具 delta，则追加 input_json_delta
+			if toolCall.Function.Arguments != "" {
+				idx := 0
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx,
+					Type:  "content_block_delta",
+					Delta: &dto.ClaudeMediaMessage{
+						Type:        "input_json_delta",
+						PartialJson: &toolCall.Function.Arguments,
+					},
+				})
+			}
+		} else {
+
+		}
+		// 判断首个响应是否存在内容（非标准的 OpenAI 响应）
+		if len(openAIResponse.Choices) > 0 {
+			reasoning := openAIResponse.Choices[0].Delta.GetReasoningContent()
+			content := openAIResponse.Choices[0].Delta.GetContentString()
+
+			if reasoning != "" {
+				if state.LastMessagesType != convmeta.LastMessageTypeThinking {
+					stopOpenBlocksAndAdvance()
+				}
+				idx := state.Index
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx,
+					Type:  "content_block_start",
+					ContentBlock: &dto.ClaudeMediaMessage{
+						Type:     "thinking",
+						Thinking: kitutil.GetPointer[string](""),
+					},
+				})
+				idx2 := idx
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx2,
+					Type:  "content_block_delta",
+					Delta: &dto.ClaudeMediaMessage{
+						Type:     "thinking_delta",
+						Thinking: &reasoning,
+					},
+				})
+				state.LastMessagesType = convmeta.LastMessageTypeThinking
+			} else if content != "" {
+				if state.LastMessagesType != convmeta.LastMessageTypeText {
+					stopOpenBlocksAndAdvance()
+				}
+				idx := state.Index
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx,
+					Type:  "content_block_start",
+					ContentBlock: &dto.ClaudeMediaMessage{
+						Type: "text",
+						Text: kitutil.GetPointer[string](""),
+					},
+				})
+				idx2 := idx
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx2,
+					Type:  "content_block_delta",
+					Delta: &dto.ClaudeMediaMessage{
+						Type: "text_delta",
+						Text: kitutil.GetPointer[string](content),
+					},
+				})
+				state.LastMessagesType = convmeta.LastMessageTypeText
+			}
+		}
+
+		// 如果首块就带 finish_reason，需要立即发送停止块
+		if len(openAIResponse.Choices) > 0 && openAIResponse.Choices[0].FinishReason != nil && *openAIResponse.Choices[0].FinishReason != "" {
+			state.FinishReason = *openAIResponse.Choices[0].FinishReason
+			stopOpenBlocks()
+			oaiUsage := openAIResponse.Usage
+			if oaiUsage == nil {
+				oaiUsage = state.Usage
+			}
+			if oaiUsage != nil {
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Type:  "message_delta",
+					Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
+					Delta: &dto.ClaudeMediaMessage{
+						StopReason: kitutil.GetPointer[string](stopReasonOpenAI2Claude(state.FinishReason)),
+					},
+				})
+			}
+			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+				Type: "message_stop",
+			})
+			state.Done = true
+		}
+		return claudeResponses
+	}
+
+	if len(openAIResponse.Choices) == 0 {
+		// Some OpenAI-compatible upstreams end with a usage-only SSE chunk.
+		oaiUsage := openAIResponse.Usage
+		if oaiUsage == nil {
+			oaiUsage = state.Usage
+		}
+		if oaiUsage != nil {
+			stopOpenBlocks()
+			stopReason := stopReasonOpenAI2Claude(state.FinishReason)
+			if stopReason == "" {
+				stopReason = "end_turn"
+			}
+			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+				Type:  "message_delta",
+				Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
+				Delta: &dto.ClaudeMediaMessage{
+					StopReason: kitutil.GetPointer[string](stopReason),
+				},
+			})
+			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+				Type: "message_stop",
+			})
+			state.Done = true
+		}
+		return claudeResponses
+	} else {
+		chosenChoice := openAIResponse.Choices[0]
+		doneChunk := chosenChoice.FinishReason != nil && *chosenChoice.FinishReason != ""
+		if doneChunk {
+			state.FinishReason = *chosenChoice.FinishReason
+			oaiUsage := openAIResponse.Usage
+			if oaiUsage == nil {
+				oaiUsage = state.Usage
+				// Some upstreams emit finish_reason first, then send a final usage-only chunk.
+				// Defer closing until usage is available so the final message_delta carries it.
+				return claudeResponses
+			}
+		}
+
+		var claudeResponse dto.ClaudeResponse
+		var isEmpty bool
+		claudeResponse.Type = "content_block_delta"
+		if len(chosenChoice.Delta.ToolCalls) > 0 {
+			toolCalls := chosenChoice.Delta.ToolCalls
+			if state.LastMessagesType != convmeta.LastMessageTypeTools {
+				stopOpenBlocksAndAdvance()
+				state.ToolCallBaseIndex = state.Index
+				state.ToolCallMaxIndexOffset = 0
+			}
+			state.LastMessagesType = convmeta.LastMessageTypeTools
+			base := state.ToolCallBaseIndex
+			maxOffset := state.ToolCallMaxIndexOffset
+
+			for i, toolCall := range toolCalls {
+				offset := 0
+				if toolCall.Index != nil {
+					offset = *toolCall.Index
+				} else {
+					offset = i
+				}
+				if offset > maxOffset {
+					maxOffset = offset
+				}
+				blockIndex := base + offset
+
+				idx := blockIndex
+				if toolCall.Function.Name != "" {
+					claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+						Index: &idx,
+						Type:  "content_block_start",
+						ContentBlock: &dto.ClaudeMediaMessage{
+							Id:    toolCall.ID,
+							Type:  "tool_use",
+							Name:  toolCall.Function.Name,
+							Input: map[string]interface{}{},
+						},
+					})
+				}
+
+				if len(toolCall.Function.Arguments) > 0 {
+					claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+						Index: &idx,
+						Type:  "content_block_delta",
+						Delta: &dto.ClaudeMediaMessage{
+							Type:        "input_json_delta",
+							PartialJson: &toolCall.Function.Arguments,
+						},
+					})
+				}
+			}
+			state.ToolCallMaxIndexOffset = maxOffset
+			state.Index = base + maxOffset
+		} else {
+			reasoning := chosenChoice.Delta.GetReasoningContent()
+			textContent := chosenChoice.Delta.GetContentString()
+			if reasoning != "" || textContent != "" {
+				if reasoning != "" {
+					if state.LastMessagesType != convmeta.LastMessageTypeThinking {
+						stopOpenBlocksAndAdvance()
+						idx := state.Index
+						claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+							Index: &idx,
+							Type:  "content_block_start",
+							ContentBlock: &dto.ClaudeMediaMessage{
+								Type:     "thinking",
+								Thinking: kitutil.GetPointer[string](""),
+							},
+						})
+					}
+					state.LastMessagesType = convmeta.LastMessageTypeThinking
+					claudeResponse.Delta = &dto.ClaudeMediaMessage{
+						Type:     "thinking_delta",
+						Thinking: &reasoning,
+					}
+				} else {
+					if state.LastMessagesType != convmeta.LastMessageTypeText {
+						stopOpenBlocksAndAdvance()
+						idx := state.Index
+						claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+							Index: &idx,
+							Type:  "content_block_start",
+							ContentBlock: &dto.ClaudeMediaMessage{
+								Type: "text",
+								Text: kitutil.GetPointer[string](""),
+							},
+						})
+					}
+					state.LastMessagesType = convmeta.LastMessageTypeText
+					claudeResponse.Delta = &dto.ClaudeMediaMessage{
+						Type: "text_delta",
+						Text: kitutil.GetPointer[string](textContent),
+					}
+				}
+			} else {
+				isEmpty = true
+			}
+		}
+
+		claudeResponse.Index = kitutil.GetPointer[int](state.Index)
+		if !isEmpty && claudeResponse.Delta != nil {
+			claudeResponses = append(claudeResponses, &claudeResponse)
+		}
+
+		if doneChunk || state.Done {
+			stopOpenBlocks()
+			oaiUsage := openAIResponse.Usage
+			if oaiUsage == nil {
+				oaiUsage = state.Usage
+			}
+			if oaiUsage != nil {
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Type:  "message_delta",
+					Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
+					Delta: &dto.ClaudeMediaMessage{
+						StopReason: kitutil.GetPointer[string](stopReasonOpenAI2Claude(state.FinishReason)),
+					},
+				})
+			}
+			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+				Type: "message_stop",
+			})
+			state.Done = true
+			return claudeResponses
+		}
+	}
+
+	return claudeResponses
+}
+
+func ResponseOpenAI2Claude(openAIResponse *dto.OpenAITextResponse, info convmeta.Meta) *dto.ClaudeResponse {
+	var stopReason string
+	contents := make([]dto.ClaudeMediaMessage, 0)
+	claudeResponse := &dto.ClaudeResponse{
+		Id:    openAIResponse.Id,
+		Type:  "message",
+		Role:  "assistant",
+		Model: openAIResponse.Model,
+	}
+	for _, choice := range openAIResponse.Choices {
+		stopReason = stopReasonOpenAI2Claude(choice.FinishReason)
+		textContent := choice.Message.StringContent()
+		toolCalls := choice.Message.ParseToolCalls()
+		if textContent != "" || len(toolCalls) == 0 {
+			claudeContent := dto.ClaudeMediaMessage{}
+			claudeContent.Type = "text"
+			claudeContent.SetText(textContent)
+			contents = append(contents, claudeContent)
+		}
+		for _, toolUse := range toolCalls {
+			claudeContent := dto.ClaudeMediaMessage{}
+			claudeContent.Type = "tool_use"
+			claudeContent.Id = toolUse.ID
+			claudeContent.Name = toolUse.Function.Name
+			mapParams := map[string]interface{}{}
+			if strings.TrimSpace(toolUse.Function.Arguments) != "" {
+				var parsed map[string]interface{}
+				if err := kitutil.Unmarshal([]byte(toolUse.Function.Arguments), &parsed); err == nil && parsed != nil {
+					mapParams = parsed
+				}
+			}
+			claudeContent.Input = mapParams
+			contents = append(contents, claudeContent)
+		}
+	}
+	claudeResponse.Content = contents
+	claudeResponse.StopReason = stopReason
+	claudeResponse.Usage = buildClaudeUsageFromOpenAIUsage(&openAIResponse.Usage)
+
+	return claudeResponse
+}
+
+func stopReasonOpenAI2Claude(reason string) string {
+	return reasonmap.OpenAIFinishReasonToClaudeStopReason(reason)
+}

relaykit/relayconvert/internal/oai_chat/to_claude_messages_resp_test.go

@@ -0,0 +1,219 @@
+package oaichat
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestResponseOpenAI2ClaudeToolUseInputIsObject(t *testing.T) {
+	tests := []struct {
+		name string
+		args string
+		want map[string]interface{}
+	}{
+		{name: "object", args: `{"q":"x"}`, want: map[string]interface{}{"q": "x"}},
+		{name: "empty", args: "", want: map[string]interface{}{}},
+		{name: "invalid", args: "{", want: map[string]interface{}{}},
+		{name: "null", args: "null", want: map[string]interface{}{}},
+		{name: "array", args: `["x"]`, want: map[string]interface{}{}},
+		{name: "string", args: `"x"`, want: map[string]interface{}{}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			msg := dto.Message{Role: "assistant"}
+			msg.SetToolCalls([]dto.ToolCallRequest{
+				{
+					ID:   "call_1",
+					Type: "function",
+					Function: dto.FunctionRequest{
+						Name:      "lookup",
+						Arguments: tt.args,
+					},
+				},
+			})
+
+			resp := ResponseOpenAI2Claude(&dto.OpenAITextResponse{
+				Id:    "chatcmpl_1",
+				Model: "gpt-test",
+				Choices: []dto.OpenAITextResponseChoice{
+					{Message: msg, FinishReason: "tool_calls"},
+				},
+			}, nil)
+
+			require.Len(t, resp.Content, 1)
+			assert.Equal(t, "tool_use", resp.Content[0].Type)
+			assert.Equal(t, tt.want, resp.Content[0].Input)
+		})
+	}
+}
+
+func TestResponseOpenAI2ClaudeUsageCarriesOpenAIBillingUsage(t *testing.T) {
+	resp := ResponseOpenAI2Claude(&dto.OpenAITextResponse{
+		Id:    "chatcmpl_1",
+		Model: "gpt-test",
+		Choices: []dto.OpenAITextResponseChoice{
+			{Message: dto.Message{Role: "assistant", Content: "hello"}, FinishReason: "stop"},
+		},
+		Usage: dto.Usage{
+			PromptTokens:     11,
+			CompletionTokens: 5,
+			TotalTokens:      16,
+		},
+	}, nil)
+
+	require.NotNil(t, resp.Usage)
+	assert.Equal(t, 11, resp.Usage.InputTokens)
+	assert.Equal(t, 5, resp.Usage.OutputTokens)
+	require.NotNil(t, resp.Usage.BillingUsage)
+	require.NotNil(t, resp.Usage.BillingUsage.OpenAIUsage)
+	assert.Equal(t, dto.BillingUsageSourceOAIChat, resp.Usage.BillingUsage.Source)
+	assert.Equal(t, dto.BillingUsageSemanticOpenAI, resp.Usage.BillingUsage.Semantic)
+	assert.Equal(t, 11, resp.Usage.BillingUsage.OpenAIUsage.PromptTokens)
+	assert.Equal(t, 5, resp.Usage.BillingUsage.OpenAIUsage.CompletionTokens)
+	assert.Equal(t, 16, resp.Usage.BillingUsage.OpenAIUsage.TotalTokens)
+	assert.Nil(t, resp.Usage.BillingUsage.OpenAIUsage.BillingUsage)
+}
+
+func TestBuildClaudeUsageFromOpenAICacheWriteUsage(t *testing.T) {
+	usage := buildClaudeUsageFromOpenAIUsage(&dto.Usage{
+		PromptTokens:     3619,
+		CompletionTokens: 36,
+		TotalTokens:      3655,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens:     2921,
+			CacheWriteTokens: 3616,
+		},
+	})
+
+	require.NotNil(t, usage)
+	// Claude semantics reports input_tokens excluding cache read/write; the
+	// overlapping unadjusted prefixes drive the remainder negative, clamp to 0.
+	assert.Equal(t, 0, usage.InputTokens)
+	assert.Equal(t, 2921, usage.CacheReadInputTokens)
+	assert.Equal(t, 3616, usage.CacheCreationInputTokens)
+	assert.Equal(t, 36, usage.OutputTokens)
+	require.NotNil(t, usage.BillingUsage)
+	require.NotNil(t, usage.BillingUsage.OpenAIUsage)
+	assert.Equal(t, dto.BillingUsageSemanticOpenAI, usage.BillingUsage.Semantic)
+	assert.Equal(t, 3616, usage.BillingUsage.OpenAIUsage.PromptTokensDetails.CacheWriteTokens)
+}
+
+func TestStreamResponseOpenAI2ClaudeClosesTextThinkingAndToolBlocks(t *testing.T) {
+	info := &convmeta.Values{
+		ClaudeConvertInfo: &convmeta.ClaudeConvertInfo{
+			LastMessagesType: convmeta.LastMessageTypeNone,
+		},
+	}
+
+	info.SendResponseCount = 1
+	textResponses := StreamResponseOpenAI2Claude(&dto.ChatCompletionsStreamResponse{
+		Id:    "chatcmpl_1",
+		Model: "gpt-test",
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{
+				Delta: dto.ChatCompletionsStreamResponseChoiceDelta{
+					Content: ptr("hello"),
+				},
+			},
+		},
+	}, info)
+	require.Len(t, textResponses, 3)
+	assert.Equal(t, "message_start", textResponses[0].Type)
+	assert.Equal(t, "content_block_start", textResponses[1].Type)
+	assert.Equal(t, 0, textResponses[1].GetIndex())
+	assert.Equal(t, "content_block_delta", textResponses[2].Type)
+
+	info.SendResponseCount = 2
+	thinkingResponses := StreamResponseOpenAI2Claude(&dto.ChatCompletionsStreamResponse{
+		Id:    "chatcmpl_1",
+		Model: "gpt-test",
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{
+				Delta: dto.ChatCompletionsStreamResponseChoiceDelta{
+					ReasoningContent: ptr("thinking"),
+				},
+			},
+		},
+	}, info)
+	require.Len(t, thinkingResponses, 3)
+	assert.Equal(t, "content_block_stop", thinkingResponses[0].Type)
+	assert.Equal(t, 0, thinkingResponses[0].GetIndex())
+	assert.Equal(t, "content_block_start", thinkingResponses[1].Type)
+	assert.Equal(t, 1, thinkingResponses[1].GetIndex())
+	assert.Equal(t, "thinking", thinkingResponses[1].ContentBlock.Type)
+	assert.Equal(t, "content_block_delta", thinkingResponses[2].Type)
+
+	info.SendResponseCount = 3
+	toolResponses := StreamResponseOpenAI2Claude(&dto.ChatCompletionsStreamResponse{
+		Id:    "chatcmpl_1",
+		Model: "gpt-test",
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{
+				Delta: dto.ChatCompletionsStreamResponseChoiceDelta{
+					ToolCalls: []dto.ToolCallResponse{
+						{
+							Index: ptr(0),
+							ID:    "call_1",
+							Type:  "function",
+							Function: dto.FunctionResponse{
+								Name:      "lookup",
+								Arguments: `{"q":"x"}`,
+							},
+						},
+					},
+				},
+			},
+		},
+	}, info)
+	require.Len(t, toolResponses, 3)
+	assert.Equal(t, "content_block_stop", toolResponses[0].Type)
+	assert.Equal(t, 1, toolResponses[0].GetIndex())
+	assert.Equal(t, "content_block_start", toolResponses[1].Type)
+	assert.Equal(t, 2, toolResponses[1].GetIndex())
+	assert.Equal(t, "tool_use", toolResponses[1].ContentBlock.Type)
+	assert.Equal(t, "content_block_delta", toolResponses[2].Type)
+
+	info.SendResponseCount = 4
+	finishResponses := StreamResponseOpenAI2Claude(&dto.ChatCompletionsStreamResponse{
+		Id:    "chatcmpl_1",
+		Model: "gpt-test",
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{FinishReason: ptr("tool_calls")},
+		},
+		Usage: &dto.Usage{
+			PromptTokens:     7,
+			CompletionTokens: 3,
+			TotalTokens:      10,
+		},
+	}, info)
+	require.Len(t, finishResponses, 3)
+	assert.Equal(t, "content_block_stop", finishResponses[0].Type)
+	assert.Equal(t, 2, finishResponses[0].GetIndex())
+	assert.Equal(t, "message_delta", finishResponses[1].Type)
+	assert.Equal(t, "tool_use", *finishResponses[1].Delta.StopReason)
+	require.NotNil(t, finishResponses[1].Usage)
+	require.NotNil(t, finishResponses[1].Usage.BillingUsage)
+	require.NotNil(t, finishResponses[1].Usage.BillingUsage.OpenAIUsage)
+	assert.Equal(t, 7, finishResponses[1].Usage.BillingUsage.OpenAIUsage.PromptTokens)
+	assert.Equal(t, 3, finishResponses[1].Usage.BillingUsage.OpenAIUsage.CompletionTokens)
+	assert.Equal(t, "message_stop", finishResponses[2].Type)
+}
+
+func TestNormalizeCacheCreationSplit(t *testing.T) {
+	cache5m, cache1h := NormalizeCacheCreationSplit(10, 3, 2)
+	assert.Equal(t, 8, cache5m)
+	assert.Equal(t, 2, cache1h)
+
+	cache5m, cache1h = NormalizeCacheCreationSplit(3, 5, 1)
+	assert.Equal(t, 5, cache5m)
+	assert.Equal(t, 1, cache1h)
+}
+
+func ptr[T any](value T) *T {
+	return &value
+}

relaykit/relayconvert/internal/oai_chat/to_gemini_chat_req.go

@@ -0,0 +1,411 @@
+package oaichat
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+
+	"context"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	relaymedia "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/media"
+	sharedgemini "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/shared/gemini"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+func OpenAIChatRequestToGeminiGenerateContent(c context.Context, textRequest dto.GeneralOpenAIRequest, info convmeta.Meta) (*dto.GeminiChatRequest, error) {
+	opts := convmeta.OptionsOf(info)
+	geminiRequest := dto.GeminiChatRequest{
+		Contents: make([]dto.GeminiChatContent, 0, len(textRequest.Messages)),
+		GenerationConfig: dto.GeminiChatGenerationConfig{
+			Temperature: textRequest.Temperature,
+		},
+	}
+
+	if textRequest.TopP != nil && *textRequest.TopP > 0 {
+		geminiRequest.GenerationConfig.TopP = kitutil.GetPointer(*textRequest.TopP)
+	}
+	if maxTokens := textRequest.GetMaxTokens(); maxTokens > 0 {
+		geminiRequest.GenerationConfig.MaxOutputTokens = kitutil.GetPointer(maxTokens)
+	}
+	if textRequest.Seed != nil && *textRequest.Seed != 0 {
+		geminiRequest.GenerationConfig.Seed = kitutil.GetPointer(int64(*textRequest.Seed))
+	}
+
+	upstreamModelName := textRequest.Model
+	if modelName := convmeta.UpstreamModelName(info); modelName != "" {
+		upstreamModelName = modelName
+	}
+
+	if opts.Gemini.SupportsImagineModel(upstreamModelName) {
+		geminiRequest.GenerationConfig.ResponseModalities = []string{
+			"TEXT",
+			"IMAGE",
+		}
+	}
+	if stopSequences := sharedgemini.ParseStopSequences(textRequest.Stop); len(stopSequences) > 0 {
+		if len(stopSequences) > 5 {
+			stopSequences = stopSequences[:5]
+		}
+		geminiRequest.GenerationConfig.StopSequences = stopSequences
+	}
+
+	adaptorWithExtraBody := false
+	if len(textRequest.ExtraBody) > 0 {
+		var extraBody map[string]interface{}
+		if err := kitutil.Unmarshal(textRequest.ExtraBody, &extraBody); err != nil {
+			return nil, fmt.Errorf("invalid extra body: %w", err)
+		}
+
+		if googleBody, ok := extraBody["google"].(map[string]interface{}); ok {
+			if !strings.HasSuffix(upstreamModelName, "-nothinking") {
+				adaptorWithExtraBody = true
+				if _, hasErrorParam := googleBody["thinkingConfig"]; hasErrorParam {
+					return nil, errors.New("extra_body.google.thinkingConfig is not supported, use extra_body.google.thinking_config instead")
+				}
+
+				if thinkingConfig, ok := googleBody["thinking_config"].(map[string]interface{}); ok {
+					if _, hasErrorParam := thinkingConfig["thinkingBudget"]; hasErrorParam {
+						return nil, errors.New("extra_body.google.thinking_config.thinkingBudget is not supported, use extra_body.google.thinking_config.thinking_budget instead")
+					}
+					var hasThinkingConfig bool
+					var tempThinkingConfig dto.GeminiThinkingConfig
+
+					if thinkingBudget, exists := thinkingConfig["thinking_budget"]; exists {
+						switch v := thinkingBudget.(type) {
+						case float64:
+							budgetInt := int(v)
+							tempThinkingConfig.ThinkingBudget = kitutil.GetPointer(budgetInt)
+							tempThinkingConfig.IncludeThoughts = budgetInt > 0
+							hasThinkingConfig = true
+						default:
+							return nil, errors.New("extra_body.google.thinking_config.thinking_budget must be an integer")
+						}
+					}
+
+					if includeThoughts, exists := thinkingConfig["include_thoughts"]; exists {
+						if v, ok := includeThoughts.(bool); ok {
+							tempThinkingConfig.IncludeThoughts = v
+							hasThinkingConfig = true
+						} else {
+							return nil, errors.New("extra_body.google.thinking_config.include_thoughts must be a boolean")
+						}
+					}
+					if thinkingLevel, exists := thinkingConfig["thinking_level"]; exists {
+						if v, ok := thinkingLevel.(string); ok {
+							tempThinkingConfig.ThinkingLevel = v
+							hasThinkingConfig = true
+						} else {
+							return nil, errors.New("extra_body.google.thinking_config.thinking_level must be a string")
+						}
+					}
+
+					if hasThinkingConfig {
+						if geminiRequest.GenerationConfig.ThinkingConfig == nil {
+							geminiRequest.GenerationConfig.ThinkingConfig = &tempThinkingConfig
+						} else {
+							if tempThinkingConfig.ThinkingBudget != nil {
+								geminiRequest.GenerationConfig.ThinkingConfig.ThinkingBudget = tempThinkingConfig.ThinkingBudget
+							}
+							geminiRequest.GenerationConfig.ThinkingConfig.IncludeThoughts = tempThinkingConfig.IncludeThoughts
+							if tempThinkingConfig.ThinkingLevel != "" {
+								geminiRequest.GenerationConfig.ThinkingConfig.ThinkingLevel = tempThinkingConfig.ThinkingLevel
+							}
+						}
+					}
+				}
+			}
+
+			if _, hasErrorParam := googleBody["imageConfig"]; hasErrorParam {
+				return nil, errors.New("extra_body.google.imageConfig is not supported, use extra_body.google.image_config instead")
+			}
+
+			if imageConfig, ok := googleBody["image_config"].(map[string]interface{}); ok {
+				if _, hasErrorParam := imageConfig["aspectRatio"]; hasErrorParam {
+					return nil, errors.New("extra_body.google.image_config.aspectRatio is not supported, use extra_body.google.image_config.aspect_ratio instead")
+				}
+				if _, hasErrorParam := imageConfig["imageSize"]; hasErrorParam {
+					return nil, errors.New("extra_body.google.image_config.imageSize is not supported, use extra_body.google.image_config.image_size instead")
+				}
+
+				geminiImageConfig := make(map[string]interface{})
+				if aspectRatio, ok := imageConfig["aspect_ratio"]; ok {
+					geminiImageConfig["aspectRatio"] = aspectRatio
+				}
+				if imageSize, ok := imageConfig["image_size"]; ok {
+					geminiImageConfig["imageSize"] = imageSize
+				}
+
+				if len(geminiImageConfig) > 0 {
+					imageConfigBytes, err := kitutil.Marshal(geminiImageConfig)
+					if err != nil {
+						return nil, fmt.Errorf("failed to marshal image_config: %w", err)
+					}
+					geminiRequest.GenerationConfig.ImageConfig = imageConfigBytes
+				}
+			}
+		}
+	}
+
+	if !adaptorWithExtraBody {
+		sharedgemini.ApplyThinkingConfig(&geminiRequest, info, textRequest)
+	}
+
+	var safetySettings []dto.GeminiChatSafetySettings
+	for _, category := range sharedgemini.SafetySettingCategories {
+		threshold := opts.Gemini.SafetySettingFor(category)
+		if threshold == "" {
+			continue
+		}
+		safetySettings = append(safetySettings, dto.GeminiChatSafetySettings{
+			Category:  category,
+			Threshold: threshold,
+		})
+	}
+	if len(safetySettings) > 0 {
+		geminiRequest.SafetySettings = safetySettings
+	}
+
+	if textRequest.Tools != nil {
+		functions := make([]dto.FunctionRequest, 0, len(textRequest.Tools))
+		googleSearch := false
+		codeExecution := false
+		urlContext := false
+		for _, tool := range textRequest.Tools {
+			if tool.Function.Name == "googleSearch" {
+				googleSearch = true
+				continue
+			}
+			if tool.Function.Name == "codeExecution" {
+				codeExecution = true
+				continue
+			}
+			if tool.Function.Name == "urlContext" {
+				urlContext = true
+				continue
+			}
+			if tool.Function.Parameters != nil {
+				if params, ok := tool.Function.Parameters.(map[string]interface{}); ok {
+					if props, hasProps := params["properties"].(map[string]interface{}); hasProps && len(props) == 0 {
+						tool.Function.Parameters = nil
+					}
+				}
+			}
+			tool.Function.Parameters = sharedgemini.CleanFunctionParameters(tool.Function.Parameters)
+			functions = append(functions, tool.Function)
+		}
+		geminiTools := geminiRequest.GetTools()
+		if codeExecution {
+			geminiTools = append(geminiTools, dto.GeminiChatTool{
+				CodeExecution: make(map[string]string),
+			})
+		}
+		if googleSearch {
+			geminiTools = append(geminiTools, dto.GeminiChatTool{
+				GoogleSearch: make(map[string]string),
+			})
+		}
+		if urlContext {
+			geminiTools = append(geminiTools, dto.GeminiChatTool{
+				URLContext: make(map[string]string),
+			})
+		}
+		if len(functions) > 0 {
+			geminiTools = append(geminiTools, dto.GeminiChatTool{
+				FunctionDeclarations: functions,
+			})
+		}
+		geminiRequest.SetTools(geminiTools)
+
+		if textRequest.ToolChoice != nil {
+			geminiRequest.ToolConfig = sharedgemini.OpenAIToolChoiceToConfig(textRequest.ToolChoice)
+		}
+	}
+
+	if textRequest.ResponseFormat != nil && (textRequest.ResponseFormat.Type == "json_schema" || textRequest.ResponseFormat.Type == "json_object") {
+		geminiRequest.GenerationConfig.ResponseMimeType = "application/json"
+
+		if len(textRequest.ResponseFormat.JsonSchema) > 0 {
+			var jsonSchema dto.FormatJsonSchema
+			if err := kitutil.Unmarshal(textRequest.ResponseFormat.JsonSchema, &jsonSchema); err == nil {
+				cleanedSchema := sharedgemini.RemoveAdditionalProperties(jsonSchema.Schema, 0)
+				geminiRequest.GenerationConfig.ResponseSchema = cleanedSchema
+			}
+		}
+	}
+
+	toolCallIDs := make(map[string]string)
+	var systemContent []string
+	for _, message := range textRequest.Messages {
+		if message.Role == "system" || message.Role == "developer" {
+			systemContent = append(systemContent, message.StringContent())
+			continue
+		}
+		if message.Role == "tool" || message.Role == "function" {
+			if len(geminiRequest.Contents) == 0 || geminiRequest.Contents[len(geminiRequest.Contents)-1].Role == "model" {
+				geminiRequest.Contents = append(geminiRequest.Contents, dto.GeminiChatContent{
+					Role: "user",
+				})
+			}
+			parts := &geminiRequest.Contents[len(geminiRequest.Contents)-1].Parts
+			name := ""
+			if message.Name != nil {
+				name = *message.Name
+			} else if val, exists := toolCallIDs[message.ToolCallId]; exists {
+				name = val
+			}
+			var contentMap map[string]interface{}
+			contentStr := message.StringContent()
+
+			if err := kitutil.Unmarshal([]byte(contentStr), &contentMap); err != nil {
+				var contentSlice []interface{}
+				if err := kitutil.Unmarshal([]byte(contentStr), &contentSlice); err == nil {
+					contentMap = map[string]interface{}{"result": contentSlice}
+				} else {
+					contentMap = map[string]interface{}{"content": contentStr}
+				}
+			}
+
+			functionResp := &dto.GeminiFunctionResponse{
+				Name:     name,
+				Response: contentMap,
+			}
+
+			*parts = append(*parts, dto.GeminiPart{
+				FunctionResponse: functionResp,
+			})
+			continue
+		}
+
+		var parts []dto.GeminiPart
+		content := dto.GeminiChatContent{
+			Role: message.Role,
+		}
+		shouldAttachThoughtSignature := (message.Role == "assistant" || message.Role == "model") && sharedgemini.ShouldAttachThoughtSignature(opts)
+		signatureAttached := false
+		if message.ToolCalls != nil {
+			for _, call := range message.ParseToolCalls() {
+				args := map[string]interface{}{}
+				if call.Function.Arguments != "" {
+					if kitutil.Unmarshal([]byte(call.Function.Arguments), &args) != nil {
+						return nil, fmt.Errorf("invalid arguments for function %s, args: %s", call.Function.Name, call.Function.Arguments)
+					}
+				}
+				toolCall := dto.GeminiPart{
+					FunctionCall: &dto.FunctionCall{
+						FunctionName: call.Function.Name,
+						Arguments:    args,
+					},
+				}
+				if shouldAttachThoughtSignature && !signatureAttached && sharedgemini.AttachFunctionCallThoughtSignature(opts, &toolCall) {
+					signatureAttached = true
+				}
+				parts = append(parts, toolCall)
+				toolCallIDs[call.ID] = call.Function.Name
+			}
+		}
+
+		openaiContent := message.ParseContent()
+		for _, part := range openaiContent {
+			if part.Type == dto.ContentTypeText {
+				if part.Text == "" {
+					continue
+				}
+				text := part.Text
+				hasMarkdownImage := false
+				for {
+					startIdx := strings.Index(text, "![")
+					if startIdx == -1 {
+						break
+					}
+					bracketIdx := strings.Index(text[startIdx:], "](data:")
+					if bracketIdx == -1 {
+						break
+					}
+					bracketIdx += startIdx
+					closeIdx := strings.Index(text[bracketIdx+2:], ")")
+					if closeIdx == -1 {
+						break
+					}
+					closeIdx += bracketIdx + 2
+
+					hasMarkdownImage = true
+					if startIdx > 0 {
+						textBefore := text[:startIdx]
+						if textBefore != "" {
+							parts = append(parts, dto.GeminiPart{
+								Text: textBefore,
+							})
+						}
+					}
+
+					dataURL := text[bracketIdx+2 : closeIdx]
+					format, base64String, err := relaymedia.DecodeBase64FileData(dataURL)
+					if err != nil {
+						return nil, fmt.Errorf("decode markdown base64 image data failed: %s", err.Error())
+					}
+					imgPart := dto.GeminiPart{
+						InlineData: &dto.GeminiInlineData{
+							MimeType: format,
+							Data:     base64String,
+						},
+					}
+					if shouldAttachThoughtSignature {
+						sharedgemini.AttachThoughtSignatureBypass(opts, &imgPart)
+					}
+					parts = append(parts, imgPart)
+					text = text[closeIdx+1:]
+				}
+				if !hasMarkdownImage {
+					parts = append(parts, dto.GeminiPart{
+						Text: part.Text,
+					})
+				}
+			} else {
+				source := part.ToFileSource()
+				if source == nil {
+					continue
+				}
+				base64Data, mimeType, err := relaymedia.ResolveBase64Data(c, source, "formatting image for Gemini")
+				if err != nil {
+					return nil, fmt.Errorf("get file data from '%s' failed: %w", source.GetIdentifier(), err)
+				}
+
+				if _, ok := sharedgemini.SupportedMimeTypes[strings.ToLower(mimeType)]; !ok {
+					return nil, fmt.Errorf("mime type is not supported by Gemini: '%s', url: '%s', supported types are: %v", mimeType, source.GetIdentifier(), sharedgemini.SupportedMimeTypesList())
+				}
+
+				parts = append(parts, dto.GeminiPart{
+					InlineData: &dto.GeminiInlineData{
+						MimeType: mimeType,
+						Data:     base64Data,
+					},
+				})
+			}
+		}
+
+		if shouldAttachThoughtSignature && !signatureAttached && len(parts) > 0 {
+			sharedgemini.AttachFirstTextThoughtSignature(opts, parts)
+		}
+
+		content.Parts = parts
+		if content.Role == "assistant" {
+			content.Role = "model"
+		}
+		if len(content.Parts) > 0 {
+			geminiRequest.Contents = append(geminiRequest.Contents, content)
+		}
+	}
+
+	if len(systemContent) > 0 {
+		geminiRequest.SystemInstructions = &dto.GeminiChatContent{
+			Parts: []dto.GeminiPart{
+				{
+					Text: strings.Join(systemContent, "\n"),
+				},
+			},
+		}
+	}
+
+	return &geminiRequest, nil
+}

relaykit/relayconvert/internal/oai_chat/to_gemini_chat_resp.go

@@ -0,0 +1,230 @@
+package oaichat
+
+import (
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+// ResponseOpenAI2Gemini 将 OpenAI 响应转换为 Gemini 格式
+func ResponseOpenAI2Gemini(openAIResponse *dto.OpenAITextResponse, info convmeta.Meta) *dto.GeminiChatResponse {
+	totalTokens := openAIResponse.TotalTokens
+	if totalTokens == 0 {
+		totalTokens = openAIResponse.PromptTokens + openAIResponse.CompletionTokens
+	}
+	geminiResponse := &dto.GeminiChatResponse{
+		Candidates:       make([]dto.GeminiChatCandidate, 0, len(openAIResponse.Choices)),
+		HasUsageMetadata: true,
+		UsageMetadata: dto.GeminiUsageMetadata{
+			PromptTokenCount:     openAIResponse.PromptTokens,
+			CandidatesTokenCount: openAIResponse.CompletionTokens,
+			TotalTokenCount:      totalTokens,
+			BillingUsage:         openAIBillingUsageFromUsage(&openAIResponse.Usage),
+		},
+	}
+	if metadata, ok := geminiBillingMetadataFromOpenAIUsage(&openAIResponse.Usage); ok {
+		geminiResponse.UsageMetadata = metadata
+	}
+
+	for _, choice := range openAIResponse.Choices {
+		candidate := dto.GeminiChatCandidate{
+			Index:         int64(choice.Index),
+			SafetyRatings: []dto.GeminiChatSafetyRating{},
+		}
+
+		// 设置结束原因
+		var finishReason string
+		switch choice.FinishReason {
+		case "stop":
+			finishReason = "STOP"
+		case "length":
+			finishReason = "MAX_TOKENS"
+		case "content_filter":
+			finishReason = "SAFETY"
+		case "tool_calls":
+			finishReason = "STOP"
+		default:
+			finishReason = "STOP"
+		}
+		candidate.FinishReason = &finishReason
+
+		// 转换消息内容
+		content := dto.GeminiChatContent{
+			Role:  "model",
+			Parts: make([]dto.GeminiPart, 0),
+		}
+
+		textContent := choice.Message.StringContent()
+		if textContent != "" {
+			part := dto.GeminiPart{
+				Text: textContent,
+			}
+			content.Parts = append(content.Parts, part)
+		}
+
+		toolCalls := choice.Message.ParseToolCalls()
+		for _, toolCall := range toolCalls {
+			var args map[string]interface{}
+			if toolCall.Function.Arguments != "" {
+				if err := kitutil.Unmarshal([]byte(toolCall.Function.Arguments), &args); err != nil {
+					args = map[string]interface{}{"arguments": toolCall.Function.Arguments}
+				}
+			} else {
+				args = make(map[string]interface{})
+			}
+
+			part := dto.GeminiPart{
+				FunctionCall: &dto.FunctionCall{
+					FunctionName: toolCall.Function.Name,
+					Arguments:    args,
+				},
+			}
+			content.Parts = append(content.Parts, part)
+		}
+
+		candidate.Content = content
+		geminiResponse.Candidates = append(geminiResponse.Candidates, candidate)
+	}
+
+	return geminiResponse
+}
+
+// StreamResponseOpenAI2Gemini 将 OpenAI 流式响应转换为 Gemini 格式
+func StreamResponseOpenAI2Gemini(openAIResponse *dto.ChatCompletionsStreamResponse, info convmeta.Meta) *dto.GeminiChatResponse {
+	// 检查是否有实际内容或结束标志
+	hasContent := false
+	hasFinishReason := false
+	for _, choice := range openAIResponse.Choices {
+		if len(choice.Delta.GetContentString()) > 0 || (choice.Delta.ToolCalls != nil && len(choice.Delta.ToolCalls) > 0) {
+			hasContent = true
+		}
+		if choice.FinishReason != nil {
+			hasFinishReason = true
+		}
+	}
+
+	// 如果没有实际内容且没有结束标志，跳过。主要针对 openai 流响应开头的空数据
+	if !hasContent && !hasFinishReason {
+		return nil
+	}
+
+	estimatePromptTokens := 0
+	if info != nil {
+		estimatePromptTokens = info.GetEstimatePromptTokens()
+	}
+	geminiResponse := &dto.GeminiChatResponse{
+		Candidates:       make([]dto.GeminiChatCandidate, 0, len(openAIResponse.Choices)),
+		HasUsageMetadata: true,
+		UsageMetadata: dto.GeminiUsageMetadata{
+			PromptTokenCount:     estimatePromptTokens,
+			CandidatesTokenCount: 0, // 流式响应中可能没有完整的 usage 信息
+			TotalTokenCount:      estimatePromptTokens,
+		},
+	}
+
+	if openAIResponse.Usage != nil {
+		geminiResponse.UsageMetadata.PromptTokenCount = openAIResponse.Usage.PromptTokens
+		geminiResponse.UsageMetadata.CandidatesTokenCount = openAIResponse.Usage.CompletionTokens
+		geminiResponse.UsageMetadata.TotalTokenCount = openAIResponse.Usage.TotalTokens
+		geminiResponse.UsageMetadata.BillingUsage = openAIBillingUsageFromUsage(openAIResponse.Usage)
+		if metadata, ok := geminiBillingMetadataFromOpenAIUsage(openAIResponse.Usage); ok {
+			geminiResponse.UsageMetadata = metadata
+		}
+	}
+
+	for _, choice := range openAIResponse.Choices {
+		candidate := dto.GeminiChatCandidate{
+			Index:         int64(choice.Index),
+			SafetyRatings: []dto.GeminiChatSafetyRating{},
+		}
+
+		// 设置结束原因
+		if choice.FinishReason != nil {
+			var finishReason string
+			switch *choice.FinishReason {
+			case "stop":
+				finishReason = "STOP"
+			case "length":
+				finishReason = "MAX_TOKENS"
+			case "content_filter":
+				finishReason = "SAFETY"
+			case "tool_calls":
+				finishReason = "STOP"
+			default:
+				finishReason = "STOP"
+			}
+			candidate.FinishReason = &finishReason
+		}
+
+		// 转换消息内容
+		content := dto.GeminiChatContent{
+			Role:  "model",
+			Parts: make([]dto.GeminiPart, 0),
+		}
+
+		// 处理工具调用
+		if choice.Delta.ToolCalls != nil {
+			for _, toolCall := range choice.Delta.ToolCalls {
+				// 解析参数
+				var args map[string]interface{}
+				if toolCall.Function.Arguments != "" {
+					if err := kitutil.Unmarshal([]byte(toolCall.Function.Arguments), &args); err != nil {
+						args = map[string]interface{}{"arguments": toolCall.Function.Arguments}
+					}
+				} else {
+					args = make(map[string]interface{})
+				}
+
+				part := dto.GeminiPart{
+					FunctionCall: &dto.FunctionCall{
+						FunctionName: toolCall.Function.Name,
+						Arguments:    args,
+					},
+				}
+				content.Parts = append(content.Parts, part)
+			}
+		} else {
+			// 处理文本内容
+			textContent := choice.Delta.GetContentString()
+			if textContent != "" {
+				part := dto.GeminiPart{
+					Text: textContent,
+				}
+				content.Parts = append(content.Parts, part)
+			}
+		}
+
+		candidate.Content = content
+		geminiResponse.Candidates = append(geminiResponse.Candidates, candidate)
+	}
+
+	return geminiResponse
+}
+
+func geminiBillingMetadataFromOpenAIUsage(usage *dto.Usage) (dto.GeminiUsageMetadata, bool) {
+	if usage == nil || usage.BillingUsage == nil || usage.BillingUsage.GeminiUsageMetadata == nil {
+		return dto.GeminiUsageMetadata{}, false
+	}
+	if usage.BillingUsage.Source != dto.BillingUsageSourceGeminiChat && usage.BillingUsage.Semantic != dto.BillingUsageSemanticGemini {
+		return dto.GeminiUsageMetadata{}, false
+	}
+	billingUsage := dto.CloneBillingUsage(usage.BillingUsage)
+	if billingUsage == nil || billingUsage.GeminiUsageMetadata == nil {
+		return dto.GeminiUsageMetadata{}, false
+	}
+	return *billingUsage.GeminiUsageMetadata, true
+}
+
+func openAIBillingUsageFromUsage(usage *dto.Usage) *dto.BillingUsage {
+	if usage == nil {
+		return nil
+	}
+	if existingBillingUsage := dto.CloneBillingUsage(usage.BillingUsage); existingBillingUsage != nil && existingBillingUsage.OpenAIUsage != nil {
+		if existingBillingUsage.Source == dto.BillingUsageSourceOAIChat ||
+			existingBillingUsage.Source == dto.BillingUsageSourceOAIResponses ||
+			existingBillingUsage.Semantic == dto.BillingUsageSemanticOpenAI {
+			return existingBillingUsage
+		}
+	}
+	return dto.NewOpenAIChatBillingUsage(usage)
+}

relaykit/relayconvert/internal/oai_chat/to_gemini_chat_resp_test.go

@@ -0,0 +1,112 @@
+package oaichat
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestResponseOpenAI2GeminiMapsTextToolFinishReasonAndUsage(t *testing.T) {
+	msg := dto.Message{
+		Role:    "assistant",
+		Content: "hello",
+	}
+	msg.SetToolCalls([]dto.ToolCallRequest{
+		{
+			ID:   "call_1",
+			Type: "function",
+			Function: dto.FunctionRequest{
+				Name:      "lookup",
+				Arguments: `{"q":"x"}`,
+			},
+		},
+	})
+
+	resp := ResponseOpenAI2Gemini(&dto.OpenAITextResponse{
+		Model: "gpt-test",
+		Choices: []dto.OpenAITextResponseChoice{
+			{
+				Index:        2,
+				Message:      msg,
+				FinishReason: "length",
+			},
+		},
+		Usage: dto.Usage{
+			PromptTokens:     11,
+			CompletionTokens: 5,
+			TotalTokens:      16,
+		},
+	}, nil)
+
+	assert.Equal(t, 11, resp.UsageMetadata.PromptTokenCount)
+	assert.Equal(t, 5, resp.UsageMetadata.CandidatesTokenCount)
+	assert.Equal(t, 16, resp.UsageMetadata.TotalTokenCount)
+	require.NotNil(t, resp.UsageMetadata.BillingUsage)
+	require.NotNil(t, resp.UsageMetadata.BillingUsage.OpenAIUsage)
+	assert.Equal(t, dto.BillingUsageSourceOAIChat, resp.UsageMetadata.BillingUsage.Source)
+	assert.Equal(t, dto.BillingUsageSemanticOpenAI, resp.UsageMetadata.BillingUsage.Semantic)
+	assert.Equal(t, 11, resp.UsageMetadata.BillingUsage.OpenAIUsage.PromptTokens)
+	assert.Equal(t, 5, resp.UsageMetadata.BillingUsage.OpenAIUsage.CompletionTokens)
+	assert.Equal(t, 16, resp.UsageMetadata.BillingUsage.OpenAIUsage.TotalTokens)
+	assert.Nil(t, resp.UsageMetadata.BillingUsage.OpenAIUsage.BillingUsage)
+	require.Len(t, resp.Candidates, 1)
+	assert.Equal(t, int64(2), resp.Candidates[0].Index)
+	require.NotNil(t, resp.Candidates[0].FinishReason)
+	assert.Equal(t, "MAX_TOKENS", *resp.Candidates[0].FinishReason)
+	require.Len(t, resp.Candidates[0].Content.Parts, 2)
+	assert.Equal(t, "hello", resp.Candidates[0].Content.Parts[0].Text)
+	require.NotNil(t, resp.Candidates[0].Content.Parts[1].FunctionCall)
+	assert.Equal(t, "lookup", resp.Candidates[0].Content.Parts[1].FunctionCall.FunctionName)
+	assert.Equal(t, map[string]interface{}{"q": "x"}, resp.Candidates[0].Content.Parts[1].FunctionCall.Arguments)
+}
+
+func TestStreamResponseOpenAI2GeminiMapsToolCallFinishReasonAndUsage(t *testing.T) {
+	resp := StreamResponseOpenAI2Gemini(&dto.ChatCompletionsStreamResponse{
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{
+				Index:        1,
+				FinishReason: geminiRespPtr("tool_calls"),
+				Delta: dto.ChatCompletionsStreamResponseChoiceDelta{
+					ToolCalls: []dto.ToolCallResponse{
+						{
+							Type: "function",
+							Function: dto.FunctionResponse{
+								Name:      "lookup",
+								Arguments: `{"q":"x"}`,
+							},
+						},
+					},
+				},
+			},
+		},
+		Usage: &dto.Usage{
+			PromptTokens:     13,
+			CompletionTokens: 8,
+			TotalTokens:      21,
+		},
+	}, &convmeta.Values{})
+
+	require.NotNil(t, resp)
+	assert.Equal(t, 13, resp.UsageMetadata.PromptTokenCount)
+	assert.Equal(t, 8, resp.UsageMetadata.CandidatesTokenCount)
+	assert.Equal(t, 21, resp.UsageMetadata.TotalTokenCount)
+	require.NotNil(t, resp.UsageMetadata.BillingUsage)
+	require.NotNil(t, resp.UsageMetadata.BillingUsage.OpenAIUsage)
+	assert.Equal(t, 13, resp.UsageMetadata.BillingUsage.OpenAIUsage.PromptTokens)
+	assert.Equal(t, 8, resp.UsageMetadata.BillingUsage.OpenAIUsage.CompletionTokens)
+	require.Len(t, resp.Candidates, 1)
+	assert.Equal(t, int64(1), resp.Candidates[0].Index)
+	require.NotNil(t, resp.Candidates[0].FinishReason)
+	assert.Equal(t, "STOP", *resp.Candidates[0].FinishReason)
+	require.Len(t, resp.Candidates[0].Content.Parts, 1)
+	require.NotNil(t, resp.Candidates[0].Content.Parts[0].FunctionCall)
+	assert.Equal(t, "lookup", resp.Candidates[0].Content.Parts[0].FunctionCall.FunctionName)
+	assert.Equal(t, map[string]interface{}{"q": "x"}, resp.Candidates[0].Content.Parts[0].FunctionCall.Arguments)
+}
+
+func geminiRespPtr[T any](value T) *T {
+	return &value
+}

relaykit/relayconvert/internal/oai_chat/to_oai_responses_req.go

@@ -0,0 +1,402 @@
+package oaichat
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/samber/lo"
+)
+
+func normalizeChatImageURLToString(v any) any {
+	switch vv := v.(type) {
+	case string:
+		return vv
+	case map[string]any:
+		if url := kitutil.Interface2String(vv["url"]); url != "" {
+			return url
+		}
+		return v
+	case dto.MessageImageUrl:
+		if vv.Url != "" {
+			return vv.Url
+		}
+		return v
+	case *dto.MessageImageUrl:
+		if vv != nil && vv.Url != "" {
+			return vv.Url
+		}
+		return v
+	default:
+		return v
+	}
+}
+
+func convertChatResponseFormatToResponsesText(reqFormat *dto.ResponseFormat) json.RawMessage {
+	if reqFormat == nil || strings.TrimSpace(reqFormat.Type) == "" {
+		return nil
+	}
+
+	format := map[string]any{
+		"type": reqFormat.Type,
+	}
+
+	if reqFormat.Type == "json_schema" && len(reqFormat.JsonSchema) > 0 {
+		var chatSchema map[string]any
+		if err := kitutil.Unmarshal(reqFormat.JsonSchema, &chatSchema); err == nil {
+			for key, value := range chatSchema {
+				if key == "type" {
+					continue
+				}
+				format[key] = value
+			}
+
+			if nested, ok := format["json_schema"].(map[string]any); ok {
+				for key, value := range nested {
+					if _, exists := format[key]; !exists {
+						format[key] = value
+					}
+				}
+				delete(format, "json_schema")
+			}
+		} else {
+			format["json_schema"] = reqFormat.JsonSchema
+		}
+	}
+
+	textRaw, _ := kitutil.Marshal(map[string]any{
+		"format": format,
+	})
+	return textRaw
+}
+
+func ChatCompletionsRequestToResponsesRequest(req *dto.GeneralOpenAIRequest) (*dto.OpenAIResponsesRequest, error) {
+	if req == nil {
+		return nil, errors.New("request is nil")
+	}
+	if req.Model == "" {
+		return nil, errors.New("model is required")
+	}
+	if lo.FromPtrOr(req.N, 1) > 1 {
+		return nil, fmt.Errorf("n>1 is not supported in responses compatibility mode")
+	}
+
+	var instructionsParts []string
+	inputItems := make([]map[string]any, 0, len(req.Messages))
+
+	for _, msg := range req.Messages {
+		role := strings.TrimSpace(msg.Role)
+		if role == "" {
+			continue
+		}
+
+		if role == "tool" || role == "function" {
+			callID := strings.TrimSpace(msg.ToolCallId)
+
+			var output any
+			if msg.Content == nil {
+				output = ""
+			} else if msg.IsStringContent() {
+				output = msg.StringContent()
+			} else {
+				if b, err := kitutil.Marshal(msg.Content); err == nil {
+					output = string(b)
+				} else {
+					output = fmt.Sprintf("%v", msg.Content)
+				}
+			}
+
+			if callID == "" {
+				inputItems = append(inputItems, map[string]any{
+					"role":    "user",
+					"content": fmt.Sprintf("[tool_output_missing_call_id] %v", output),
+				})
+				continue
+			}
+
+			inputItems = append(inputItems, map[string]any{
+				"type":    "function_call_output",
+				"call_id": callID,
+				"output":  output,
+			})
+			continue
+		}
+
+		// Prefer mapping system/developer messages into `instructions`.
+		if role == "system" || role == "developer" {
+			if msg.Content == nil {
+				continue
+			}
+			if msg.IsStringContent() {
+				if s := strings.TrimSpace(msg.StringContent()); s != "" {
+					instructionsParts = append(instructionsParts, s)
+				}
+				continue
+			}
+			parts := msg.ParseContent()
+			var sb strings.Builder
+			for _, part := range parts {
+				if part.Type == dto.ContentTypeText && strings.TrimSpace(part.Text) != "" {
+					if sb.Len() > 0 {
+						sb.WriteString("\n")
+					}
+					sb.WriteString(part.Text)
+				}
+			}
+			if s := strings.TrimSpace(sb.String()); s != "" {
+				instructionsParts = append(instructionsParts, s)
+			}
+			continue
+		}
+
+		item := map[string]any{
+			"role": role,
+		}
+
+		if msg.Content == nil {
+			item["content"] = ""
+			inputItems = append(inputItems, item)
+
+			if role == "assistant" {
+				for _, tc := range msg.ParseToolCalls() {
+					if strings.TrimSpace(tc.ID) == "" {
+						continue
+					}
+					if tc.Type != "" && tc.Type != "function" {
+						continue
+					}
+					name := strings.TrimSpace(tc.Function.Name)
+					if name == "" {
+						continue
+					}
+					inputItems = append(inputItems, map[string]any{
+						"type":      "function_call",
+						"call_id":   tc.ID,
+						"name":      name,
+						"arguments": tc.Function.Arguments,
+					})
+				}
+			}
+			continue
+		}
+
+		if msg.IsStringContent() {
+			item["content"] = msg.StringContent()
+			inputItems = append(inputItems, item)
+
+			if role == "assistant" {
+				for _, tc := range msg.ParseToolCalls() {
+					if strings.TrimSpace(tc.ID) == "" {
+						continue
+					}
+					if tc.Type != "" && tc.Type != "function" {
+						continue
+					}
+					name := strings.TrimSpace(tc.Function.Name)
+					if name == "" {
+						continue
+					}
+					inputItems = append(inputItems, map[string]any{
+						"type":      "function_call",
+						"call_id":   tc.ID,
+						"name":      name,
+						"arguments": tc.Function.Arguments,
+					})
+				}
+			}
+			continue
+		}
+
+		parts := msg.ParseContent()
+		contentParts := make([]map[string]any, 0, len(parts))
+		for _, part := range parts {
+			switch part.Type {
+			case dto.ContentTypeText:
+				textType := "input_text"
+				if role == "assistant" {
+					textType = "output_text"
+				}
+				contentParts = append(contentParts, map[string]any{
+					"type": textType,
+					"text": part.Text,
+				})
+			case dto.ContentTypeImageURL:
+				contentParts = append(contentParts, map[string]any{
+					"type":      "input_image",
+					"image_url": normalizeChatImageURLToString(part.ImageUrl),
+				})
+			case dto.ContentTypeInputAudio:
+				contentParts = append(contentParts, map[string]any{
+					"type":        "input_audio",
+					"input_audio": part.InputAudio,
+				})
+			case dto.ContentTypeFile:
+				contentParts = append(contentParts, map[string]any{
+					"type": "input_file",
+					"file": part.File,
+				})
+			case dto.ContentTypeVideoUrl:
+				contentParts = append(contentParts, map[string]any{
+					"type":      "input_video",
+					"video_url": part.VideoUrl,
+				})
+			default:
+				contentParts = append(contentParts, map[string]any{
+					"type": part.Type,
+				})
+			}
+		}
+		item["content"] = contentParts
+		inputItems = append(inputItems, item)
+
+		if role == "assistant" {
+			for _, tc := range msg.ParseToolCalls() {
+				if strings.TrimSpace(tc.ID) == "" {
+					continue
+				}
+				if tc.Type != "" && tc.Type != "function" {
+					continue
+				}
+				name := strings.TrimSpace(tc.Function.Name)
+				if name == "" {
+					continue
+				}
+				inputItems = append(inputItems, map[string]any{
+					"type":      "function_call",
+					"call_id":   tc.ID,
+					"name":      name,
+					"arguments": tc.Function.Arguments,
+				})
+			}
+		}
+	}
+
+	inputRaw, err := kitutil.Marshal(inputItems)
+	if err != nil {
+		return nil, err
+	}
+
+	var instructionsRaw json.RawMessage
+	if len(instructionsParts) > 0 {
+		instructions := strings.Join(instructionsParts, "\n\n")
+		instructionsRaw, _ = kitutil.Marshal(instructions)
+	}
+
+	var toolsRaw json.RawMessage
+	if req.Tools != nil {
+		tools := make([]map[string]any, 0, len(req.Tools))
+		for _, tool := range req.Tools {
+			switch tool.Type {
+			case "function":
+				tools = append(tools, map[string]any{
+					"type":        "function",
+					"name":        tool.Function.Name,
+					"description": tool.Function.Description,
+					"parameters":  tool.Function.Parameters,
+				})
+			default:
+				// Best-effort: keep original tool shape for unknown types.
+				var m map[string]any
+				if b, err := kitutil.Marshal(tool); err == nil {
+					_ = kitutil.Unmarshal(b, &m)
+				}
+				if len(m) == 0 {
+					m = map[string]any{"type": tool.Type}
+				}
+				tools = append(tools, m)
+			}
+		}
+		toolsRaw, _ = kitutil.Marshal(tools)
+	}
+
+	var toolChoiceRaw json.RawMessage
+	if req.ToolChoice != nil {
+		switch v := req.ToolChoice.(type) {
+		case string:
+			toolChoiceRaw, _ = kitutil.Marshal(v)
+		default:
+			var m map[string]any
+			if b, err := kitutil.Marshal(v); err == nil {
+				_ = kitutil.Unmarshal(b, &m)
+			}
+			if m == nil {
+				toolChoiceRaw, _ = kitutil.Marshal(v)
+			} else if t, _ := m["type"].(string); t == "function" {
+				// Chat: {"type":"function","function":{"name":"..."}}
+				// Responses: {"type":"function","name":"..."}
+				if name, ok := m["name"].(string); ok && name != "" {
+					toolChoiceRaw, _ = kitutil.Marshal(map[string]any{
+						"type": "function",
+						"name": name,
+					})
+				} else if fn, ok := m["function"].(map[string]any); ok {
+					if name, ok := fn["name"].(string); ok && name != "" {
+						toolChoiceRaw, _ = kitutil.Marshal(map[string]any{
+							"type": "function",
+							"name": name,
+						})
+					} else {
+						toolChoiceRaw, _ = kitutil.Marshal(v)
+					}
+				} else {
+					toolChoiceRaw, _ = kitutil.Marshal(v)
+				}
+			} else {
+				toolChoiceRaw, _ = kitutil.Marshal(v)
+			}
+		}
+	}
+
+	var parallelToolCallsRaw json.RawMessage
+	if req.ParallelTooCalls != nil {
+		parallelToolCallsRaw, _ = kitutil.Marshal(*req.ParallelTooCalls)
+	}
+
+	textRaw := convertChatResponseFormatToResponsesText(req.ResponseFormat)
+
+	maxOutputTokens := lo.FromPtrOr(req.MaxTokens, uint(0))
+	maxCompletionTokens := lo.FromPtrOr(req.MaxCompletionTokens, uint(0))
+	if maxCompletionTokens > maxOutputTokens {
+		maxOutputTokens = maxCompletionTokens
+	}
+	// OpenAI Responses API rejects max_output_tokens < 16 when explicitly provided.
+	//if maxOutputTokens > 0 && maxOutputTokens < 16 {
+	//	maxOutputTokens = 16
+	//}
+
+	var topP *float64
+	if req.TopP != nil {
+		topP = kitutil.GetPointer(lo.FromPtr(req.TopP))
+	}
+
+	out := &dto.OpenAIResponsesRequest{
+		Model:             req.Model,
+		Input:             inputRaw,
+		Instructions:      instructionsRaw,
+		Stream:            req.Stream,
+		Temperature:       req.Temperature,
+		Text:              textRaw,
+		ToolChoice:        toolChoiceRaw,
+		Tools:             toolsRaw,
+		TopP:              topP,
+		User:              req.User,
+		ParallelToolCalls: parallelToolCallsRaw,
+		Store:             req.Store,
+		Metadata:          req.Metadata,
+	}
+	if req.MaxTokens != nil || req.MaxCompletionTokens != nil {
+		out.MaxOutputTokens = lo.ToPtr(maxOutputTokens)
+	}
+
+	if req.ReasoningEffort != "" {
+		out.Reasoning = &dto.Reasoning{
+			Effort:  req.ReasoningEffort,
+			Summary: "detailed",
+		}
+	}
+
+	return out, nil
+}

relaykit/relayconvert/internal/oai_chat/to_oai_responses_req_test.go

@@ -0,0 +1,62 @@
+package oaichat
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/samber/lo"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+func TestChatCompletionsRequestToResponsesRequestInstructionsAndTools(t *testing.T) {
+	req := &dto.GeneralOpenAIRequest{
+		Model: "gpt-test",
+		N:     lo.ToPtr(1),
+		Messages: []dto.Message{
+			{Role: "system", Content: "system rules"},
+			{Role: "developer", Content: "developer rules"},
+			{Role: "user", Content: []any{
+				map[string]any{"type": "text", "text": "look"},
+				map[string]any{"type": "image_url", "image_url": map[string]any{"url": "https://example.test/a.png"}},
+			}},
+			assistantMessageWithTool("partial text", "call_1", "lookup", `{"q":"x"}`),
+			{Role: "tool", ToolCallId: "call_1", Content: "tool result"},
+		},
+	}
+
+	got, err := ChatCompletionsRequestToResponsesRequest(req)
+	require.NoError(t, err)
+
+	assert.Equal(t, "gpt-test", got.Model)
+	assert.Equal(t, `"system rules\n\ndeveloper rules"`, string(got.Instructions))
+	assert.Equal(t, "input_image", gjson.GetBytes(got.Input, "0.content.1.type").String())
+	assert.Equal(t, "function_call", gjson.GetBytes(got.Input, "2.type").String())
+	assert.Equal(t, "call_1", gjson.GetBytes(got.Input, "2.call_id").String())
+	assert.Equal(t, "function_call_output", gjson.GetBytes(got.Input, "3.type").String())
+}
+
+func TestChatCompletionsRequestToResponsesRequestRejectsMultipleChoices(t *testing.T) {
+	_, err := ChatCompletionsRequestToResponsesRequest(&dto.GeneralOpenAIRequest{
+		Model: "gpt-test",
+		N:     lo.ToPtr(2),
+	})
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "n>1")
+}
+
+func assistantMessageWithTool(content string, id string, name string, args string) dto.Message {
+	msg := dto.Message{Role: "assistant", Content: content}
+	msg.SetToolCalls([]dto.ToolCallRequest{
+		{
+			ID:   id,
+			Type: "function",
+			Function: dto.FunctionRequest{
+				Name:      name,
+				Arguments: args,
+			},
+		},
+	})
+	return msg
+}

relaykit/relayconvert/internal/oai_chat/to_oai_responses_resp.go

@@ -0,0 +1,232 @@
+package oaichat
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+const (
+	chatFinishReasonLength        = "length"
+	chatFinishReasonContentFilter = "content_filter"
+
+	responsesEventCreated                  = "response.created"
+	responsesEventCompleted                = "response.completed"
+	responsesEventIncomplete               = "response.incomplete"
+	responsesEventOutputTextDelta          = "response.output_text.delta"
+	responsesEventOutputItemAdded          = "response.output_item.added"
+	responsesEventOutputItemDone           = "response.output_item.done"
+	responsesEventFunctionArgsDelta        = "response.function_call_arguments.delta"
+	responsesEventFunctionArgsDone         = "response.function_call_arguments.done"
+	responsesEventReasoningSummaryDelta    = "response.reasoning_summary_text.delta"
+	responsesEventReasoningSummaryDone     = "response.reasoning_summary_text.done"
+	responsesOutputTypeFunctionCall        = "function_call"
+	responsesOutputTypeMessage             = "message"
+	responsesOutputTypeReasoning           = "reasoning"
+	responsesIncompleteReasonContentFilter = "content_filter"
+	responsesIncompleteReasonMaxTokens     = "max_output_tokens"
+)
+
+func ChatCompletionsResponseToResponsesResponse(resp *dto.OpenAITextResponse, id string) (*dto.OpenAIResponsesResponse, *dto.Usage, error) {
+	if resp == nil {
+		return nil, nil, errors.New("response is nil")
+	}
+
+	usage := UsageFromChatUsage(&resp.Usage)
+	out := &dto.OpenAIResponsesResponse{
+		ID:        id,
+		Object:    "response",
+		CreatedAt: chatCreatedAt(resp.Created),
+		Status:    []byte(`"completed"`),
+		Model:     resp.Model,
+		Output:    make([]dto.ResponsesOutput, 0),
+		Usage:     usage,
+	}
+
+	if len(resp.Choices) == 0 {
+		return out, usage, nil
+	}
+
+	choice := resp.Choices[0]
+	if status, details := ResponsesStatusFromChatFinishReason(choice.FinishReason); status != "" {
+		out.Status = []byte(fmt.Sprintf("%q", status))
+		out.IncompleteDetails = details
+	}
+
+	if text := choice.Message.StringContent(); text != "" {
+		out.Output = append(out.Output, dto.ResponsesOutput{
+			Type:   responsesOutputTypeMessage,
+			ID:     fmt.Sprintf("%s_msg_0", id),
+			Status: responseOutputStatus(out),
+			Role:   "assistant",
+			Content: []dto.ResponsesOutputContent{
+				{
+					Type:        "output_text",
+					Text:        text,
+					Annotations: []interface{}{},
+				},
+			},
+		})
+	}
+	if reasoning := choice.Message.GetReasoningContent(); reasoning != "" {
+		out.Output = append(out.Output, dto.ResponsesOutput{
+			Type:   responsesOutputTypeReasoning,
+			ID:     fmt.Sprintf("%s_reasoning_0", id),
+			Status: responseOutputStatus(out),
+			Content: []dto.ResponsesOutputContent{
+				{
+					Type: "summary_text",
+					Text: reasoning,
+				},
+			},
+		})
+	}
+
+	for i, toolCall := range choice.Message.ParseToolCalls() {
+		toolOutput, err := chatToolCallToResponsesOutput(toolCall, id, i, responseOutputStatus(out))
+		if err != nil {
+			return nil, nil, err
+		}
+		out.Output = append(out.Output, toolOutput)
+	}
+
+	return out, usage, nil
+}
+
+func ResponsesStatusFromChatFinishReason(finishReason string) (string, *dto.IncompleteDetails) {
+	switch strings.TrimSpace(finishReason) {
+	case chatFinishReasonLength:
+		return "incomplete", &dto.IncompleteDetails{Reason: responsesIncompleteReasonMaxTokens}
+	case chatFinishReasonContentFilter:
+		return "incomplete", &dto.IncompleteDetails{Reason: responsesIncompleteReasonContentFilter}
+	default:
+		return "completed", nil
+	}
+}
+
+func UsageFromChatUsage(src *dto.Usage) *dto.Usage {
+	usage := &dto.Usage{}
+	if src == nil {
+		return usage
+	}
+	usage.UsageSemantic = src.UsageSemantic
+	usage.UsageSource = src.UsageSource
+	usage.BillingUsage = dto.CloneBillingUsage(src.BillingUsage)
+	if usage.BillingUsage == nil {
+		usage.BillingUsage = dto.NewOpenAIChatBillingUsage(src)
+	}
+	usage.Cost = src.Cost
+	if src.PromptTokens != 0 {
+		usage.PromptTokens = src.PromptTokens
+		usage.InputTokens = src.PromptTokens
+	}
+	if src.CompletionTokens != 0 {
+		usage.CompletionTokens = src.CompletionTokens
+		usage.OutputTokens = src.CompletionTokens
+	}
+	if src.TotalTokens != 0 {
+		usage.TotalTokens = src.TotalTokens
+	} else {
+		usage.TotalTokens = usage.InputTokens + usage.OutputTokens
+	}
+	if src.PromptTokensDetails.CachedTokens != 0 ||
+		src.PromptTokensDetails.ImageTokens != 0 ||
+		src.PromptTokensDetails.AudioTokens != 0 ||
+		src.PromptTokensDetails.CachedCreationTokens != 0 ||
+		src.PromptTokensDetails.CacheWriteTokens != 0 ||
+		src.PromptTokensDetails.TextTokens != 0 {
+		details := src.PromptTokensDetails
+		usage.InputTokensDetails = &details
+	}
+	if src.CompletionTokenDetails.ReasoningTokens != 0 ||
+		src.CompletionTokenDetails.TextTokens != 0 ||
+		src.CompletionTokenDetails.AudioTokens != 0 ||
+		src.CompletionTokenDetails.ImageTokens != 0 {
+		usage.CompletionTokenDetails = src.CompletionTokenDetails
+	}
+	usage.ClaudeCacheCreation5mTokens = src.ClaudeCacheCreation5mTokens
+	usage.ClaudeCacheCreation1hTokens = src.ClaudeCacheCreation1hTokens
+	return usage
+}
+
+func responseOutputStatus(resp *dto.OpenAIResponsesResponse) string {
+	if resp == nil || responseStatusString(resp) != "incomplete" {
+		return "completed"
+	}
+	return "incomplete"
+}
+
+func responseStatusString(resp *dto.OpenAIResponsesResponse) string {
+	if resp == nil || len(resp.Status) == 0 {
+		return ""
+	}
+	var status string
+	_ = kitutil.Unmarshal(resp.Status, &status)
+	return strings.TrimSpace(status)
+}
+
+func chatToolCallToResponsesOutput(toolCall dto.ToolCallRequest, responseID string, index int, status string) (dto.ResponsesOutput, error) {
+	callID := strings.TrimSpace(toolCall.ID)
+	if callID == "" {
+		callID = fmt.Sprintf("%s_call_%d", responseID, index)
+	}
+	if toolCall.Type == "" || toolCall.Type == "function" {
+		return dto.ResponsesOutput{
+			Type:      responsesOutputTypeFunctionCall,
+			ID:        callID,
+			Status:    status,
+			CallId:    callID,
+			Name:      toolCall.Function.Name,
+			Arguments: chatArgumentsRawMessage(toolCall.Function.Arguments),
+		}, nil
+	}
+	return dto.ResponsesOutput{
+		Type:      toolCall.Type,
+		ID:        callID,
+		Status:    status,
+		CallId:    callID,
+		Arguments: toolCall.Custom,
+	}, nil
+}
+
+func chatArgumentsRawMessage(arguments string) []byte {
+	raw, err := kitutil.Marshal(arguments)
+	if err != nil {
+		return []byte(`""`)
+	}
+	return raw
+}
+
+func chatCreatedAt(created any) int {
+	switch v := created.(type) {
+	case int:
+		return v
+	case int64:
+		return int(v)
+	case float64:
+		return int(v)
+	case float32:
+		return int(v)
+	case string:
+		if parsed := kitutil.String2Int(v); parsed != 0 {
+			return parsed
+		}
+	}
+	return int(time.Now().Unix())
+}
+
+func responsesStreamEvent(eventType string, payload dto.ResponsesStreamResponse) ChatToResponsesStreamEvent {
+	payload.Type = eventType
+	return ChatToResponsesStreamEvent{
+		Type:    eventType,
+		Payload: payload,
+	}
+}
+
+func intPtr(v int) *int {
+	return &v
+}

relaykit/relayconvert/internal/oai_chat/to_oai_responses_resp_test.go

@@ -0,0 +1,140 @@
+package oaichat
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/samber/lo"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestChatCompletionsResponseToResponsesPreservesTextToolCallsAndUsage(t *testing.T) {
+	chat := &dto.OpenAITextResponse{
+		Id:      "chatcmpl_1",
+		Model:   "gpt-test",
+		Created: 456,
+		Choices: []dto.OpenAITextResponseChoice{
+			{
+				Message:      assistantMessageWithTool("I will call.", "call_1", "lookup", `{"q":"x"}`),
+				FinishReason: "tool_calls",
+			},
+		},
+		Usage: dto.Usage{PromptTokens: 3, CompletionTokens: 5, TotalTokens: 8},
+	}
+
+	resp, usage, err := ChatCompletionsResponseToResponsesResponse(chat, "resp_1")
+	require.NoError(t, err)
+	require.NotNil(t, usage)
+
+	assert.Equal(t, "resp_1", resp.ID)
+	assert.Equal(t, "response", resp.Object)
+	assert.Equal(t, `"completed"`, string(resp.Status))
+	assert.Equal(t, 3, resp.Usage.InputTokens)
+	assert.Equal(t, 5, resp.Usage.OutputTokens)
+	require.Len(t, resp.Output, 2)
+	assert.Equal(t, responsesOutputTypeMessage, resp.Output[0].Type)
+	assert.Equal(t, "I will call.", resp.Output[0].Content[0].Text)
+	assert.Equal(t, responsesOutputTypeFunctionCall, resp.Output[1].Type)
+	assert.Equal(t, "call_1", resp.Output[1].CallId)
+	assert.Equal(t, "lookup", resp.Output[1].Name)
+	assert.Equal(t, `"{\"q\":\"x\"}"`, string(resp.Output[1].Arguments))
+}
+
+func TestChatCompletionsResponseToResponsesMapsIncompleteFinishReasons(t *testing.T) {
+	tests := []struct {
+		name         string
+		finishReason string
+		wantReason   string
+	}{
+		{name: "length", finishReason: "length", wantReason: responsesIncompleteReasonMaxTokens},
+		{name: "content filter", finishReason: "content_filter", wantReason: responsesIncompleteReasonContentFilter},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			resp, _, err := ChatCompletionsResponseToResponsesResponse(&dto.OpenAITextResponse{
+				Id:    "chatcmpl_1",
+				Model: "gpt-test",
+				Choices: []dto.OpenAITextResponseChoice{
+					{
+						Message:      dto.Message{Role: "assistant", Content: "partial"},
+						FinishReason: tt.finishReason,
+					},
+				},
+			}, "resp_1")
+			require.NoError(t, err)
+
+			assert.Equal(t, `"incomplete"`, string(resp.Status))
+			require.NotNil(t, resp.IncompleteDetails)
+			assert.Equal(t, tt.wantReason, resp.IncompleteDetails.Reason)
+			require.Len(t, resp.Output, 1)
+			assert.Equal(t, "incomplete", resp.Output[0].Status)
+		})
+	}
+}
+
+func TestChatCompletionsStreamToResponsesEventsAggregatesUsageAndToolArgs(t *testing.T) {
+	state := NewChatToResponsesStreamState("resp_1", "gpt-test")
+	state.Created = 123
+	toolIndex := 0
+
+	var events []ChatToResponsesStreamEvent
+	events = append(events, mustResponsesEventsFromChatChunk(t, state, &dto.ChatCompletionsStreamResponse{
+		Id:      "chatcmpl_1",
+		Model:   "gpt-test",
+		Created: 123,
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{Index: 0, Delta: dto.ChatCompletionsStreamResponseChoiceDelta{Role: "assistant"}},
+		},
+	})...)
+	events = append(events, mustResponsesEventsFromChatChunk(t, state, &dto.ChatCompletionsStreamResponse{
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{Index: 0, Delta: dto.ChatCompletionsStreamResponseChoiceDelta{Content: lo.ToPtr("hello")}},
+		},
+	})...)
+	events = append(events, mustResponsesEventsFromChatChunk(t, state, &dto.ChatCompletionsStreamResponse{
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{Index: 0, Delta: dto.ChatCompletionsStreamResponseChoiceDelta{ToolCalls: []dto.ToolCallResponse{
+				{Index: &toolIndex, ID: "call_1", Type: "function", Function: dto.FunctionResponse{Name: "lookup"}},
+			}}},
+		},
+	})...)
+	events = append(events, mustResponsesEventsFromChatChunk(t, state, &dto.ChatCompletionsStreamResponse{
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{Index: 0, Delta: dto.ChatCompletionsStreamResponseChoiceDelta{ToolCalls: []dto.ToolCallResponse{
+				{Index: &toolIndex, Function: dto.FunctionResponse{Arguments: `{"q":"x"}`}},
+			}}},
+		},
+	})...)
+	finishReason := "tool_calls"
+	events = append(events, mustResponsesEventsFromChatChunk(t, state, &dto.ChatCompletionsStreamResponse{
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{Index: 0, FinishReason: &finishReason},
+		},
+	})...)
+	events = append(events, mustResponsesEventsFromChatChunk(t, state, &dto.ChatCompletionsStreamResponse{
+		Usage: &dto.Usage{PromptTokens: 2, CompletionTokens: 4, TotalTokens: 6},
+	})...)
+	events = append(events, FinalizeChatCompletionsStreamToResponses(state)...)
+
+	require.Len(t, events, 10)
+	assert.Equal(t, responsesEventCreated, events[0].Type)
+	assert.Equal(t, responsesEventOutputTextDelta, events[2].Type)
+	assert.Equal(t, "hello", events[2].Payload.Delta)
+	assert.Equal(t, responsesEventFunctionArgsDelta, events[4].Type)
+	assert.Equal(t, `{"q":"x"}`, events[4].Payload.Delta)
+	assert.Equal(t, responsesEventCompleted, events[9].Type)
+	require.NotNil(t, events[9].Payload.Response)
+	assert.Equal(t, 6, events[9].Payload.Response.Usage.TotalTokens)
+	require.Len(t, events[9].Payload.Response.Output, 2)
+	assert.Equal(t, "hello", events[9].Payload.Response.Output[0].Content[0].Text)
+	assert.Equal(t, `"{\"q\":\"x\"}"`, string(events[9].Payload.Response.Output[1].Arguments))
+}
+
+func mustResponsesEventsFromChatChunk(t *testing.T, state *ChatToResponsesStreamState, chunk *dto.ChatCompletionsStreamResponse) []ChatToResponsesStreamEvent {
+	t.Helper()
+	events, err := ChatCompletionsStreamChunkToResponsesEvents(chunk, state)
+	require.NoError(t, err)
+	return events
+}

relaykit/relayconvert/internal/oai_chat/to_oai_responses_stream_resp.go

@@ -0,0 +1,417 @@
+package oaichat
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+)
+
+type ChatToResponsesStreamEvent struct {
+	Type    string
+	Payload dto.ResponsesStreamResponse
+}
+
+type ChatToResponsesStreamState struct {
+	ID      string
+	Model   string
+	Created int64
+	Usage   *dto.Usage
+
+	status            string
+	incompleteDetails *dto.IncompleteDetails
+	sentCreated       bool
+	textOutputIndex   int
+	textStarted       bool
+	textDone          bool
+	reasoningIndex    int
+	reasoningStarted  bool
+	reasoningDone     bool
+	finalized         bool
+	nextOutputIndex   int
+	toolsByIndex      map[int]*chatToResponsesStreamTool
+	outputOrder       []chatToResponsesOutputRef
+	text              strings.Builder
+	reasoning         strings.Builder
+}
+
+type chatToResponsesStreamTool struct {
+	ChatIndex   int
+	OutputIndex int
+	ID          string
+	Name        string
+	Arguments   strings.Builder
+	Done        bool
+}
+
+type chatToResponsesOutputRef struct {
+	Kind      string
+	ToolIndex int
+}
+
+func NewChatToResponsesStreamState(id string, model string) *ChatToResponsesStreamState {
+	return &ChatToResponsesStreamState{
+		ID:              id,
+		Model:           model,
+		Created:         time.Now().Unix(),
+		Usage:           &dto.Usage{},
+		status:          "completed",
+		textOutputIndex: -1,
+		reasoningIndex:  -1,
+		toolsByIndex:    make(map[int]*chatToResponsesStreamTool),
+	}
+}
+
+func ChatCompletionsStreamChunkToResponsesEvents(chunk *dto.ChatCompletionsStreamResponse, state *ChatToResponsesStreamState) ([]ChatToResponsesStreamEvent, error) {
+	if chunk == nil || state == nil {
+		return nil, nil
+	}
+	if state.ID == "" {
+		state.ID = chunk.Id
+	}
+	if state.Model == "" {
+		state.Model = chunk.Model
+	}
+	if state.Created == 0 {
+		state.Created = chunk.Created
+	}
+	if chunk.Usage != nil {
+		state.Usage = UsageFromChatUsage(chunk.Usage)
+	}
+
+	events := make([]ChatToResponsesStreamEvent, 0)
+	if !state.sentCreated {
+		state.sentCreated = true
+		events = append(events, responsesStreamEvent(responsesEventCreated, dto.ResponsesStreamResponse{
+			Type:     responsesEventCreated,
+			Response: state.createdResponse(),
+		}))
+	}
+	for _, choice := range chunk.Choices {
+		if choice.Delta.GetReasoningContent() != "" {
+			events = append(events, state.appendReasoningDelta(choice.Delta.GetReasoningContent())...)
+		}
+		if choice.Delta.GetContentString() != "" {
+			events = append(events, state.appendTextDelta(choice.Delta.GetContentString())...)
+		}
+		for _, toolCall := range choice.Delta.ToolCalls {
+			toolEvents, err := state.appendToolCallDelta(toolCall)
+			if err != nil {
+				return nil, err
+			}
+			events = append(events, toolEvents...)
+		}
+		if choice.FinishReason != nil && strings.TrimSpace(*choice.FinishReason) != "" {
+			state.applyFinishReason(*choice.FinishReason)
+			events = append(events, state.doneDeltaEvents()...)
+		}
+	}
+	return events, nil
+}
+
+func FinalizeChatCompletionsStreamToResponses(state *ChatToResponsesStreamState) []ChatToResponsesStreamEvent {
+	if state == nil || state.finalized {
+		return nil
+	}
+	events := state.doneDeltaEvents()
+	state.finalized = true
+	resp := state.finalResponse()
+	eventType := responsesEventCompleted
+	if state.status == "incomplete" {
+		eventType = responsesEventIncomplete
+	}
+	events = append(events, responsesStreamEvent(eventType, dto.ResponsesStreamResponse{
+		Type:     eventType,
+		Response: resp,
+	}))
+	return events
+}
+
+func (s *ChatToResponsesStreamState) UsageText() string {
+	if s == nil {
+		return ""
+	}
+	return s.text.String()
+}
+
+func (s *ChatToResponsesStreamState) appendTextDelta(delta string) []ChatToResponsesStreamEvent {
+	events := make([]ChatToResponsesStreamEvent, 0, 2)
+	if !s.textStarted {
+		s.textStarted = true
+		s.textOutputIndex = s.nextIndex("message", -1)
+		events = append(events, responsesStreamEvent(responsesEventOutputItemAdded, dto.ResponsesStreamResponse{
+			Type:        responsesEventOutputItemAdded,
+			OutputIndex: intPtr(s.textOutputIndex),
+			Item: &dto.ResponsesOutput{
+				Type:    responsesOutputTypeMessage,
+				ID:      s.messageID(),
+				Status:  "in_progress",
+				Role:    "assistant",
+				Content: []dto.ResponsesOutputContent{},
+			},
+		}))
+	}
+	s.text.WriteString(delta)
+	events = append(events, responsesStreamEvent(responsesEventOutputTextDelta, dto.ResponsesStreamResponse{
+		Type:         responsesEventOutputTextDelta,
+		OutputIndex:  intPtr(s.textOutputIndex),
+		ContentIndex: intPtr(0),
+		Delta:        delta,
+		ItemID:       s.messageID(),
+	}))
+	return events
+}
+
+func (s *ChatToResponsesStreamState) appendReasoningDelta(delta string) []ChatToResponsesStreamEvent {
+	events := make([]ChatToResponsesStreamEvent, 0, 2)
+	if !s.reasoningStarted {
+		s.reasoningStarted = true
+		s.reasoningIndex = s.nextIndex("reasoning", -1)
+		events = append(events, responsesStreamEvent(responsesEventOutputItemAdded, dto.ResponsesStreamResponse{
+			Type:        responsesEventOutputItemAdded,
+			OutputIndex: intPtr(s.reasoningIndex),
+			Item: &dto.ResponsesOutput{
+				Type:    responsesOutputTypeReasoning,
+				ID:      s.reasoningID(),
+				Status:  "in_progress",
+				Content: []dto.ResponsesOutputContent{},
+			},
+		}))
+	}
+	s.reasoning.WriteString(delta)
+	events = append(events, responsesStreamEvent(responsesEventReasoningSummaryDelta, dto.ResponsesStreamResponse{
+		Type:         responsesEventReasoningSummaryDelta,
+		OutputIndex:  intPtr(s.reasoningIndex),
+		SummaryIndex: intPtr(0),
+		Delta:        delta,
+		ItemID:       s.reasoningID(),
+	}))
+	return events
+}
+
+func (s *ChatToResponsesStreamState) appendToolCallDelta(toolCall dto.ToolCallResponse) ([]ChatToResponsesStreamEvent, error) {
+	chatIndex := 0
+	if toolCall.Index != nil {
+		chatIndex = *toolCall.Index
+	}
+	tool := s.toolsByIndex[chatIndex]
+	events := make([]ChatToResponsesStreamEvent, 0, 2)
+	if tool == nil {
+		tool = &chatToResponsesStreamTool{
+			ChatIndex:   chatIndex,
+			OutputIndex: s.nextIndex("tool", chatIndex),
+			ID:          strings.TrimSpace(toolCall.ID),
+			Name:        strings.TrimSpace(toolCall.Function.Name),
+		}
+		if tool.ID == "" {
+			tool.ID = fmt.Sprintf("%s_call_%d", s.ID, chatIndex)
+		}
+		s.toolsByIndex[chatIndex] = tool
+		events = append(events, responsesStreamEvent(responsesEventOutputItemAdded, dto.ResponsesStreamResponse{
+			Type:        responsesEventOutputItemAdded,
+			OutputIndex: intPtr(tool.OutputIndex),
+			ItemID:      tool.ID,
+			Item: &dto.ResponsesOutput{
+				Type:      responsesOutputTypeFunctionCall,
+				ID:        tool.ID,
+				Status:    "in_progress",
+				CallId:    tool.ID,
+				Name:      tool.Name,
+				Arguments: []byte(`""`),
+			},
+		}))
+	}
+	if strings.TrimSpace(toolCall.ID) != "" {
+		tool.ID = strings.TrimSpace(toolCall.ID)
+	}
+	if strings.TrimSpace(toolCall.Function.Name) != "" {
+		tool.Name = strings.TrimSpace(toolCall.Function.Name)
+	}
+	if toolCall.Function.Arguments != "" {
+		tool.Arguments.WriteString(toolCall.Function.Arguments)
+		events = append(events, responsesStreamEvent(responsesEventFunctionArgsDelta, dto.ResponsesStreamResponse{
+			Type:        responsesEventFunctionArgsDelta,
+			OutputIndex: intPtr(tool.OutputIndex),
+			ItemID:      tool.ID,
+			Delta:       toolCall.Function.Arguments,
+		}))
+	}
+	return events, nil
+}
+
+func (s *ChatToResponsesStreamState) doneDeltaEvents() []ChatToResponsesStreamEvent {
+	events := make([]ChatToResponsesStreamEvent, 0)
+	status := s.outputStatus()
+	if s.textStarted && !s.textDone {
+		s.textDone = true
+		events = append(events, responsesStreamEvent("response.output_text.done", dto.ResponsesStreamResponse{
+			Type:         "response.output_text.done",
+			OutputIndex:  intPtr(s.textOutputIndex),
+			ContentIndex: intPtr(0),
+			ItemID:       s.messageID(),
+		}))
+		events = append(events, responsesStreamEvent(responsesEventOutputItemDone, dto.ResponsesStreamResponse{
+			Type:        responsesEventOutputItemDone,
+			OutputIndex: intPtr(s.textOutputIndex),
+			Item:        s.messageOutput(status),
+		}))
+	}
+	if s.reasoningStarted && !s.reasoningDone {
+		s.reasoningDone = true
+		events = append(events, responsesStreamEvent(responsesEventReasoningSummaryDone, dto.ResponsesStreamResponse{
+			Type:         responsesEventReasoningSummaryDone,
+			OutputIndex:  intPtr(s.reasoningIndex),
+			SummaryIndex: intPtr(0),
+			ItemID:       s.reasoningID(),
+			Part: &dto.ResponsesReasoningSummaryPart{
+				Type: "summary_text",
+				Text: s.reasoning.String(),
+			},
+		}))
+		events = append(events, responsesStreamEvent(responsesEventOutputItemDone, dto.ResponsesStreamResponse{
+			Type:        responsesEventOutputItemDone,
+			OutputIndex: intPtr(s.reasoningIndex),
+			Item:        s.reasoningOutput(status),
+		}))
+	}
+	for _, tool := range s.sortedTools() {
+		if tool.Done {
+			continue
+		}
+		tool.Done = true
+		events = append(events, responsesStreamEvent(responsesEventFunctionArgsDone, dto.ResponsesStreamResponse{
+			Type:        responsesEventFunctionArgsDone,
+			OutputIndex: intPtr(tool.OutputIndex),
+			ItemID:      tool.ID,
+		}))
+		events = append(events, responsesStreamEvent(responsesEventOutputItemDone, dto.ResponsesStreamResponse{
+			Type:        responsesEventOutputItemDone,
+			OutputIndex: intPtr(tool.OutputIndex),
+			Item:        s.toolOutput(tool, status),
+		}))
+	}
+	return events
+}
+
+func (s *ChatToResponsesStreamState) applyFinishReason(finishReason string) {
+	if status, details := ResponsesStatusFromChatFinishReason(finishReason); status != "" {
+		s.status = status
+		s.incompleteDetails = details
+	}
+}
+
+func (s *ChatToResponsesStreamState) finalResponse() *dto.OpenAIResponsesResponse {
+	output := make([]dto.ResponsesOutput, 0, len(s.outputOrder))
+	status := s.outputStatus()
+	for _, ref := range s.outputOrder {
+		switch ref.Kind {
+		case "message":
+			output = append(output, *s.messageOutput(status))
+		case "reasoning":
+			output = append(output, *s.reasoningOutput(status))
+		case "tool":
+			if tool := s.toolsByIndex[ref.ToolIndex]; tool != nil {
+				output = append(output, *s.toolOutput(tool, status))
+			}
+		}
+	}
+	return &dto.OpenAIResponsesResponse{
+		ID:                s.ID,
+		Object:            "response",
+		CreatedAt:         int(s.Created),
+		Status:            []byte(fmt.Sprintf("%q", s.status)),
+		IncompleteDetails: s.incompleteDetails,
+		Model:             s.Model,
+		Output:            output,
+		Usage:             s.Usage,
+	}
+}
+
+func (s *ChatToResponsesStreamState) createdResponse() *dto.OpenAIResponsesResponse {
+	return &dto.OpenAIResponsesResponse{
+		ID:        s.ID,
+		Object:    "response",
+		CreatedAt: int(s.Created),
+		Status:    []byte(`"in_progress"`),
+		Model:     s.Model,
+		Output:    []dto.ResponsesOutput{},
+	}
+}
+
+func (s *ChatToResponsesStreamState) nextIndex(kind string, toolIndex int) int {
+	index := s.nextOutputIndex
+	s.nextOutputIndex++
+	s.outputOrder = append(s.outputOrder, chatToResponsesOutputRef{Kind: kind, ToolIndex: toolIndex})
+	return index
+}
+
+func (s *ChatToResponsesStreamState) sortedTools() []*chatToResponsesStreamTool {
+	indexes := make([]int, 0, len(s.toolsByIndex))
+	for index := range s.toolsByIndex {
+		indexes = append(indexes, index)
+	}
+	sort.Ints(indexes)
+	tools := make([]*chatToResponsesStreamTool, 0, len(indexes))
+	for _, index := range indexes {
+		tools = append(tools, s.toolsByIndex[index])
+	}
+	return tools
+}
+
+func (s *ChatToResponsesStreamState) outputStatus() string {
+	if s.status == "incomplete" {
+		return "incomplete"
+	}
+	return "completed"
+}
+
+func (s *ChatToResponsesStreamState) messageID() string {
+	return fmt.Sprintf("%s_msg_0", s.ID)
+}
+
+func (s *ChatToResponsesStreamState) reasoningID() string {
+	return fmt.Sprintf("%s_reasoning_0", s.ID)
+}
+
+func (s *ChatToResponsesStreamState) messageOutput(status string) *dto.ResponsesOutput {
+	return &dto.ResponsesOutput{
+		Type:   responsesOutputTypeMessage,
+		ID:     s.messageID(),
+		Status: status,
+		Role:   "assistant",
+		Content: []dto.ResponsesOutputContent{
+			{
+				Type:        "output_text",
+				Text:        s.text.String(),
+				Annotations: []interface{}{},
+			},
+		},
+	}
+}
+
+func (s *ChatToResponsesStreamState) reasoningOutput(status string) *dto.ResponsesOutput {
+	return &dto.ResponsesOutput{
+		Type:   responsesOutputTypeReasoning,
+		ID:     s.reasoningID(),
+		Status: status,
+		Content: []dto.ResponsesOutputContent{
+			{
+				Type: "summary_text",
+				Text: s.reasoning.String(),
+			},
+		},
+	}
+}
+
+func (s *ChatToResponsesStreamState) toolOutput(tool *chatToResponsesStreamTool, status string) *dto.ResponsesOutput {
+	return &dto.ResponsesOutput{
+		Type:      responsesOutputTypeFunctionCall,
+		ID:        tool.ID,
+		Status:    status,
+		CallId:    tool.ID,
+		Name:      tool.Name,
+		Arguments: chatArgumentsRawMessage(tool.Arguments.String()),
+	}
+}

relaykit/relayconvert/internal/oai_responses/req_helpers.go

@@ -0,0 +1,263 @@
+package oairesponses
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/QuantumNous/new-api/relaykit/types"
+)
+
+func openAIResponsesRequestFromAny(request any) (*dto.OpenAIResponsesRequest, error) {
+	responsesRequest, ok := request.(*dto.OpenAIResponsesRequest)
+	if !ok {
+		if value, ok := request.(dto.OpenAIResponsesRequest); ok {
+			responsesRequest = &value
+		}
+	}
+	if responsesRequest == nil {
+		return nil, fmt.Errorf("expected OpenAI responses request, got %T", request)
+	}
+	return responsesRequest, nil
+}
+
+func OpenAIResponsesRequestFromAny(request any) (*dto.OpenAIResponsesRequest, error) {
+	return openAIResponsesRequestFromAny(request)
+}
+
+func responsesInputItems(raw []byte) ([]map[string]any, error) {
+	if !rawJSONPresent(raw) {
+		return nil, nil
+	}
+
+	switch kitutil.GetJsonType(raw) {
+	case "string":
+		input, err := responsesJSONString(raw)
+		if err != nil {
+			return nil, fmt.Errorf("invalid input string: %w", err)
+		}
+		return []map[string]any{
+			{
+				"role":    "user",
+				"content": input,
+			},
+		}, nil
+	case "array":
+		var items []map[string]any
+		if err := kitutil.Unmarshal(raw, &items); err != nil {
+			return nil, fmt.Errorf("invalid input array: %w", err)
+		}
+		return items, nil
+	default:
+		return nil, fmt.Errorf("unsupported responses input type %q", kitutil.GetJsonType(raw))
+	}
+}
+
+func InputItems(raw []byte) ([]map[string]any, error) {
+	return responsesInputItems(raw)
+}
+
+func responsesContentParts(content any) ([]map[string]any, error) {
+	switch typed := content.(type) {
+	case nil:
+		return nil, nil
+	case string:
+		return []map[string]any{{"type": "input_text", "text": typed}}, nil
+	case []map[string]any:
+		return typed, nil
+	case []any:
+		parts := make([]map[string]any, 0, len(typed))
+		for _, item := range typed {
+			switch part := item.(type) {
+			case string:
+				parts = append(parts, map[string]any{"type": "input_text", "text": part})
+			case map[string]any:
+				parts = append(parts, part)
+			default:
+				raw, err := kitutil.Marshal(part)
+				if err != nil {
+					return nil, err
+				}
+				parts = append(parts, map[string]any{"type": "input_text", "text": string(raw)})
+			}
+		}
+		return parts, nil
+	default:
+		raw, err := kitutil.Marshal(typed)
+		if err != nil {
+			return nil, err
+		}
+		return []map[string]any{{"type": "input_text", "text": string(raw)}}, nil
+	}
+}
+
+func ContentParts(content any) ([]map[string]any, error) {
+	return responsesContentParts(content)
+}
+
+func responsesRequestFunctionDeclarations(raw []byte) ([]dto.FunctionRequest, error) {
+	if !rawJSONPresent(raw) {
+		return nil, nil
+	}
+
+	var tools []map[string]any
+	if err := kitutil.Unmarshal(raw, &tools); err != nil {
+		return nil, fmt.Errorf("invalid tools: %w", err)
+	}
+
+	functions := make([]dto.FunctionRequest, 0, len(tools))
+	for _, tool := range tools {
+		if strings.TrimSpace(kitutil.Interface2String(tool["type"])) != "function" {
+			continue
+		}
+		name := strings.TrimSpace(kitutil.Interface2String(tool["name"]))
+		if name == "" {
+			continue
+		}
+		functions = append(functions, dto.FunctionRequest{
+			Name:        name,
+			Description: kitutil.Interface2String(tool["description"]),
+			Parameters:  tool["parameters"],
+		})
+	}
+	return functions, nil
+}
+
+func RequestFunctionDeclarations(raw []byte) ([]dto.FunctionRequest, error) {
+	return responsesRequestFunctionDeclarations(raw)
+}
+
+func responsesReasoningEffort(req *dto.OpenAIResponsesRequest) string {
+	if req == nil || req.Reasoning == nil {
+		return ""
+	}
+	return req.Reasoning.Effort
+}
+
+func ReasoningEffort(req *dto.OpenAIResponsesRequest) string {
+	return responsesReasoningEffort(req)
+}
+
+func responsesObjectValue(value any, fallbackKey string) map[string]any {
+	switch typed := value.(type) {
+	case nil:
+		return map[string]any{}
+	case map[string]any:
+		return typed
+	case string:
+		var object map[string]any
+		if err := kitutil.Unmarshal([]byte(typed), &object); err == nil {
+			return object
+		}
+		var array []any
+		if err := kitutil.Unmarshal([]byte(typed), &array); err == nil {
+			return map[string]any{fallbackKey: array}
+		}
+		return map[string]any{fallbackKey: typed}
+	case []any:
+		return map[string]any{fallbackKey: typed}
+	default:
+		return map[string]any{fallbackKey: typed}
+	}
+}
+
+func ObjectValue(value any, fallbackKey string) map[string]any {
+	return responsesObjectValue(value, fallbackKey)
+}
+
+func responsesGeminiResponseMap(value any) map[string]interface{} {
+	switch typed := value.(type) {
+	case nil:
+		return map[string]interface{}{}
+	case map[string]any:
+		return typed
+	case string:
+		var object map[string]interface{}
+		if err := kitutil.Unmarshal([]byte(typed), &object); err == nil {
+			return object
+		}
+		var array []interface{}
+		if err := kitutil.Unmarshal([]byte(typed), &array); err == nil {
+			return map[string]interface{}{"result": array}
+		}
+		return map[string]interface{}{"content": typed}
+	case []any:
+		return map[string]interface{}{"result": typed}
+	default:
+		return map[string]interface{}{"content": typed}
+	}
+}
+
+func GeminiResponseMap(value any) map[string]interface{} {
+	return responsesGeminiResponseMap(value)
+}
+
+func responsesParallelToolCalls(raw []byte) *bool {
+	if !rawJSONPresent(raw) || kitutil.GetJsonType(raw) != "boolean" {
+		return nil
+	}
+	var parallelToolCalls bool
+	if err := kitutil.Unmarshal(raw, &parallelToolCalls); err != nil {
+		return nil
+	}
+	return &parallelToolCalls
+}
+
+func ParallelToolCalls(raw []byte) *bool {
+	return responsesParallelToolCalls(raw)
+}
+
+func ContentPartToFileSource(part map[string]any) types.FileSource {
+	partType := strings.TrimSpace(kitutil.Interface2String(part["type"]))
+	var data string
+	var mimeType string
+
+	switch partType {
+	case "input_image":
+		data, mimeType = responsesPartDataAndMime(part, "image_url", "url")
+	case "input_file":
+		data, mimeType = responsesPartDataAndMime(part, "file", "file_data", "file_url", "url")
+	case "input_audio":
+		data, mimeType = responsesPartDataAndMime(part, "input_audio", "data", "url")
+		if mimeType == "" {
+			if payload, ok := part["input_audio"].(map[string]any); ok {
+				if format := strings.TrimSpace(kitutil.Interface2String(payload["format"])); format != "" {
+					mimeType = "audio/" + format
+				}
+			}
+		}
+	case "input_video":
+		data, mimeType = responsesPartDataAndMime(part, "video_url", "url")
+	}
+	if data == "" {
+		return nil
+	}
+	return types.NewFileSourceFromData(data, mimeType)
+}
+
+func responsesPartDataAndMime(part map[string]any, keys ...string) (string, string) {
+	mimeType := strings.TrimSpace(kitutil.Interface2String(part["mime_type"]))
+	for _, key := range keys {
+		value, ok := part[key]
+		if !ok {
+			continue
+		}
+		switch typed := value.(type) {
+		case string:
+			if typed != "" {
+				return typed, mimeType
+			}
+		case map[string]any:
+			if mimeType == "" {
+				mimeType = strings.TrimSpace(kitutil.Interface2String(typed["mime_type"]))
+			}
+			for _, nestedKey := range []string{"url", "file_data", "file_url", "data"} {
+				if data := strings.TrimSpace(kitutil.Interface2String(typed[nestedKey])); data != "" {
+					return data, mimeType
+				}
+			}
+		}
+	}
+	return "", mimeType
+}

relaykit/relayconvert/internal/oai_responses/to_claude_messages_req.go

@@ -0,0 +1,329 @@
+package oairesponses
+
+import (
+	"fmt"
+	"strings"
+
+	"context"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	relaymedia "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/media"
+	sharedclaude "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/shared/claude"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+func convertOpenAIResponsesRequestToClaudeMessages(c context.Context, info convmeta.Meta, request any) (any, error) {
+	responsesRequest, err := OpenAIResponsesRequestFromAny(request)
+	if err != nil {
+		return nil, err
+	}
+	return OpenAIResponsesRequestToClaudeMessages(c, info, responsesRequest)
+}
+
+func OpenAIResponsesRequestToClaudeMessages(c context.Context, info convmeta.Meta, req *dto.OpenAIResponsesRequest) (*dto.ClaudeRequest, error) {
+	if req == nil {
+		return nil, fmt.Errorf("request is nil")
+	}
+	if req.Model == "" {
+		return nil, fmt.Errorf("model is required")
+	}
+	if err := ValidateRequestChatUnsupportedFields(req); err != nil {
+		return nil, err
+	}
+
+	claudeRequest := &dto.ClaudeRequest{
+		Model:       req.Model,
+		Temperature: req.Temperature,
+		TopP:        req.TopP,
+		Stream:      req.Stream,
+	}
+	if req.MaxOutputTokens != nil && *req.MaxOutputTokens > 0 {
+		claudeRequest.MaxTokens = kitutil.GetPointer(*req.MaxOutputTokens)
+	}
+	if claudeRequest.MaxTokens == nil || *claudeRequest.MaxTokens == 0 {
+		if defaultMaxTokens, configured := convmeta.OptionsOf(info).Claude.DefaultMaxTokensFor(req.Model); configured {
+			value := uint(defaultMaxTokens)
+			claudeRequest.MaxTokens = &value
+		}
+	}
+
+	functions, err := RequestFunctionDeclarations(req.Tools)
+	if err != nil {
+		return nil, err
+	}
+	if len(functions) > 0 {
+		claudeRequest.Tools = responsesFunctionDeclarationsToClaudeTools(functions)
+	}
+
+	toolChoice, err := RequestToolChoiceToChat(req.ToolChoice)
+	if err != nil {
+		return nil, err
+	}
+	if toolChoice != nil || RawJSONPresent(req.ParallelToolCalls) {
+		claudeRequest.ToolChoice = sharedclaude.MapOpenAIToolChoice(toolChoice, ParallelToolCalls(req.ParallelToolCalls))
+	}
+	applyResponsesReasoningToClaude(req, claudeRequest)
+
+	systemMessages := make([]dto.ClaudeMediaMessage, 0)
+	if RawJSONPresent(req.Instructions) {
+		instructions, err := JSONString(req.Instructions)
+		if err != nil {
+			return nil, fmt.Errorf("invalid instructions: %w", err)
+		}
+		if strings.TrimSpace(instructions) != "" {
+			systemMessages = append(systemMessages, dto.ClaudeMediaMessage{
+				Type: "text",
+				Text: kitutil.GetPointer(instructions),
+			})
+		}
+	}
+
+	inputItems, err := InputItems(req.Input)
+	if err != nil {
+		return nil, err
+	}
+	for _, item := range inputItems {
+		itemType := strings.TrimSpace(kitutil.Interface2String(item["type"]))
+		switch itemType {
+		case ResponsesInputTypeFunctionCall:
+			claudeRequest.Messages = appendClaudeToolUse(claudeRequest.Messages, responsesFunctionCallItemToClaudeToolUse(item, "arguments"))
+		case ResponsesInputTypeCustomToolCall:
+			claudeRequest.Messages = appendClaudeToolUse(claudeRequest.Messages, responsesFunctionCallItemToClaudeToolUse(item, "input"))
+		case ResponsesInputTypeFunctionCallOutput, ResponsesInputTypeCustomToolOutput:
+			claudeRequest.Messages = appendClaudeToolResult(claudeRequest.Messages, responsesFunctionOutputItemToClaudeToolResult(item))
+		default:
+			role := responsesClaudeRole(item)
+			parts, err := responsesInputContentToClaudeMediaMessages(c, item["content"])
+			if err != nil {
+				return nil, err
+			}
+			if role == "system" {
+				systemMessages = append(systemMessages, parts...)
+				continue
+			}
+			if len(parts) == 0 {
+				parts = []dto.ClaudeMediaMessage{
+					{
+						Type: "text",
+						Text: kitutil.GetPointer("..."),
+					},
+				}
+			}
+			claudeRequest.Messages = append(claudeRequest.Messages, dto.ClaudeMessage{
+				Role:    role,
+				Content: parts,
+			})
+		}
+	}
+
+	if len(systemMessages) > 0 {
+		claudeRequest.System = systemMessages
+	}
+	claudeRequest.Messages = ensureClaudeMessagesStartWithUser(claudeRequest.Messages)
+	// Checked last so every injection path has had its chance to satisfy the
+	// required field.
+	if claudeRequest.MaxTokens == nil {
+		return nil, sharedclaude.ErrMissingMaxTokens
+	}
+	return claudeRequest, nil
+}
+
+func responsesFunctionDeclarationsToClaudeTools(functions []dto.FunctionRequest) []any {
+	tools := make([]any, 0, len(functions))
+	for _, function := range functions {
+		tools = append(tools, &dto.Tool{
+			Name:        function.Name,
+			Description: function.Description,
+			InputSchema: responsesFunctionParametersToClaudeInputSchema(function.Parameters),
+		})
+	}
+	return tools
+}
+
+func responsesFunctionParametersToClaudeInputSchema(parameters any) map[string]interface{} {
+	if params, ok := parameters.(map[string]any); ok {
+		schema := make(map[string]interface{}, len(params))
+		for key, value := range params {
+			schema[key] = value
+		}
+		if schema["type"] == nil {
+			schema["type"] = "object"
+		}
+		if schema["properties"] == nil {
+			schema["properties"] = map[string]interface{}{}
+		}
+		return schema
+	}
+	return map[string]interface{}{
+		"type":       "object",
+		"properties": map[string]interface{}{},
+	}
+}
+
+func applyResponsesReasoningToClaude(req *dto.OpenAIResponsesRequest, claudeRequest *dto.ClaudeRequest) {
+	effort := ReasoningEffort(req)
+	switch effort {
+	case "low":
+		claudeRequest.Thinking = &dto.Thinking{
+			Type:         "enabled",
+			BudgetTokens: kitutil.GetPointer(1280),
+		}
+	case "medium":
+		claudeRequest.Thinking = &dto.Thinking{
+			Type:         "enabled",
+			BudgetTokens: kitutil.GetPointer(2048),
+		}
+	case "high":
+		claudeRequest.Thinking = &dto.Thinking{
+			Type:         "enabled",
+			BudgetTokens: kitutil.GetPointer(4096),
+		}
+	}
+}
+
+func responsesInputContentToClaudeMediaMessages(c context.Context, content any) ([]dto.ClaudeMediaMessage, error) {
+	contentParts, err := ContentParts(content)
+	if err != nil {
+		return nil, err
+	}
+
+	parts := make([]dto.ClaudeMediaMessage, 0, len(contentParts))
+	for _, contentPart := range contentParts {
+		partType := strings.TrimSpace(kitutil.Interface2String(contentPart["type"]))
+		switch partType {
+		case "input_text", "output_text", "text":
+			text := kitutil.Interface2String(contentPart["text"])
+			if text != "" {
+				parts = append(parts, dto.ClaudeMediaMessage{
+					Type: "text",
+					Text: kitutil.GetPointer(text),
+				})
+			}
+		case "input_image", "input_file", "input_audio", "input_video":
+			source := ContentPartToFileSource(contentPart)
+			if source == nil {
+				continue
+			}
+			base64Data, mimeType, err := relaymedia.ResolveBase64Data(c, source, "formatting Responses input for Claude")
+			if err != nil {
+				return nil, fmt.Errorf("get file data failed: %s", err.Error())
+			}
+			claudePart := dto.ClaudeMediaMessage{
+				Source: &dto.ClaudeMessageSource{
+					Type:      "base64",
+					MediaType: mimeType,
+					Data:      base64Data,
+				},
+			}
+			if strings.HasPrefix(mimeType, "application/pdf") {
+				claudePart.Type = "document"
+			} else {
+				claudePart.Type = "image"
+			}
+			parts = append(parts, claudePart)
+		}
+	}
+	return parts, nil
+}
+
+func responsesFunctionCallItemToClaudeToolUse(item map[string]any, inputKey string) dto.ClaudeMediaMessage {
+	return dto.ClaudeMediaMessage{
+		Type:  "tool_use",
+		Id:    CallID(item),
+		Name:  strings.TrimSpace(kitutil.Interface2String(item["name"])),
+		Input: ObjectValue(item[inputKey], inputKey),
+	}
+}
+
+func responsesFunctionOutputItemToClaudeToolResult(item map[string]any) dto.ClaudeMediaMessage {
+	return dto.ClaudeMediaMessage{
+		Type:      "tool_result",
+		ToolUseId: CallID(item),
+		Content:   responsesToolOutputValue(item["output"]),
+	}
+}
+
+func responsesToolOutputValue(value any) any {
+	if value == nil {
+		return ""
+	}
+	return value
+}
+
+func appendClaudeToolUse(messages []dto.ClaudeMessage, toolUse dto.ClaudeMediaMessage) []dto.ClaudeMessage {
+	if len(messages) > 0 && messages[len(messages)-1].Role == "assistant" {
+		last := messages[len(messages)-1]
+		parts := claudeMessageContentParts(last.Content)
+		parts = append(parts, toolUse)
+		last.Content = parts
+		messages[len(messages)-1] = last
+		return messages
+	}
+	return append(messages, dto.ClaudeMessage{
+		Role:    "assistant",
+		Content: []dto.ClaudeMediaMessage{toolUse},
+	})
+}
+
+func appendClaudeToolResult(messages []dto.ClaudeMessage, toolResult dto.ClaudeMediaMessage) []dto.ClaudeMessage {
+	if len(messages) > 0 && messages[len(messages)-1].Role == "user" {
+		last := messages[len(messages)-1]
+		parts := claudeMessageContentParts(last.Content)
+		parts = append(parts, toolResult)
+		last.Content = parts
+		messages[len(messages)-1] = last
+		return messages
+	}
+	return append(messages, dto.ClaudeMessage{
+		Role:    "user",
+		Content: []dto.ClaudeMediaMessage{toolResult},
+	})
+}
+
+func claudeMessageContentParts(content any) []dto.ClaudeMediaMessage {
+	switch typed := content.(type) {
+	case []dto.ClaudeMediaMessage:
+		return typed
+	case string:
+		if typed == "" {
+			return nil
+		}
+		return []dto.ClaudeMediaMessage{
+			{
+				Type: "text",
+				Text: kitutil.GetPointer(typed),
+			},
+		}
+	default:
+		parts, _ := kitutil.Any2Type[[]dto.ClaudeMediaMessage](content)
+		return parts
+	}
+}
+
+func responsesClaudeRole(item map[string]any) string {
+	switch strings.TrimSpace(kitutil.Interface2String(item["role"])) {
+	case "assistant":
+		return "assistant"
+	case "system", "developer":
+		return "system"
+	default:
+		return "user"
+	}
+}
+
+func ensureClaudeMessagesStartWithUser(messages []dto.ClaudeMessage) []dto.ClaudeMessage {
+	if len(messages) == 0 || messages[0].Role == "user" {
+		return messages
+	}
+	return append([]dto.ClaudeMessage{
+		{
+			Role: "user",
+			Content: []dto.ClaudeMediaMessage{
+				{
+					Type: "text",
+					Text: kitutil.GetPointer("..."),
+				},
+			},
+		},
+	}, messages...)
+}

relaykit/relayconvert/internal/oai_responses/to_gemini_chat_req.go

@@ -0,0 +1,309 @@
+package oairesponses
+
+import (
+	"fmt"
+	"strings"
+
+	"context"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	relaymedia "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/media"
+	sharedgemini "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/shared/gemini"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+func convertOpenAIResponsesRequestToGeminiChat(c context.Context, info convmeta.Meta, request any) (any, error) {
+	responsesRequest, err := OpenAIResponsesRequestFromAny(request)
+	if err != nil {
+		return nil, err
+	}
+
+	prepared, err := PrepareOpenAIResponsesRequest(*responsesRequest)
+	if err != nil {
+		return nil, err
+	}
+	return OpenAIResponsesRequestToGeminiChat(c, &prepared, info)
+}
+
+func OpenAIResponsesRequestToGeminiChat(c context.Context, req *dto.OpenAIResponsesRequest, info convmeta.Meta) (*dto.GeminiChatRequest, error) {
+	opts := convmeta.OptionsOf(info)
+	if req == nil {
+		return nil, fmt.Errorf("request is nil")
+	}
+	if req.Model == "" {
+		return nil, fmt.Errorf("model is required")
+	}
+	if err := ValidateRequestChatUnsupportedFields(req); err != nil {
+		return nil, err
+	}
+
+	geminiRequest := &dto.GeminiChatRequest{
+		GenerationConfig: dto.GeminiChatGenerationConfig{
+			Temperature: req.Temperature,
+		},
+	}
+	if req.TopP != nil && *req.TopP > 0 {
+		geminiRequest.GenerationConfig.TopP = kitutil.GetPointer(*req.TopP)
+	}
+	if req.MaxOutputTokens != nil && *req.MaxOutputTokens > 0 {
+		geminiRequest.GenerationConfig.MaxOutputTokens = kitutil.GetPointer(*req.MaxOutputTokens)
+	}
+
+	upstreamModelName := req.Model
+	if modelName := convmeta.UpstreamModelName(info); modelName != "" {
+		upstreamModelName = modelName
+	}
+	if opts.Gemini.SupportsImagineModel(upstreamModelName) {
+		geminiRequest.GenerationConfig.ResponseModalities = []string{"TEXT", "IMAGE"}
+	}
+	if err := applyResponsesTextToGemini(req.Text, geminiRequest); err != nil {
+		return nil, err
+	}
+	sharedgemini.ApplyThinkingConfig(geminiRequest, info, dto.GeneralOpenAIRequest{
+		Model:               req.Model,
+		MaxCompletionTokens: req.MaxOutputTokens,
+		ReasoningEffort:     ReasoningEffort(req),
+	})
+
+	var safetySettings []dto.GeminiChatSafetySettings
+	for _, category := range sharedgemini.SafetySettingCategories {
+		threshold := opts.Gemini.SafetySettingFor(category)
+		if threshold == "" {
+			continue
+		}
+		safetySettings = append(safetySettings, dto.GeminiChatSafetySettings{
+			Category:  category,
+			Threshold: threshold,
+		})
+	}
+	if len(safetySettings) > 0 {
+		geminiRequest.SafetySettings = safetySettings
+	}
+
+	functions, err := RequestFunctionDeclarations(req.Tools)
+	if err != nil {
+		return nil, err
+	}
+	for i := range functions {
+		if params, ok := functions[i].Parameters.(map[string]interface{}); ok {
+			if props, hasProps := params["properties"].(map[string]interface{}); hasProps && len(props) == 0 {
+				functions[i].Parameters = nil
+				continue
+			}
+		}
+		functions[i].Parameters = sharedgemini.CleanFunctionParameters(functions[i].Parameters)
+	}
+	if len(functions) > 0 {
+		geminiRequest.SetTools([]dto.GeminiChatTool{
+			{FunctionDeclarations: functions},
+		})
+	}
+
+	toolChoice, err := RequestToolChoiceToChat(req.ToolChoice)
+	if err != nil {
+		return nil, err
+	}
+	if toolChoice != nil {
+		geminiRequest.ToolConfig = sharedgemini.OpenAIToolChoiceToConfig(toolChoice)
+	}
+
+	systemTexts := make([]string, 0)
+	if RawJSONPresent(req.Instructions) {
+		instructions, err := JSONString(req.Instructions)
+		if err != nil {
+			return nil, fmt.Errorf("invalid instructions: %w", err)
+		}
+		if strings.TrimSpace(instructions) != "" {
+			systemTexts = append(systemTexts, instructions)
+		}
+	}
+
+	inputItems, err := InputItems(req.Input)
+	if err != nil {
+		return nil, err
+	}
+	callNames := make(map[string]string)
+	for _, item := range inputItems {
+		itemType := strings.TrimSpace(kitutil.Interface2String(item["type"]))
+		switch itemType {
+		case ResponsesInputTypeFunctionCall:
+			part, callID, err := responsesFunctionCallItemToGeminiPart(item)
+			if err != nil {
+				return nil, err
+			}
+			sharedgemini.AttachFunctionCallThoughtSignature(opts, &part)
+			if callID != "" {
+				callNames[callID] = part.FunctionCall.FunctionName
+			}
+			appendGeminiContentPart(geminiRequest, "model", part)
+		case ResponsesInputTypeFunctionCallOutput:
+			part := responsesFunctionOutputItemToGeminiPart(item, callNames)
+			appendGeminiContentPart(geminiRequest, "user", part)
+		default:
+			role := responsesGeminiRole(item)
+			parts, err := responsesInputContentToGeminiParts(c, item["content"])
+			if err != nil {
+				return nil, err
+			}
+			if role == "system" {
+				for _, part := range parts {
+					if part.Text != "" {
+						systemTexts = append(systemTexts, part.Text)
+					}
+				}
+				continue
+			}
+			if len(parts) > 0 {
+				geminiRequest.Contents = append(geminiRequest.Contents, dto.GeminiChatContent{
+					Role:  role,
+					Parts: parts,
+				})
+			}
+		}
+	}
+
+	if len(systemTexts) > 0 {
+		geminiRequest.SystemInstructions = &dto.GeminiChatContent{
+			Parts: []dto.GeminiPart{{Text: strings.Join(systemTexts, "\n")}},
+		}
+	}
+
+	return geminiRequest, nil
+}
+
+func applyResponsesTextToGemini(raw []byte, geminiRequest *dto.GeminiChatRequest) error {
+	responseFormat, err := RequestTextToChatResponseFormat(raw)
+	if err != nil {
+		return err
+	}
+	if responseFormat == nil || (responseFormat.Type != "json_schema" && responseFormat.Type != "json_object") {
+		return nil
+	}
+
+	geminiRequest.GenerationConfig.ResponseMimeType = "application/json"
+	if len(responseFormat.JsonSchema) == 0 {
+		return nil
+	}
+
+	var jsonSchema dto.FormatJsonSchema
+	if err := kitutil.Unmarshal(responseFormat.JsonSchema, &jsonSchema); err != nil {
+		return nil
+	}
+	geminiRequest.GenerationConfig.ResponseSchema = sharedgemini.RemoveAdditionalProperties(jsonSchema.Schema, 0)
+	return nil
+}
+
+func responsesInputContentToGeminiParts(c context.Context, content any) ([]dto.GeminiPart, error) {
+	contentParts, err := ContentParts(content)
+	if err != nil {
+		return nil, err
+	}
+
+	parts := make([]dto.GeminiPart, 0, len(contentParts))
+	for _, contentPart := range contentParts {
+		nextParts, err := responsesContentPartToGeminiParts(c, contentPart)
+		if err != nil {
+			return nil, err
+		}
+		parts = append(parts, nextParts...)
+	}
+	return parts, nil
+}
+
+func responsesContentPartToGeminiParts(c context.Context, part map[string]any) ([]dto.GeminiPart, error) {
+	partType := strings.TrimSpace(kitutil.Interface2String(part["type"]))
+	switch partType {
+	case "input_text", "output_text", "text":
+		text := kitutil.Interface2String(part["text"])
+		if text == "" {
+			return nil, nil
+		}
+		return []dto.GeminiPart{{Text: text}}, nil
+	case "input_image", "input_file", "input_audio", "input_video":
+		source := ContentPartToFileSource(part)
+		if source == nil {
+			return nil, nil
+		}
+		base64Data, mimeType, err := relaymedia.ResolveBase64Data(c, source, "formatting Responses input for Gemini")
+		if err != nil {
+			return nil, fmt.Errorf("get file data from '%s' failed: %w", source.GetIdentifier(), err)
+		}
+		if _, ok := sharedgemini.SupportedMimeTypes[strings.ToLower(mimeType)]; !ok {
+			return nil, fmt.Errorf("mime type is not supported by Gemini: '%s', url: '%s', supported types are: %v", mimeType, source.GetIdentifier(), sharedgemini.SupportedMimeTypesList())
+		}
+		return []dto.GeminiPart{
+			{
+				InlineData: &dto.GeminiInlineData{
+					MimeType: mimeType,
+					Data:     base64Data,
+				},
+			},
+		}, nil
+	default:
+		return nil, nil
+	}
+}
+
+func responsesFunctionCallItemToGeminiPart(item map[string]any) (dto.GeminiPart, string, error) {
+	name := strings.TrimSpace(kitutil.Interface2String(item["name"]))
+	if name == "" {
+		return dto.GeminiPart{}, "", fmt.Errorf("function_call item is missing name")
+	}
+	callID := CallID(item)
+	return dto.GeminiPart{
+		FunctionCall: &dto.FunctionCall{
+			FunctionName: name,
+			Arguments:    ObjectValue(item["arguments"], "arguments"),
+		},
+	}, callID, nil
+}
+
+func responsesFunctionOutputItemToGeminiPart(item map[string]any, callNames map[string]string) dto.GeminiPart {
+	callID := CallID(item)
+	name := strings.TrimSpace(kitutil.Interface2String(item["name"]))
+	if name == "" {
+		name = callNames[callID]
+	}
+	return dto.GeminiPart{
+		FunctionResponse: &dto.GeminiFunctionResponse{
+			Name:     name,
+			Response: GeminiResponseMap(item["output"]),
+		},
+	}
+}
+
+func appendGeminiContentPart(req *dto.GeminiChatRequest, role string, part dto.GeminiPart) {
+	if len(req.Contents) > 0 && req.Contents[len(req.Contents)-1].Role == role {
+		if role == "model" && part.FunctionCall != nil {
+			parts := req.Contents[len(req.Contents)-1].Parts
+			insertAt := 0
+			for insertAt < len(parts) && parts[insertAt].FunctionCall != nil {
+				insertAt++
+			}
+			parts = append(parts, dto.GeminiPart{})
+			copy(parts[insertAt+1:], parts[insertAt:])
+			parts[insertAt] = part
+			req.Contents[len(req.Contents)-1].Parts = parts
+			return
+		}
+		req.Contents[len(req.Contents)-1].Parts = append(req.Contents[len(req.Contents)-1].Parts, part)
+		return
+	}
+	req.Contents = append(req.Contents, dto.GeminiChatContent{
+		Role:  role,
+		Parts: []dto.GeminiPart{part},
+	})
+}
+
+func responsesGeminiRole(item map[string]any) string {
+	switch strings.TrimSpace(kitutil.Interface2String(item["role"])) {
+	case "assistant":
+		return "model"
+	case "system", "developer":
+		return "system"
+	case "model":
+		return "model"
+	default:
+		return "user"
+	}
+}

relaykit/relayconvert/internal/oai_responses/to_gemini_chat_req_preprocess.go

@@ -0,0 +1,101 @@
+package oairesponses
+
+import (
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+const (
+	geminiResponsesInputTypeCustomToolCall       = "custom_tool_call"
+	geminiResponsesInputTypeCustomToolCallOutput = "custom_tool_call_output"
+	geminiResponsesInputTypeFunctionCallOutput   = "function_call_output"
+)
+
+const (
+	ResponsesInputTypeCustomToolCallOutput = geminiResponsesInputTypeCustomToolCallOutput
+)
+
+func PrepareOpenAIResponsesRequest(request dto.OpenAIResponsesRequest) (dto.OpenAIResponsesRequest, error) {
+	tools, err := filterGeminiResponsesTools(request.Tools)
+	if err != nil {
+		return request, err
+	}
+	request.Tools = tools
+
+	input, err := filterGeminiResponsesInput(request.Input)
+	if err != nil {
+		return request, err
+	}
+	request.Input = input
+
+	return request, nil
+}
+
+func filterGeminiResponsesTools(raw []byte) ([]byte, error) {
+	if !geminiRawJSONPresent(raw) || kitutil.GetJsonType(raw) != "array" {
+		return raw, nil
+	}
+
+	var tools []map[string]any
+	if err := kitutil.Unmarshal(raw, &tools); err != nil {
+		return nil, err
+	}
+
+	filtered := make([]map[string]any, 0, len(tools))
+	for _, tool := range tools {
+		if strings.TrimSpace(kitutil.Interface2String(tool["type"])) != "function" {
+			continue
+		}
+		filtered = append(filtered, tool)
+	}
+	if len(filtered) == 0 {
+		return nil, nil
+	}
+	return kitutil.Marshal(filtered)
+}
+
+func filterGeminiResponsesInput(raw []byte) ([]byte, error) {
+	if !geminiRawJSONPresent(raw) || kitutil.GetJsonType(raw) != "array" {
+		return raw, nil
+	}
+
+	var items []map[string]any
+	if err := kitutil.Unmarshal(raw, &items); err != nil {
+		return nil, err
+	}
+
+	skippedCustomCallIDs := make(map[string]struct{})
+	for _, item := range items {
+		if strings.TrimSpace(kitutil.Interface2String(item["type"])) != geminiResponsesInputTypeCustomToolCall {
+			continue
+		}
+		if callID := strings.TrimSpace(kitutil.Interface2String(item["call_id"])); callID != "" {
+			skippedCustomCallIDs[callID] = struct{}{}
+		}
+	}
+
+	filtered := make([]map[string]any, 0, len(items))
+	for _, item := range items {
+		itemType := strings.TrimSpace(kitutil.Interface2String(item["type"]))
+		switch itemType {
+		case geminiResponsesInputTypeCustomToolCall, geminiResponsesInputTypeCustomToolCallOutput:
+			continue
+		case geminiResponsesInputTypeFunctionCallOutput:
+			if _, ok := skippedCustomCallIDs[strings.TrimSpace(kitutil.Interface2String(item["call_id"]))]; ok {
+				continue
+			}
+		}
+		filtered = append(filtered, item)
+	}
+
+	return kitutil.Marshal(filtered)
+}
+
+func geminiRawJSONPresent(raw []byte) bool {
+	if len(raw) == 0 {
+		return false
+	}
+	return kitutil.GetJsonType(raw) != "null"
+}

relaykit/relayconvert/internal/oai_responses/to_oai_chat_req.go

@@ -0,0 +1,553 @@
+package oairesponses
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+const (
+	responsesInputTypeFunctionCall       = "function_call"
+	responsesInputTypeFunctionCallOutput = "function_call_output"
+	responsesInputTypeCustomToolCall     = "custom_tool_call"
+	responsesInputTypeCustomToolOutput   = "custom_tool_call_output"
+)
+
+const (
+	ResponsesInputTypeFunctionCall       = responsesInputTypeFunctionCall
+	ResponsesInputTypeFunctionCallOutput = responsesInputTypeFunctionCallOutput
+	ResponsesInputTypeCustomToolCall     = responsesInputTypeCustomToolCall
+	ResponsesInputTypeCustomToolOutput   = responsesInputTypeCustomToolOutput
+)
+
+func ResponsesRequestToChatCompletionsRequest(req *dto.OpenAIResponsesRequest) (*dto.GeneralOpenAIRequest, error) {
+	if req == nil {
+		return nil, errors.New("request is nil")
+	}
+	if req.Model == "" {
+		return nil, errors.New("model is required")
+	}
+	if err := validateResponsesRequestChatUnsupportedFields(req); err != nil {
+		return nil, err
+	}
+
+	messages, err := responsesRequestMessagesToChat(req)
+	if err != nil {
+		return nil, err
+	}
+
+	tools, err := responsesRequestToolsToChat(req.Tools)
+	if err != nil {
+		return nil, err
+	}
+
+	toolChoice, err := responsesRequestToolChoiceToChat(req.ToolChoice)
+	if err != nil {
+		return nil, err
+	}
+
+	responseFormat, err := responsesRequestTextToChatResponseFormat(req.Text)
+	if err != nil {
+		return nil, err
+	}
+
+	out := &dto.GeneralOpenAIRequest{
+		Model:                req.Model,
+		Messages:             messages,
+		Stream:               req.Stream,
+		StreamOptions:        req.StreamOptions,
+		MaxCompletionTokens:  req.MaxOutputTokens,
+		Temperature:          req.Temperature,
+		TopP:                 req.TopP,
+		TopLogProbs:          req.TopLogProbs,
+		ResponseFormat:       responseFormat,
+		Tools:                tools,
+		ToolChoice:           toolChoice,
+		User:                 req.User,
+		Store:                req.Store,
+		Metadata:             req.Metadata,
+		SafetyIdentifier:     req.SafetyIdentifier,
+		PromptCacheRetention: req.PromptCacheRetention,
+		EnableThinking:       req.EnableThinking,
+	}
+
+	if req.Reasoning != nil {
+		out.ReasoningEffort = req.Reasoning.Effort
+	}
+	if req.ServiceTier != "" {
+		out.ServiceTier, _ = kitutil.Marshal(req.ServiceTier)
+	}
+	if len(req.ParallelToolCalls) > 0 && kitutil.GetJsonType(req.ParallelToolCalls) == "boolean" {
+		var parallelToolCalls bool
+		if err := kitutil.Unmarshal(req.ParallelToolCalls, &parallelToolCalls); err == nil {
+			out.ParallelTooCalls = &parallelToolCalls
+		}
+	}
+	if len(req.PromptCacheKey) > 0 && kitutil.GetJsonType(req.PromptCacheKey) == "string" {
+		var promptCacheKey string
+		if err := kitutil.Unmarshal(req.PromptCacheKey, &promptCacheKey); err == nil {
+			out.PromptCacheKey = promptCacheKey
+		}
+	}
+
+	return out, nil
+}
+
+func validateResponsesRequestChatUnsupportedFields(req *dto.OpenAIResponsesRequest) error {
+	unsupported := make([]string, 0, 4)
+	if rawJSONPresent(req.Conversation) {
+		unsupported = append(unsupported, "conversation")
+	}
+	if strings.TrimSpace(req.PreviousResponseID) != "" {
+		unsupported = append(unsupported, "previous_response_id")
+	}
+	if rawJSONPresent(req.Prompt) {
+		unsupported = append(unsupported, "prompt")
+	}
+	if rawJSONPresent(req.ContextManagement) {
+		unsupported = append(unsupported, "context_management")
+	}
+	if len(unsupported) > 0 {
+		return fmt.Errorf("responses to chat conversion does not support stateful fields: %s", strings.Join(unsupported, ", "))
+	}
+	return nil
+}
+
+func ValidateRequestChatUnsupportedFields(req *dto.OpenAIResponsesRequest) error {
+	return validateResponsesRequestChatUnsupportedFields(req)
+}
+
+func responsesRequestMessagesToChat(req *dto.OpenAIResponsesRequest) ([]dto.Message, error) {
+	messages := make([]dto.Message, 0)
+	if rawJSONPresent(req.Instructions) {
+		instructions, err := responsesJSONString(req.Instructions)
+		if err != nil {
+			return nil, fmt.Errorf("invalid instructions: %w", err)
+		}
+		if strings.TrimSpace(instructions) != "" {
+			messages = append(messages, dto.Message{Role: "system", Content: instructions})
+		}
+	}
+
+	if !rawJSONPresent(req.Input) {
+		return messages, nil
+	}
+
+	switch kitutil.GetJsonType(req.Input) {
+	case "string":
+		input, err := responsesJSONString(req.Input)
+		if err != nil {
+			return nil, fmt.Errorf("invalid input string: %w", err)
+		}
+		messages = append(messages, dto.Message{Role: "user", Content: input})
+		return messages, nil
+	case "array":
+		var items []map[string]any
+		if err := kitutil.Unmarshal(req.Input, &items); err != nil {
+			return nil, fmt.Errorf("invalid input array: %w", err)
+		}
+		for _, item := range items {
+			nextMessages, err := responsesInputItemToChatMessages(item, messages)
+			if err != nil {
+				return nil, err
+			}
+			messages = nextMessages
+		}
+		return messages, nil
+	default:
+		return nil, fmt.Errorf("unsupported responses input type %q", kitutil.GetJsonType(req.Input))
+	}
+}
+
+func responsesInputItemToChatMessages(item map[string]any, messages []dto.Message) ([]dto.Message, error) {
+	itemType := strings.TrimSpace(kitutil.Interface2String(item["type"]))
+	switch itemType {
+	case responsesInputTypeFunctionCall:
+		toolCall, err := responsesFunctionCallItemToChatToolCall(item)
+		if err != nil {
+			return nil, err
+		}
+		return appendToolCallToLastAssistant(messages, toolCall), nil
+	case responsesInputTypeCustomToolCall:
+		toolCall, err := responsesCustomToolCallItemToChatToolCall(item)
+		if err != nil {
+			return nil, err
+		}
+		return appendToolCallToLastAssistant(messages, toolCall), nil
+	case responsesInputTypeFunctionCallOutput:
+		callID := strings.TrimSpace(kitutil.Interface2String(item["call_id"]))
+		content := responseToolOutputToChatContent(item["output"])
+		return append(messages, dto.Message{Role: "tool", ToolCallId: callID, Content: content}), nil
+	}
+
+	role := strings.TrimSpace(kitutil.Interface2String(item["role"]))
+	if role == "" {
+		role = "user"
+	}
+	content, err := responsesInputContentToChatContent(item["content"])
+	if err != nil {
+		return nil, err
+	}
+	return append(messages, dto.Message{Role: role, Content: content}), nil
+}
+
+func responsesInputContentToChatContent(content any) (any, error) {
+	if content == nil {
+		return "", nil
+	}
+
+	switch value := content.(type) {
+	case string:
+		return value, nil
+	case []any:
+		return responsesContentPartsToChatContent(value)
+	case []map[string]any:
+		parts := make([]any, 0, len(value))
+		for _, part := range value {
+			parts = append(parts, part)
+		}
+		return responsesContentPartsToChatContent(parts)
+	default:
+		return content, nil
+	}
+}
+
+func responsesContentPartsToChatContent(parts []any) (any, error) {
+	chatParts := make([]any, 0, len(parts))
+	var textOnly strings.Builder
+	onlyText := true
+
+	for _, rawPart := range parts {
+		part, ok := rawPart.(map[string]any)
+		if !ok {
+			onlyText = false
+			chatParts = append(chatParts, rawPart)
+			continue
+		}
+
+		partType := strings.TrimSpace(kitutil.Interface2String(part["type"]))
+		switch partType {
+		case "input_text", "output_text", "text":
+			text := kitutil.Interface2String(part["text"])
+			textOnly.WriteString(text)
+			chatParts = append(chatParts, map[string]any{
+				"type": dto.ContentTypeText,
+				"text": text,
+			})
+		case "input_image":
+			onlyText = false
+			chatParts = append(chatParts, map[string]any{
+				"type":      dto.ContentTypeImageURL,
+				"image_url": responsesImagePartToChatImageURL(part),
+			})
+		case "input_file":
+			onlyText = false
+			chatParts = append(chatParts, map[string]any{
+				"type": dto.ContentTypeFile,
+				"file": responsesFilePartToChatFile(part),
+			})
+		case "input_audio":
+			onlyText = false
+			chatParts = append(chatParts, map[string]any{
+				"type":        dto.ContentTypeInputAudio,
+				"input_audio": responsesPartPayload(part, "input_audio"),
+			})
+		case "input_video":
+			onlyText = false
+			chatParts = append(chatParts, map[string]any{
+				"type":      dto.ContentTypeVideoUrl,
+				"video_url": responsesVideoPartToChatVideoURL(part),
+			})
+		default:
+			onlyText = false
+			chatParts = append(chatParts, part)
+		}
+	}
+
+	if onlyText {
+		return textOnly.String(), nil
+	}
+	return chatParts, nil
+}
+
+func responsesFunctionCallItemToChatToolCall(item map[string]any) (dto.ToolCallRequest, error) {
+	name := strings.TrimSpace(kitutil.Interface2String(item["name"]))
+	if name == "" {
+		return dto.ToolCallRequest{}, errors.New("function_call item is missing name")
+	}
+	return dto.ToolCallRequest{
+		ID:   responsesCallID(item),
+		Type: "function",
+		Function: dto.FunctionRequest{
+			Name:      name,
+			Arguments: responsesArgumentsString(item["arguments"]),
+		},
+	}, nil
+}
+
+func responsesCustomToolCallItemToChatToolCall(item map[string]any) (dto.ToolCallRequest, error) {
+	raw, err := kitutil.Marshal(item)
+	if err != nil {
+		return dto.ToolCallRequest{}, err
+	}
+	return dto.ToolCallRequest{
+		ID:     responsesCallID(item),
+		Type:   dto.CustomType,
+		Custom: raw,
+		Function: dto.FunctionRequest{
+			Name:      strings.TrimSpace(kitutil.Interface2String(item["name"])),
+			Arguments: responsesArgumentsString(item["input"]),
+		},
+	}, nil
+}
+
+func appendToolCallToLastAssistant(messages []dto.Message, toolCall dto.ToolCallRequest) []dto.Message {
+	if len(messages) == 0 || messages[len(messages)-1].Role != "assistant" {
+		messages = append(messages, dto.Message{Role: "assistant"})
+	}
+
+	idx := len(messages) - 1
+	toolCalls := messages[idx].ParseToolCalls()
+	toolCalls = append(toolCalls, toolCall)
+	toolCallsRaw, _ := kitutil.Marshal(toolCalls)
+	messages[idx].ToolCalls = toolCallsRaw
+	return messages
+}
+
+func responsesRequestToolsToChat(raw json.RawMessage) ([]dto.ToolCallRequest, error) {
+	if !rawJSONPresent(raw) {
+		return nil, nil
+	}
+
+	var tools []map[string]any
+	if err := kitutil.Unmarshal(raw, &tools); err != nil {
+		return nil, fmt.Errorf("invalid tools: %w", err)
+	}
+
+	out := make([]dto.ToolCallRequest, 0, len(tools))
+	for _, tool := range tools {
+		toolType := strings.TrimSpace(kitutil.Interface2String(tool["type"]))
+		if toolType == "function" {
+			out = append(out, dto.ToolCallRequest{
+				Type: "function",
+				Function: dto.FunctionRequest{
+					Name:        strings.TrimSpace(kitutil.Interface2String(tool["name"])),
+					Description: kitutil.Interface2String(tool["description"]),
+					Parameters:  tool["parameters"],
+				},
+			})
+			continue
+		}
+
+		rawTool, err := kitutil.Marshal(tool)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, dto.ToolCallRequest{
+			Type:   toolType,
+			Custom: rawTool,
+		})
+	}
+	return out, nil
+}
+
+func responsesRequestToolChoiceToChat(raw json.RawMessage) (any, error) {
+	if !rawJSONPresent(raw) {
+		return nil, nil
+	}
+	if kitutil.GetJsonType(raw) == "string" {
+		var choice string
+		if err := kitutil.Unmarshal(raw, &choice); err != nil {
+			return nil, fmt.Errorf("invalid tool_choice: %w", err)
+		}
+		return choice, nil
+	}
+
+	var choice map[string]any
+	if err := kitutil.Unmarshal(raw, &choice); err != nil {
+		return nil, fmt.Errorf("invalid tool_choice: %w", err)
+	}
+	if kitutil.Interface2String(choice["type"]) == "function" {
+		name := strings.TrimSpace(kitutil.Interface2String(choice["name"]))
+		if name != "" {
+			return map[string]any{
+				"type": "function",
+				"function": map[string]any{
+					"name": name,
+				},
+			}, nil
+		}
+	}
+	return choice, nil
+}
+
+func RequestToolChoiceToChat(raw json.RawMessage) (any, error) {
+	return responsesRequestToolChoiceToChat(raw)
+}
+
+func responsesRequestTextToChatResponseFormat(raw json.RawMessage) (*dto.ResponseFormat, error) {
+	if !rawJSONPresent(raw) {
+		return nil, nil
+	}
+
+	var textConfig map[string]any
+	if err := kitutil.Unmarshal(raw, &textConfig); err != nil {
+		return nil, fmt.Errorf("invalid text config: %w", err)
+	}
+	format, ok := textConfig["format"].(map[string]any)
+	if !ok {
+		return nil, nil
+	}
+
+	formatType := strings.TrimSpace(kitutil.Interface2String(format["type"]))
+	if formatType == "" {
+		return nil, nil
+	}
+
+	out := &dto.ResponseFormat{Type: formatType}
+	if formatType == "json_schema" {
+		schemaRaw, err := kitutil.Marshal(format)
+		if err != nil {
+			return nil, err
+		}
+		out.JsonSchema = schemaRaw
+	}
+	return out, nil
+}
+
+func RequestTextToChatResponseFormat(raw json.RawMessage) (*dto.ResponseFormat, error) {
+	return responsesRequestTextToChatResponseFormat(raw)
+}
+
+func responsesImagePartToChatImageURL(part map[string]any) any {
+	if imageURL, ok := part["image_url"]; ok {
+		return imageURL
+	}
+	imageURL := map[string]any{}
+	for _, key := range []string{"url", "file_id", "detail"} {
+		if value, ok := part[key]; ok {
+			imageURL[key] = value
+		}
+	}
+	if len(imageURL) == 0 {
+		return part
+	}
+	return imageURL
+}
+
+func responsesFilePartToChatFile(part map[string]any) any {
+	if file, ok := part["file"]; ok {
+		return file
+	}
+	file := map[string]any{}
+	for _, key := range []string{"file_id", "file_data", "filename", "file_url"} {
+		if value, ok := part[key]; ok {
+			file[key] = value
+		}
+	}
+	if len(file) == 0 {
+		return part
+	}
+	return file
+}
+
+func responsesVideoPartToChatVideoURL(part map[string]any) any {
+	if videoURL, ok := part["video_url"]; ok {
+		if videoURLMap, ok := videoURL.(map[string]any); ok {
+			if url := kitutil.Interface2String(videoURLMap["url"]); url != "" {
+				return url
+			}
+		}
+		return videoURL
+	}
+	if url := kitutil.Interface2String(part["url"]); url != "" {
+		return url
+	}
+	return responsesPartPayload(part, "video_url")
+}
+
+func responsesPartPayload(part map[string]any, key string) any {
+	if value, ok := part[key]; ok {
+		return value
+	}
+	payload := make(map[string]any, len(part))
+	for k, value := range part {
+		if k == "type" {
+			continue
+		}
+		payload[k] = value
+	}
+	return payload
+}
+
+func responsesCallID(item map[string]any) string {
+	callID := strings.TrimSpace(kitutil.Interface2String(item["call_id"]))
+	if callID != "" {
+		return callID
+	}
+	return strings.TrimSpace(kitutil.Interface2String(item["id"]))
+}
+
+func CallID(item map[string]any) string {
+	return responsesCallID(item)
+}
+
+func responsesArgumentsString(value any) string {
+	switch v := value.(type) {
+	case nil:
+		return ""
+	case string:
+		return v
+	default:
+		raw, err := kitutil.Marshal(v)
+		if err != nil {
+			return kitutil.Interface2String(v)
+		}
+		return string(raw)
+	}
+}
+
+func responseToolOutputToChatContent(value any) any {
+	switch v := value.(type) {
+	case nil:
+		return ""
+	case string:
+		return v
+	default:
+		raw, err := kitutil.Marshal(v)
+		if err != nil {
+			return fmt.Sprintf("%v", v)
+		}
+		return string(raw)
+	}
+}
+
+func responsesJSONString(raw json.RawMessage) (string, error) {
+	if kitutil.GetJsonType(raw) != "string" {
+		return string(raw), nil
+	}
+	var value string
+	if err := kitutil.Unmarshal(raw, &value); err != nil {
+		return "", err
+	}
+	return value, nil
+}
+
+func rawJSONPresent(raw json.RawMessage) bool {
+	if len(raw) == 0 {
+		return false
+	}
+	return kitutil.GetJsonType(raw) != "null"
+}
+
+func JSONString(raw json.RawMessage) (string, error) {
+	return responsesJSONString(raw)
+}
+
+func RawJSONPresent(raw json.RawMessage) bool {
+	return rawJSONPresent(raw)
+}

relaykit/relayconvert/internal/oai_responses/to_oai_chat_req_test.go

@@ -0,0 +1,270 @@
+package oairesponses
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/samber/lo"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+func TestResponsesRequestToChatCompletionsRequestInstructionsAndScalarInput(t *testing.T) {
+	stream := true
+	temperature := 0.0
+	topP := 0.9
+	maxOutputTokens := uint(128)
+	parallelToolCalls := true
+
+	got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+		Model:                "gpt-test",
+		Instructions:         mustRawMessage(t, "system rules"),
+		Input:                mustRawMessage(t, "hello"),
+		Stream:               &stream,
+		StreamOptions:        &dto.StreamOptions{IncludeUsage: true},
+		MaxOutputTokens:      &maxOutputTokens,
+		Temperature:          &temperature,
+		TopP:                 &topP,
+		User:                 mustRawMessage(t, "user-1"),
+		Store:                mustRawMessage(t, false),
+		Metadata:             mustRawMessage(t, map[string]any{"trace": "abc"}),
+		ParallelToolCalls:    mustRawMessage(t, parallelToolCalls),
+		PromptCacheKey:       mustRawMessage(t, "cache-key"),
+		PromptCacheRetention: mustRawMessage(t, "24h"),
+		Reasoning:            &dto.Reasoning{Effort: "medium"},
+	})
+	require.NoError(t, err)
+
+	assert.Equal(t, "gpt-test", got.Model)
+	require.Len(t, got.Messages, 2)
+	assert.Equal(t, dto.Message{Role: "system", Content: "system rules"}, got.Messages[0])
+	assert.Equal(t, dto.Message{Role: "user", Content: "hello"}, got.Messages[1])
+	assert.Same(t, &stream, got.Stream)
+	require.NotNil(t, got.StreamOptions)
+	assert.True(t, got.StreamOptions.IncludeUsage)
+	assert.Equal(t, maxOutputTokens, lo.FromPtr(got.MaxCompletionTokens))
+	assert.Equal(t, 0.0, lo.FromPtr(got.Temperature))
+	assert.Equal(t, 0.9, lo.FromPtr(got.TopP))
+	assert.True(t, lo.FromPtr(got.ParallelTooCalls))
+	assert.Equal(t, "cache-key", got.PromptCacheKey)
+	assert.Equal(t, "medium", got.ReasoningEffort)
+	assert.Equal(t, `"user-1"`, string(got.User))
+	assert.Equal(t, `false`, string(got.Store))
+	assert.Equal(t, "abc", gjson.GetBytes(got.Metadata, "trace").String())
+}
+
+func TestResponsesRequestToChatCompletionsRequestMultimodalInput(t *testing.T) {
+	got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+		Model: "gpt-test",
+		Input: mustRawMessage(t, []map[string]any{
+			{
+				"role": "user",
+				"content": []map[string]any{
+					{"type": "input_text", "text": "look"},
+					{"type": "input_image", "image_url": "https://example.test/a.png", "detail": "low"},
+					{"type": "input_file", "file_id": "file_1", "filename": "a.txt"},
+					{"type": "input_audio", "input_audio": map[string]any{"data": "abc", "format": "wav"}},
+					{"type": "input_video", "video_url": map[string]any{"url": "https://example.test/v.mp4"}},
+				},
+			},
+		}),
+	})
+	require.NoError(t, err)
+
+	require.Len(t, got.Messages, 1)
+	assert.Equal(t, "user", got.Messages[0].Role)
+	parts := got.Messages[0].ParseContent()
+	require.Len(t, parts, 5)
+	assert.Equal(t, dto.ContentTypeText, parts[0].Type)
+	assert.Equal(t, "look", parts[0].Text)
+	assert.Equal(t, dto.ContentTypeImageURL, parts[1].Type)
+	assert.Equal(t, "https://example.test/a.png", parts[1].GetImageMedia().Url)
+	assert.Equal(t, dto.ContentTypeFile, parts[2].Type)
+	assert.Equal(t, "file_1", parts[2].GetFile().FileId)
+	assert.Equal(t, dto.ContentTypeInputAudio, parts[3].Type)
+	assert.Equal(t, "wav", parts[3].GetInputAudio().Format)
+	assert.Equal(t, dto.ContentTypeVideoUrl, parts[4].Type)
+	assert.Equal(t, "https://example.test/v.mp4", parts[4].GetVideoUrl().Url)
+}
+
+func TestResponsesRequestToChatCompletionsRequestAssistantTextAndFunctionCallCoexist(t *testing.T) {
+	got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+		Model: "gpt-test",
+		Input: mustRawMessage(t, []map[string]any{
+			{
+				"role": "assistant",
+				"content": []map[string]any{
+					{"type": "output_text", "text": "I will call."},
+				},
+			},
+			{
+				"type":      "function_call",
+				"call_id":   "call_1",
+				"name":      "lookup",
+				"arguments": map[string]any{"q": "x"},
+			},
+			{
+				"type":    "function_call_output",
+				"call_id": "call_1",
+				"output":  map[string]any{"ok": true},
+			},
+		}),
+	})
+	require.NoError(t, err)
+
+	require.Len(t, got.Messages, 2)
+	assert.Equal(t, "assistant", got.Messages[0].Role)
+	assert.Equal(t, "I will call.", got.Messages[0].StringContent())
+	toolCalls := got.Messages[0].ParseToolCalls()
+	require.Len(t, toolCalls, 1)
+	assert.Equal(t, "call_1", toolCalls[0].ID)
+	assert.Equal(t, "function", toolCalls[0].Type)
+	assert.Equal(t, "lookup", toolCalls[0].Function.Name)
+	assert.JSONEq(t, `{"q":"x"}`, toolCalls[0].Function.Arguments)
+	assert.Equal(t, "tool", got.Messages[1].Role)
+	assert.Equal(t, "call_1", got.Messages[1].ToolCallId)
+	assert.JSONEq(t, `{"ok":true}`, got.Messages[1].StringContent())
+}
+
+func TestResponsesRequestToChatCompletionsRequestOnlyFunctionCallCreatesAssistant(t *testing.T) {
+	got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+		Model: "gpt-test",
+		Input: mustRawMessage(t, []map[string]any{
+			{
+				"type":      "function_call",
+				"call_id":   "call_1",
+				"name":      "lookup",
+				"arguments": `{"q":"x"}`,
+			},
+		}),
+	})
+	require.NoError(t, err)
+
+	require.Len(t, got.Messages, 1)
+	assert.Equal(t, "assistant", got.Messages[0].Role)
+	assert.Nil(t, got.Messages[0].Content)
+	toolCalls := got.Messages[0].ParseToolCalls()
+	require.Len(t, toolCalls, 1)
+	assert.Equal(t, `{"q":"x"}`, toolCalls[0].Function.Arguments)
+}
+
+func TestResponsesRequestToChatCompletionsRequestToolsToolChoiceAndTextFormat(t *testing.T) {
+	got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+		Model: "gpt-test",
+		Input: mustRawMessage(t, "hello"),
+		Tools: mustRawMessage(t, []map[string]any{
+			{
+				"type":        "function",
+				"name":        "lookup",
+				"description": "Lookup data",
+				"parameters": map[string]any{
+					"type": "object",
+					"properties": map[string]any{
+						"q": map[string]any{"type": "string"},
+					},
+				},
+			},
+		}),
+		ToolChoice: mustRawMessage(t, map[string]any{
+			"type": "function",
+			"name": "lookup",
+		}),
+		Text: mustRawMessage(t, map[string]any{
+			"format": map[string]any{
+				"type":   "json_schema",
+				"name":   "answer",
+				"schema": map[string]any{"type": "object"},
+				"strict": true,
+			},
+		}),
+	})
+	require.NoError(t, err)
+
+	require.Len(t, got.Tools, 1)
+	assert.Equal(t, "function", got.Tools[0].Type)
+	assert.Equal(t, "lookup", got.Tools[0].Function.Name)
+	assert.Equal(t, "Lookup data", got.Tools[0].Function.Description)
+	assert.Equal(t, "object", got.Tools[0].Function.Parameters.(map[string]any)["type"])
+	assert.Equal(t, map[string]any{
+		"type": "function",
+		"function": map[string]any{
+			"name": "lookup",
+		},
+	}, got.ToolChoice)
+	require.NotNil(t, got.ResponseFormat)
+	assert.Equal(t, "json_schema", got.ResponseFormat.Type)
+	assert.Equal(t, "answer", gjson.GetBytes(got.ResponseFormat.JsonSchema, "name").String())
+	assert.True(t, gjson.GetBytes(got.ResponseFormat.JsonSchema, "strict").Bool())
+}
+
+func TestResponsesRequestToChatCompletionsRequestCustomToolCallPreservesRawShape(t *testing.T) {
+	got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+		Model: "gpt-test",
+		Input: mustRawMessage(t, []map[string]any{
+			{
+				"type":    "custom_tool_call",
+				"call_id": "call_custom",
+				"name":    "apply_patch",
+				"input":   "patch body",
+			},
+		}),
+	})
+	require.NoError(t, err)
+
+	require.Len(t, got.Messages, 1)
+	toolCalls := got.Messages[0].ParseToolCalls()
+	require.Len(t, toolCalls, 1)
+	assert.Equal(t, dto.CustomType, toolCalls[0].Type)
+	assert.Equal(t, "call_custom", toolCalls[0].ID)
+	assert.Equal(t, "apply_patch", toolCalls[0].Function.Name)
+	assert.Equal(t, "patch body", toolCalls[0].Function.Arguments)
+	assert.Equal(t, "custom_tool_call", gjson.GetBytes(toolCalls[0].Custom, "type").String())
+	assert.Equal(t, "patch body", gjson.GetBytes(toolCalls[0].Custom, "input").String())
+}
+
+func TestResponsesRequestToChatCompletionsRequestRejectsStatefulFields(t *testing.T) {
+	tests := []struct {
+		name string
+		req  *dto.OpenAIResponsesRequest
+		want string
+	}{
+		{
+			name: "conversation",
+			req:  &dto.OpenAIResponsesRequest{Model: "gpt-test", Conversation: mustRawMessage(t, "conv_1")},
+			want: "conversation",
+		},
+		{
+			name: "previous response",
+			req:  &dto.OpenAIResponsesRequest{Model: "gpt-test", PreviousResponseID: "resp_1"},
+			want: "previous_response_id",
+		},
+		{
+			name: "prompt",
+			req:  &dto.OpenAIResponsesRequest{Model: "gpt-test", Prompt: mustRawMessage(t, map[string]any{"id": "pmpt_1"})},
+			want: "prompt",
+		},
+		{
+			name: "context management",
+			req:  &dto.OpenAIResponsesRequest{Model: "gpt-test", ContextManagement: mustRawMessage(t, map[string]any{"type": "auto"})},
+			want: "context_management",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			_, err := ResponsesRequestToChatCompletionsRequest(tt.req)
+			require.Error(t, err)
+			assert.Contains(t, err.Error(), tt.want)
+			assert.Contains(t, err.Error(), "stateful fields")
+		})
+	}
+}
+
+func mustRawMessage(t *testing.T, value any) []byte {
+	t.Helper()
+	raw, err := kitutil.Marshal(value)
+	require.NoError(t, err)
+	return raw
+}

relaykit/relayconvert/internal/oai_responses/to_oai_chat_resp.go

@@ -0,0 +1,290 @@
+package oairesponses
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+const (
+	responsesEventCreated                  = "response.created"
+	responsesEventCompleted                = "response.completed"
+	responsesEventDone                     = "response.done"
+	responsesEventIncomplete               = "response.incomplete"
+	responsesEventFailed                   = "response.failed"
+	responsesEventError                    = "response.error"
+	responsesEventOutputTextDelta          = "response.output_text.delta"
+	responsesEventOutputItemAdded          = "response.output_item.added"
+	responsesEventOutputItemDone           = "response.output_item.done"
+	responsesEventFunctionArgsDelta        = "response.function_call_arguments.delta"
+	responsesEventFunctionArgsDone         = "response.function_call_arguments.done"
+	responsesEventCustomToolInputDelta     = "response.custom_tool_call_input.delta"
+	responsesEventCustomToolInputDone      = "response.custom_tool_call_input.done"
+	responsesEventReasoningSummaryDelta    = "response.reasoning_summary_text.delta"
+	responsesEventReasoningSummaryDone     = "response.reasoning_summary_text.done"
+	responsesEventReasoningTextDelta       = "response.reasoning_text.delta"
+	responsesEventReasoningTextDone        = "response.reasoning_text.done"
+	responsesOutputTypeFunctionCall        = "function_call"
+	responsesOutputTypeCustomToolCall      = "custom_tool_call"
+	responsesOutputTypeMessage             = "message"
+	responsesOutputTypeReasoning           = "reasoning"
+	responsesIncompleteReasonContentFilter = "content_filter"
+	responsesIncompleteReasonMaxTokens     = "max_output_tokens"
+)
+
+func ResponsesFinishReasonFromStatus(resp *dto.OpenAIResponsesResponse) (string, bool) {
+	if resp == nil {
+		return "", false
+	}
+
+	status := responseStatusString(resp)
+	if status != "incomplete" {
+		return "", false
+	}
+
+	reason := ""
+	if resp.IncompleteDetails != nil {
+		reason = strings.TrimSpace(resp.IncompleteDetails.Reason)
+	}
+	if reason == responsesIncompleteReasonContentFilter {
+		return "content_filter", true
+	}
+	return "length", true
+}
+
+func ResponsesResponseToChatCompletionsResponse(resp *dto.OpenAIResponsesResponse, id string) (*dto.OpenAITextResponse, *dto.Usage, error) {
+	if resp == nil {
+		return nil, nil, errors.New("response is nil")
+	}
+
+	text := ExtractOutputTextFromResponses(resp)
+	reasoning := ExtractReasoningTextFromResponses(resp)
+
+	usage := UsageFromResponsesUsage(resp.Usage)
+
+	created := resp.CreatedAt
+
+	var toolCalls []dto.ToolCallResponse
+	if len(resp.Output) > 0 {
+		for _, out := range resp.Output {
+			if !isResponsesToolOutputType(out.Type) {
+				continue
+			}
+			name := strings.TrimSpace(out.Name)
+			if name == "" {
+				continue
+			}
+			callId := strings.TrimSpace(out.CallId)
+			if callId == "" {
+				callId = strings.TrimSpace(out.ID)
+			}
+			toolCalls = append(toolCalls, dto.ToolCallResponse{
+				ID:   callId,
+				Type: "function",
+				Function: dto.FunctionResponse{
+					Name:      name,
+					Arguments: out.ArgumentsString(),
+				},
+			})
+		}
+	}
+
+	finishReason := "stop"
+	if mappedReason, ok := ResponsesFinishReasonFromStatus(resp); ok {
+		finishReason = mappedReason
+	} else if len(toolCalls) > 0 {
+		finishReason = "tool_calls"
+	}
+
+	msg := dto.Message{
+		Role:    "assistant",
+		Content: text,
+	}
+	if reasoning != "" {
+		msg.ReasoningContent = &reasoning
+	}
+	if len(toolCalls) > 0 {
+		msg.SetToolCalls(toolCalls)
+	}
+
+	out := &dto.OpenAITextResponse{
+		Id:      id,
+		Object:  "chat.completion",
+		Created: created,
+		Model:   resp.Model,
+		Choices: []dto.OpenAITextResponseChoice{
+			{
+				Index:        0,
+				Message:      msg,
+				FinishReason: finishReason,
+			},
+		},
+		Usage: *usage,
+	}
+
+	return out, usage, nil
+}
+
+func UsageFromResponsesUsage(src *dto.Usage) *dto.Usage {
+	usage := &dto.Usage{}
+	if src == nil {
+		return usage
+	}
+	usage.UsageSemantic = src.UsageSemantic
+	usage.UsageSource = src.UsageSource
+	usage.BillingUsage = dto.CloneBillingUsage(src.BillingUsage)
+	if usage.BillingUsage == nil {
+		usage.BillingUsage = dto.NewOpenAIResponsesBillingUsage(src)
+	}
+	usage.Cost = src.Cost
+	if src.InputTokens != 0 {
+		usage.PromptTokens = src.InputTokens
+		usage.InputTokens = src.InputTokens
+	}
+	if src.OutputTokens != 0 {
+		usage.CompletionTokens = src.OutputTokens
+		usage.OutputTokens = src.OutputTokens
+	}
+	if src.TotalTokens != 0 {
+		usage.TotalTokens = src.TotalTokens
+	} else {
+		usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	}
+	if src.InputTokensDetails != nil {
+		usage.PromptTokensDetails.CachedTokens = src.InputTokensDetails.CachedTokens
+		usage.PromptTokensDetails.CachedCreationTokens = src.InputTokensDetails.CachedCreationTokens
+		usage.PromptTokensDetails.CacheWriteTokens = src.InputTokensDetails.CacheWriteTokens
+		usage.PromptTokensDetails.TextTokens = src.InputTokensDetails.TextTokens
+		usage.PromptTokensDetails.ImageTokens = src.InputTokensDetails.ImageTokens
+		usage.PromptTokensDetails.AudioTokens = src.InputTokensDetails.AudioTokens
+	}
+	if src.CompletionTokenDetails.ReasoningTokens != 0 ||
+		src.CompletionTokenDetails.TextTokens != 0 ||
+		src.CompletionTokenDetails.AudioTokens != 0 ||
+		src.CompletionTokenDetails.ImageTokens != 0 {
+		usage.CompletionTokenDetails.ReasoningTokens = src.CompletionTokenDetails.ReasoningTokens
+		usage.CompletionTokenDetails.TextTokens = src.CompletionTokenDetails.TextTokens
+		usage.CompletionTokenDetails.AudioTokens = src.CompletionTokenDetails.AudioTokens
+		usage.CompletionTokenDetails.ImageTokens = src.CompletionTokenDetails.ImageTokens
+	}
+	usage.ClaudeCacheCreation5mTokens = src.ClaudeCacheCreation5mTokens
+	usage.ClaudeCacheCreation1hTokens = src.ClaudeCacheCreation1hTokens
+	return usage
+}
+
+func ExtractOutputTextFromResponses(resp *dto.OpenAIResponsesResponse) string {
+	if resp == nil || len(resp.Output) == 0 {
+		return ""
+	}
+
+	var sb strings.Builder
+
+	// Prefer assistant message outputs.
+	for _, out := range resp.Output {
+		if out.Type != "message" {
+			continue
+		}
+		if out.Role != "" && out.Role != "assistant" {
+			continue
+		}
+		for _, c := range out.Content {
+			if c.Type == "output_text" && c.Text != "" {
+				sb.WriteString(c.Text)
+			}
+		}
+	}
+	if sb.Len() > 0 {
+		return sb.String()
+	}
+	for _, out := range resp.Output {
+		for _, c := range out.Content {
+			if c.Text != "" {
+				sb.WriteString(c.Text)
+			}
+		}
+	}
+	return sb.String()
+}
+
+func ExtractReasoningTextFromResponses(resp *dto.OpenAIResponsesResponse) string {
+	if resp == nil || len(resp.Output) == 0 {
+		return ""
+	}
+
+	var sb strings.Builder
+	for _, out := range resp.Output {
+		if out.Type != responsesOutputTypeReasoning {
+			continue
+		}
+		for _, c := range out.Content {
+			if c.Text != "" {
+				sb.WriteString(c.Text)
+			}
+		}
+	}
+	return sb.String()
+}
+
+func responseStatusString(resp *dto.OpenAIResponsesResponse) string {
+	if resp == nil || len(resp.Status) == 0 {
+		return ""
+	}
+	var status string
+	_ = kitutil.Unmarshal(resp.Status, &status)
+	return strings.TrimSpace(status)
+}
+
+func ensureIncompleteResponse(resp *dto.OpenAIResponsesResponse) *dto.OpenAIResponsesResponse {
+	if resp == nil {
+		resp = &dto.OpenAIResponsesResponse{}
+	}
+	if len(resp.Status) == 0 {
+		resp.Status = []byte(`"incomplete"`)
+	}
+	return resp
+}
+
+func isResponsesToolOutputType(outputType string) bool {
+	return outputType == responsesOutputTypeFunctionCall || outputType == responsesOutputTypeCustomToolCall
+}
+
+func responseStreamEventItemID(event *dto.ResponsesStreamResponse) string {
+	if event == nil {
+		return ""
+	}
+	if event.Item != nil {
+		if itemID := strings.TrimSpace(event.Item.ID); itemID != "" {
+			return itemID
+		}
+	}
+	return strings.TrimSpace(event.ItemID)
+}
+
+func fallbackToolKey(itemID string, callID string, outputIndex *int) string {
+	if outputIndex != nil {
+		return fmt.Sprintf("output:%d", *outputIndex)
+	}
+	if strings.TrimSpace(itemID) != "" {
+		return "item:" + strings.TrimSpace(itemID)
+	}
+	if strings.TrimSpace(callID) != "" {
+		return "call:" + strings.TrimSpace(callID)
+	}
+	return ""
+}
+
+func fallbackCallID(event *dto.ResponsesStreamResponse) string {
+	if event == nil {
+		return ""
+	}
+	if strings.TrimSpace(event.ItemID) != "" {
+		return strings.TrimSpace(event.ItemID)
+	}
+	if event.OutputIndex != nil {
+		return fmt.Sprintf("call_output_%d", *event.OutputIndex)
+	}
+	return ""
+}

relaykit/relayconvert/internal/oai_responses/to_oai_chat_resp_test.go

@@ -0,0 +1,448 @@
+package oairesponses
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestResponsesResponseToChatCompletionsPreservesTextAndToolCalls(t *testing.T) {
+	resp := &dto.OpenAIResponsesResponse{
+		ID:        "resp_1",
+		CreatedAt: 123,
+		Model:     "gpt-test",
+		Status:    []byte(`"completed"`),
+		Output: []dto.ResponsesOutput{
+			{
+				Type: responsesOutputTypeMessage,
+				Role: "assistant",
+				Content: []dto.ResponsesOutputContent{
+					{Type: "output_text", Text: "I will call a tool."},
+				},
+			},
+			{
+				Type:      responsesOutputTypeFunctionCall,
+				ID:        "fc_1",
+				CallId:    "call_1",
+				Name:      "lookup",
+				Arguments: []byte(`{"q":"x"}`),
+			},
+		},
+		Usage: &dto.Usage{InputTokens: 3, OutputTokens: 4, TotalTokens: 7},
+	}
+
+	chat, usage, err := ResponsesResponseToChatCompletionsResponse(resp, "chatcmpl_1")
+	require.NoError(t, err)
+	require.NotNil(t, usage)
+
+	require.Len(t, chat.Choices, 1)
+	assert.Equal(t, "tool_calls", chat.Choices[0].FinishReason)
+	assert.Equal(t, "I will call a tool.", chat.Choices[0].Message.StringContent())
+	toolCalls := chat.Choices[0].Message.ParseToolCalls()
+	require.Len(t, toolCalls, 1)
+	assert.Equal(t, "call_1", toolCalls[0].ID)
+	assert.Equal(t, "lookup", toolCalls[0].Function.Name)
+	assert.Equal(t, `{"q":"x"}`, toolCalls[0].Function.Arguments)
+	assert.Equal(t, 7, usage.TotalTokens)
+}
+
+func TestResponsesResponseToChatCompletionsPreservesReasoningSummary(t *testing.T) {
+	resp := &dto.OpenAIResponsesResponse{
+		ID:     "resp_1",
+		Model:  "gpt-test",
+		Status: []byte(`"completed"`),
+		Output: []dto.ResponsesOutput{
+			{
+				Type: responsesOutputTypeReasoning,
+				Content: []dto.ResponsesOutputContent{
+					{Type: "summary_text", Text: "first summary"},
+					{Type: "summary_text", Text: "\n\nsecond summary"},
+				},
+			},
+			{
+				Type: responsesOutputTypeMessage,
+				Role: "assistant",
+				Content: []dto.ResponsesOutputContent{
+					{Type: "output_text", Text: "final"},
+				},
+			},
+		},
+	}
+
+	chat, _, err := ResponsesResponseToChatCompletionsResponse(resp, "chatcmpl_1")
+	require.NoError(t, err)
+	assert.Equal(t, "first summary\n\nsecond summary", chat.Choices[0].Message.GetReasoningContent())
+	assert.Equal(t, "final", chat.Choices[0].Message.StringContent())
+}
+
+func TestResponsesFinishReasonFromIncompleteStatus(t *testing.T) {
+	tests := []struct {
+		name   string
+		reason string
+		want   string
+	}{
+		{name: "max output", reason: responsesIncompleteReasonMaxTokens, want: "length"},
+		{name: "content filter", reason: responsesIncompleteReasonContentFilter, want: "content_filter"},
+		{name: "unknown", reason: "other", want: "length"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, ok := ResponsesFinishReasonFromStatus(&dto.OpenAIResponsesResponse{
+				Status:            []byte(`"incomplete"`),
+				IncompleteDetails: &dto.IncompleteDetails{Reason: tt.reason},
+			})
+			require.True(t, ok)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestResponsesStreamEventToChatChunksUsesOutputIndexForToolArguments(t *testing.T) {
+	state := newTestResponsesStreamState()
+	outputIndex := 1
+
+	var chunks []dto.ChatCompletionsStreamResponse
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{Type: responsesEventCreated})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{Type: responsesEventOutputTextDelta, Delta: "text before tool"})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventFunctionArgsDelta,
+		OutputIndex: &outputIndex,
+		Delta:       `{"cmd":"ls"}`,
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		Item: &dto.ResponsesOutput{
+			Type:   responsesOutputTypeFunctionCall,
+			ID:     "fc_1",
+			CallId: "call_1",
+			Name:   "exec",
+		},
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type: responsesEventCompleted,
+		Response: &dto.OpenAIResponsesResponse{
+			Status: []byte(`"completed"`),
+			Usage:  &dto.Usage{InputTokens: 1, OutputTokens: 2, TotalTokens: 3},
+		},
+	})...)
+
+	require.Len(t, chunks, 4)
+	assert.Equal(t, "assistant", chunks[0].Choices[0].Delta.Role)
+	assert.Equal(t, "text before tool", chunks[1].Choices[0].Delta.GetContentString())
+	tool := chunks[2].Choices[0].Delta.ToolCalls[0]
+	require.NotNil(t, tool.Index)
+	assert.Equal(t, 0, *tool.Index)
+	assert.Equal(t, "call_1", tool.ID)
+	assert.Equal(t, "exec", tool.Function.Name)
+	assert.Equal(t, `{"cmd":"ls"}`, tool.Function.Arguments)
+	require.NotNil(t, chunks[3].Choices[0].FinishReason)
+	assert.Equal(t, "tool_calls", *chunks[3].Choices[0].FinishReason)
+	assert.Equal(t, 3, state.Usage.TotalTokens)
+}
+
+func TestResponsesStreamEventToChatChunksDoesNotDuplicatePendingArgsWithOutputIndexAndItemID(t *testing.T) {
+	state := newTestResponsesStreamState()
+	outputIndex := 1
+
+	var chunks []dto.ChatCompletionsStreamResponse
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{Type: responsesEventCreated})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventFunctionArgsDelta,
+		OutputIndex: &outputIndex,
+		ItemID:      "fc_1",
+		Delta:       `{"q":"x"}`,
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		ItemID:      "fc_1",
+		Item: &dto.ResponsesOutput{
+			Type:   responsesOutputTypeFunctionCall,
+			ID:     "fc_1",
+			CallId: "call_1",
+			Name:   "lookup",
+		},
+	})...)
+
+	require.Len(t, chunks, 2)
+	tool := chunks[1].Choices[0].Delta.ToolCalls[0]
+	assert.Equal(t, "call_1", tool.ID)
+	assert.Equal(t, "lookup", tool.Function.Name)
+	assert.Equal(t, `{"q":"x"}`, tool.Function.Arguments)
+	assert.Empty(t, state.pendingArgsByOutputIndex)
+	assert.Empty(t, state.pendingArgsByItemID)
+}
+
+func TestResponsesStreamEventToChatChunksDrainsItemOnlyPendingArgsWhenOutputIndexArrives(t *testing.T) {
+	state := newTestResponsesStreamState()
+	outputIndex := 1
+
+	var chunks []dto.ChatCompletionsStreamResponse
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{Type: responsesEventCreated})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:   responsesEventFunctionArgsDelta,
+		ItemID: "fc_1",
+		Delta:  `{"q":"x"}`,
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		ItemID:      "fc_1",
+		Item: &dto.ResponsesOutput{
+			Type:   responsesOutputTypeFunctionCall,
+			CallId: "call_1",
+			Name:   "lookup",
+		},
+	})...)
+
+	require.Len(t, chunks, 2)
+	tool := chunks[1].Choices[0].Delta.ToolCalls[0]
+	assert.Equal(t, "call_1", tool.ID)
+	assert.Equal(t, "lookup", tool.Function.Name)
+	assert.Equal(t, `{"q":"x"}`, tool.Function.Arguments)
+	assert.Empty(t, state.pendingArgsByOutputIndex)
+	assert.Empty(t, state.pendingArgsByItemID)
+}
+
+func TestResponsesStreamEventToChatChunksCustomToolAndReasoning(t *testing.T) {
+	state := newTestResponsesStreamState()
+	outputIndex := 0
+
+	chunks := mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:  responsesEventReasoningTextDelta,
+		Delta: "thinking",
+	})
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		Item: &dto.ResponsesOutput{
+			Type:   responsesOutputTypeCustomToolCall,
+			ID:     "ct_1",
+			CallId: "call_custom",
+			Name:   "apply_patch",
+		},
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventCustomToolInputDelta,
+		OutputIndex: &outputIndex,
+		Delta:       "patch body",
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type: responsesEventIncomplete,
+		Response: &dto.OpenAIResponsesResponse{
+			IncompleteDetails: &dto.IncompleteDetails{Reason: responsesIncompleteReasonContentFilter},
+		},
+	})...)
+
+	require.Len(t, chunks, 5)
+	assert.Equal(t, "thinking", chunks[1].Choices[0].Delta.GetReasoningContent())
+	assert.Equal(t, "apply_patch", chunks[2].Choices[0].Delta.ToolCalls[0].Function.Name)
+	assert.Equal(t, "patch body", chunks[3].Choices[0].Delta.ToolCalls[0].Function.Arguments)
+	require.NotNil(t, chunks[4].Choices[0].FinishReason)
+	assert.Equal(t, "content_filter", *chunks[4].Choices[0].FinishReason)
+}
+
+func TestResponsesStreamEventToChatChunksUsesTerminalDoneOutput(t *testing.T) {
+	state := newTestResponsesStreamState()
+	chunks := mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type: responsesEventDone,
+		Response: &dto.OpenAIResponsesResponse{
+			Status: []byte(`"completed"`),
+			Output: []dto.ResponsesOutput{
+				{
+					Type: responsesOutputTypeMessage,
+					Role: "assistant",
+					Content: []dto.ResponsesOutputContent{
+						{Type: "output_text", Text: "terminal text"},
+					},
+				},
+				{
+					Type:      responsesOutputTypeFunctionCall,
+					ID:        "fc_1",
+					CallId:    "call_1",
+					Name:      "lookup",
+					Arguments: []byte(`{"q":"x"}`),
+				},
+			},
+		},
+	})
+
+	require.Len(t, chunks, 4)
+	assert.Equal(t, "assistant", chunks[0].Choices[0].Delta.Role)
+	assert.Equal(t, "terminal text", chunks[1].Choices[0].Delta.GetContentString())
+	tool := chunks[2].Choices[0].Delta.ToolCalls[0]
+	assert.Equal(t, "lookup", tool.Function.Name)
+	assert.Equal(t, `{"q":"x"}`, tool.Function.Arguments)
+	require.NotNil(t, chunks[3].Choices[0].FinishReason)
+	assert.Equal(t, "tool_calls", *chunks[3].Choices[0].FinishReason)
+}
+
+func TestResponsesStreamEventToChatChunksDoesNotResendToolOnTerminalOutput(t *testing.T) {
+	state := newTestResponsesStreamState()
+	outputIndex := 0
+
+	var chunks []dto.ChatCompletionsStreamResponse
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{Type: responsesEventCreated})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		Item: &dto.ResponsesOutput{
+			Type:   responsesOutputTypeFunctionCall,
+			ID:     "fc_1",
+			CallId: "call_1",
+			Name:   "lookup",
+		},
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type:        responsesEventFunctionArgsDelta,
+		OutputIndex: &outputIndex,
+		Delta:       `{"q":"x"}`,
+	})...)
+	chunks = append(chunks, mustStreamChunks(t, state, &dto.ResponsesStreamResponse{
+		Type: responsesEventCompleted,
+		Response: &dto.OpenAIResponsesResponse{
+			Status: []byte(`"completed"`),
+			Output: []dto.ResponsesOutput{
+				{
+					Type:      responsesOutputTypeFunctionCall,
+					ID:        "fc_1",
+					CallId:    "call_1",
+					Name:      "lookup",
+					Arguments: []byte(`{"q":"x"}`),
+				},
+			},
+		},
+	})...)
+
+	totalArgs := ""
+	toolIndexes := map[int]bool{}
+	var finishReason string
+	for _, chunk := range chunks {
+		for _, choice := range chunk.Choices {
+			for _, tc := range choice.Delta.ToolCalls {
+				require.NotNil(t, tc.Index)
+				toolIndexes[*tc.Index] = true
+				totalArgs += tc.Function.Arguments
+			}
+			if choice.FinishReason != nil {
+				finishReason = *choice.FinishReason
+			}
+		}
+	}
+
+	assert.Equal(t, map[int]bool{0: true}, toolIndexes)
+	assert.Equal(t, `{"q":"x"}`, totalArgs)
+	assert.Equal(t, "tool_calls", finishReason)
+}
+
+func TestFinalizeResponsesToChatStreamFlushesPendingDeltaOnlyArguments(t *testing.T) {
+	state := newTestResponsesStreamState()
+	outputIndex := 2
+	_, err := ResponsesStreamEventToChatChunks(&dto.ResponsesStreamResponse{
+		Type:        responsesEventFunctionArgsDelta,
+		OutputIndex: &outputIndex,
+		Delta:       `{"pending":true}`,
+	}, state)
+	require.NoError(t, err)
+
+	chunks := FinalizeResponsesToChatStream(state)
+	require.Len(t, chunks, 3)
+	tool := chunks[1].Choices[0].Delta.ToolCalls[0]
+	assert.Equal(t, "call_output_2", tool.ID)
+	assert.Equal(t, `{"pending":true}`, tool.Function.Arguments)
+	require.NotNil(t, chunks[2].Choices[0].FinishReason)
+	assert.Equal(t, "tool_calls", *chunks[2].Choices[0].FinishReason)
+}
+
+func TestResponsesStreamEventToChatChunksFailedEventReturnsError(t *testing.T) {
+	_, err := ResponsesStreamEventToChatChunks(&dto.ResponsesStreamResponse{Type: responsesEventFailed}, newTestResponsesStreamState())
+	require.Error(t, err)
+}
+
+func TestResponsesBufferedAccumulatorSupplementsEmptyTerminalOutput(t *testing.T) {
+	acc := NewResponsesBufferedAccumulator()
+	outputIndex := 1
+	acc.ProcessEvent(&dto.ResponsesStreamResponse{Type: responsesEventOutputTextDelta, Delta: "buffered text"})
+	acc.ProcessEvent(&dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		Item: &dto.ResponsesOutput{
+			Type:   responsesOutputTypeFunctionCall,
+			ID:     "fc_1",
+			CallId: "call_1",
+			Name:   "lookup",
+		},
+	})
+	acc.ProcessEvent(&dto.ResponsesStreamResponse{
+		Type:        responsesEventFunctionArgsDelta,
+		OutputIndex: &outputIndex,
+		Delta:       `{"q":"x"}`,
+	})
+
+	resp := &dto.OpenAIResponsesResponse{
+		Status: []byte(`"completed"`),
+		Model:  "gpt-test",
+	}
+	acc.SupplementResponseOutput(resp)
+
+	chat, _, err := ResponsesResponseToChatCompletionsResponse(resp, "chatcmpl_1")
+	require.NoError(t, err)
+	assert.Equal(t, "buffered text", chat.Choices[0].Message.StringContent())
+	toolCalls := chat.Choices[0].Message.ParseToolCalls()
+	require.Len(t, toolCalls, 1)
+	assert.Equal(t, `{"q":"x"}`, toolCalls[0].Function.Arguments)
+}
+
+func TestResponsesBufferedAccumulatorDoesNotDuplicatePendingArgsWithOutputIndexAndItemID(t *testing.T) {
+	acc := NewResponsesBufferedAccumulator()
+	outputIndex := 1
+	acc.ProcessEvent(&dto.ResponsesStreamResponse{
+		Type:        responsesEventFunctionArgsDelta,
+		OutputIndex: &outputIndex,
+		ItemID:      "fc_1",
+		Delta:       `{"q":"x"}`,
+	})
+	acc.ProcessEvent(&dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		ItemID:      "fc_1",
+		Item: &dto.ResponsesOutput{
+			Type:   responsesOutputTypeFunctionCall,
+			ID:     "fc_1",
+			CallId: "call_1",
+			Name:   "lookup",
+		},
+	})
+
+	resp := &dto.OpenAIResponsesResponse{
+		Status: []byte(`"completed"`),
+		Model:  "gpt-test",
+	}
+	acc.SupplementResponseOutput(resp)
+
+	chat, _, err := ResponsesResponseToChatCompletionsResponse(resp, "chatcmpl_1")
+	require.NoError(t, err)
+	toolCalls := chat.Choices[0].Message.ParseToolCalls()
+	require.Len(t, toolCalls, 1)
+	assert.Equal(t, `{"q":"x"}`, toolCalls[0].Function.Arguments)
+	assert.Empty(t, acc.pendingByOutputIndex)
+	assert.Empty(t, acc.pendingByItemID)
+}
+
+func newTestResponsesStreamState() *ResponsesToChatStreamState {
+	state := NewResponsesToChatStreamState("gpt-test", false)
+	state.ID = "chatcmpl_test"
+	state.Created = 123
+	return state
+}
+
+func mustStreamChunks(t *testing.T, state *ResponsesToChatStreamState, event *dto.ResponsesStreamResponse) []dto.ChatCompletionsStreamResponse {
+	t.Helper()
+	chunks, err := ResponsesStreamEventToChatChunks(event, state)
+	require.NoError(t, err)
+	return chunks
+}

relaykit/relayconvert/internal/oai_responses/to_oai_chat_stream_resp.go

@@ -0,0 +1,719 @@
+package oairesponses
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+type ResponsesToChatStreamState struct {
+	ID           string
+	Model        string
+	Created      int64
+	IncludeUsage bool
+
+	Usage *dto.Usage
+
+	sentStart                  bool
+	finalized                  bool
+	hasSentText                bool
+	sawToolCall                bool
+	hasSentReasoning           bool
+	needsReasoningSummaryBreak bool
+	nextToolIndex              int
+	toolByKey                  map[string]*responsesStreamTool
+	outputIndexToKey           map[int]string
+	itemIDToKey                map[string]string
+	callIDToKey                map[string]string
+	pendingArgsByOutputIndex   map[int]string
+	pendingArgsByItemID        map[string]string
+	usageText                  strings.Builder
+}
+
+type responsesStreamTool struct {
+	Key        string
+	CallID     string
+	ItemID     string
+	Name       string
+	Arguments  string
+	Index      int
+	Sent       bool
+	NameSent   bool
+	ArgsSentAt int
+}
+
+func NewResponsesToChatStreamState(model string, includeUsage bool) *ResponsesToChatStreamState {
+	return &ResponsesToChatStreamState{
+		Model:                    model,
+		Created:                  time.Now().Unix(),
+		IncludeUsage:             includeUsage,
+		Usage:                    &dto.Usage{},
+		toolByKey:                make(map[string]*responsesStreamTool),
+		outputIndexToKey:         make(map[int]string),
+		itemIDToKey:              make(map[string]string),
+		callIDToKey:              make(map[string]string),
+		pendingArgsByOutputIndex: make(map[int]string),
+		pendingArgsByItemID:      make(map[string]string),
+	}
+}
+
+func (s *ResponsesToChatStreamState) UsageText() string {
+	if s == nil {
+		return ""
+	}
+	return s.usageText.String()
+}
+
+func ResponsesStreamEventToChatChunks(event *dto.ResponsesStreamResponse, state *ResponsesToChatStreamState) ([]dto.ChatCompletionsStreamResponse, error) {
+	if event == nil || state == nil {
+		return nil, nil
+	}
+
+	switch event.Type {
+	case responsesEventCreated:
+		state.applyResponseMetadata(event.Response)
+		return state.ensureStart(), nil
+	case responsesEventReasoningSummaryDelta, responsesEventReasoningTextDelta:
+		return state.reasoningDelta(event.Delta), nil
+	case responsesEventReasoningSummaryDone, responsesEventReasoningTextDone:
+		if state.hasSentReasoning {
+			state.needsReasoningSummaryBreak = true
+		}
+		return nil, nil
+	case responsesEventOutputTextDelta:
+		return state.textDelta(event.Delta), nil
+	case responsesEventOutputItemAdded, responsesEventOutputItemDone:
+		if event.Item == nil || !isResponsesToolOutputType(event.Item.Type) {
+			return nil, nil
+		}
+		return state.toolItem(event), nil
+	case responsesEventFunctionArgsDelta, responsesEventCustomToolInputDelta:
+		return state.toolArgumentsDelta(event), nil
+	case responsesEventFunctionArgsDone, responsesEventCustomToolInputDone:
+		return state.flushPendingTool(event), nil
+	case responsesEventCompleted, responsesEventDone, responsesEventIncomplete:
+		response := event.Response
+		if event.Type == responsesEventIncomplete {
+			response = ensureIncompleteResponse(response)
+		}
+		state.applyResponseMetadata(response)
+		chunks := state.terminalOutputChunks(response)
+		chunks = append(chunks, state.finalize(response)...)
+		return chunks, nil
+	case responsesEventFailed, responsesEventError:
+		return nil, fmt.Errorf("responses stream error: %s", event.Type)
+	default:
+		return nil, nil
+	}
+}
+
+func FinalizeResponsesToChatStream(state *ResponsesToChatStreamState) []dto.ChatCompletionsStreamResponse {
+	if state == nil {
+		return nil
+	}
+	return state.finalize(nil)
+}
+
+func (s *ResponsesToChatStreamState) applyResponseMetadata(response *dto.OpenAIResponsesResponse) {
+	if response == nil {
+		return
+	}
+	if response.ID != "" && s.ID == "" {
+		s.ID = response.ID
+	}
+	if response.Model != "" {
+		s.Model = response.Model
+	}
+	if response.CreatedAt != 0 {
+		s.Created = int64(response.CreatedAt)
+	}
+	if response.Usage != nil {
+		s.Usage = UsageFromResponsesUsage(response.Usage)
+	}
+}
+
+func (s *ResponsesToChatStreamState) ensureStart() []dto.ChatCompletionsStreamResponse {
+	if s.sentStart {
+		return nil
+	}
+	s.sentStart = true
+	return []dto.ChatCompletionsStreamResponse{s.makeChunk(dto.ChatCompletionsStreamResponseChoiceDelta{
+		Role:    "assistant",
+		Content: kitutil.GetPointer(""),
+	}, nil)}
+}
+
+func (s *ResponsesToChatStreamState) textDelta(delta string) []dto.ChatCompletionsStreamResponse {
+	if delta == "" {
+		return nil
+	}
+	s.usageText.WriteString(delta)
+	s.hasSentText = true
+	chunks := s.ensureStart()
+	chunks = append(chunks, s.makeChunk(dto.ChatCompletionsStreamResponseChoiceDelta{
+		Content: &delta,
+	}, nil))
+	return chunks
+}
+
+func (s *ResponsesToChatStreamState) terminalOutputChunks(response *dto.OpenAIResponsesResponse) []dto.ChatCompletionsStreamResponse {
+	if s == nil || response == nil || len(response.Output) == 0 {
+		return nil
+	}
+
+	var chunks []dto.ChatCompletionsStreamResponse
+	for i := range response.Output {
+		out := &response.Output[i]
+		switch {
+		case out.Type == responsesOutputTypeMessage && !s.hasSentText:
+			var text strings.Builder
+			for _, c := range out.Content {
+				if c.Type == "output_text" && c.Text != "" {
+					text.WriteString(c.Text)
+				}
+			}
+			chunks = append(chunks, s.textDelta(text.String())...)
+		case out.Type == responsesOutputTypeReasoning && !s.hasSentReasoning:
+			var reasoning strings.Builder
+			for _, c := range out.Content {
+				if c.Text != "" {
+					reasoning.WriteString(c.Text)
+				}
+			}
+			chunks = append(chunks, s.reasoningDelta(reasoning.String())...)
+		case isResponsesToolOutputType(out.Type):
+			chunks = append(chunks, s.toolItem(&dto.ResponsesStreamResponse{Item: out})...)
+		}
+	}
+	return chunks
+}
+
+func (s *ResponsesToChatStreamState) reasoningDelta(delta string) []dto.ChatCompletionsStreamResponse {
+	if delta == "" {
+		return nil
+	}
+	if s.needsReasoningSummaryBreak {
+		if strings.HasPrefix(delta, "\n\n") {
+			s.needsReasoningSummaryBreak = false
+		} else if strings.HasPrefix(delta, "\n") {
+			delta = "\n" + delta
+			s.needsReasoningSummaryBreak = false
+		} else {
+			delta = "\n\n" + delta
+			s.needsReasoningSummaryBreak = false
+		}
+	}
+	s.usageText.WriteString(delta)
+	chunks := s.ensureStart()
+	chunks = append(chunks, s.makeChunk(dto.ChatCompletionsStreamResponseChoiceDelta{
+		ReasoningContent: &delta,
+	}, nil))
+	s.hasSentReasoning = true
+	return chunks
+}
+
+func (s *ResponsesToChatStreamState) toolItem(event *dto.ResponsesStreamResponse) []dto.ChatCompletionsStreamResponse {
+	tool := s.ensureToolForEvent(event)
+	if tool == nil {
+		return nil
+	}
+	args := event.Item.ArgumentsString()
+	if args != "" {
+		tool.Arguments = args
+	}
+	return s.toolDelta(tool, "")
+}
+
+func (s *ResponsesToChatStreamState) toolArgumentsDelta(event *dto.ResponsesStreamResponse) []dto.ChatCompletionsStreamResponse {
+	if event.Delta == "" {
+		return nil
+	}
+	tool := s.findToolForEvent(event)
+	if tool == nil {
+		if event.OutputIndex != nil {
+			s.pendingArgsByOutputIndex[*event.OutputIndex] += event.Delta
+		} else if itemID := strings.TrimSpace(event.ItemID); itemID != "" {
+			s.pendingArgsByItemID[itemID] += event.Delta
+		}
+		return nil
+	}
+	tool.Arguments += event.Delta
+	return s.toolDelta(tool, event.Delta)
+}
+
+func (s *ResponsesToChatStreamState) flushPendingTool(event *dto.ResponsesStreamResponse) []dto.ChatCompletionsStreamResponse {
+	tool := s.findToolForEvent(event)
+	if tool == nil {
+		tool = s.ensureFallbackToolForEvent(event)
+	}
+	if tool == nil {
+		return nil
+	}
+	return s.toolDelta(tool, "")
+}
+
+func (s *ResponsesToChatStreamState) ensureToolForEvent(event *dto.ResponsesStreamResponse) *responsesStreamTool {
+	if event == nil || event.Item == nil {
+		return nil
+	}
+	key := s.keyForEvent(event)
+	if key == "" {
+		key = fallbackToolKey(event.Item.ID, event.Item.CallId, event.OutputIndex)
+	}
+	if key == "" {
+		return nil
+	}
+
+	tool := s.toolByKey[key]
+	if tool == nil {
+		if itemID := responseStreamEventItemID(event); itemID != "" {
+			if existingKey := s.itemIDToKey[itemID]; existingKey != "" {
+				tool = s.toolByKey[existingKey]
+			}
+		}
+		if tool == nil {
+			if callID := strings.TrimSpace(event.Item.CallId); callID != "" {
+				if existingKey := s.callIDToKey[callID]; existingKey != "" {
+					tool = s.toolByKey[existingKey]
+				}
+			}
+		}
+		if tool != nil {
+			s.toolByKey[key] = tool
+		}
+	}
+	if tool == nil {
+		tool = &responsesStreamTool{Key: key, Index: s.nextToolIndex}
+		s.nextToolIndex++
+		s.toolByKey[key] = tool
+	}
+
+	if event.OutputIndex != nil {
+		s.outputIndexToKey[*event.OutputIndex] = key
+		if pending := s.pendingArgsByOutputIndex[*event.OutputIndex]; pending != "" {
+			tool.Arguments += pending
+			delete(s.pendingArgsByOutputIndex, *event.OutputIndex)
+		}
+	}
+	if itemID := responseStreamEventItemID(event); itemID != "" {
+		tool.ItemID = itemID
+		s.itemIDToKey[itemID] = key
+		if pending := s.pendingArgsByItemID[itemID]; pending != "" {
+			tool.Arguments += pending
+			delete(s.pendingArgsByItemID, itemID)
+		}
+	}
+	if callID := strings.TrimSpace(event.Item.CallId); callID != "" {
+		tool.CallID = callID
+		s.callIDToKey[callID] = key
+	} else if tool.CallID == "" {
+		tool.CallID = strings.TrimSpace(event.Item.ID)
+	}
+	if name := strings.TrimSpace(event.Item.Name); name != "" {
+		tool.Name = name
+	}
+	return tool
+}
+
+func (s *ResponsesToChatStreamState) findToolForEvent(event *dto.ResponsesStreamResponse) *responsesStreamTool {
+	if event == nil {
+		return nil
+	}
+	if event.OutputIndex != nil {
+		if key := s.outputIndexToKey[*event.OutputIndex]; key != "" {
+			return s.toolByKey[key]
+		}
+	}
+	if itemID := strings.TrimSpace(event.ItemID); itemID != "" {
+		if key := s.itemIDToKey[itemID]; key != "" {
+			return s.toolByKey[key]
+		}
+	}
+	if event.Item != nil {
+		if key := s.keyForEvent(event); key != "" {
+			return s.toolByKey[key]
+		}
+	}
+	return nil
+}
+
+func (s *ResponsesToChatStreamState) ensureFallbackToolForEvent(event *dto.ResponsesStreamResponse) *responsesStreamTool {
+	if event == nil {
+		return nil
+	}
+	key := ""
+	if event.OutputIndex != nil {
+		key = fmt.Sprintf("output:%d", *event.OutputIndex)
+	}
+	if key == "" && strings.TrimSpace(event.ItemID) != "" {
+		key = "item:" + strings.TrimSpace(event.ItemID)
+	}
+	if key == "" {
+		return nil
+	}
+	tool := s.toolByKey[key]
+	if tool == nil {
+		tool = &responsesStreamTool{
+			Key:    key,
+			Index:  s.nextToolIndex,
+			CallID: fallbackCallID(event),
+		}
+		s.nextToolIndex++
+		s.toolByKey[key] = tool
+	}
+	if event.OutputIndex != nil {
+		s.outputIndexToKey[*event.OutputIndex] = key
+		if pending := s.pendingArgsByOutputIndex[*event.OutputIndex]; pending != "" {
+			tool.Arguments += pending
+			delete(s.pendingArgsByOutputIndex, *event.OutputIndex)
+		}
+	}
+	if itemID := responseStreamEventItemID(event); itemID != "" {
+		tool.ItemID = itemID
+		s.itemIDToKey[itemID] = key
+		if pending := s.pendingArgsByItemID[itemID]; pending != "" {
+			tool.Arguments += pending
+			delete(s.pendingArgsByItemID, itemID)
+		}
+	}
+	return tool
+}
+
+func (s *ResponsesToChatStreamState) toolDelta(tool *responsesStreamTool, explicitDelta string) []dto.ChatCompletionsStreamResponse {
+	if tool == nil {
+		return nil
+	}
+
+	argsDelta := explicitDelta
+	if argsDelta == "" && len(tool.Arguments) > tool.ArgsSentAt {
+		argsDelta = tool.Arguments[tool.ArgsSentAt:]
+	}
+	if tool.Sent && argsDelta == "" && (tool.Name == "" || tool.NameSent) {
+		return nil
+	}
+
+	chunks := s.ensureStart()
+	callID := strings.TrimSpace(tool.CallID)
+	if callID == "" {
+		callID = tool.Key
+	}
+	responseTool := dto.ToolCallResponse{
+		ID:   callID,
+		Type: "function",
+		Function: dto.FunctionResponse{
+			Arguments: argsDelta,
+		},
+	}
+	responseTool.SetIndex(tool.Index)
+	if !tool.NameSent && tool.Name != "" {
+		responseTool.Function.Name = tool.Name
+		tool.NameSent = true
+	}
+	if !tool.Sent {
+		tool.Sent = true
+	}
+	if argsDelta != "" {
+		tool.ArgsSentAt += len(argsDelta)
+		s.usageText.WriteString(argsDelta)
+	}
+	if responseTool.Function.Name != "" {
+		s.usageText.WriteString(responseTool.Function.Name)
+	}
+
+	chunks = append(chunks, s.makeChunk(dto.ChatCompletionsStreamResponseChoiceDelta{
+		ToolCalls: []dto.ToolCallResponse{responseTool},
+	}, nil))
+	s.sawToolCall = true
+	return chunks
+}
+
+func (s *ResponsesToChatStreamState) finalize(response *dto.OpenAIResponsesResponse) []dto.ChatCompletionsStreamResponse {
+	if s.finalized {
+		return nil
+	}
+	s.finalized = true
+
+	chunks := s.flushAllPendingTools()
+	chunks = append(chunks, s.ensureStart()...)
+
+	finishReason := "stop"
+	if mappedReason, ok := ResponsesFinishReasonFromStatus(response); ok {
+		finishReason = mappedReason
+	} else if s.sawToolCall {
+		finishReason = "tool_calls"
+	}
+	chunks = append(chunks, s.makeChunk(dto.ChatCompletionsStreamResponseChoiceDelta{}, &finishReason))
+	if s.IncludeUsage && s.Usage != nil {
+		chunks = append(chunks, dto.ChatCompletionsStreamResponse{
+			Id:      s.ID,
+			Object:  "chat.completion.chunk",
+			Created: s.Created,
+			Model:   s.Model,
+			Choices: make([]dto.ChatCompletionsStreamResponseChoice, 0),
+			Usage:   s.Usage,
+		})
+	}
+	return chunks
+}
+
+func (s *ResponsesToChatStreamState) flushAllPendingTools() []dto.ChatCompletionsStreamResponse {
+	keys := make([]string, 0, len(s.toolByKey)+len(s.pendingArgsByOutputIndex)+len(s.pendingArgsByItemID))
+	seen := make(map[string]bool)
+	for key := range s.toolByKey {
+		keys = append(keys, key)
+		seen[key] = true
+	}
+	for outputIndex := range s.pendingArgsByOutputIndex {
+		key := fmt.Sprintf("output:%d", outputIndex)
+		if !seen[key] {
+			keys = append(keys, key)
+			seen[key] = true
+		}
+	}
+	for itemID := range s.pendingArgsByItemID {
+		key := "item:" + itemID
+		if !seen[key] {
+			keys = append(keys, key)
+			seen[key] = true
+		}
+	}
+	sort.Strings(keys)
+
+	var chunks []dto.ChatCompletionsStreamResponse
+	for _, key := range keys {
+		tool := s.toolByKey[key]
+		if tool == nil {
+			callID := strings.TrimPrefix(key, "item:")
+			if strings.HasPrefix(key, "output:") {
+				callID = "call_output_" + strings.TrimPrefix(key, "output:")
+			}
+			tool = &responsesStreamTool{
+				Key:    key,
+				Index:  s.nextToolIndex,
+				CallID: callID,
+			}
+			s.nextToolIndex++
+			s.toolByKey[key] = tool
+		}
+		if strings.HasPrefix(key, "output:") {
+			var outputIndex int
+			if _, err := fmt.Sscanf(key, "output:%d", &outputIndex); err == nil {
+				tool.Arguments += s.pendingArgsByOutputIndex[outputIndex]
+				delete(s.pendingArgsByOutputIndex, outputIndex)
+			}
+		}
+		if strings.HasPrefix(key, "item:") {
+			itemID := strings.TrimPrefix(key, "item:")
+			tool.Arguments += s.pendingArgsByItemID[itemID]
+			delete(s.pendingArgsByItemID, itemID)
+		}
+		chunks = append(chunks, s.toolDelta(tool, "")...)
+	}
+	return chunks
+}
+
+func (s *ResponsesToChatStreamState) makeChunk(delta dto.ChatCompletionsStreamResponseChoiceDelta, finishReason *string) dto.ChatCompletionsStreamResponse {
+	return dto.ChatCompletionsStreamResponse{
+		Id:      s.ID,
+		Object:  "chat.completion.chunk",
+		Created: s.Created,
+		Model:   s.Model,
+		Choices: []dto.ChatCompletionsStreamResponseChoice{
+			{
+				Index:        0,
+				Delta:        delta,
+				FinishReason: finishReason,
+			},
+		},
+	}
+}
+
+func (s *ResponsesToChatStreamState) keyForEvent(event *dto.ResponsesStreamResponse) string {
+	if event == nil {
+		return ""
+	}
+	if event.OutputIndex != nil {
+		return fmt.Sprintf("output:%d", *event.OutputIndex)
+	}
+	if event.Item != nil {
+		if itemID := strings.TrimSpace(event.Item.ID); itemID != "" {
+			return "item:" + itemID
+		}
+		if callID := strings.TrimSpace(event.Item.CallId); callID != "" {
+			return "call:" + callID
+		}
+	}
+	if itemID := strings.TrimSpace(event.ItemID); itemID != "" {
+		return "item:" + itemID
+	}
+	return ""
+}
+
+type ResponsesBufferedAccumulator struct {
+	text                 strings.Builder
+	reasoning            strings.Builder
+	tools                []*responsesBufferedTool
+	outputIndexToToolIdx map[int]int
+	itemIDToToolIdx      map[string]int
+	pendingByOutputIndex map[int]string
+	pendingByItemID      map[string]string
+}
+
+type responsesBufferedTool struct {
+	CallID    string
+	ItemID    string
+	Name      string
+	Arguments strings.Builder
+}
+
+func NewResponsesBufferedAccumulator() *ResponsesBufferedAccumulator {
+	return &ResponsesBufferedAccumulator{
+		outputIndexToToolIdx: make(map[int]int),
+		itemIDToToolIdx:      make(map[string]int),
+		pendingByOutputIndex: make(map[int]string),
+		pendingByItemID:      make(map[string]string),
+	}
+}
+
+func (a *ResponsesBufferedAccumulator) ProcessEvent(event *dto.ResponsesStreamResponse) {
+	if a == nil || event == nil {
+		return
+	}
+	switch event.Type {
+	case responsesEventOutputTextDelta:
+		a.text.WriteString(event.Delta)
+	case responsesEventReasoningSummaryDelta, responsesEventReasoningTextDelta:
+		a.reasoning.WriteString(event.Delta)
+	case responsesEventOutputItemAdded, responsesEventOutputItemDone:
+		if event.Item != nil && isResponsesToolOutputType(event.Item.Type) {
+			tool := a.ensureTool(event)
+			if args := event.Item.ArgumentsString(); args != "" {
+				tool.Arguments.Reset()
+				tool.Arguments.WriteString(args)
+			}
+		}
+	case responsesEventFunctionArgsDelta, responsesEventCustomToolInputDelta:
+		if idx, ok := a.findToolIndex(event); ok {
+			a.tools[idx].Arguments.WriteString(event.Delta)
+			return
+		}
+		if event.OutputIndex != nil {
+			a.pendingByOutputIndex[*event.OutputIndex] += event.Delta
+		} else if itemID := strings.TrimSpace(event.ItemID); itemID != "" {
+			a.pendingByItemID[itemID] += event.Delta
+		}
+	}
+}
+
+func (a *ResponsesBufferedAccumulator) SupplementResponseOutput(resp *dto.OpenAIResponsesResponse) {
+	if a == nil || resp == nil || len(resp.Output) > 0 {
+		return
+	}
+	resp.Output = a.BuildOutput()
+}
+
+func (a *ResponsesBufferedAccumulator) BuildOutput() []dto.ResponsesOutput {
+	if a == nil {
+		return nil
+	}
+	out := make([]dto.ResponsesOutput, 0, 2+len(a.tools))
+	if a.reasoning.Len() > 0 {
+		out = append(out, dto.ResponsesOutput{
+			Type: responsesOutputTypeReasoning,
+			Content: []dto.ResponsesOutputContent{
+				{Type: "summary_text", Text: a.reasoning.String()},
+			},
+		})
+	}
+	if a.text.Len() > 0 {
+		out = append(out, dto.ResponsesOutput{
+			Type: responsesOutputTypeMessage,
+			Role: "assistant",
+			Content: []dto.ResponsesOutputContent{
+				{Type: "output_text", Text: a.text.String()},
+			},
+		})
+	}
+	for _, tool := range a.tools {
+		if tool == nil {
+			continue
+		}
+		argsRaw, _ := kitutil.Marshal(tool.Arguments.String())
+		out = append(out, dto.ResponsesOutput{
+			Type:      responsesOutputTypeFunctionCall,
+			ID:        tool.ItemID,
+			CallId:    tool.CallID,
+			Name:      tool.Name,
+			Arguments: argsRaw,
+		})
+	}
+	return out
+}
+
+func (a *ResponsesBufferedAccumulator) ensureTool(event *dto.ResponsesStreamResponse) *responsesBufferedTool {
+	if idx, ok := a.findToolIndex(event); ok {
+		tool := a.tools[idx]
+		a.applyToolMetadata(tool, event)
+		return tool
+	}
+	tool := &responsesBufferedTool{}
+	a.applyToolMetadata(tool, event)
+	idx := len(a.tools)
+	a.tools = append(a.tools, tool)
+	if event.OutputIndex != nil {
+		a.outputIndexToToolIdx[*event.OutputIndex] = idx
+		if pending := a.pendingByOutputIndex[*event.OutputIndex]; pending != "" {
+			tool.Arguments.WriteString(pending)
+			delete(a.pendingByOutputIndex, *event.OutputIndex)
+		}
+	}
+	if tool.ItemID != "" {
+		a.itemIDToToolIdx[tool.ItemID] = idx
+		if pending := a.pendingByItemID[tool.ItemID]; pending != "" {
+			tool.Arguments.WriteString(pending)
+			delete(a.pendingByItemID, tool.ItemID)
+		}
+	}
+	return tool
+}
+
+func (a *ResponsesBufferedAccumulator) applyToolMetadata(tool *responsesBufferedTool, event *dto.ResponsesStreamResponse) {
+	if tool == nil || event == nil || event.Item == nil {
+		return
+	}
+	if itemID := strings.TrimSpace(event.Item.ID); itemID != "" {
+		tool.ItemID = itemID
+	}
+	if callID := strings.TrimSpace(event.Item.CallId); callID != "" {
+		tool.CallID = callID
+	} else if tool.CallID == "" {
+		tool.CallID = strings.TrimSpace(event.Item.ID)
+	}
+	if name := strings.TrimSpace(event.Item.Name); name != "" {
+		tool.Name = name
+	}
+}
+
+func (a *ResponsesBufferedAccumulator) findToolIndex(event *dto.ResponsesStreamResponse) (int, bool) {
+	if event == nil {
+		return 0, false
+	}
+	if event.OutputIndex != nil {
+		if idx, ok := a.outputIndexToToolIdx[*event.OutputIndex]; ok {
+			return idx, true
+		}
+	}
+	itemID := strings.TrimSpace(event.ItemID)
+	if itemID == "" && event.Item != nil {
+		itemID = strings.TrimSpace(event.Item.ID)
+	}
+	if itemID != "" {
+		idx, ok := a.itemIDToToolIdx[itemID]
+		return idx, ok
+	}
+	return 0, false
+}

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

@@ -0,0 +1,9 @@
+package claude
+
+func NormalizeCacheCreationSplit(totalTokens int, tokens5m int, tokens1h int) (int, int) {
+	remainder := totalTokens - tokens5m - tokens1h
+	if remainder < 0 {
+		remainder = 0
+	}
+	return tokens5m + remainder, tokens1h
+}

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

@@ -0,0 +1,10 @@
+package claude
+
+import "errors"
+
+// ErrMissingMaxTokens is returned when an OpenAI-format request carries no
+// usable max_tokens and no Options.Claude.DefaultMaxTokens hook is
+// configured. The Claude Messages API rejects requests without max_tokens
+// (400 "max_tokens: Field required"), so conversion fails loudly instead of
+// emitting a request the upstream is guaranteed to refuse.
+var ErrMissingMaxTokens = errors.New("claude messages request requires max_tokens: set max_tokens on the request or configure Options.Claude.DefaultMaxTokens")

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

@@ -0,0 +1,46 @@
+package claude
+
+import "github.com/QuantumNous/new-api/relaykit/dto"
+
+func MapOpenAIToolChoice(toolChoice any, parallelToolCalls *bool) *dto.ClaudeToolChoice {
+	var claudeToolChoice *dto.ClaudeToolChoice
+
+	if toolChoiceStr, ok := toolChoice.(string); ok {
+		switch toolChoiceStr {
+		case "auto":
+			claudeToolChoice = &dto.ClaudeToolChoice{
+				Type: "auto",
+			}
+		case "required":
+			claudeToolChoice = &dto.ClaudeToolChoice{
+				Type: "any",
+			}
+		case "none":
+			claudeToolChoice = &dto.ClaudeToolChoice{
+				Type: "none",
+			}
+		}
+	} else if toolChoiceMap, ok := toolChoice.(map[string]interface{}); ok {
+		if function, ok := toolChoiceMap["function"].(map[string]interface{}); ok {
+			if toolName, ok := function["name"].(string); ok {
+				claudeToolChoice = &dto.ClaudeToolChoice{
+					Type: "tool",
+					Name: toolName,
+				}
+			}
+		}
+	}
+
+	if parallelToolCalls != nil {
+		if claudeToolChoice == nil {
+			claudeToolChoice = &dto.ClaudeToolChoice{
+				Type: "auto",
+			}
+		}
+		if claudeToolChoice.Type != "none" {
+			claudeToolChoice.DisableParallelToolUse = !*parallelToolCalls
+		}
+	}
+
+	return claudeToolChoice
+}

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

@@ -0,0 +1,267 @@
+package gemini
+
+import (
+	"strconv"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/convmeta"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/reasoning"
+)
+
+var SupportedMimeTypes = map[string]bool{
+	"application/pdf": true,
+	"audio/mpeg":      true,
+	"audio/mp3":       true,
+	"audio/wav":       true,
+	"image/png":       true,
+	"image/jpeg":      true,
+	"image/jpg":       true,
+	"image/webp":      true,
+	"image/heic":      true,
+	"image/heif":      true,
+	"text/plain":      true,
+	"video/mov":       true,
+	"video/mpeg":      true,
+	"video/mp4":       true,
+	"video/mpg":       true,
+	"video/avi":       true,
+	"video/wmv":       true,
+	"video/mpegps":    true,
+	"video/flv":       true,
+}
+
+var SafetySettingCategories = []string{
+	"HARM_CATEGORY_HARASSMENT",
+	"HARM_CATEGORY_HATE_SPEECH",
+	"HARM_CATEGORY_SEXUALLY_EXPLICIT",
+	"HARM_CATEGORY_DANGEROUS_CONTENT",
+}
+
+const ThoughtSignatureBypassValue = "context_engineering_is_the_way_to_go"
+
+const (
+	pro25MinBudget       = 128
+	pro25MaxBudget       = 32768
+	flash25MaxBudget     = 24576
+	flash25LiteMinBudget = 512
+	flash25LiteMaxBudget = 24576
+)
+
+func ShouldAttachThoughtSignature(opts *convmeta.Options) bool {
+	return opts != nil && opts.Gemini.FunctionCallThoughtSignatureEnabled
+}
+
+func AttachThoughtSignatureBypass(opts *convmeta.Options, part *dto.GeminiPart) bool {
+	if part == nil || len(part.ThoughtSignature) > 0 || !ShouldAttachThoughtSignature(opts) {
+		return false
+	}
+	part.ThoughtSignature = []byte(strconv.Quote(ThoughtSignatureBypassValue))
+	return true
+}
+
+func AttachFunctionCallThoughtSignature(opts *convmeta.Options, part *dto.GeminiPart) bool {
+	if part == nil || !HasFunctionCallContent(part.FunctionCall) {
+		return false
+	}
+	return AttachThoughtSignatureBypass(opts, part)
+}
+
+func AttachFirstTextThoughtSignature(opts *convmeta.Options, parts []dto.GeminiPart) bool {
+	if !ShouldAttachThoughtSignature(opts) {
+		return false
+	}
+	for i := range parts {
+		if parts[i].Text != "" && len(parts[i].ThoughtSignature) == 0 {
+			parts[i].ThoughtSignature = []byte(strconv.Quote(ThoughtSignatureBypassValue))
+			return true
+		}
+	}
+	return false
+}
+
+func ApplyThinkingConfig(geminiRequest *dto.GeminiChatRequest, info convmeta.Meta, oaiRequest ...dto.GeneralOpenAIRequest) {
+	opts := convmeta.OptionsOf(info)
+	if geminiRequest == nil || info == nil || !opts.Gemini.ThinkingAdapterEnabled {
+		return
+	}
+
+	modelName := convmeta.UpstreamModelName(info)
+	isNew25Pro := strings.HasPrefix(modelName, "gemini-2.5-pro") &&
+		!strings.HasPrefix(modelName, "gemini-2.5-pro-preview-05-06") &&
+		!strings.HasPrefix(modelName, "gemini-2.5-pro-preview-03-25")
+
+	if strings.Contains(modelName, "-thinking-") {
+		parts := strings.SplitN(modelName, "-thinking-", 2)
+		if len(parts) == 2 && parts[1] != "" {
+			if budgetTokens, err := strconv.Atoi(parts[1]); err == nil {
+				clampedBudget := clampThinkingBudget(modelName, budgetTokens)
+				geminiRequest.GenerationConfig.ThinkingConfig = &dto.GeminiThinkingConfig{
+					ThinkingBudget:  kitutil.GetPointer(clampedBudget),
+					IncludeThoughts: true,
+				}
+			}
+		}
+	} else if strings.HasSuffix(modelName, "-thinking") {
+		unsupportedModels := []string{
+			"gemini-2.5-pro-preview-05-06",
+			"gemini-2.5-pro-preview-03-25",
+		}
+		isUnsupported := false
+		for _, unsupportedModel := range unsupportedModels {
+			if strings.HasPrefix(modelName, unsupportedModel) {
+				isUnsupported = true
+				break
+			}
+		}
+
+		if isUnsupported {
+			geminiRequest.GenerationConfig.ThinkingConfig = &dto.GeminiThinkingConfig{
+				IncludeThoughts: true,
+			}
+		} else {
+			geminiRequest.GenerationConfig.ThinkingConfig = &dto.GeminiThinkingConfig{
+				IncludeThoughts: true,
+			}
+			if geminiRequest.GenerationConfig.MaxOutputTokens != nil && *geminiRequest.GenerationConfig.MaxOutputTokens > 0 {
+				budgetTokens := opts.Gemini.ThinkingAdapterBudgetTokensPercentage * float64(*geminiRequest.GenerationConfig.MaxOutputTokens)
+				clampedBudget := clampThinkingBudget(modelName, int(budgetTokens))
+				geminiRequest.GenerationConfig.ThinkingConfig.ThinkingBudget = kitutil.GetPointer(clampedBudget)
+			} else if len(oaiRequest) > 0 {
+				geminiRequest.GenerationConfig.ThinkingConfig.ThinkingBudget = kitutil.GetPointer(clampThinkingBudgetByEffort(modelName, oaiRequest[0].ReasoningEffort))
+			}
+		}
+	} else if strings.HasSuffix(modelName, "-nothinking") {
+		if !isNew25Pro {
+			geminiRequest.GenerationConfig.ThinkingConfig = &dto.GeminiThinkingConfig{
+				ThinkingBudget: kitutil.GetPointer(0),
+			}
+		}
+	} else if _, level, ok := reasoning.TrimEffortSuffix(modelName); ok && level != "" {
+		geminiRequest.GenerationConfig.ThinkingConfig = &dto.GeminiThinkingConfig{
+			IncludeThoughts: true,
+			ThinkingLevel:   level,
+		}
+		info.SetReasoningEffort(level)
+	}
+}
+
+func ParseStopSequences(stop any) []string {
+	if stop == nil {
+		return nil
+	}
+
+	switch v := stop.(type) {
+	case string:
+		if v != "" {
+			return []string{v}
+		}
+	case []string:
+		return v
+	case []interface{}:
+		sequences := make([]string, 0, len(v))
+		for _, item := range v {
+			if str, ok := item.(string); ok && str != "" {
+				sequences = append(sequences, str)
+			}
+		}
+		return sequences
+	}
+	return nil
+}
+
+func HasFunctionCallContent(call *dto.FunctionCall) bool {
+	if call == nil {
+		return false
+	}
+	if strings.TrimSpace(call.FunctionName) != "" {
+		return true
+	}
+
+	switch v := call.Arguments.(type) {
+	case nil:
+		return false
+	case string:
+		return strings.TrimSpace(v) != ""
+	case map[string]interface{}:
+		return len(v) > 0
+	case []interface{}:
+		return len(v) > 0
+	default:
+		return true
+	}
+}
+
+func SupportedMimeTypesList() []string {
+	keys := make([]string, 0, len(SupportedMimeTypes))
+	for key := range SupportedMimeTypes {
+		keys = append(keys, key)
+	}
+	return keys
+}
+
+func isNew25ProModel(modelName string) bool {
+	return strings.HasPrefix(modelName, "gemini-2.5-pro") &&
+		!strings.HasPrefix(modelName, "gemini-2.5-pro-preview-05-06") &&
+		!strings.HasPrefix(modelName, "gemini-2.5-pro-preview-03-25")
+}
+
+func is25FlashLiteModel(modelName string) bool {
+	return strings.HasPrefix(modelName, "gemini-2.5-flash-lite")
+}
+
+func clampThinkingBudget(modelName string, budget int) int {
+	isNew25Pro := isNew25ProModel(modelName)
+	is25FlashLite := is25FlashLiteModel(modelName)
+
+	if is25FlashLite {
+		if budget < flash25LiteMinBudget {
+			return flash25LiteMinBudget
+		}
+		if budget > flash25LiteMaxBudget {
+			return flash25LiteMaxBudget
+		}
+	} else if isNew25Pro {
+		if budget < pro25MinBudget {
+			return pro25MinBudget
+		}
+		if budget > pro25MaxBudget {
+			return pro25MaxBudget
+		}
+	} else {
+		if budget < 0 {
+			return 0
+		}
+		if budget > flash25MaxBudget {
+			return flash25MaxBudget
+		}
+	}
+	return budget
+}
+
+func clampThinkingBudgetByEffort(modelName string, effort string) int {
+	isNew25Pro := isNew25ProModel(modelName)
+	is25FlashLite := is25FlashLiteModel(modelName)
+
+	maxBudget := 0
+	if is25FlashLite {
+		maxBudget = flash25LiteMaxBudget
+	}
+	if isNew25Pro {
+		maxBudget = pro25MaxBudget
+	} else {
+		maxBudget = flash25MaxBudget
+	}
+	switch effort {
+	case "high":
+		maxBudget = maxBudget * 80 / 100
+	case "medium":
+		maxBudget = maxBudget * 50 / 100
+	case "low":
+		maxBudget = maxBudget * 20 / 100
+	case "minimal":
+		maxBudget = maxBudget * 5 / 100
+	}
+	return clampThinkingBudget(modelName, maxBudget)
+}

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

@@ -0,0 +1,256 @@
+package gemini
+
+import (
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+)
+
+var geminiOpenAPISchemaAllowedFields = map[string]struct{}{
+	"anyOf":            {},
+	"default":          {},
+	"description":      {},
+	"enum":             {},
+	"example":          {},
+	"format":           {},
+	"items":            {},
+	"maxItems":         {},
+	"maxLength":        {},
+	"maxProperties":    {},
+	"maximum":          {},
+	"minItems":         {},
+	"minLength":        {},
+	"minProperties":    {},
+	"minimum":          {},
+	"nullable":         {},
+	"pattern":          {},
+	"properties":       {},
+	"propertyOrdering": {},
+	"required":         {},
+	"title":            {},
+	"type":             {},
+}
+
+const geminiFunctionSchemaMaxDepth = 64
+
+func CleanFunctionParameters(params interface{}) interface{} {
+	return cleanGeminiFunctionParametersWithDepth(params, 0)
+}
+
+func cleanGeminiFunctionParametersWithDepth(params interface{}, depth int) interface{} {
+	if params == nil {
+		return nil
+	}
+
+	if depth >= geminiFunctionSchemaMaxDepth {
+		return cleanGeminiFunctionParametersShallow(params)
+	}
+
+	switch v := params.(type) {
+	case map[string]interface{}:
+		cleanedMap := make(map[string]interface{}, len(v))
+		for key, val := range v {
+			if _, ok := geminiOpenAPISchemaAllowedFields[key]; ok {
+				cleanedMap[key] = val
+			}
+		}
+
+		normalizeGeminiSchemaTypeAndNullable(cleanedMap)
+
+		if props, ok := cleanedMap["properties"].(map[string]interface{}); ok && props != nil {
+			cleanedProps := make(map[string]interface{})
+			for propName, propValue := range props {
+				cleanedProps[propName] = cleanGeminiFunctionParametersWithDepth(propValue, depth+1)
+			}
+			cleanedMap["properties"] = cleanedProps
+		}
+
+		if items, ok := cleanedMap["items"].(map[string]interface{}); ok && items != nil {
+			cleanedMap["items"] = cleanGeminiFunctionParametersWithDepth(items, depth+1)
+		}
+		if itemsArray, ok := cleanedMap["items"].([]interface{}); ok && len(itemsArray) > 0 {
+			cleanedMap["items"] = cleanGeminiFunctionParametersWithDepth(itemsArray[0], depth+1)
+		}
+
+		if nested, ok := cleanedMap["anyOf"].([]interface{}); ok && nested != nil {
+			cleanedNested := make([]interface{}, len(nested))
+			for i, item := range nested {
+				cleanedNested[i] = cleanGeminiFunctionParametersWithDepth(item, depth+1)
+			}
+			cleanedMap["anyOf"] = cleanedNested
+		}
+
+		return cleanedMap
+	case []interface{}:
+		cleanedArray := make([]interface{}, len(v))
+		for i, item := range v {
+			cleanedArray[i] = cleanGeminiFunctionParametersWithDepth(item, depth+1)
+		}
+		return cleanedArray
+	default:
+		return params
+	}
+}
+
+func cleanGeminiFunctionParametersShallow(params interface{}) interface{} {
+	switch v := params.(type) {
+	case map[string]interface{}:
+		cleanedMap := make(map[string]interface{}, len(v))
+		for key, val := range v {
+			if _, ok := geminiOpenAPISchemaAllowedFields[key]; ok {
+				cleanedMap[key] = val
+			}
+		}
+		normalizeGeminiSchemaTypeAndNullable(cleanedMap)
+		delete(cleanedMap, "properties")
+		delete(cleanedMap, "items")
+		delete(cleanedMap, "anyOf")
+		return cleanedMap
+	case []interface{}:
+		return []interface{}{}
+	default:
+		return params
+	}
+}
+
+func normalizeGeminiSchemaTypeAndNullable(schema map[string]interface{}) {
+	rawType, ok := schema["type"]
+	if !ok || rawType == nil {
+		return
+	}
+
+	normalize := func(t string) (string, bool) {
+		switch strings.ToLower(strings.TrimSpace(t)) {
+		case "object":
+			return "OBJECT", false
+		case "array":
+			return "ARRAY", false
+		case "string":
+			return "STRING", false
+		case "integer":
+			return "INTEGER", false
+		case "number":
+			return "NUMBER", false
+		case "boolean":
+			return "BOOLEAN", false
+		case "null":
+			return "", true
+		default:
+			return t, false
+		}
+	}
+
+	switch typed := rawType.(type) {
+	case string:
+		normalized, isNull := normalize(typed)
+		if isNull {
+			schema["nullable"] = true
+			delete(schema, "type")
+			return
+		}
+		schema["type"] = normalized
+	case []interface{}:
+		nullable := false
+		var chosen string
+		for _, item := range typed {
+			if value, ok := item.(string); ok {
+				normalized, isNull := normalize(value)
+				if isNull {
+					nullable = true
+					continue
+				}
+				if chosen == "" {
+					chosen = normalized
+				}
+			}
+		}
+		if nullable {
+			schema["nullable"] = true
+		}
+		if chosen != "" {
+			schema["type"] = chosen
+		} else {
+			delete(schema, "type")
+		}
+	}
+}
+
+func RemoveAdditionalProperties(schema interface{}, depth int) interface{} {
+	if depth >= 5 {
+		return schema
+	}
+
+	value, ok := schema.(map[string]interface{})
+	if !ok || len(value) == 0 {
+		return schema
+	}
+	delete(value, "title")
+	delete(value, "$schema")
+	if typeVal, exists := value["type"]; !exists || (typeVal != "object" && typeVal != "array") {
+		return schema
+	}
+	switch value["type"] {
+	case "object":
+		delete(value, "additionalProperties")
+		if properties, ok := value["properties"].(map[string]interface{}); ok {
+			for key, nested := range properties {
+				properties[key] = RemoveAdditionalProperties(nested, depth+1)
+			}
+		}
+		for _, field := range []string{"allOf", "anyOf", "oneOf"} {
+			if nested, ok := value[field].([]interface{}); ok {
+				for i, item := range nested {
+					nested[i] = RemoveAdditionalProperties(item, depth+1)
+				}
+			}
+		}
+	case "array":
+		if items, ok := value["items"].(map[string]interface{}); ok {
+			value["items"] = RemoveAdditionalProperties(items, depth+1)
+		}
+	}
+
+	return value
+}
+
+func OpenAIToolChoiceToConfig(toolChoice any) *dto.ToolConfig {
+	if toolChoice == nil {
+		return nil
+	}
+
+	if toolChoiceStr, ok := toolChoice.(string); ok {
+		config := &dto.ToolConfig{
+			FunctionCallingConfig: &dto.FunctionCallingConfig{},
+		}
+		switch toolChoiceStr {
+		case "auto":
+			config.FunctionCallingConfig.Mode = "AUTO"
+		case "none":
+			config.FunctionCallingConfig.Mode = "NONE"
+		case "required":
+			config.FunctionCallingConfig.Mode = "ANY"
+		default:
+			config.FunctionCallingConfig.Mode = "AUTO"
+		}
+		return config
+	}
+
+	if toolChoiceMap, ok := toolChoice.(map[string]interface{}); ok {
+		if toolChoiceMap["type"] == "function" {
+			config := &dto.ToolConfig{
+				FunctionCallingConfig: &dto.FunctionCallingConfig{
+					Mode: "ANY",
+				},
+			}
+			if function, ok := toolChoiceMap["function"].(map[string]interface{}); ok {
+				if name, ok := function["name"].(string); ok && name != "" {
+					config.FunctionCallingConfig.AllowedFunctionNames = []string{name}
+				}
+			}
+			return config
+		}
+		return nil
+	}
+
+	return nil
+}