feat: publish v1.0.0-rc.36-custom.1

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]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{"result": array}
+		}
+		return map[string]any{"content": typed}
+	case []any:
+		return map[string]any{"result": typed}
+	default:
+		return map[string]any{"content": typed}
+	}
+}
+
+func GeminiResponseMap(value any) map[string]any {
+	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 +1,304 @@
+package oairesponses
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"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"
+)
+
+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)
+	}
+	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))
+	}
+	sourceReasoning, err := reasoning.FromOpenAIResponses(req)
+	if err != nil {
+		return nil, reasoning.AsClientError(err)
+	}
+	if err := sharedclaude.ApplyReasoning(c, claudeRequest, info, sourceReasoning, true); err != nil {
+		return nil, reasoning.AsClientError(err)
+	}
+	if claudeRequest.MaxTokens == nil {
+		if defaultMaxTokens, configured := convmeta.OptionsOf(info).Claude.DefaultMaxTokensFor(claudeRequest.Model); configured {
+			value := uint(defaultMaxTokens)
+			claudeRequest.MaxTokens = &value
+		}
+	}
+
+	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:
+			sourceRole := strings.TrimSpace(kitutil.Interface2String(item["role"]))
+			role := responsesClaudeRole(sourceRole)
+			parts, err := responsesInputContentToClaudeMediaMessages(c, item["content"])
+			if err != nil {
+				return nil, err
+			}
+			if sourceRole == "" && len(parts) == 0 {
+				continue
+			}
+			if role == "system" {
+				for _, part := range parts {
+					if part.Type == "text" {
+						systemMessages = append(systemMessages, part)
+					}
+				}
+				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
+	}
+	if len(claudeRequest.Messages) > 0 || len(systemMessages) > 0 {
+		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: sharedclaude.FunctionParametersToInputSchema(function.Parameters),
+		})
+	}
+	return tools
+}
+
+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(role string) string {
+	switch 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_claude_messages_resp.go

@@ -0,0 +1,156 @@
+package oairesponses
+
+import (
+	"encoding/json"
+	"errors"
+	"strings"
+	"unicode/utf8"
+
+	"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"
+)
+
+func ResponsesResponseToClaudeMessagesResponse(resp *dto.OpenAIResponsesResponse) (*dto.ClaudeResponse, *dto.Usage, error) {
+	if resp == nil {
+		return nil, nil, errors.New("response is nil")
+	}
+
+	usage := UsageFromResponsesUsage(resp.Usage)
+	claudeResponse := &dto.ClaudeResponse{
+		Id:    resp.ID,
+		Type:  "message",
+		Role:  "assistant",
+		Model: resp.Model,
+		Usage: sharedclaude.UsageFromOpenAI(usage),
+	}
+	sawToolCall := false
+	for index := range resp.Output {
+		output := resp.Output[index]
+		if output.Type == responsesOutputTypeMessage && output.Role != "" && output.Role != "assistant" {
+			continue
+		}
+		switch output.Type {
+		case responsesOutputTypeReasoning:
+			if thinking := reasoningOutputText(&output); thinking != "" {
+				claudeResponse.Content = append(claudeResponse.Content, dto.ClaudeMediaMessage{
+					Type:     "thinking",
+					Thinking: kitutil.GetPointer(thinking),
+				})
+			}
+		case responsesOutputTypeMessage:
+			for _, content := range output.Content {
+				if content.Type != "output_text" {
+					continue
+				}
+				block := dto.ClaudeMediaMessage{Type: "text", Text: kitutil.GetPointer(content.Text)}
+				if citations := responsesAnnotationsToClaude(content.Annotations, content.Text); len(citations) > 0 {
+					block.Citations, _ = kitutil.Marshal(citations)
+				}
+				claudeResponse.Content = append(claudeResponse.Content, block)
+			}
+		case responsesOutputTypeFunctionCall, responsesOutputTypeCustomToolCall:
+			sawToolCall = true
+			callID := strings.TrimSpace(output.CallId)
+			if callID == "" {
+				callID = strings.TrimSpace(output.ID)
+			}
+			claudeResponse.Content = append(claudeResponse.Content, dto.ClaudeMediaMessage{
+				Type:  "tool_use",
+				Id:    callID,
+				Name:  output.Name,
+				Input: responsesArgumentsToClaudeInput(output.ArgumentsString()),
+			})
+		}
+	}
+	if len(claudeResponse.Content) == 0 {
+		claudeResponse.Content = []dto.ClaudeMediaMessage{{Type: "text", Text: kitutil.GetPointer("")}}
+	}
+	claudeResponse.StopReason = responsesClaudeStopReason(resp, sawToolCall)
+	return claudeResponse, usage, nil
+}
+
+func responsesArgumentsToClaudeInput(arguments string) map[string]any {
+	input := make(map[string]any)
+	if strings.TrimSpace(arguments) == "" {
+		return input
+	}
+	if err := kitutil.Unmarshal([]byte(arguments), &input); err == nil && input != nil {
+		return input
+	}
+	return map[string]any{"input": arguments}
+}
+
+func responsesClaudeStopReason(resp *dto.OpenAIResponsesResponse, sawToolCall bool) string {
+	if finishReason, ok := ResponsesFinishReasonFromStatus(resp); ok {
+		return reasonmap.OpenAIFinishReasonToClaudeStopReason(finishReason)
+	}
+	if sawToolCall {
+		return "tool_use"
+	}
+	return "end_turn"
+}
+
+func responsesAnnotationsToClaude(annotations []any, text string) []json.RawMessage {
+	citations := make([]json.RawMessage, 0, len(annotations))
+	for _, rawAnnotation := range annotations {
+		annotation, err := kitutil.Any2Type[map[string]any](rawAnnotation)
+		if err != nil || strings.TrimSpace(kitutil.Interface2String(annotation["type"])) != "url_citation" {
+			continue
+		}
+		citation := annotation
+		if nested, ok := annotation["url_citation"].(map[string]any); ok {
+			citation = nested
+		}
+		url := strings.TrimSpace(kitutil.Interface2String(citation["url"]))
+		if url == "" {
+			continue
+		}
+		converted := map[string]any{
+			"type":  "web_search_result_location",
+			"url":   url,
+			"title": strings.TrimSpace(kitutil.Interface2String(citation["title"])),
+		}
+		if citedText := kitutil.Interface2String(citation["cited_text"]); citedText != "" {
+			converted["cited_text"] = citedText
+		} else if citedText := responsesCitedText(text, citation); citedText != "" {
+			converted["cited_text"] = citedText
+		}
+		if encryptedIndex := kitutil.Interface2String(citation["encrypted_index"]); encryptedIndex != "" {
+			converted["encrypted_index"] = encryptedIndex
+		}
+		if converted["title"] == "" {
+			delete(converted, "title")
+		}
+		encoded, err := kitutil.Marshal(converted)
+		if err == nil {
+			citations = append(citations, encoded)
+		}
+	}
+	return citations
+}
+
+func responsesCitedText(text string, citation map[string]any) string {
+	start, startOK := responsesAnnotationIndex(citation["start_index"])
+	end, endOK := responsesAnnotationIndex(citation["end_index"])
+	if !startOK || !endOK || start < 0 || end <= start || end > utf8.RuneCountInString(text) {
+		return ""
+	}
+	runes := []rune(text)
+	return string(runes[start:end])
+}
+
+func responsesAnnotationIndex(value any) (int, bool) {
+	switch number := value.(type) {
+	case float64:
+		return int(number), number >= 0 && number == float64(int(number))
+	case int:
+		return number, number >= 0
+	case json.Number:
+		parsed, err := number.Int64()
+		return int(parsed), err == nil && parsed >= 0
+	default:
+		return 0, false
+	}
+}

relaykit/relayconvert/internal/oai_responses/to_claude_messages_stream_resp.go

@@ -0,0 +1,500 @@
+package oairesponses
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	sharedclaude "github.com/QuantumNous/new-api/relaykit/relayconvert/internal/shared/claude"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+)
+
+const responsesEventOutputTextDone = "response.output_text.done"
+
+type ResponsesToClaudeStreamState struct {
+	ID    string
+	Model string
+	Usage *dto.Usage
+
+	sentMessageStart bool
+	done             bool
+	sawToolCall      bool
+	nextBlockIndex   int
+	blocks           []*responsesClaudeStreamBlock
+	byOutputIndex    map[int]*responsesClaudeStreamBlock
+	byItemID         map[string]*responsesClaudeStreamBlock
+	lastByKind       map[string]*responsesClaudeStreamBlock
+	usageText        strings.Builder
+}
+
+type responsesClaudeStreamBlock struct {
+	Index               int
+	Kind                string
+	ItemID              string
+	CallID              string
+	Name                string
+	Started             bool
+	Stopped             bool
+	Value               strings.Builder
+	SentBytes           int
+	AnnotationCount     int
+	NeedsReasoningBreak bool
+}
+
+func NewResponsesToClaudeStreamState(id string, model string) *ResponsesToClaudeStreamState {
+	return &ResponsesToClaudeStreamState{
+		ID:            strings.TrimSpace(id),
+		Model:         strings.TrimSpace(model),
+		byOutputIndex: make(map[int]*responsesClaudeStreamBlock),
+		byItemID:      make(map[string]*responsesClaudeStreamBlock),
+		lastByKind:    make(map[string]*responsesClaudeStreamBlock),
+	}
+}
+
+func (s *ResponsesToClaudeStreamState) UsageText() string {
+	if s == nil {
+		return ""
+	}
+	return s.usageText.String()
+}
+
+func (s *ResponsesToClaudeStreamState) Done() bool {
+	return s != nil && s.done
+}
+
+func (s *ResponsesToClaudeStreamState) SetUsage(usage *dto.Usage) {
+	if s != nil && usage != nil {
+		s.Usage = usage
+	}
+}
+
+func (s *ResponsesToClaudeStreamState) StreamUsage() *dto.Usage {
+	if s == nil {
+		return nil
+	}
+	return s.Usage
+}
+
+func (s *ResponsesToClaudeStreamState) SetStreamUsage(usage *dto.Usage) {
+	s.SetUsage(usage)
+}
+
+func (s *ResponsesToClaudeStreamState) ConvertChunk(event *dto.ResponsesStreamResponse, estimatedInputTokens int) ([]*dto.ClaudeResponse, *dto.Usage, error) {
+	if s == nil {
+		return nil, nil, nil
+	}
+	if event == nil || s.done {
+		return nil, s.Usage, nil
+	}
+
+	s.applyResponseMetadata(event.Response)
+	switch event.Type {
+	case responsesEventCreated:
+		return s.ensureMessageStart(estimatedInputTokens), s.Usage, nil
+	case responsesEventReasoningSummaryDelta, responsesEventReasoningTextDelta:
+		block, err := s.ensureBlock(event, "thinking")
+		if err != nil {
+			return nil, s.Usage, err
+		}
+		delta := event.Delta
+		if block.NeedsReasoningBreak && delta != "" {
+			delta = separatedResponsesDelta(delta)
+			block.NeedsReasoningBreak = false
+		}
+		return s.appendDelta(block, delta, estimatedInputTokens), s.Usage, nil
+	case responsesEventReasoningSummaryDone, responsesEventReasoningTextDone:
+		block, err := s.ensureBlock(event, "thinking")
+		if err != nil {
+			return nil, s.Usage, err
+		}
+		var responses []*dto.ClaudeResponse
+		if event.Text != nil {
+			responses = append(responses, s.mergeFinalValue(block, *event.Text, estimatedInputTokens)...)
+		}
+		if block.Value.Len() > 0 {
+			block.NeedsReasoningBreak = true
+		}
+		return responses, s.Usage, nil
+	case responsesEventOutputTextDelta:
+		block, err := s.ensureBlock(event, "text")
+		if err != nil {
+			return nil, s.Usage, err
+		}
+		return s.appendDelta(block, event.Delta, estimatedInputTokens), s.Usage, nil
+	case responsesEventOutputTextDone:
+		block, err := s.ensureBlock(event, "text")
+		if err != nil {
+			return nil, s.Usage, err
+		}
+		if event.Text == nil {
+			return nil, s.Usage, nil
+		}
+		return s.mergeFinalValue(block, *event.Text, estimatedInputTokens), s.Usage, nil
+	case responsesEventOutputTextAnnotationAdded:
+		block, err := s.ensureBlock(event, "text")
+		if err != nil {
+			return nil, s.Usage, err
+		}
+		var annotation any
+		if err := kitutil.Unmarshal(event.Annotation, &annotation); err != nil {
+			return nil, s.Usage, fmt.Errorf("invalid Responses stream annotation: %w", err)
+		}
+		return s.appendAnnotations(block, []any{annotation}, estimatedInputTokens, false), s.Usage, nil
+	case responsesEventOutputItemAdded, responsesEventOutputItemDone:
+		responses, err := s.applyOutputItem(event, estimatedInputTokens, event.Type == responsesEventOutputItemDone)
+		return responses, s.Usage, err
+	case responsesEventFunctionArgsDelta, responsesEventCustomToolInputDelta:
+		block, err := s.ensureBlock(event, "tool_use")
+		if err != nil {
+			return nil, s.Usage, err
+		}
+		return s.appendDelta(block, event.Delta, estimatedInputTokens), s.Usage, nil
+	case responsesEventFunctionArgsDone, responsesEventCustomToolInputDone:
+		block, err := s.ensureBlock(event, "tool_use")
+		if err != nil {
+			return nil, s.Usage, err
+		}
+		if event.Arguments == nil {
+			return nil, s.Usage, nil
+		}
+		return s.mergeFinalValue(block, *event.Arguments, estimatedInputTokens), s.Usage, nil
+	case responsesEventCompleted, responsesEventDone, responsesEventIncomplete:
+		responses, err := s.finish(event.Response, estimatedInputTokens)
+		return responses, s.Usage, err
+	case responsesEventFailed, responsesEventError:
+		message := strings.TrimSpace(event.Message)
+		if message == "" {
+			message = event.Type
+		}
+		return nil, s.Usage, fmt.Errorf("responses stream error: %s", message)
+	default:
+		return nil, s.Usage, nil
+	}
+}
+
+func (s *ResponsesToClaudeStreamState) Finalize(estimatedInputTokens int) ([]*dto.ClaudeResponse, error) {
+	if s == nil || s.done {
+		return nil, nil
+	}
+	return s.finish(nil, estimatedInputTokens)
+}
+
+func (s *ResponsesToClaudeStreamState) applyResponseMetadata(response *dto.OpenAIResponsesResponse) {
+	if s == nil || response == nil {
+		return
+	}
+	if response.ID != "" {
+		s.ID = response.ID
+	}
+	if response.Model != "" {
+		s.Model = response.Model
+	}
+	if response.Usage != nil {
+		s.Usage = dto.MergeUsageNonZero(s.Usage, UsageFromResponsesUsage(response.Usage))
+	}
+}
+
+func (s *ResponsesToClaudeStreamState) ensureMessageStart(estimatedInputTokens int) []*dto.ClaudeResponse {
+	if s.sentMessageStart {
+		return nil
+	}
+	s.sentMessageStart = true
+	inputTokens := estimatedInputTokens
+	if s.Usage != nil {
+		if usage := sharedclaude.UsageFromOpenAI(s.Usage); usage != nil {
+			inputTokens = usage.InputTokens
+		}
+	}
+	message := &dto.ClaudeMediaMessage{
+		Id:    s.ID,
+		Type:  "message",
+		Role:  "assistant",
+		Model: s.Model,
+		Usage: &dto.ClaudeUsage{InputTokens: inputTokens},
+	}
+	message.SetContent(make([]any, 0))
+	return []*dto.ClaudeResponse{{Type: "message_start", Message: message}}
+}
+
+func (s *ResponsesToClaudeStreamState) ensureBlock(event *dto.ResponsesStreamResponse, kind string) (*responsesClaudeStreamBlock, error) {
+	block := s.findBlock(event)
+	if block == nil {
+		if last := s.lastByKind[kind]; last != nil && !last.Stopped && event.OutputIndex == nil && responseStreamEventItemID(event) == "" {
+			block = last
+		}
+	}
+	if block == nil {
+		block = &responsesClaudeStreamBlock{Index: s.nextBlockIndex, Kind: kind}
+		s.nextBlockIndex++
+		s.blocks = append(s.blocks, block)
+	}
+	if block.Kind == "" {
+		block.Kind = kind
+	}
+	if block.Kind != kind {
+		return nil, fmt.Errorf("Responses output item changed from %s to %s", block.Kind, kind)
+	}
+	s.applyBlockMetadata(block, event)
+	s.lastByKind[kind] = block
+	return block, nil
+}
+
+func (s *ResponsesToClaudeStreamState) findBlock(event *dto.ResponsesStreamResponse) *responsesClaudeStreamBlock {
+	if event == nil {
+		return nil
+	}
+	if event.OutputIndex != nil {
+		if block := s.byOutputIndex[*event.OutputIndex]; block != nil {
+			return block
+		}
+	}
+	if itemID := responseStreamEventItemID(event); itemID != "" {
+		return s.byItemID[itemID]
+	}
+	return nil
+}
+
+func (s *ResponsesToClaudeStreamState) applyBlockMetadata(block *responsesClaudeStreamBlock, event *dto.ResponsesStreamResponse) {
+	if block == nil || event == nil {
+		return
+	}
+	if event.OutputIndex != nil {
+		s.byOutputIndex[*event.OutputIndex] = block
+	}
+	if itemID := responseStreamEventItemID(event); itemID != "" {
+		block.ItemID = itemID
+		s.byItemID[itemID] = block
+	}
+	if event.Item == nil {
+		return
+	}
+	if callID := strings.TrimSpace(event.Item.CallId); callID != "" {
+		block.CallID = callID
+	} else if block.CallID == "" {
+		block.CallID = strings.TrimSpace(event.Item.ID)
+	}
+	if name := strings.TrimSpace(event.Item.Name); name != "" {
+		block.Name = name
+	}
+}
+
+func (s *ResponsesToClaudeStreamState) startBlock(block *responsesClaudeStreamBlock, estimatedInputTokens int) []*dto.ClaudeResponse {
+	if block == nil || block.Started || block.Stopped {
+		return nil
+	}
+	var content dto.ClaudeMediaMessage
+	switch block.Kind {
+	case "text":
+		content = dto.ClaudeMediaMessage{Type: "text", Text: kitutil.GetPointer("")}
+	case "thinking":
+		content = dto.ClaudeMediaMessage{Type: "thinking", Thinking: kitutil.GetPointer("")}
+	case "tool_use":
+		if block.Name == "" {
+			return nil
+		}
+		callID := block.CallID
+		if callID == "" {
+			callID = block.ItemID
+		}
+		content = dto.ClaudeMediaMessage{Type: "tool_use", Id: callID, Name: block.Name, Input: map[string]any{}}
+		s.sawToolCall = true
+	default:
+		return nil
+	}
+	block.Started = true
+	responses := s.ensureMessageStart(estimatedInputTokens)
+	index := block.Index
+	responses = append(responses, &dto.ClaudeResponse{Type: "content_block_start", Index: &index, ContentBlock: &content})
+	return responses
+}
+
+func (s *ResponsesToClaudeStreamState) appendDelta(block *responsesClaudeStreamBlock, delta string, estimatedInputTokens int) []*dto.ClaudeResponse {
+	if block == nil || block.Stopped || delta == "" {
+		return nil
+	}
+	block.Value.WriteString(delta)
+	return s.flushBlock(block, estimatedInputTokens)
+}
+
+func (s *ResponsesToClaudeStreamState) mergeFinalValue(block *responsesClaudeStreamBlock, finalValue string, estimatedInputTokens int) []*dto.ClaudeResponse {
+	if block == nil || block.Stopped {
+		return nil
+	}
+	current := block.Value.String()
+	if current == "" {
+		block.Value.WriteString(finalValue)
+	} else if strings.HasPrefix(finalValue, current) {
+		block.Value.WriteString(finalValue[len(current):])
+	}
+	return s.flushBlock(block, estimatedInputTokens)
+}
+
+func (s *ResponsesToClaudeStreamState) flushBlock(block *responsesClaudeStreamBlock, estimatedInputTokens int) []*dto.ClaudeResponse {
+	if block == nil || block.Stopped {
+		return nil
+	}
+	responses := s.startBlock(block, estimatedInputTokens)
+	if !block.Started {
+		return responses
+	}
+	value := block.Value.String()
+	if block.SentBytes >= len(value) {
+		return responses
+	}
+	delta := value[block.SentBytes:]
+	block.SentBytes = len(value)
+	s.usageText.WriteString(delta)
+	index := block.Index
+	media := &dto.ClaudeMediaMessage{}
+	switch block.Kind {
+	case "text":
+		media.Type = "text_delta"
+		media.Text = &delta
+	case "thinking":
+		media.Type = "thinking_delta"
+		media.Thinking = &delta
+	case "tool_use":
+		media.Type = "input_json_delta"
+		media.PartialJson = &delta
+	}
+	responses = append(responses, &dto.ClaudeResponse{Type: "content_block_delta", Index: &index, Delta: media})
+	return responses
+}
+
+func (s *ResponsesToClaudeStreamState) stopBlock(block *responsesClaudeStreamBlock, estimatedInputTokens int) []*dto.ClaudeResponse {
+	if block == nil || block.Stopped {
+		return nil
+	}
+	responses := s.flushBlock(block, estimatedInputTokens)
+	responses = append(responses, s.startBlock(block, estimatedInputTokens)...)
+	if !block.Started {
+		return responses
+	}
+	block.Stopped = true
+	index := block.Index
+	return append(responses, &dto.ClaudeResponse{Type: "content_block_stop", Index: &index})
+}
+
+func (s *ResponsesToClaudeStreamState) applyOutputItem(event *dto.ResponsesStreamResponse, estimatedInputTokens int, stop bool) ([]*dto.ClaudeResponse, error) {
+	if event == nil || event.Item == nil {
+		return nil, nil
+	}
+	item := event.Item
+	var kind string
+	switch item.Type {
+	case responsesOutputTypeReasoning:
+		kind = "thinking"
+	case responsesOutputTypeMessage:
+		if item.Role != "" && item.Role != "assistant" {
+			return nil, nil
+		}
+		kind = "text"
+	case responsesOutputTypeFunctionCall, responsesOutputTypeCustomToolCall:
+		kind = "tool_use"
+	default:
+		return nil, nil
+	}
+	block, err := s.ensureBlock(event, kind)
+	if err != nil {
+		return nil, err
+	}
+	var responses []*dto.ClaudeResponse
+	switch kind {
+	case "thinking":
+		responses = append(responses, s.mergeFinalValue(block, reasoningOutputText(item), estimatedInputTokens)...)
+	case "text":
+		var text strings.Builder
+		var annotations []any
+		for _, content := range item.Content {
+			if content.Type != "output_text" {
+				continue
+			}
+			text.WriteString(content.Text)
+			annotations = append(annotations, content.Annotations...)
+		}
+		responses = append(responses, s.mergeFinalValue(block, text.String(), estimatedInputTokens)...)
+		responses = append(responses, s.appendAnnotations(block, annotations, estimatedInputTokens, true)...)
+	case "tool_use":
+		responses = append(responses, s.mergeFinalValue(block, item.ArgumentsString(), estimatedInputTokens)...)
+	}
+	if stop {
+		responses = append(responses, s.stopBlock(block, estimatedInputTokens)...)
+	}
+	return responses, nil
+}
+
+func (s *ResponsesToClaudeStreamState) appendAnnotations(block *responsesClaudeStreamBlock, annotations []any, estimatedInputTokens int, snapshot bool) []*dto.ClaudeResponse {
+	if block == nil || block.Kind != "text" || block.Stopped || len(annotations) == 0 {
+		return nil
+	}
+	remaining := annotations
+	if snapshot {
+		if len(annotations) <= block.AnnotationCount {
+			return nil
+		}
+		remaining = annotations[block.AnnotationCount:]
+		block.AnnotationCount = len(annotations)
+	} else {
+		block.AnnotationCount += len(annotations)
+	}
+	citations := responsesAnnotationsToClaude(remaining, block.Value.String())
+	if len(citations) == 0 {
+		return nil
+	}
+	responses := s.startBlock(block, estimatedInputTokens)
+	index := block.Index
+	for _, citation := range citations {
+		responses = append(responses, &dto.ClaudeResponse{
+			Type:  "content_block_delta",
+			Index: &index,
+			Delta: &dto.ClaudeMediaMessage{Type: "citations_delta", Citation: citation},
+		})
+	}
+	return responses
+}
+
+func (s *ResponsesToClaudeStreamState) finish(response *dto.OpenAIResponsesResponse, estimatedInputTokens int) ([]*dto.ClaudeResponse, error) {
+	if s.done {
+		return nil, nil
+	}
+	s.applyResponseMetadata(response)
+	responses := make([]*dto.ClaudeResponse, 0)
+	if response != nil {
+		for outputIndex := range response.Output {
+			index := outputIndex
+			item := response.Output[outputIndex]
+			event := &dto.ResponsesStreamResponse{OutputIndex: &index, ItemID: item.ID, Item: &item}
+			itemResponses, err := s.applyOutputItem(event, estimatedInputTokens, true)
+			if err != nil {
+				return nil, err
+			}
+			responses = append(responses, itemResponses...)
+		}
+	}
+	for _, block := range s.blocks {
+		responses = append(responses, s.stopBlock(block, estimatedInputTokens)...)
+	}
+	responses = append(responses, s.ensureMessageStart(estimatedInputTokens)...)
+	stopReason := responsesClaudeStopReason(response, s.sawToolCall)
+	usage := sharedclaude.UsageFromOpenAI(s.Usage)
+	responses = append(responses,
+		&dto.ClaudeResponse{
+			Type:  "message_delta",
+			Usage: usage,
+			Delta: &dto.ClaudeMediaMessage{StopReason: &stopReason},
+		},
+		&dto.ClaudeResponse{Type: "message_stop"},
+	)
+	s.done = true
+	return responses, nil
+}
+
+func separatedResponsesDelta(delta string) string {
+	if strings.HasPrefix(delta, "\n\n") {
+		return delta
+	}
+	if strings.HasPrefix(delta, "\n") {
+		return "\n" + delta
+	}
+	return "\n\n" + delta
+}

relaykit/relayconvert/internal/oai_responses/to_claude_messages_stream_resp_test.go

@@ -0,0 +1,130 @@
+package oairesponses
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestResponsesToClaudeStreamDoesNotRepeatBlocksFromDoneAndCompletedEvents(t *testing.T) {
+	state := NewResponsesToClaudeStreamState("", "")
+	arguments := `{"q":"x"}`
+	argumentRaw, err := kitutil.Marshal(arguments)
+	require.NoError(t, err)
+	statusRaw, err := kitutil.Marshal("completed")
+	require.NoError(t, err)
+
+	reasoningItem := dto.ResponsesOutput{
+		Type:    responsesOutputTypeReasoning,
+		ID:      "rs_1",
+		Summary: []dto.ResponsesReasoningSummaryPart{{Type: "summary_text", Text: "plan"}},
+	}
+	messageItem := dto.ResponsesOutput{
+		Type:    responsesOutputTypeMessage,
+		ID:      "msg_1",
+		Role:    "assistant",
+		Content: []dto.ResponsesOutputContent{{Type: "output_text", Text: "hello"}},
+	}
+	toolItem := dto.ResponsesOutput{
+		Type:      responsesOutputTypeFunctionCall,
+		ID:        "fc_1",
+		CallId:    "call_1",
+		Name:      "lookup",
+		Arguments: argumentRaw,
+	}
+
+	events := []*dto.ResponsesStreamResponse{
+		{Type: responsesEventCreated, Response: &dto.OpenAIResponsesResponse{ID: "resp_1", Model: "gpt-test"}},
+		{Type: responsesEventOutputItemAdded, OutputIndex: kitutil.GetPointer(0), ItemID: reasoningItem.ID, Item: &dto.ResponsesOutput{Type: reasoningItem.Type, ID: reasoningItem.ID}},
+		{Type: responsesEventReasoningSummaryDelta, OutputIndex: kitutil.GetPointer(0), ItemID: reasoningItem.ID, Delta: "plan"},
+		{Type: responsesEventReasoningSummaryDone, OutputIndex: kitutil.GetPointer(0), ItemID: reasoningItem.ID, Text: kitutil.GetPointer("plan")},
+		{Type: responsesEventOutputItemDone, OutputIndex: kitutil.GetPointer(0), ItemID: reasoningItem.ID, Item: &reasoningItem},
+		{Type: responsesEventOutputItemAdded, OutputIndex: kitutil.GetPointer(1), ItemID: messageItem.ID, Item: &dto.ResponsesOutput{Type: messageItem.Type, ID: messageItem.ID, Role: "assistant"}},
+		{Type: responsesEventOutputTextDelta, OutputIndex: kitutil.GetPointer(1), ItemID: messageItem.ID, Delta: "hello"},
+		{Type: responsesEventOutputTextDone, OutputIndex: kitutil.GetPointer(1), ItemID: messageItem.ID, Text: kitutil.GetPointer("hello")},
+		{Type: responsesEventOutputItemDone, OutputIndex: kitutil.GetPointer(1), ItemID: messageItem.ID, Item: &messageItem},
+		{Type: responsesEventOutputItemAdded, OutputIndex: kitutil.GetPointer(2), ItemID: toolItem.ID, Item: &dto.ResponsesOutput{Type: toolItem.Type, ID: toolItem.ID, CallId: toolItem.CallId, Name: toolItem.Name}},
+		{Type: responsesEventFunctionArgsDelta, OutputIndex: kitutil.GetPointer(2), ItemID: toolItem.ID, Delta: `{"q":`},
+		{Type: responsesEventFunctionArgsDelta, OutputIndex: kitutil.GetPointer(2), ItemID: toolItem.ID, Delta: `"x"}`},
+		{Type: responsesEventFunctionArgsDone, OutputIndex: kitutil.GetPointer(2), ItemID: toolItem.ID, Arguments: &arguments},
+		{Type: responsesEventOutputItemDone, OutputIndex: kitutil.GetPointer(2), ItemID: toolItem.ID, Item: &toolItem},
+		{
+			Type: responsesEventCompleted,
+			Response: &dto.OpenAIResponsesResponse{
+				ID:     "resp_1",
+				Model:  "gpt-test",
+				Status: statusRaw,
+				Output: []dto.ResponsesOutput{reasoningItem, messageItem, toolItem},
+				Usage:  &dto.Usage{InputTokens: 11, OutputTokens: 7, TotalTokens: 18},
+			},
+		},
+	}
+
+	var output []*dto.ClaudeResponse
+	for _, event := range events {
+		converted, _, err := state.ConvertChunk(event, 9)
+		require.NoError(t, err)
+		output = append(output, converted...)
+	}
+
+	starts := responsesOfType(output, "content_block_start")
+	stops := responsesOfType(output, "content_block_stop")
+	require.Len(t, responsesOfType(output, "message_start"), 1)
+	require.Len(t, starts, 3)
+	require.Len(t, stops, 3)
+	require.Len(t, responsesOfType(output, "message_delta"), 1)
+	require.Len(t, responsesOfType(output, "message_stop"), 1)
+	assert.Equal(t, []int{0, 1, 2}, []int{starts[0].GetIndex(), starts[1].GetIndex(), starts[2].GetIndex()})
+	assert.Equal(t, []string{"thinking", "text", "tool_use"}, []string{starts[0].ContentBlock.Type, starts[1].ContentBlock.Type, starts[2].ContentBlock.Type})
+	assert.Equal(t, "plan", joinedClaudeDeltas(output, "thinking_delta"))
+	assert.Equal(t, "hello", joinedClaudeDeltas(output, "text_delta"))
+	assert.Equal(t, arguments, joinedClaudeDeltas(output, "input_json_delta"))
+	messageDelta := responsesOfType(output, "message_delta")[0]
+	require.NotNil(t, messageDelta.Delta.StopReason)
+	assert.Equal(t, "tool_use", *messageDelta.Delta.StopReason)
+
+	finalized, err := state.Finalize(9)
+	require.NoError(t, err)
+	assert.Empty(t, finalized)
+	repeated, _, err := state.ConvertChunk(events[len(events)-1], 9)
+	require.NoError(t, err)
+	assert.Empty(t, repeated)
+}
+
+func responsesOfType(responses []*dto.ClaudeResponse, responseType string) []*dto.ClaudeResponse {
+	filtered := make([]*dto.ClaudeResponse, 0)
+	for _, response := range responses {
+		if response != nil && response.Type == responseType {
+			filtered = append(filtered, response)
+		}
+	}
+	return filtered
+}
+
+func joinedClaudeDeltas(responses []*dto.ClaudeResponse, deltaType string) string {
+	var result strings.Builder
+	for _, response := range responses {
+		if response == nil || response.Type != "content_block_delta" || response.Delta == nil || response.Delta.Type != deltaType {
+			continue
+		}
+		switch deltaType {
+		case "thinking_delta":
+			if response.Delta.Thinking != nil {
+				result.WriteString(*response.Delta.Thinking)
+			}
+		case "text_delta":
+			if response.Delta.Text != nil {
+				result.WriteString(*response.Delta.Text)
+			}
+		case "input_json_delta":
+			if response.Delta.PartialJson != nil {
+				result.WriteString(*response.Delta.PartialJson)
+			}
+		}
+	}
+	return result.String()
+}

relaykit/relayconvert/internal/oai_responses/to_gemini_chat_req.go

@@ -0,0 +1,331 @@
+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"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/reasoning"
+)
+
+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 {
+		geminiRequest.GenerationConfig.TopP = kitutil.GetPointer(*req.TopP)
+	}
+	if req.MaxOutputTokens != nil {
+		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
+	}
+	reasoningIntent, err := reasoning.FromOpenAIResponses(req)
+	if err != nil {
+		return nil, reasoning.AsClientError(err)
+	}
+	var reasoningPivot dto.GeneralOpenAIRequest
+	if err := reasoning.ApplyToOpenAIChat(&reasoningPivot, reasoningIntent); err != nil {
+		return nil, reasoning.AsClientError(err)
+	}
+	reasoningPivot.Model = req.Model
+	reasoningPivot.MaxCompletionTokens = req.MaxOutputTokens
+	if err := sharedgemini.ApplyThinkingConfig(geminiRequest, info, reasoningPivot); err != nil {
+		return nil, reasoning.AsClientError(err)
+	}
+
+	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]any); ok {
+			if props, hasProps := params["properties"].(map[string]any); 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, err := responsesFunctionOutputItemToGeminiPart(item, callNames)
+			if err != nil {
+				return nil, err
+			}
+			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{
+			ID:           callID,
+			FunctionName: name,
+			Arguments:    ObjectValue(item["arguments"], "arguments"),
+		},
+	}, callID, nil
+}
+
+func responsesFunctionOutputItemToGeminiPart(item map[string]any, callNames map[string]string) (dto.GeminiPart, error) {
+	callID := CallID(item)
+	name := strings.TrimSpace(kitutil.Interface2String(item["name"]))
+	if name == "" {
+		name = callNames[callID]
+	}
+	response := &dto.GeminiFunctionResponse{
+		Name:     name,
+		Response: GeminiResponseMap(item["output"]),
+	}
+	if callID != "" {
+		id, err := kitutil.Marshal(callID)
+		if err != nil {
+			return dto.GeminiPart{}, fmt.Errorf("failed to marshal function response ID: %w", err)
+		}
+		response.ID = id
+	}
+	return dto.GeminiPart{
+		FunctionResponse: response,
+	}, nil
+}
+
+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 +1,577 @@
+package oairesponses
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	kitutil "github.com/QuantumNous/new-api/relaykit/relayconvert/kitutil"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert/reasoning"
+)
+
+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,
+		ThinkingBudget:       req.ThinkingBudget,
+	}
+
+	out.FrequencyPenalty, err = responsesRawFloat(req.FrequencyPenalty)
+	if err != nil {
+		return nil, fmt.Errorf("invalid frequency_penalty: %w", err)
+	}
+	out.PresencePenalty, err = responsesRawFloat(req.PresencePenalty)
+	if err != nil {
+		return nil, fmt.Errorf("invalid presence_penalty: %w", err)
+	}
+
+	if reasoningIntent, err := reasoning.FromOpenAIResponses(req); err != nil {
+		return nil, reasoning.AsClientError(err)
+	} else if err := reasoning.ApplyToOpenAIChat(out, reasoningIntent); err != nil {
+		return nil, reasoning.AsClientError(err)
+	}
+	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 responsesRawFloat(raw json.RawMessage) (*float64, error) {
+	if !rawJSONPresent(raw) {
+		return nil, nil
+	}
+	var value float64
+	if err := kitutil.Unmarshal(raw, &value); err != nil {
+		return nil, err
+	}
+	return &value, nil
+}
+
+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 +1,358 @@
+package oairesponses
+
+import (
+	"encoding/json"
+	"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 TestResponsesRequestToChatCompletionsRequestPreservesQwenThinkingBudget(t *testing.T) {
+	tests := []struct {
+		name   string
+		budget json.RawMessage
+		want   int64
+	}{
+		{name: "positive budget", budget: json.RawMessage(`128`), want: 128},
+		{name: "zero budget", budget: json.RawMessage(`0`), want: 0},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+				Model:          "qwen-plus",
+				Input:          mustRawMessage(t, "hello"),
+				EnableThinking: json.RawMessage(`true`),
+				ThinkingBudget: tt.budget,
+			})
+			require.NoError(t, err)
+			assert.Equal(t, tt.budget, got.ThinkingBudget)
+
+			encoded, err := kitutil.Marshal(got)
+			require.NoError(t, err)
+
+			assert.True(t, gjson.GetBytes(encoded, "enable_thinking").Bool())
+			value := gjson.GetBytes(encoded, "thinking_budget")
+			assert.True(t, value.Exists())
+			assert.Equal(t, tt.want, value.Int())
+		})
+	}
+}
+
+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 TestResponsesRequestToChatCompletionsRequestPreservesPenalties(t *testing.T) {
+	tests := []struct {
+		name          string
+		frequencyRaw  json.RawMessage
+		frequencyWant *float64
+		presenceRaw   json.RawMessage
+		presenceWant  *float64
+	}{
+		{
+			name:          "positive values",
+			frequencyRaw:  json.RawMessage(`0.5`),
+			frequencyWant: lo.ToPtr(0.5),
+			presenceRaw:   json.RawMessage(`1.5`),
+			presenceWant:  lo.ToPtr(1.5),
+		},
+		{
+			name:          "explicit zero values",
+			frequencyRaw:  json.RawMessage(`0.0`),
+			frequencyWant: lo.ToPtr(0.0),
+			presenceRaw:   json.RawMessage(`0.0`),
+			presenceWant:  lo.ToPtr(0.0),
+		},
+		{
+			name:         "unset stays nil",
+			frequencyRaw: nil,
+			presenceRaw:  nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+				Model:            "gpt-test",
+				Input:            mustRawMessage(t, "hello"),
+				FrequencyPenalty: tt.frequencyRaw,
+				PresencePenalty:  tt.presenceRaw,
+			})
+			require.NoError(t, err)
+
+			assert.Equal(t, tt.frequencyWant, got.FrequencyPenalty)
+			assert.Equal(t, tt.presenceWant, got.PresencePenalty)
+		})
+	}
+}
+
+func TestResponsesRequestToChatCompletionsRequestRejectsMalformedPenalty(t *testing.T) {
+	_, err := ResponsesRequestToChatCompletionsRequest(&dto.OpenAIResponsesRequest{
+		Model:            "gpt-test",
+		Input:            mustRawMessage(t, "hello"),
+		FrequencyPenalty: json.RawMessage(`"not-a-number"`),
+	})
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "frequency_penalty")
+}
+
+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 +1,399 @@
+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"
+	responsesEventOutputTextAnnotationAdded = "response.output_text.annotation.added"
+	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 annotations, err := responsesAnnotationsToChat(resp); err != nil {
+		return nil, nil, err
+	} else if len(annotations) > 0 {
+		msg.Annotations = annotations
+	}
+	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 responsesAnnotationsToChat(resp *dto.OpenAIResponsesResponse) ([]byte, error) {
+	annotations := make([]any, 0)
+	for _, output := range resp.Output {
+		if output.Type != responsesOutputTypeMessage {
+			continue
+		}
+		for _, content := range output.Content {
+			for _, annotation := range content.Annotations {
+				converted, err := responseAnnotationToChat(annotation)
+				if err != nil {
+					return nil, err
+				}
+				annotations = append(annotations, converted)
+			}
+		}
+	}
+	if len(annotations) == 0 {
+		return nil, nil
+	}
+	return kitutil.Marshal(annotations)
+}
+
+func responseAnnotationToChat(annotation any) (map[string]any, error) {
+	value, ok := annotation.(map[string]any)
+	if !ok {
+		converted, err := kitutil.Any2Type[map[string]any](annotation)
+		if err != nil {
+			return nil, fmt.Errorf("invalid Responses annotation: %w", err)
+		}
+		value = converted
+	}
+	if strings.TrimSpace(kitutil.Interface2String(value["type"])) != "url_citation" {
+		return value, nil
+	}
+	citation := make(map[string]any, len(value)-1)
+	for key, item := range value {
+		if key != "type" {
+			citation[key] = item
+		}
+	}
+	return map[string]any{
+		"type":         "url_citation",
+		"url_citation": citation,
+	}, nil
+}
+
+func UsageFromResponsesUsage(src *dto.Usage) *dto.Usage {
+	return usageFromResponsesUsage(src, true)
+}
+
+// NormalizeResponsesUsage maps Responses usage into the shared accounting
+// shape without creating a BillingUsage snapshot. Native Responses handlers
+// use it so passthrough traffic preserves an existing snapshot but does not
+// introduce a conversion sidecar solely for local settlement.
+func NormalizeResponsesUsage(src *dto.Usage) *dto.Usage {
+	return usageFromResponsesUsage(src, false)
+}
+
+func usageFromResponsesUsage(src *dto.Usage, createBillingSnapshot bool) *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 && createBillingSnapshot {
+		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
+		}
+		var outputText strings.Builder
+		for _, c := range out.Content {
+			if c.Type == "output_text" && c.Text != "" {
+				outputText.WriteString(c.Text)
+			}
+		}
+		appendSeparatedText(&sb, outputText.String())
+	}
+	if sb.Len() > 0 {
+		return sb.String()
+	}
+	for _, out := range resp.Output {
+		var outputText strings.Builder
+		for _, c := range out.Content {
+			if c.Text != "" {
+				outputText.WriteString(c.Text)
+			}
+		}
+		appendSeparatedText(&sb, outputText.String())
+	}
+	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
+		}
+		appendSeparatedText(&sb, reasoningOutputText(&out))
+	}
+	return sb.String()
+}
+
+func reasoningOutputText(output *dto.ResponsesOutput) string {
+	if output == nil {
+		return ""
+	}
+	var text strings.Builder
+	hasContentText := false
+	for _, part := range output.Content {
+		if part.Text != "" {
+			hasContentText = true
+			break
+		}
+	}
+	if hasContentText {
+		for _, part := range output.Content {
+			appendSeparatedText(&text, part.Text)
+		}
+		return text.String()
+	}
+	for _, part := range output.Summary {
+		appendSeparatedText(&text, part.Text)
+	}
+	return text.String()
+}
+
+func appendSeparatedText(builder *strings.Builder, text string) {
+	if builder == nil || text == "" {
+		return
+	}
+	if builder.Len() > 0 {
+		current := builder.String()
+		trailingNewlines := 0
+		for index := len(current) - 1; index >= 0 && trailingNewlines < 2 && current[index] == '\n'; index-- {
+			trailingNewlines++
+		}
+		leadingNewlines := 0
+		for leadingNewlines < len(text) && leadingNewlines < 2 && text[leadingNewlines] == '\n' {
+			leadingNewlines++
+		}
+		for missing := 2 - trailingNewlines - leadingNewlines; missing > 0; missing-- {
+			builder.WriteByte('\n')
+		}
+	}
+	builder.WriteString(text)
+}
+
+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 +1,547 @@
+package oairesponses
+
+import (
+	"strings"
+	"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,
+				Summary: []dto.ResponsesReasoningSummaryPart{
+					{Type: "summary_text", Text: "first summary"},
+					{Type: "summary_text", Text: "second 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 TestResponsesResponseToChatCompletionsSeparatesInterleavedOutputItems(t *testing.T) {
+	resp := &dto.OpenAIResponsesResponse{
+		ID:     "resp_1",
+		Model:  "gpt-test",
+		Status: []byte(`"completed"`),
+		Output: interleavedReasoningAndTextOutput(),
+	}
+
+	chat, _, err := ResponsesResponseToChatCompletionsResponse(resp, "chatcmpl_1")
+	require.NoError(t, err)
+	assert.Equal(t, "**Planning file inspection**\n\n**Clarifying environment task requirements**", chat.Choices[0].Message.GetReasoningContent())
+	assert.Equal(t, "I’ll inspect the starter repository.\n\nWhat would you like me to build?", 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"}`),
+				},
+			},
+		},
+	})...)
+
+	var totalArgs strings.Builder
+	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.WriteString(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.String())
+	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 TestResponsesBufferedAccumulatorPreservesInterleavedReasoningAndTextItems(t *testing.T) {
+	acc := NewResponsesBufferedAccumulator()
+	events := []dto.ResponsesStreamResponse{
+		bufferedOutputItemAdded(0, "rs_1", responsesOutputTypeReasoning),
+		{Type: responsesEventReasoningSummaryDelta, OutputIndex: intPointer(0), ItemID: "rs_1", Delta: "**Planning file inspection**"},
+		bufferedOutputItemAdded(1, "msg_1", responsesOutputTypeMessage),
+		{Type: responsesEventOutputTextDelta, OutputIndex: intPointer(1), ItemID: "msg_1", Delta: "I’ll inspect the starter repository."},
+		bufferedOutputItemAdded(2, "rs_2", responsesOutputTypeReasoning),
+		{Type: responsesEventReasoningSummaryDelta, OutputIndex: intPointer(2), ItemID: "rs_2", Delta: "**Clarifying environment task requirements**"},
+		bufferedOutputItemAdded(3, "msg_2", responsesOutputTypeMessage),
+		{Type: responsesEventOutputTextDelta, OutputIndex: intPointer(3), ItemID: "msg_2", Delta: "What would you like me to build?"},
+	}
+	for index := range events {
+		acc.ProcessEvent(&events[index])
+	}
+
+	output := acc.BuildOutput()
+	require.Len(t, output, 4)
+	assert.Equal(t, []string{
+		responsesOutputTypeReasoning,
+		responsesOutputTypeMessage,
+		responsesOutputTypeReasoning,
+		responsesOutputTypeMessage,
+	}, []string{output[0].Type, output[1].Type, output[2].Type, output[3].Type})
+	assert.Equal(t, "**Planning file inspection**", output[0].Summary[0].Text)
+	assert.Equal(t, "I’ll inspect the starter repository.", output[1].Content[0].Text)
+	assert.Equal(t, "**Clarifying environment task requirements**", output[2].Summary[0].Text)
+	assert.Equal(t, "What would you like me to build?", output[3].Content[0].Text)
+}
+
+func TestResponsesStreamTerminalOutputPreservesInterleavedReasoningAndTextItems(t *testing.T) {
+	state := newTestResponsesStreamState()
+	chunks, err := ResponsesStreamEventToChatChunks(&dto.ResponsesStreamResponse{
+		Type: responsesEventCompleted,
+		Response: &dto.OpenAIResponsesResponse{
+			Status: []byte(`"completed"`),
+			Output: interleavedReasoningAndTextOutput(),
+		},
+	}, state)
+	require.NoError(t, err)
+
+	var deltas []string
+	for _, chunk := range chunks {
+		if len(chunk.Choices) == 0 {
+			continue
+		}
+		delta := chunk.Choices[0].Delta
+		if delta.ReasoningContent != nil {
+			deltas = append(deltas, "thinking:"+*delta.ReasoningContent)
+		}
+		if delta.Content != nil && *delta.Content != "" {
+			deltas = append(deltas, "text:"+*delta.Content)
+		}
+	}
+	assert.Equal(t, []string{
+		"thinking:**Planning file inspection**",
+		"text:I’ll inspect the starter repository.",
+		"thinking:**Clarifying environment task requirements**",
+		"text:What would you like me to build?",
+	}, deltas)
+}
+
+func bufferedOutputItemAdded(outputIndex int, itemID string, itemType string) dto.ResponsesStreamResponse {
+	return dto.ResponsesStreamResponse{
+		Type:        responsesEventOutputItemAdded,
+		OutputIndex: &outputIndex,
+		ItemID:      itemID,
+		Item:        &dto.ResponsesOutput{ID: itemID, Type: itemType},
+	}
+}
+
+func interleavedReasoningAndTextOutput() []dto.ResponsesOutput {
+	return []dto.ResponsesOutput{
+		{Type: responsesOutputTypeReasoning, Summary: []dto.ResponsesReasoningSummaryPart{{Type: "summary_text", Text: "**Planning file inspection**"}}},
+		{Type: responsesOutputTypeMessage, Role: "assistant", Content: []dto.ResponsesOutputContent{{Type: "output_text", Text: "I’ll inspect the starter repository."}}},
+		{Type: responsesOutputTypeReasoning, Summary: []dto.ResponsesReasoningSummaryPart{{Type: "summary_text", Text: "**Clarifying environment task requirements**"}}},
+		{Type: responsesOutputTypeMessage, Role: "assistant", Content: []dto.ResponsesOutputContent{{Type: "output_text", Text: "What would you like me to build?"}}},
+	}
+}
+
+func intPointer(value int) *int {
+	return &value
+}
+
+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 +1,934 @@
+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
+	sentAnnotationCount        int
+	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) StreamUsage() *dto.Usage {
+	if s == nil {
+		return nil
+	}
+	return s.Usage
+}
+
+func (s *ResponsesToChatStreamState) SetStreamUsage(usage *dto.Usage) {
+	if s != nil && usage != nil {
+		s.Usage = usage
+	}
+}
+
+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 responsesEventOutputTextAnnotationAdded:
+		return state.annotationRawDelta(event.Annotation)
+	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, err := state.terminalOutputChunks(response)
+		if err != nil {
+			return nil, err
+		}
+		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 = dto.MergeUsageNonZero(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, error) {
+	if s == nil || response == nil || len(response.Output) == 0 {
+		return nil, nil
+	}
+
+	var chunks []dto.ChatCompletionsStreamResponse
+	hadSentText := s.hasSentText
+	hadSentReasoning := s.hasSentReasoning
+	annotationOffset := 0
+	for i := range response.Output {
+		out := &response.Output[i]
+		switch {
+		case out.Type == responsesOutputTypeMessage && !hadSentText:
+			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())...)
+			annotationChunks, err := s.remainingAnnotationChunks(out, annotationOffset)
+			if err != nil {
+				return nil, err
+			}
+			chunks = append(chunks, annotationChunks...)
+			annotationOffset += responsesOutputAnnotationCount(out)
+		case out.Type == responsesOutputTypeMessage:
+			annotationChunks, err := s.remainingAnnotationChunks(out, annotationOffset)
+			if err != nil {
+				return nil, err
+			}
+			chunks = append(chunks, annotationChunks...)
+			annotationOffset += responsesOutputAnnotationCount(out)
+		case out.Type == responsesOutputTypeReasoning && !hadSentReasoning:
+			chunks = append(chunks, s.reasoningDelta(reasoningOutputText(out))...)
+		case isResponsesToolOutputType(out.Type):
+			chunks = append(chunks, s.toolItem(&dto.ResponsesStreamResponse{Item: out})...)
+		}
+	}
+	return chunks, nil
+}
+
+func (s *ResponsesToChatStreamState) annotationRawDelta(raw []byte) ([]dto.ChatCompletionsStreamResponse, error) {
+	if len(raw) == 0 {
+		return nil, nil
+	}
+	var annotation map[string]any
+	if err := kitutil.Unmarshal(raw, &annotation); err != nil {
+		return nil, fmt.Errorf("invalid Responses stream annotation: %w", err)
+	}
+	return s.annotationDelta(annotation)
+}
+
+func (s *ResponsesToChatStreamState) annotationDelta(annotation any) ([]dto.ChatCompletionsStreamResponse, error) {
+	converted, err := responseAnnotationToChat(annotation)
+	if err != nil {
+		return nil, err
+	}
+	raw, err := kitutil.Marshal([]any{converted})
+	if err != nil {
+		return nil, fmt.Errorf("marshal Chat annotation: %w", err)
+	}
+	s.sentAnnotationCount++
+	chunks := s.ensureStart()
+	chunks = append(chunks, s.makeChunk(dto.ChatCompletionsStreamResponseChoiceDelta{
+		Annotations: raw,
+	}, nil))
+	return chunks, nil
+}
+
+func (s *ResponsesToChatStreamState) remainingAnnotationChunks(output *dto.ResponsesOutput, offset int) ([]dto.ChatCompletionsStreamResponse, error) {
+	if output == nil {
+		return nil, nil
+	}
+	annotations := make([]any, 0)
+	for _, content := range output.Content {
+		annotations = append(annotations, content.Annotations...)
+	}
+	start := max(s.sentAnnotationCount-offset, 0)
+	if start >= len(annotations) {
+		return nil, nil
+	}
+	var chunks []dto.ChatCompletionsStreamResponse
+	for _, annotation := range annotations[start:] {
+		converted, err := s.annotationDelta(annotation)
+		if err != nil {
+			return nil, err
+		}
+		chunks = append(chunks, converted...)
+	}
+	return chunks, nil
+}
+
+func responsesOutputAnnotationCount(output *dto.ResponsesOutput) int {
+	count := 0
+	for _, content := range output.Content {
+		count += len(content.Annotations)
+	}
+	return count
+}
+
+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 after, ok := strings.CutPrefix(key, "output:"); ok {
+				callID = "call_output_" + after
+			}
+			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 after, ok := strings.CutPrefix(key, "item:"); ok {
+			itemID := after
+			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 {
+	items                []*responsesBufferedItem
+	outputIndexToItemIdx map[int]int
+	itemIDToItemIdx      map[string]int
+	lastUnindexedItemIdx int
+	tools                []*responsesBufferedTool
+	outputIndexToToolIdx map[int]int
+	itemIDToToolIdx      map[string]int
+	pendingByOutputIndex map[int]string
+	pendingByItemID      map[string]string
+}
+
+type responsesBufferedItem struct {
+	Type                string
+	ID                  string
+	Text                strings.Builder
+	Annotations         []any
+	ToolIndex           int
+	NeedsReasoningBreak bool
+}
+
+type responsesBufferedTool struct {
+	CallID    string
+	ItemID    string
+	Name      string
+	Arguments strings.Builder
+}
+
+func NewResponsesBufferedAccumulator() *ResponsesBufferedAccumulator {
+	return &ResponsesBufferedAccumulator{
+		outputIndexToItemIdx: make(map[int]int),
+		itemIDToItemIdx:      make(map[string]int),
+		lastUnindexedItemIdx: -1,
+		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:
+		item := a.ensureItem(event, responsesOutputTypeMessage)
+		item.Text.WriteString(event.Delta)
+	case responsesEventOutputTextAnnotationAdded:
+		item := a.ensureItem(event, responsesOutputTypeMessage)
+		var annotation any
+		if err := kitutil.Unmarshal(event.Annotation, &annotation); err == nil && annotation != nil {
+			item.Annotations = append(item.Annotations, annotation)
+		}
+	case responsesEventReasoningSummaryDelta, responsesEventReasoningTextDelta:
+		item := a.ensureItem(event, responsesOutputTypeReasoning)
+		if item.NeedsReasoningBreak {
+			appendSeparatedText(&item.Text, event.Delta)
+			item.NeedsReasoningBreak = false
+		} else {
+			item.Text.WriteString(event.Delta)
+		}
+	case responsesEventReasoningSummaryDone, responsesEventReasoningTextDone:
+		item := a.ensureItem(event, responsesOutputTypeReasoning)
+		if item.Text.Len() == 0 && event.Text != nil {
+			item.Text.WriteString(*event.Text)
+		}
+		if item.Text.Len() > 0 {
+			item.NeedsReasoningBreak = true
+		}
+	case responsesEventOutputItemAdded, responsesEventOutputItemDone:
+		if event.Item == nil {
+			return
+		}
+		switch {
+		case event.Item.Type == responsesOutputTypeReasoning:
+			item := a.ensureItem(event, event.Item.Type)
+			if item.Text.Len() == 0 {
+				item.Text.WriteString(reasoningOutputText(event.Item))
+			}
+		case event.Item.Type == responsesOutputTypeMessage:
+			item := a.ensureItem(event, event.Item.Type)
+			seedText := item.Text.Len() == 0
+			seedAnnotations := len(item.Annotations) == 0
+			for _, content := range event.Item.Content {
+				if content.Type == "output_text" {
+					if seedText {
+						item.Text.WriteString(content.Text)
+					}
+					if seedAnnotations {
+						item.Annotations = append(item.Annotations, content.Annotations...)
+					}
+				}
+			}
+		case 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, len(a.items))
+	for _, item := range a.items {
+		if item == nil {
+			continue
+		}
+		switch item.Type {
+		case responsesOutputTypeReasoning:
+			if item.Text.Len() == 0 {
+				continue
+			}
+			out = append(out, dto.ResponsesOutput{
+				Type: item.Type,
+				ID:   item.ID,
+				Summary: []dto.ResponsesReasoningSummaryPart{
+					{Type: "summary_text", Text: item.Text.String()},
+				},
+			})
+		case responsesOutputTypeMessage:
+			if item.Text.Len() == 0 {
+				continue
+			}
+			out = append(out, dto.ResponsesOutput{
+				Type: item.Type,
+				ID:   item.ID,
+				Role: "assistant",
+				Content: []dto.ResponsesOutputContent{
+					{Type: "output_text", Text: item.Text.String(), Annotations: item.Annotations},
+				},
+			})
+		case responsesOutputTypeFunctionCall, responsesOutputTypeCustomToolCall:
+			if item.ToolIndex < 0 || item.ToolIndex >= len(a.tools) || a.tools[item.ToolIndex] == nil {
+				continue
+			}
+			tool := a.tools[item.ToolIndex]
+			argsRaw, _ := kitutil.Marshal(tool.Arguments.String())
+			out = append(out, dto.ResponsesOutput{
+				Type:      item.Type,
+				ID:        tool.ItemID,
+				CallId:    tool.CallID,
+				Name:      tool.Name,
+				Arguments: argsRaw,
+			})
+		}
+	}
+	return out
+}
+
+func (a *ResponsesBufferedAccumulator) ensureItem(event *dto.ResponsesStreamResponse, itemType string) *responsesBufferedItem {
+	if idx, ok := a.findItemIndex(event); ok {
+		item := a.items[idx]
+		if item.Type == "" {
+			item.Type = itemType
+		}
+		a.applyItemMetadata(idx, item, event)
+		return item
+	}
+	if event != nil && event.OutputIndex == nil && responseStreamEventItemID(event) == "" && a.lastUnindexedItemIdx >= 0 {
+		item := a.items[a.lastUnindexedItemIdx]
+		if item != nil && item.Type == itemType {
+			return item
+		}
+	}
+	item := &responsesBufferedItem{Type: itemType, ToolIndex: -1}
+	idx := len(a.items)
+	a.items = append(a.items, item)
+	a.lastUnindexedItemIdx = idx
+	a.applyItemMetadata(idx, item, event)
+	return item
+}
+
+func (a *ResponsesBufferedAccumulator) applyItemMetadata(idx int, item *responsesBufferedItem, event *dto.ResponsesStreamResponse) {
+	if item == nil || event == nil {
+		return
+	}
+	if event.OutputIndex != nil {
+		a.outputIndexToItemIdx[*event.OutputIndex] = idx
+	}
+	if itemID := responseStreamEventItemID(event); itemID != "" {
+		item.ID = itemID
+		a.itemIDToItemIdx[itemID] = idx
+	}
+}
+
+func (a *ResponsesBufferedAccumulator) findItemIndex(event *dto.ResponsesStreamResponse) (int, bool) {
+	if event == nil {
+		return 0, false
+	}
+	if event.OutputIndex != nil {
+		if idx, ok := a.outputIndexToItemIdx[*event.OutputIndex]; ok {
+			return idx, true
+		}
+	}
+	if itemID := responseStreamEventItemID(event); itemID != "" {
+		idx, ok := a.itemIDToItemIdx[itemID]
+		return idx, ok
+	}
+	return 0, false
+}
+
+func (a *ResponsesBufferedAccumulator) ensureTool(event *dto.ResponsesStreamResponse) *responsesBufferedTool {
+	if idx, ok := a.findToolIndex(event); ok {
+		tool := a.tools[idx]
+		a.applyToolMetadata(tool, event)
+		item := a.ensureItem(event, event.Item.Type)
+		item.ToolIndex = idx
+		return tool
+	}
+	tool := &responsesBufferedTool{}
+	a.applyToolMetadata(tool, event)
+	idx := len(a.tools)
+	a.tools = append(a.tools, tool)
+	item := a.ensureItem(event, event.Item.Type)
+	item.ToolIndex = idx
+	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
+}