feat: support Responses to Chat (#5787)

* fix(openai): harden Chat-to-Responses compatibility

Add a shared Responses-to-Chat stream state machine and use it from the OpenAI relay path. Preserve assistant text alongside tool calls, bind tool argument deltas by output_index, map incomplete finish reasons, support reasoning/custom tool events, and buffer upstream SSE for non-stream Chat clients.

Add deterministic service tests and relay SSE tests for the conversion path.

Related to #5745.

* refactor: rename openaicompat to relayconvert for improved clarity

* feat(gemini): support responses request conversion

* feat: add responses to chat conversion support

* fix: harden responses chat conversion edge cases

relay/channel/gemini/adaptor.go

@@ -12,6 +12,7 @@ import (
 	"github.com/QuantumNous/new-api/relay/channel/openai"
 	relaycommon "github.com/QuantumNous/new-api/relay/common"
 	"github.com/QuantumNous/new-api/relay/constant"
+	"github.com/QuantumNous/new-api/service/relayconvert"
 	"github.com/QuantumNous/new-api/setting/model_setting"
 	"github.com/QuantumNous/new-api/setting/reasoning"
 	"github.com/QuantumNous/new-api/types"
@@ -238,8 +239,17 @@ func (a *Adaptor) ConvertEmbeddingRequest(c *gin.Context, info *relaycommon.Rela
 }
 
 func (a *Adaptor) ConvertOpenAIResponsesRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.OpenAIResponsesRequest) (any, error) {
-	// TODO implement me
-	return nil, errors.New("not implemented")
+	request, err := preprocessGeminiOpenAIResponsesRequest(request)
+	if err != nil {
+		return nil, err
+	}
+
+	chatRequest, err := relayconvert.ResponsesRequestToChatCompletionsRequest(&request)
+	if err != nil {
+		return nil, err
+	}
+
+	return a.ConvertOpenAIRequest(c, info, chatRequest)
 }
 
 func (a *Adaptor) DoRequest(c *gin.Context, info *relaycommon.RelayInfo, requestBody io.Reader) (any, error) {
@@ -247,6 +257,13 @@ func (a *Adaptor) DoRequest(c *gin.Context, info *relaycommon.RelayInfo, request
 }
 
 func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycommon.RelayInfo) (usage any, err *types.NewAPIError) {
+	if info.RelayMode == constant.RelayModeResponses {
+		if info.IsStream {
+			return GeminiResponsesStreamHandler(c, info, resp)
+		}
+		return GeminiResponsesHandler(c, info, resp)
+	}
+
 	if info.RelayMode == constant.RelayModeGemini {
 		if strings.Contains(info.RequestURLPath, ":embedContent") ||
 			strings.Contains(info.RequestURLPath, ":batchEmbedContents") {

relay/channel/gemini/adaptor_responses.go

@@ -0,0 +1,99 @@
+package gemini
+
+import (
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/dto"
+)
+
+const (
+	geminiResponsesInputTypeCustomToolCall       = "custom_tool_call"
+	geminiResponsesInputTypeCustomToolCallOutput = "custom_tool_call_output"
+	geminiResponsesInputTypeFunctionCallOutput   = "function_call_output"
+)
+
+func preprocessGeminiOpenAIResponsesRequest(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) || common.GetJsonType(raw) != "array" {
+		return raw, nil
+	}
+
+	var tools []map[string]any
+	if err := common.Unmarshal(raw, &tools); err != nil {
+		return nil, err
+	}
+
+	filtered := make([]map[string]any, 0, len(tools))
+	for _, tool := range tools {
+		if strings.TrimSpace(common.Interface2String(tool["type"])) != "function" {
+			// TODO: Support Responses custom/freeform tools when Gemini has a safe equivalent representation.
+			continue
+		}
+		filtered = append(filtered, tool)
+	}
+	if len(filtered) == 0 {
+		return nil, nil
+	}
+	return common.Marshal(filtered)
+}
+
+func filterGeminiResponsesInput(raw []byte) ([]byte, error) {
+	if !geminiRawJSONPresent(raw) || common.GetJsonType(raw) != "array" {
+		return raw, nil
+	}
+
+	var items []map[string]any
+	if err := common.Unmarshal(raw, &items); err != nil {
+		return nil, err
+	}
+
+	skippedCustomCallIDs := make(map[string]struct{})
+	for _, item := range items {
+		if strings.TrimSpace(common.Interface2String(item["type"])) != geminiResponsesInputTypeCustomToolCall {
+			continue
+		}
+		if callID := strings.TrimSpace(common.Interface2String(item["call_id"])); callID != "" {
+			skippedCustomCallIDs[callID] = struct{}{}
+		}
+	}
+
+	filtered := make([]map[string]any, 0, len(items))
+	for _, item := range items {
+		itemType := strings.TrimSpace(common.Interface2String(item["type"]))
+		switch itemType {
+		case geminiResponsesInputTypeCustomToolCall, geminiResponsesInputTypeCustomToolCallOutput:
+			// TODO: Support Responses custom/freeform tool calls once Gemini can preserve their semantics.
+			continue
+		case geminiResponsesInputTypeFunctionCallOutput:
+			if _, ok := skippedCustomCallIDs[strings.TrimSpace(common.Interface2String(item["call_id"]))]; ok {
+				continue
+			}
+		}
+		filtered = append(filtered, item)
+	}
+
+	return common.Marshal(filtered)
+}
+
+func geminiRawJSONPresent(raw []byte) bool {
+	if len(raw) == 0 {
+		return false
+	}
+	return common.GetJsonType(raw) != "null"
+}

relay/channel/gemini/adaptor_responses_test.go

@@ -0,0 +1,176 @@
+package gemini
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/dto"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertOpenAIResponsesRequestToGeminiInstructionsAndInput(t *testing.T) {
+	got := mustConvertResponsesToGemini(t, dto.OpenAIResponsesRequest{
+		Model:        "gemini-test",
+		Instructions: mustGeminiRawMessage(t, "system rules"),
+		Input:        mustGeminiRawMessage(t, "hello"),
+	})
+
+	require.NotNil(t, got.SystemInstructions)
+	require.Len(t, got.SystemInstructions.Parts, 1)
+	assert.Equal(t, "system rules", got.SystemInstructions.Parts[0].Text)
+	require.Len(t, got.Contents, 1)
+	assert.Equal(t, "user", got.Contents[0].Role)
+	require.Len(t, got.Contents[0].Parts, 1)
+	assert.Equal(t, "hello", got.Contents[0].Parts[0].Text)
+}
+
+func TestConvertOpenAIResponsesRequestToGeminiFunctionToolAndChoice(t *testing.T) {
+	got := mustConvertResponsesToGemini(t, dto.OpenAIResponsesRequest{
+		Model: "gemini-test",
+		Input: mustGeminiRawMessage(t, "lookup weather"),
+		Tools: mustGeminiRawMessage(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"},
+					},
+				},
+			},
+			{"type": "custom", "name": "freeform"},
+		}),
+		ToolChoice: mustGeminiRawMessage(t, map[string]any{
+			"type": "function",
+			"name": "lookup",
+		}),
+	})
+
+	tools := got.GetTools()
+	require.Len(t, tools, 1)
+	assert.Equal(t, "lookup", gjson.GetBytes(got.Tools, "0.functionDeclarations.0.name").String())
+	assert.Equal(t, "Lookup data", gjson.GetBytes(got.Tools, "0.functionDeclarations.0.description").String())
+	require.NotNil(t, got.ToolConfig)
+	require.NotNil(t, got.ToolConfig.FunctionCallingConfig)
+	assert.Equal(t, dto.FunctionCallingConfigMode("ANY"), got.ToolConfig.FunctionCallingConfig.Mode)
+	assert.Equal(t, []string{"lookup"}, got.ToolConfig.FunctionCallingConfig.AllowedFunctionNames)
+}
+
+func TestConvertOpenAIResponsesRequestToGeminiFunctionCallConversation(t *testing.T) {
+	got := mustConvertResponsesToGemini(t, dto.OpenAIResponsesRequest{
+		Model: "gemini-test",
+		Input: mustGeminiRawMessage(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},
+			},
+		}),
+		Tools: mustGeminiRawMessage(t, []map[string]any{
+			{"type": "function", "name": "lookup", "parameters": map[string]any{"type": "object"}},
+		}),
+	})
+
+	require.Len(t, got.Contents, 2)
+	assert.Equal(t, "model", got.Contents[0].Role)
+	require.Len(t, got.Contents[0].Parts, 2)
+	require.NotNil(t, got.Contents[0].Parts[0].FunctionCall)
+	assert.Equal(t, "lookup", got.Contents[0].Parts[0].FunctionCall.FunctionName)
+	assert.Equal(t, map[string]interface{}{"q": "x"}, got.Contents[0].Parts[0].FunctionCall.Arguments)
+	assert.Equal(t, "I will call.", got.Contents[0].Parts[1].Text)
+
+	assert.Equal(t, "user", got.Contents[1].Role)
+	require.Len(t, got.Contents[1].Parts, 1)
+	require.NotNil(t, got.Contents[1].Parts[0].FunctionResponse)
+	assert.Equal(t, "lookup", got.Contents[1].Parts[0].FunctionResponse.Name)
+	assert.Equal(t, map[string]interface{}{"ok": true}, got.Contents[1].Parts[0].FunctionResponse.Response)
+}
+
+func TestConvertOpenAIResponsesRequestToGeminiSkipsCustomToolCalls(t *testing.T) {
+	got := mustConvertResponsesToGemini(t, dto.OpenAIResponsesRequest{
+		Model: "gemini-test",
+		Input: mustGeminiRawMessage(t, []map[string]any{
+			{
+				"role": "assistant",
+				"content": []map[string]any{
+					{"type": "output_text", "text": "before custom"},
+				},
+			},
+			{
+				"type":    "custom_tool_call",
+				"call_id": "call_custom",
+				"name":    "apply_patch",
+				"input":   "patch body",
+			},
+			{
+				"type":    "custom_tool_call_output",
+				"call_id": "call_custom",
+				"output":  "ok",
+			},
+			{
+				"type":    "function_call_output",
+				"call_id": "call_custom",
+				"output":  "legacy custom output",
+			},
+			{
+				"role":    "user",
+				"content": "next turn",
+			},
+		}),
+		Tools: mustGeminiRawMessage(t, []map[string]any{
+			{"type": "custom", "name": "apply_patch"},
+			{"type": "unknown", "name": "unknown"},
+		}),
+	})
+
+	assert.Empty(t, got.GetTools())
+	require.Len(t, got.Contents, 2)
+	assert.Equal(t, "model", got.Contents[0].Role)
+	require.Len(t, got.Contents[0].Parts, 1)
+	assert.Equal(t, "before custom", got.Contents[0].Parts[0].Text)
+	assert.Nil(t, got.Contents[0].Parts[0].FunctionCall)
+
+	assert.Equal(t, "user", got.Contents[1].Role)
+	require.Len(t, got.Contents[1].Parts, 1)
+	assert.Equal(t, "next turn", got.Contents[1].Parts[0].Text)
+	assert.Nil(t, got.Contents[1].Parts[0].FunctionResponse)
+}
+
+func mustConvertResponsesToGemini(t *testing.T, req dto.OpenAIResponsesRequest) *dto.GeminiChatRequest {
+	t.Helper()
+	info := &relaycommon.RelayInfo{
+		OriginModelName: req.Model,
+		ChannelMeta: &relaycommon.ChannelMeta{
+			UpstreamModelName: req.Model,
+		},
+	}
+	got, err := (&Adaptor{}).ConvertOpenAIResponsesRequest(nil, info, req)
+	require.NoError(t, err)
+	geminiReq, ok := got.(*dto.GeminiChatRequest)
+	require.True(t, ok)
+	return geminiReq
+}
+
+func mustGeminiRawMessage(t *testing.T, value any) []byte {
+	t.Helper()
+	raw, err := common.Marshal(value)
+	require.NoError(t, err)
+	return raw
+}

relay/channel/gemini/relay_responses.go

@@ -0,0 +1,162 @@
+package gemini
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/service"
+	"github.com/QuantumNous/new-api/service/relayconvert"
+	"github.com/QuantumNous/new-api/types"
+	"github.com/gin-gonic/gin"
+)
+
+func GeminiResponsesHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	defer service.CloseResponseBodyGracefully(resp)
+
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	logger.LogDebug(c, "Gemini responses response body: %s", responseBody)
+
+	var geminiResponse dto.GeminiChatResponse
+	if err := common.Unmarshal(responseBody, &geminiResponse); err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if len(geminiResponse.Candidates) == 0 {
+		usage := buildUsageFromGeminiMetadata(geminiResponse.UsageMetadata, info.GetEstimatePromptTokens())
+		if geminiResponse.PromptFeedback != nil && geminiResponse.PromptFeedback.BlockReason != nil {
+			common.SetContextKey(c, constant.ContextKeyAdminRejectReason, fmt.Sprintf("gemini_block_reason=%s", *geminiResponse.PromptFeedback.BlockReason))
+			return &usage, types.NewOpenAIError(
+				errors.New("request blocked by Gemini API: "+*geminiResponse.PromptFeedback.BlockReason),
+				types.ErrorCodePromptBlocked,
+				http.StatusBadRequest,
+			)
+		}
+		common.SetContextKey(c, constant.ContextKeyAdminRejectReason, "gemini_empty_candidates")
+		return &usage, types.NewOpenAIError(
+			errors.New("empty response from Gemini API"),
+			types.ErrorCodeEmptyResponse,
+			http.StatusInternalServerError,
+		)
+	}
+
+	chatResp := responseGeminiChat2OpenAI(c, &geminiResponse)
+	chatResp.Model = info.UpstreamModelName
+	usage := buildUsageFromGeminiMetadata(geminiResponse.UsageMetadata, info.GetEstimatePromptTokens())
+	chatResp.Usage = usage
+
+	responsesResp, responsesUsage, err := service.ChatCompletionsResponseToResponsesResponse(chatResp, helper.GetResponseID(c))
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if responsesUsage == nil || responsesUsage.TotalTokens == 0 {
+		responsesResp.Usage = relayconvert.UsageFromChatUsage(&usage)
+	}
+
+	responseBody, err = common.Marshal(responsesResp)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeJsonMarshalFailed, http.StatusInternalServerError)
+	}
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+	return &usage, nil
+}
+
+func GeminiResponsesStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	responseID := helper.GetResponseID(c)
+	created := common.GetTimestamp()
+	state := relayconvert.NewChatToResponsesStreamState(responseID, info.UpstreamModelName)
+	state.Created = created
+	finishReason := constant.FinishReasonStop
+	toolCallIndexByChoice := make(map[int]map[string]int)
+	nextToolCallIndexByChoice := make(map[int]int)
+	var streamErr *types.NewAPIError
+
+	sendEvent := func(event relayconvert.ChatToResponsesStreamEvent) bool {
+		data, err := common.Marshal(event.Payload)
+		if err != nil {
+			streamErr = types.NewOpenAIError(err, types.ErrorCodeJsonMarshalFailed, http.StatusInternalServerError)
+			return false
+		}
+		helper.ResponseChunkData(c, dto.ResponsesStreamResponse{Type: event.Type}, string(data))
+		return true
+	}
+	sendChunk := func(chunk *dto.ChatCompletionsStreamResponse) bool {
+		events, err := relayconvert.ChatCompletionsStreamChunkToResponsesEvents(chunk, state)
+		if err != nil {
+			streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+			return false
+		}
+		for _, event := range events {
+			if !sendEvent(event) {
+				return false
+			}
+		}
+		return true
+	}
+
+	usage, err := geminiStreamHandler(c, info, resp, func(data string, geminiResponse *dto.GeminiChatResponse) bool {
+		response, isStop := streamResponseGeminiChat2OpenAI(geminiResponse)
+		response.Id = responseID
+		response.Created = created
+		response.Model = info.UpstreamModelName
+
+		if response.IsToolCall() {
+			finishReason = constant.FinishReasonToolCalls
+		}
+		for choiceIdx := range response.Choices {
+			choiceKey := response.Choices[choiceIdx].Index
+			for toolIdx := range response.Choices[choiceIdx].Delta.ToolCalls {
+				tool := &response.Choices[choiceIdx].Delta.ToolCalls[toolIdx]
+				if tool.ID == "" {
+					continue
+				}
+				indexByID := toolCallIndexByChoice[choiceKey]
+				if indexByID == nil {
+					indexByID = make(map[string]int)
+					toolCallIndexByChoice[choiceKey] = indexByID
+				}
+				if idx, ok := indexByID[tool.ID]; ok {
+					tool.SetIndex(idx)
+					continue
+				}
+				idx := nextToolCallIndexByChoice[choiceKey]
+				nextToolCallIndexByChoice[choiceKey] = idx + 1
+				indexByID[tool.ID] = idx
+				tool.SetIndex(idx)
+			}
+		}
+
+		if !sendChunk(response) {
+			return false
+		}
+		if isStop {
+			return sendChunk(helper.GenerateStopResponse(responseID, created, info.UpstreamModelName, finishReason))
+		}
+		return true
+	})
+	if err != nil {
+		return usage, err
+	}
+	if streamErr != nil {
+		return nil, streamErr
+	}
+
+	if usage != nil {
+		state.Usage = relayconvert.UsageFromChatUsage(usage)
+	}
+	for _, event := range relayconvert.FinalizeChatCompletionsStreamToResponses(state) {
+		if !sendEvent(event) {
+			return nil, streamErr
+		}
+	}
+	return usage, nil
+}

relay/channel/gemini/relay_responses_test.go

@@ -0,0 +1,205 @@
+package gemini
+
+import (
+	"bytes"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	relayconstant "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/QuantumNous/new-api/types"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestGeminiResponsesHandlerReturnsOpenAIResponsesJSON(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	c.Set(common.RequestIdKey, "gemini-responses-test")
+
+	info := newGeminiResponsesRelayInfo(false)
+	payload := dto.GeminiChatResponse{
+		Candidates: []dto.GeminiChatCandidate{
+			{
+				Content: dto.GeminiChatContent{
+					Role: "model",
+					Parts: []dto.GeminiPart{
+						{Text: "hello"},
+					},
+				},
+			},
+		},
+		UsageMetadata: dto.GeminiUsageMetadata{
+			PromptTokenCount:     2,
+			CandidatesTokenCount: 3,
+			TotalTokenCount:      5,
+		},
+	}
+	body, err := common.Marshal(payload)
+	require.NoError(t, err)
+
+	usage, newAPIError := GeminiResponsesHandler(c, info, &http.Response{
+		Body: io.NopCloser(bytes.NewReader(body)),
+	})
+	require.Nil(t, newAPIError)
+	require.NotNil(t, usage)
+	assert.Equal(t, 2, usage.PromptTokens)
+	assert.Equal(t, 3, usage.CompletionTokens)
+
+	got := recorder.Body.String()
+	assert.Contains(t, got, `"object":"response"`)
+	assert.Contains(t, got, `"status":"completed"`)
+	assert.Contains(t, got, `"type":"output_text"`)
+	assert.Contains(t, got, `"text":"hello"`)
+	assert.Contains(t, got, `"input_tokens":2`)
+	assert.Contains(t, got, `"output_tokens":3`)
+	assert.NotContains(t, got, `"choices"`)
+	assert.NotContains(t, got, `"candidates"`)
+}
+
+func TestGeminiResponsesHandlerClosesBodyOnReadError(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	c.Set(common.RequestIdKey, "gemini-responses-read-error-test")
+
+	body := &failingReadCloser{}
+	usage, newAPIError := GeminiResponsesHandler(c, newGeminiResponsesRelayInfo(false), &http.Response{Body: body})
+
+	require.Nil(t, usage)
+	require.NotNil(t, newAPIError)
+	assert.True(t, body.closed)
+}
+
+func TestGeminiResponsesStreamHandlerReturnsOpenAIResponsesSSE(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	c.Set(common.RequestIdKey, "gemini-responses-stream-test")
+
+	oldStreamingTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 300
+	t.Cleanup(func() { constant.StreamingTimeout = oldStreamingTimeout })
+
+	info := newGeminiResponsesRelayInfo(true)
+	first := dto.GeminiChatResponse{
+		Candidates: []dto.GeminiChatCandidate{
+			{
+				Content: dto.GeminiChatContent{
+					Role: "model",
+					Parts: []dto.GeminiPart{
+						{Text: "hello"},
+					},
+				},
+			},
+		},
+		UsageMetadata: dto.GeminiUsageMetadata{
+			PromptTokenCount:     2,
+			CandidatesTokenCount: 3,
+			TotalTokenCount:      5,
+		},
+	}
+	stop := "STOP"
+	final := dto.GeminiChatResponse{
+		Candidates: []dto.GeminiChatCandidate{
+			{
+				FinishReason: &stop,
+				Content: dto.GeminiChatContent{
+					Role:  "model",
+					Parts: []dto.GeminiPart{{Text: ""}},
+				},
+			},
+		},
+		UsageMetadata: dto.GeminiUsageMetadata{
+			PromptTokenCount:     2,
+			CandidatesTokenCount: 3,
+			TotalTokenCount:      5,
+		},
+	}
+	firstData, err := common.Marshal(first)
+	require.NoError(t, err)
+	finalData, err := common.Marshal(final)
+	require.NoError(t, err)
+	streamBody := strings.Join([]string{
+		"data: " + string(firstData),
+		"",
+		"data: " + string(finalData),
+		"",
+		"data: [DONE]",
+		"",
+	}, "\n")
+
+	usage, newAPIError := GeminiResponsesStreamHandler(c, info, &http.Response{
+		Body: io.NopCloser(strings.NewReader(streamBody)),
+	})
+	require.Nil(t, newAPIError)
+	require.NotNil(t, usage)
+	assert.Equal(t, 5, usage.TotalTokens)
+
+	got := recorder.Body.String()
+	assert.Equal(t, "text/event-stream", recorder.Header().Get("Content-Type"))
+	assert.Contains(t, got, `event: response.created`)
+	assert.Contains(t, got, `event: response.output_text.delta`)
+	assert.Contains(t, got, `"delta":"hello"`)
+	assert.Contains(t, got, `event: response.completed`)
+	assert.Contains(t, got, `"input_tokens":2`)
+	assert.Contains(t, got, `"output_tokens":3`)
+	assert.NotContains(t, got, `"choices"`)
+	assert.NotContains(t, got, `"candidates"`)
+	requireOrderedGeminiResponsesSubstrings(t, got,
+		`event: response.created`,
+		`event: response.output_item.added`,
+		`event: response.output_text.delta`,
+		`event: response.output_text.done`,
+		`event: response.completed`,
+	)
+}
+
+func newGeminiResponsesRelayInfo(isStream bool) *relaycommon.RelayInfo {
+	return &relaycommon.RelayInfo{
+		IsStream:        isStream,
+		RelayMode:       relayconstant.RelayModeResponses,
+		RelayFormat:     types.RelayFormatOpenAIResponses,
+		RequestURLPath:  "/v1/responses",
+		DisablePing:     true,
+		OriginModelName: "gemini-test",
+		ChannelMeta: &relaycommon.ChannelMeta{
+			UpstreamModelName: "gemini-test",
+		},
+	}
+}
+
+type failingReadCloser struct {
+	closed bool
+}
+
+func (r *failingReadCloser) Read([]byte) (int, error) {
+	return 0, errors.New("read failed")
+}
+
+func (r *failingReadCloser) Close() error {
+	r.closed = true
+	return nil
+}
+
+func requireOrderedGeminiResponsesSubstrings(t *testing.T, s string, parts ...string) {
+	t.Helper()
+	offset := 0
+	for _, part := range parts {
+		idx := strings.Index(s[offset:], part)
+		require.NotEqualf(t, -1, idx, "missing %q after byte offset %d", part, offset)
+		offset += idx + len(part)
+	}
+}