Make gateway Open Responses compliant

internal/api/types.go

@@ -1,81 +1,307 @@
 package api
 
 import (
+	"encoding/json"
 	"errors"
 	"fmt"
 )
 
-// ResponseRequest models the Open Responses create request payload.
+// ============================================================
+// Request Types (CreateResponseBody)
+// ============================================================
+
+// ResponseRequest models the OpenResponses CreateResponseBody.
 type ResponseRequest struct {
 	Model              string            `json:"model"`
-	Provider           string            `json:"provider,omitempty"`
-	MaxOutputTokens    int               `json:"max_output_tokens,omitempty"`
+	Input              InputUnion        `json:"input"`
+	Instructions       *string           `json:"instructions,omitempty"`
+	MaxOutputTokens    *int              `json:"max_output_tokens,omitempty"`
 	Metadata           map[string]string `json:"metadata,omitempty"`
-	Input              []Message         `json:"input"`
 	Stream             bool              `json:"stream,omitempty"`
-	PreviousResponseID string            `json:"previous_response_id,omitempty"`
+	PreviousResponseID *string           `json:"previous_response_id,omitempty"`
+	Temperature        *float64          `json:"temperature,omitempty"`
+	TopP               *float64          `json:"top_p,omitempty"`
+	FrequencyPenalty   *float64          `json:"frequency_penalty,omitempty"`
+	PresencePenalty    *float64          `json:"presence_penalty,omitempty"`
+	TopLogprobs        *int              `json:"top_logprobs,omitempty"`
+	Truncation         *string           `json:"truncation,omitempty"`
+	ToolChoice         json.RawMessage   `json:"tool_choice,omitempty"`
+	Tools              json.RawMessage   `json:"tools,omitempty"`
+	ParallelToolCalls  *bool             `json:"parallel_tool_calls,omitempty"`
+	Store              *bool             `json:"store,omitempty"`
+	Text               json.RawMessage   `json:"text,omitempty"`
+	Reasoning          json.RawMessage   `json:"reasoning,omitempty"`
+	Include            []string          `json:"include,omitempty"`
+	ServiceTier        *string           `json:"service_tier,omitempty"`
+	Background         *bool             `json:"background,omitempty"`
+	StreamOptions      json.RawMessage   `json:"stream_options,omitempty"`
+	MaxToolCalls       *int              `json:"max_tool_calls,omitempty"`
+
+	// Non-spec extension: allows client to select a specific provider.
+	Provider string `json:"provider,omitempty"`
 }
 
-// Message captures user, assistant, or system roles.
+// InputUnion handles the polymorphic "input" field: string or []InputItem.
+type InputUnion struct {
+	String *string
+	Items  []InputItem
+}
+
+func (u *InputUnion) UnmarshalJSON(data []byte) error {
+	if string(data) == "null" {
+		return nil
+	}
+	var s string
+	if err := json.Unmarshal(data, &s); err == nil {
+		u.String = &s
+		return nil
+	}
+	var items []InputItem
+	if err := json.Unmarshal(data, &items); err == nil {
+		u.Items = items
+		return nil
+	}
+	return fmt.Errorf("input must be a string or array of items")
+}
+
+func (u InputUnion) MarshalJSON() ([]byte, error) {
+	if u.String != nil {
+		return json.Marshal(*u.String)
+	}
+	if u.Items != nil {
+		return json.Marshal(u.Items)
+	}
+	return []byte("null"), nil
+}
+
+// InputItem is a discriminated union on "type".
+// Valid types: message, item_reference, function_call, function_call_output, reasoning.
+type InputItem struct {
+	Type      string          `json:"type"`
+	Role      string          `json:"role,omitempty"`
+	Content   json.RawMessage `json:"content,omitempty"`
+	ID        string          `json:"id,omitempty"`
+	CallID    string          `json:"call_id,omitempty"`
+	Name      string          `json:"name,omitempty"`
+	Arguments string          `json:"arguments,omitempty"`
+	Output    string          `json:"output,omitempty"`
+	Status    string          `json:"status,omitempty"`
+}
+
+// ============================================================
+// Internal Types (providers + conversation store)
+// ============================================================
+
+// Message is the normalized internal message representation.
 type Message struct {
 	Role    string         `json:"role"`
 	Content []ContentBlock `json:"content"`
 }
 
-// ContentBlock represents a typed content element (text, data, tool call, etc.).
+// ContentBlock is a typed content element.
 type ContentBlock struct {
 	Type string `json:"type"`
 	Text string `json:"text,omitempty"`
 }
 
-// Response is a simplified Open Responses response payload.
+// NormalizeInput converts the request Input into messages for providers.
+// Does NOT include instructions (the server prepends those separately).
+func (r *ResponseRequest) NormalizeInput() []Message {
+	if r.Input.String != nil {
+		return []Message{{
+			Role:    "user",
+			Content: []ContentBlock{{Type: "input_text", Text: *r.Input.String}},
+		}}
+	}
+
+	var msgs []Message
+	for _, item := range r.Input.Items {
+		switch item.Type {
+		case "message", "":
+			msg := Message{Role: item.Role}
+			if item.Content != nil {
+				var s string
+				if err := json.Unmarshal(item.Content, &s); err == nil {
+					contentType := "input_text"
+					if item.Role == "assistant" {
+						contentType = "output_text"
+					}
+					msg.Content = []ContentBlock{{Type: contentType, Text: s}}
+				} else {
+					var blocks []ContentBlock
+					_ = json.Unmarshal(item.Content, &blocks)
+					msg.Content = blocks
+				}
+			}
+			msgs = append(msgs, msg)
+		case "function_call_output":
+			msgs = append(msgs, Message{
+				Role:    "tool",
+				Content: []ContentBlock{{Type: "input_text", Text: item.Output}},
+			})
+		}
+	}
+	return msgs
+}
+
+// ============================================================
+// Response Types (ResponseResource)
+// ============================================================
+
+// Response is the spec-compliant ResponseResource.
 type Response struct {
-	ID       string    `json:"id"`
-	Object   string    `json:"object"`
-	Created  int64     `json:"created"`
-	Model    string    `json:"model"`
-	Provider string    `json:"provider"`
-	Output   []Message `json:"output"`
-	Usage    Usage     `json:"usage"`
+	ID                 string             `json:"id"`
+	Object             string             `json:"object"`
+	CreatedAt          int64              `json:"created_at"`
+	CompletedAt        *int64             `json:"completed_at"`
+	Status             string             `json:"status"`
+	IncompleteDetails  *IncompleteDetails `json:"incomplete_details"`
+	Model              string             `json:"model"`
+	PreviousResponseID *string            `json:"previous_response_id"`
+	Instructions       *string            `json:"instructions"`
+	Output             []OutputItem       `json:"output"`
+	Error              *ResponseError     `json:"error"`
+	Tools              json.RawMessage    `json:"tools"`
+	ToolChoice         json.RawMessage    `json:"tool_choice"`
+	Truncation         string             `json:"truncation"`
+	ParallelToolCalls  bool               `json:"parallel_tool_calls"`
+	Text               json.RawMessage    `json:"text"`
+	TopP               float64            `json:"top_p"`
+	PresencePenalty    float64            `json:"presence_penalty"`
+	FrequencyPenalty   float64            `json:"frequency_penalty"`
+	TopLogprobs        int                `json:"top_logprobs"`
+	Temperature        float64            `json:"temperature"`
+	Reasoning          json.RawMessage    `json:"reasoning"`
+	Usage              *Usage             `json:"usage"`
+	MaxOutputTokens    *int               `json:"max_output_tokens"`
+	MaxToolCalls       *int               `json:"max_tool_calls"`
+	Store              bool               `json:"store"`
+	Background         bool               `json:"background"`
+	ServiceTier        string             `json:"service_tier"`
+	Metadata           map[string]string  `json:"metadata"`
+	SafetyIdentifier   *string            `json:"safety_identifier"`
+	PromptCacheKey     *string            `json:"prompt_cache_key"`
+
+	// Non-spec extension
+	Provider string `json:"provider,omitempty"`
 }
 
-// Usage captures token accounting.
+// OutputItem represents a typed item in the response output.
+type OutputItem struct {
+	ID      string        `json:"id"`
+	Type    string        `json:"type"`
+	Status  string        `json:"status"`
+	Role    string        `json:"role,omitempty"`
+	Content []ContentPart `json:"content,omitempty"`
+}
+
+// ContentPart is a content block within an output item.
+type ContentPart struct {
+	Type        string       `json:"type"`
+	Text        string       `json:"text"`
+	Annotations []Annotation `json:"annotations"`
+}
+
+// Annotation on output text content.
+type Annotation struct {
+	Type string `json:"type"`
+}
+
+// IncompleteDetails explains why a response is incomplete.
+type IncompleteDetails struct {
+	Reason string `json:"reason"`
+}
+
+// ResponseError describes an error in the response.
+type ResponseError struct {
+	Type    string  `json:"type"`
+	Message string  `json:"message"`
+	Code    *string `json:"code"`
+}
+
+// ============================================================
+// Usage Types
+// ============================================================
+
+// Usage captures token accounting with sub-details.
 type Usage struct {
-	InputTokens  int `json:"input_tokens"`
-	OutputTokens int `json:"output_tokens"`
-	TotalTokens  int `json:"total_tokens"`
+	InputTokens         int                 `json:"input_tokens"`
+	OutputTokens        int                 `json:"output_tokens"`
+	TotalTokens         int                 `json:"total_tokens"`
+	InputTokensDetails  InputTokensDetails  `json:"input_tokens_details"`
+	OutputTokensDetails OutputTokensDetails `json:"output_tokens_details"`
 }
 
-// StreamChunk represents a single Server-Sent Event in a streaming response.
-type StreamChunk struct {
-	ID       string         `json:"id,omitempty"`
-	Object   string         `json:"object"`
-	Created  int64          `json:"created,omitempty"`
-	Model    string         `json:"model,omitempty"`
-	Provider string         `json:"provider,omitempty"`
-	Delta    *StreamDelta   `json:"delta,omitempty"`
-	Usage    *Usage         `json:"usage,omitempty"`
-	Done     bool           `json:"done,omitempty"`
+// InputTokensDetails breaks down input token usage.
+type InputTokensDetails struct {
+	CachedTokens int `json:"cached_tokens"`
 }
 
-// StreamDelta represents incremental content in a stream chunk.
-type StreamDelta struct {
-	Role    string         `json:"role,omitempty"`
-	Content []ContentBlock `json:"content,omitempty"`
+// OutputTokensDetails breaks down output token usage.
+type OutputTokensDetails struct {
+	ReasoningTokens int `json:"reasoning_tokens"`
 }
 
+// ============================================================
+// Streaming Types
+// ============================================================
+
+// StreamEvent represents a single SSE event in the streaming response.
+// Fields are selectively populated based on the event Type.
+type StreamEvent struct {
+	Type           string       `json:"type"`
+	SequenceNumber int          `json:"sequence_number"`
+	Response       *Response    `json:"response,omitempty"`
+	OutputIndex    *int         `json:"output_index,omitempty"`
+	Item           *OutputItem  `json:"item,omitempty"`
+	ItemID         string       `json:"item_id,omitempty"`
+	ContentIndex   *int         `json:"content_index,omitempty"`
+	Part           *ContentPart `json:"part,omitempty"`
+	Delta          string       `json:"delta,omitempty"`
+	Text           string       `json:"text,omitempty"`
+}
+
+// ============================================================
+// Provider Result Types (internal, not exposed via HTTP)
+// ============================================================
+
+// ProviderResult is returned by Provider.Generate.
+type ProviderResult struct {
+	ID    string
+	Model string
+	Text  string
+	Usage Usage
+}
+
+// ProviderStreamDelta is sent through the stream channel.
+type ProviderStreamDelta struct {
+	ID    string
+	Model string
+	Text  string
+	Done  bool
+	Usage *Usage
+}
+
+// ============================================================
+// Models Endpoint Types
+// ============================================================
+
 // ModelInfo describes a single model available through the gateway.
 type ModelInfo struct {
 	ID       string `json:"id"`
 	Provider string `json:"provider"`
 }
 
-// ModelsResponse is returned by the GET /v1/models endpoint.
+// ModelsResponse is returned by GET /v1/models.
 type ModelsResponse struct {
 	Object string      `json:"object"`
 	Data   []ModelInfo `json:"data"`
 }
 
+// ============================================================
+// Validation
+// ============================================================
+
 // Validate performs basic structural validation.
 func (r *ResponseRequest) Validate() error {
 	if r == nil {
@@ -84,16 +310,8 @@ func (r *ResponseRequest) Validate() error {
 	if r.Model == "" {
 		return errors.New("model is required")
 	}
-	if len(r.Input) == 0 {
-		return errors.New("input messages are required")
-	}
-	for i, msg := range r.Input {
-		if msg.Role == "" {
-			return fmt.Errorf("input[%d] role is required", i)
-		}
-		if len(msg.Content) == 0 {
-			return fmt.Errorf("input[%d] content is required", i)
-		}
+	if r.Input.String == nil && len(r.Input.Items) == 0 {
+		return errors.New("input is required")
 	}
 	return nil
 }

internal/providers/anthropic/anthropic.go

@@ -3,7 +3,6 @@ package anthropic
 import (
 	"context"
 	"fmt"
-	"time"
 
 	"github.com/anthropics/anthropic-sdk-go"
 	"github.com/anthropics/anthropic-sdk-go/option"
@@ -60,8 +59,8 @@ func NewAzure(azureCfg config.AzureAnthropicConfig) *Provider {
 
 func (p *Provider) Name() string { return Name }
 
-// Generate routes the Open Responses request to Anthropic's API.
-func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api.Response, error) {
+// Generate routes the request to Anthropic's API.
+func (p *Provider) Generate(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (*api.ProviderResult, error) {
 	if p.cfg.APIKey == "" {
 		return nil, fmt.Errorf("anthropic api key missing")
 	}
@@ -69,13 +68,11 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 		return nil, fmt.Errorf("anthropic client not initialized")
 	}
 
-	model := chooseModel(req.Model, p.cfg.Model)
-
-	// Convert Open Responses messages to Anthropic format
-	messages := make([]anthropic.MessageParam, 0, len(req.Input))
+	// Convert messages to Anthropic format
+	anthropicMsgs := make([]anthropic.MessageParam, 0, len(messages))
 	var system string
 
-	for _, msg := range req.Input {
+	for _, msg := range messages {
 		var content string
 		for _, block := range msg.Content {
 			if block.Type == "input_text" || block.Type == "output_text" {
@@ -85,19 +82,24 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 
 		switch msg.Role {
 		case "user":
-			messages = append(messages, anthropic.NewUserMessage(anthropic.NewTextBlock(content)))
+			anthropicMsgs = append(anthropicMsgs, anthropic.NewUserMessage(anthropic.NewTextBlock(content)))
 		case "assistant":
-			messages = append(messages, anthropic.NewAssistantMessage(anthropic.NewTextBlock(content)))
-		case "system":
+			anthropicMsgs = append(anthropicMsgs, anthropic.NewAssistantMessage(anthropic.NewTextBlock(content)))
+		case "system", "developer":
 			system = content
 		}
 	}
 
 	// Build request params
+	maxTokens := int64(4096)
+	if req.MaxOutputTokens != nil {
+		maxTokens = int64(*req.MaxOutputTokens)
+	}
+
 	params := anthropic.MessageNewParams{
-		Model:     anthropic.Model(model),
-		Messages:  messages,
-		MaxTokens: int64(4096),
+		Model:     anthropic.Model(req.Model),
+		Messages:  anthropicMsgs,
+		MaxTokens: maxTokens,
 	}
 
 	if system != "" {
@@ -107,36 +109,31 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 		params.System = systemBlocks
 	}
 
+	if req.Temperature != nil {
+		params.Temperature = anthropic.Float(*req.Temperature)
+	}
+	if req.TopP != nil {
+		params.TopP = anthropic.Float(*req.TopP)
+	}
+
 	// Call Anthropic API
 	resp, err := p.client.Messages.New(ctx, params)
 	if err != nil {
 		return nil, fmt.Errorf("anthropic api error: %w", err)
 	}
 
-	// Convert Anthropic response to Open Responses format
-	output := make([]api.Message, 0, 1)
+	// Extract text from response
 	var text string
-	
 	for _, block := range resp.Content {
 		if block.Type == "text" {
 			text += block.Text
 		}
 	}
 
-	output = append(output, api.Message{
-		Role: "assistant",
-		Content: []api.ContentBlock{
-			{Type: "output_text", Text: text},
-		},
-	})
-
-	return &api.Response{
-		ID:       resp.ID,
-		Object:   "response",
-		Created:  time.Now().Unix(),
-		Model:    string(resp.Model),
-		Provider: Name,
-		Output:   output,
+	return &api.ProviderResult{
+		ID:    resp.ID,
+		Model: string(resp.Model),
+		Text:  text,
 		Usage: api.Usage{
 			InputTokens:  int(resp.Usage.InputTokens),
 			OutputTokens: int(resp.Usage.OutputTokens),
@@ -146,12 +143,12 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 }
 
 // GenerateStream handles streaming requests to Anthropic.
-func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest) (<-chan *api.StreamChunk, <-chan error) {
-	chunkChan := make(chan *api.StreamChunk)
+func (p *Provider) GenerateStream(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (<-chan *api.ProviderStreamDelta, <-chan error) {
+	deltaChan := make(chan *api.ProviderStreamDelta)
 	errChan := make(chan error, 1)
 
 	go func() {
-		defer close(chunkChan)
+		defer close(deltaChan)
 		defer close(errChan)
 
 		if p.cfg.APIKey == "" {
@@ -163,35 +160,38 @@ func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest)
 			return
 		}
 
-		model := chooseModel(req.Model, p.cfg.Model)
-
-		// Convert messages
-		messages := make([]anthropic.MessageParam, 0, len(req.Input))
+		// Convert messages to Anthropic format
+		anthropicMsgs := make([]anthropic.MessageParam, 0, len(messages))
 		var system string
 
-		for _, msg := range req.Input {
+		for _, msg := range messages {
 			var content string
 			for _, block := range msg.Content {
 				if block.Type == "input_text" || block.Type == "output_text" {
 					content += block.Text
 				}
-				}
+			}
 
-				switch msg.Role {
-				case "user":
-				messages = append(messages, anthropic.NewUserMessage(anthropic.NewTextBlock(content)))
-				case "assistant":
-				messages = append(messages, anthropic.NewAssistantMessage(anthropic.NewTextBlock(content)))
-				case "system":
+			switch msg.Role {
+			case "user":
+				anthropicMsgs = append(anthropicMsgs, anthropic.NewUserMessage(anthropic.NewTextBlock(content)))
+			case "assistant":
+				anthropicMsgs = append(anthropicMsgs, anthropic.NewAssistantMessage(anthropic.NewTextBlock(content)))
+			case "system", "developer":
 				system = content
-				}
-				}
+			}
+		}
+
+		// Build params
+		maxTokens := int64(4096)
+		if req.MaxOutputTokens != nil {
+			maxTokens = int64(*req.MaxOutputTokens)
+		}
 
-				// Build params
 		params := anthropic.MessageNewParams{
-			Model:     anthropic.Model(model),
-			Messages:  messages,
-			MaxTokens: int64(4096),
+			Model:     anthropic.Model(req.Model),
+			Messages:  anthropicMsgs,
+			MaxTokens: maxTokens,
 		}
 
 		if system != "" {
@@ -201,6 +201,13 @@ func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest)
 			params.System = systemBlocks
 		}
 
+		if req.Temperature != nil {
+			params.Temperature = anthropic.Float(*req.Temperature)
+		}
+		if req.TopP != nil {
+			params.TopP = anthropic.Float(*req.TopP)
+		}
+
 		// Create stream
 		stream := p.client.Messages.NewStreaming(ctx, params)
 
@@ -208,35 +215,14 @@ func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest)
 		for stream.Next() {
 			event := stream.Current()
 
-			delta := &api.StreamDelta{}
-			var text string
-			
-			// Handle different event types
 			if event.Type == "content_block_delta" && event.Delta.Type == "text_delta" {
-				text = event.Delta.Text
-				delta.Content = []api.ContentBlock{
-					{Type: "output_text", Text: text},
+				select {
+				case deltaChan <- &api.ProviderStreamDelta{Text: event.Delta.Text}:
+				case <-ctx.Done():
+					errChan <- ctx.Err()
+					return
 				}
 			}
-			
-			if event.Type == "message_start" {
-				delta.Role = "assistant"
-			}
-
-			streamChunk := &api.StreamChunk{
-				Object:   "response.chunk",
-				Created:  time.Now().Unix(),
-				Model:    string(model),
-				Provider: Name,
-				Delta:    delta,
-			}
-
-			select {
-			case chunkChan <- streamChunk:
-			case <-ctx.Done():
-				errChan <- ctx.Err()
-				return
-			}
 		}
 
 		if err := stream.Err(); err != nil {
@@ -244,15 +230,15 @@ func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest)
 			return
 		}
 
-		// Send final chunk
+		// Send final delta
 		select {
-		case chunkChan <- &api.StreamChunk{Object: "response.chunk", Done: true}:
+		case deltaChan <- &api.ProviderStreamDelta{Done: true}:
 		case <-ctx.Done():
 			errChan <- ctx.Err()
 		}
 	}()
 
-	return chunkChan, errChan
+	return deltaChan, errChan
 }
 
 func chooseModel(requested, defaultModel string) string {

internal/providers/google/google.go

@@ -3,7 +3,6 @@ package google
 import (
 	"context"
 	"fmt"
-	"time"
 
 	"github.com/google/uuid"
 	"google.golang.org/genai"
@@ -41,8 +40,8 @@ func New(cfg config.ProviderConfig) *Provider {
 
 func (p *Provider) Name() string { return Name }
 
-// Generate routes the Open Responses request to Gemini.
-func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api.Response, error) {
+// Generate routes the request to Gemini and returns a ProviderResult.
+func (p *Provider) Generate(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (*api.ProviderResult, error) {
 	if p.cfg.APIKey == "" {
 		return nil, fmt.Errorf("google api key missing")
 	}
@@ -50,14 +49,112 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 		return nil, fmt.Errorf("google client not initialized")
 	}
 
-	model := chooseModel(req.Model, p.cfg.Model)
+	model := req.Model
 
-	// Convert Open Responses messages to Gemini format
+	contents, systemText := convertMessages(messages)
+
+	config := buildConfig(systemText, req)
+
+	resp, err := p.client.Models.GenerateContent(ctx, model, contents, config)
+	if err != nil {
+		return nil, fmt.Errorf("google api error: %w", err)
+	}
+
+	var text string
+	if len(resp.Candidates) > 0 && resp.Candidates[0].Content != nil {
+		for _, part := range resp.Candidates[0].Content.Parts {
+			if part != nil {
+				text += part.Text
+			}
+		}
+	}
+
+	var inputTokens, outputTokens int
+	if resp.UsageMetadata != nil {
+		inputTokens = int(resp.UsageMetadata.PromptTokenCount)
+		outputTokens = int(resp.UsageMetadata.CandidatesTokenCount)
+	}
+
+	return &api.ProviderResult{
+		ID:    uuid.NewString(),
+		Model: model,
+		Text:  text,
+		Usage: api.Usage{
+			InputTokens:  inputTokens,
+			OutputTokens: outputTokens,
+			TotalTokens:  inputTokens + outputTokens,
+		},
+	}, nil
+}
+
+// GenerateStream handles streaming requests to Google.
+func (p *Provider) GenerateStream(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (<-chan *api.ProviderStreamDelta, <-chan error) {
+	deltaChan := make(chan *api.ProviderStreamDelta)
+	errChan := make(chan error, 1)
+
+	go func() {
+		defer close(deltaChan)
+		defer close(errChan)
+
+		if p.cfg.APIKey == "" {
+			errChan <- fmt.Errorf("google api key missing")
+			return
+		}
+		if p.client == nil {
+			errChan <- fmt.Errorf("google client not initialized")
+			return
+		}
+
+		model := req.Model
+
+		contents, systemText := convertMessages(messages)
+
+		config := buildConfig(systemText, req)
+
+		stream := p.client.Models.GenerateContentStream(ctx, model, contents, config)
+
+		for resp, err := range stream {
+			if err != nil {
+				errChan <- fmt.Errorf("google stream error: %w", err)
+				return
+			}
+
+			var text string
+			if len(resp.Candidates) > 0 && resp.Candidates[0].Content != nil {
+				for _, part := range resp.Candidates[0].Content.Parts {
+					if part != nil {
+						text += part.Text
+					}
+				}
+			}
+
+			if text != "" {
+				select {
+				case deltaChan <- &api.ProviderStreamDelta{Text: text}:
+				case <-ctx.Done():
+					errChan <- ctx.Err()
+					return
+				}
+			}
+		}
+
+		select {
+		case deltaChan <- &api.ProviderStreamDelta{Done: true}:
+		case <-ctx.Done():
+			errChan <- ctx.Err()
+		}
+	}()
+
+	return deltaChan, errChan
+}
+
+// convertMessages splits messages into Gemini contents and system text.
+func convertMessages(messages []api.Message) ([]*genai.Content, string) {
 	var contents []*genai.Content
 	var systemText string
 
-	for _, msg := range req.Input {
-		if msg.Role == "system" {
+	for _, msg := range messages {
+		if msg.Role == "system" || msg.Role == "developer" {
 			for _, block := range msg.Content {
 				if block.Type == "input_text" || block.Type == "output_text" {
 					systemText += block.Text
@@ -84,176 +181,41 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 		})
 	}
 
-	// Build config with system instruction if present
-	var config *genai.GenerateContentConfig
-	if systemText != "" {
-		config = &genai.GenerateContentConfig{
-			SystemInstruction: &genai.Content{
-				Parts: []*genai.Part{genai.NewPartFromText(systemText)},
-			},
-		}
-	}
-
-	// Generate content
-	resp, err := p.client.Models.GenerateContent(ctx, model, contents, config)
-	if err != nil {
-		return nil, fmt.Errorf("google api error: %w", err)
-	}
-
-	// Convert Gemini response to Open Responses format
-	output := make([]api.Message, 0, 1)
-	var text string
-	
-	if len(resp.Candidates) > 0 && resp.Candidates[0].Content != nil {
-		for _, part := range resp.Candidates[0].Content.Parts {
-			if part != nil {
-				text += part.Text
-			}
-		}
-	}
-	
-	output = append(output, api.Message{
-		Role: "assistant",
-		Content: []api.ContentBlock{
-			{Type: "output_text", Text: text},
-		},
-	})
-
-	// Extract usage info if available
-	var inputTokens, outputTokens int
-	if resp.UsageMetadata != nil {
-		inputTokens = int(resp.UsageMetadata.PromptTokenCount)
-		outputTokens = int(resp.UsageMetadata.CandidatesTokenCount)
-	}
-
-	return &api.Response{
-		ID:       uuid.NewString(),
-		Object:   "response",
-		Created:  time.Now().Unix(),
-		Model:    model,
-		Provider: Name,
-		Output:   output,
-		Usage: api.Usage{
-			InputTokens:  inputTokens,
-			OutputTokens: outputTokens,
-			TotalTokens:  inputTokens + outputTokens,
-		},
-	}, nil
+	return contents, systemText
 }
 
-// GenerateStream handles streaming requests to Google.
-func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest) (<-chan *api.StreamChunk, <-chan error) {
-	chunkChan := make(chan *api.StreamChunk)
-	errChan := make(chan error, 1)
+// buildConfig constructs a GenerateContentConfig from system text and request params.
+func buildConfig(systemText string, req *api.ResponseRequest) *genai.GenerateContentConfig {
+	var cfg *genai.GenerateContentConfig
 
-	go func() {
-		defer close(chunkChan)
-		defer close(errChan)
+	needsCfg := systemText != "" || req.MaxOutputTokens != nil || req.Temperature != nil || req.TopP != nil
+	if !needsCfg {
+		return nil
+	}
 
-		if p.cfg.APIKey == "" {
-			errChan <- fmt.Errorf("google api key missing")
-			return
-		}
-		if p.client == nil {
-			errChan <- fmt.Errorf("google client not initialized")
-			return
+	cfg = &genai.GenerateContentConfig{}
+
+	if systemText != "" {
+		cfg.SystemInstruction = &genai.Content{
+			Parts: []*genai.Part{genai.NewPartFromText(systemText)},
 		}
+	}
 
-		model := chooseModel(req.Model, p.cfg.Model)
+	if req.MaxOutputTokens != nil {
+		cfg.MaxOutputTokens = int32(*req.MaxOutputTokens)
+	}
 
-		// Convert messages
-		var contents []*genai.Content
-		var systemText string
+	if req.Temperature != nil {
+		t := float32(*req.Temperature)
+		cfg.Temperature = &t
+	}
 
-		for _, msg := range req.Input {
-			if msg.Role == "system" {
-				for _, block := range msg.Content {
-					if block.Type == "input_text" || block.Type == "output_text" {
-						systemText += block.Text
-					}
-				}
-				continue
-			}
+	if req.TopP != nil {
+		tp := float32(*req.TopP)
+		cfg.TopP = &tp
+	}
 
-			var parts []*genai.Part
-			for _, block := range msg.Content {
-				if block.Type == "input_text" || block.Type == "output_text" {
-					parts = append(parts, genai.NewPartFromText(block.Text))
-				}
-			}
-			
-			role := "user"
-			if msg.Role == "assistant" || msg.Role == "model" {
-				role = "model"
-			}
-			
-			contents = append(contents, &genai.Content{
-				Role:  role,
-				Parts: parts,
-			})
-		}
-
-		// Build config with system instruction if present
-		var config *genai.GenerateContentConfig
-		if systemText != "" {
-			config = &genai.GenerateContentConfig{
-				SystemInstruction: &genai.Content{
-					Parts: []*genai.Part{genai.NewPartFromText(systemText)},
-				},
-			}
-		}
-
-		// Create stream
-		stream := p.client.Models.GenerateContentStream(ctx, model, contents, config)
-
-		// Process stream
-		for resp, err := range stream {
-			if err != nil {
-				errChan <- fmt.Errorf("google stream error: %w", err)
-				return
-			}
-
-			var text string
-			if len(resp.Candidates) > 0 && resp.Candidates[0].Content != nil {
-				for _, part := range resp.Candidates[0].Content.Parts {
-					if part != nil {
-						text += part.Text
-					}
-				}
-			}
-
-			delta := &api.StreamDelta{}
-			if text != "" {
-				delta.Content = []api.ContentBlock{
-					{Type: "output_text", Text: text},
-				}
-			}
-
-			streamChunk := &api.StreamChunk{
-				Object:   "response.chunk",
-				Created:  time.Now().Unix(),
-				Model:    model,
-				Provider: Name,
-				Delta:    delta,
-			}
-
-			select {
-			case chunkChan <- streamChunk:
-			case <-ctx.Done():
-				errChan <- ctx.Err()
-				return
-			}
-		}
-
-		// Send final chunk
-		select {
-		case chunkChan <- &api.StreamChunk{Object: "response.chunk", Done: true}:
-		case <-ctx.Done():
-			errChan <- ctx.Err()
-		}
-	}()
-
-	return chunkChan, errChan
+	return cfg
 }
 
 func chooseModel(requested, defaultModel string) string {

internal/providers/openai/openai.go

@@ -3,7 +3,6 @@ package openai
 import (
 	"context"
 	"fmt"
-	"time"
 
 	"github.com/openai/openai-go"
 	"github.com/openai/openai-go/azure"
@@ -64,8 +63,8 @@ func NewAzure(azureCfg config.AzureOpenAIConfig) *Provider {
 // Name returns the provider identifier.
 func (p *Provider) Name() string { return Name }
 
-// Generate routes the Open Responses request to OpenAI.
-func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api.Response, error) {
+// Generate routes the request to OpenAI and returns a ProviderResult.
+func (p *Provider) Generate(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (*api.ProviderResult, error) {
 	if p.cfg.APIKey == "" {
 		return nil, fmt.Errorf("openai api key missing")
 	}
@@ -73,11 +72,9 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 		return nil, fmt.Errorf("openai client not initialized")
 	}
 
-	model := chooseModel(req.Model, p.cfg.Model)
-
-	// Convert Open Responses messages to OpenAI format
-	messages := make([]openai.ChatCompletionMessageParamUnion, 0, len(req.Input))
-	for _, msg := range req.Input {
+	// Convert messages to OpenAI format
+	oaiMessages := make([]openai.ChatCompletionMessageParamUnion, 0, len(messages))
+	for _, msg := range messages {
 		var content string
 		for _, block := range msg.Content {
 			if block.Type == "input_text" || block.Type == "output_text" {
@@ -87,41 +84,45 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 
 		switch msg.Role {
 		case "user":
-			messages = append(messages, openai.UserMessage(content))
+			oaiMessages = append(oaiMessages, openai.UserMessage(content))
 		case "assistant":
-			messages = append(messages, openai.AssistantMessage(content))
+			oaiMessages = append(oaiMessages, openai.AssistantMessage(content))
 		case "system":
-			messages = append(messages, openai.SystemMessage(content))
+			oaiMessages = append(oaiMessages, openai.SystemMessage(content))
+		case "developer":
+			oaiMessages = append(oaiMessages, openai.SystemMessage(content))
 		}
 	}
 
+	params := openai.ChatCompletionNewParams{
+		Model:    openai.ChatModel(req.Model),
+		Messages: oaiMessages,
+	}
+	if req.MaxOutputTokens != nil {
+		params.MaxTokens = openai.Int(int64(*req.MaxOutputTokens))
+	}
+	if req.Temperature != nil {
+		params.Temperature = openai.Float(*req.Temperature)
+	}
+	if req.TopP != nil {
+		params.TopP = openai.Float(*req.TopP)
+	}
+
 	// Call OpenAI API
-	resp, err := p.client.Chat.Completions.New(ctx, openai.ChatCompletionNewParams{
-		Model:    openai.ChatModel(model),
-		Messages: messages,
-	})
+	resp, err := p.client.Chat.Completions.New(ctx, params)
 	if err != nil {
 		return nil, fmt.Errorf("openai api error: %w", err)
 	}
 
-	// Convert OpenAI response to Open Responses format
-	output := make([]api.Message, 0, len(resp.Choices))
+	var combinedText string
 	for _, choice := range resp.Choices {
-		output = append(output, api.Message{
-			Role: "assistant",
-			Content: []api.ContentBlock{
-				{Type: "output_text", Text: choice.Message.Content},
-			},
-		})
+		combinedText += choice.Message.Content
 	}
 
-	return &api.Response{
-		ID:       resp.ID,
-		Object:   "response",
-		Created:  time.Now().Unix(),
-		Model:    resp.Model,
-		Provider: Name,
-		Output:   output,
+	return &api.ProviderResult{
+		ID:    resp.ID,
+		Model: resp.Model,
+		Text:  combinedText,
 		Usage: api.Usage{
 			InputTokens:  int(resp.Usage.PromptTokens),
 			OutputTokens: int(resp.Usage.CompletionTokens),
@@ -131,12 +132,12 @@ func (p *Provider) Generate(ctx context.Context, req *api.ResponseRequest) (*api
 }
 
 // GenerateStream handles streaming requests to OpenAI.
-func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest) (<-chan *api.StreamChunk, <-chan error) {
-	chunkChan := make(chan *api.StreamChunk)
+func (p *Provider) GenerateStream(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (<-chan *api.ProviderStreamDelta, <-chan error) {
+	deltaChan := make(chan *api.ProviderStreamDelta)
 	errChan := make(chan error, 1)
 
 	go func() {
-		defer close(chunkChan)
+		defer close(deltaChan)
 		defer close(errChan)
 
 		if p.cfg.APIKey == "" {
@@ -148,62 +149,60 @@ func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest)
 			return
 		}
 
-		model := chooseModel(req.Model, p.cfg.Model)
-
-		// Convert messages
-		messages := make([]openai.ChatCompletionMessageParamUnion, 0, len(req.Input))
-		for _, msg := range req.Input {
+		// Convert messages to OpenAI format
+		oaiMessages := make([]openai.ChatCompletionMessageParamUnion, 0, len(messages))
+		for _, msg := range messages {
 			var content string
 			for _, block := range msg.Content {
 				if block.Type == "input_text" || block.Type == "output_text" {
 					content += block.Text
 				}
-				}
+			}
 
-				switch msg.Role {
-				case "user":
-				messages = append(messages, openai.UserMessage(content))
-				case "assistant":
-				messages = append(messages, openai.AssistantMessage(content))
-				case "system":
-				messages = append(messages, openai.SystemMessage(content))
-				}
-				}
+			switch msg.Role {
+			case "user":
+				oaiMessages = append(oaiMessages, openai.UserMessage(content))
+			case "assistant":
+				oaiMessages = append(oaiMessages, openai.AssistantMessage(content))
+			case "system":
+				oaiMessages = append(oaiMessages, openai.SystemMessage(content))
+			case "developer":
+				oaiMessages = append(oaiMessages, openai.SystemMessage(content))
+			}
+		}
 
-				// Create streaming request
-		stream := p.client.Chat.Completions.NewStreaming(ctx, openai.ChatCompletionNewParams{
-			Model:    openai.ChatModel(model),
-			Messages: messages,
-		})
+		params := openai.ChatCompletionNewParams{
+			Model:    openai.ChatModel(req.Model),
+			Messages: oaiMessages,
+		}
+		if req.MaxOutputTokens != nil {
+			params.MaxTokens = openai.Int(int64(*req.MaxOutputTokens))
+		}
+		if req.Temperature != nil {
+			params.Temperature = openai.Float(*req.Temperature)
+		}
+		if req.TopP != nil {
+			params.TopP = openai.Float(*req.TopP)
+		}
+
+		// Create streaming request
+		stream := p.client.Chat.Completions.NewStreaming(ctx, params)
 
 		// Process stream
 		for stream.Next() {
 			chunk := stream.Current()
 
 			for _, choice := range chunk.Choices {
-				delta := &api.StreamDelta{}
-				
-				if choice.Delta.Role != "" {
-					delta.Role = string(choice.Delta.Role)
-				}
-				
-				if choice.Delta.Content != "" {
-					delta.Content = []api.ContentBlock{
-						{Type: "output_text", Text: choice.Delta.Content},
-					}
-				}
-
-				streamChunk := &api.StreamChunk{
-					ID:       chunk.ID,
-					Object:   "response.chunk",
-					Created:  time.Now().Unix(),
-					Model:    chunk.Model,
-					Provider: Name,
-					Delta:    delta,
+				if choice.Delta.Content == "" {
+					continue
 				}
 
 				select {
-				case chunkChan <- streamChunk:
+				case deltaChan <- &api.ProviderStreamDelta{
+					ID:    chunk.ID,
+					Model: chunk.Model,
+					Text:  choice.Delta.Content,
+				}:
 				case <-ctx.Done():
 					errChan <- ctx.Err()
 					return
@@ -216,15 +215,15 @@ func (p *Provider) GenerateStream(ctx context.Context, req *api.ResponseRequest)
 			return
 		}
 
-		// Send final chunk
+		// Send final delta
 		select {
-		case chunkChan <- &api.StreamChunk{Object: "response.chunk", Done: true}:
+		case deltaChan <- &api.ProviderStreamDelta{Done: true}:
 		case <-ctx.Done():
 			errChan <- ctx.Err()
 		}
 	}()
 
-	return chunkChan, errChan
+	return deltaChan, errChan
 }
 
 func chooseModel(requested, defaultModel string) string {

internal/providers/providers.go

@@ -14,8 +14,8 @@ import (
 // Provider represents a unified interface that each LLM provider must implement.
 type Provider interface {
 	Name() string
-	Generate(ctx context.Context, req *api.ResponseRequest) (*api.Response, error)
-	GenerateStream(ctx context.Context, req *api.ResponseRequest) (<-chan *api.StreamChunk, <-chan error)
+	Generate(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (*api.ProviderResult, error)
+	GenerateStream(ctx context.Context, messages []api.Message, req *api.ResponseRequest) (<-chan *api.ProviderStreamDelta, <-chan error)
 }
 
 // Registry keeps track of registered providers and model-to-provider mappings.

internal/server/server.go

@@ -5,6 +5,10 @@ import (
 	"fmt"
 	"log"
 	"net/http"
+	"strings"
+	"time"
+
+	"github.com/google/uuid"
 
 	"github.com/yourusername/go-llm-gateway/internal/api"
 	"github.com/yourusername/go-llm-gateway/internal/conversation"
@@ -74,16 +78,34 @@ func (s *GatewayServer) handleResponses(w http.ResponseWriter, r *http.Request)
 		return
 	}
 
-	// Build full message history
-	messages := s.buildMessageHistory(&req)
-	if messages == nil {
-		http.Error(w, "conversation not found", http.StatusNotFound)
-		return
+	// Normalize input to internal messages
+	inputMsgs := req.NormalizeInput()
+
+	// Build full message history from previous conversation
+	var historyMsgs []api.Message
+	if req.PreviousResponseID != nil && *req.PreviousResponseID != "" {
+		conv, ok := s.convs.Get(*req.PreviousResponseID)
+		if !ok {
+			http.Error(w, "conversation not found", http.StatusNotFound)
+			return
+		}
+		historyMsgs = conv.Messages
 	}
 
-	// Update request with full history for provider
-	fullReq := req
-	fullReq.Input = messages
+	// Combined messages for conversation storage (history + new input, no instructions)
+	storeMsgs := make([]api.Message, 0, len(historyMsgs)+len(inputMsgs))
+	storeMsgs = append(storeMsgs, historyMsgs...)
+	storeMsgs = append(storeMsgs, inputMsgs...)
+
+	// Build provider messages: instructions + history + input
+	var providerMsgs []api.Message
+	if req.Instructions != nil && *req.Instructions != "" {
+		providerMsgs = append(providerMsgs, api.Message{
+			Role:    "developer",
+			Content: []api.ContentBlock{{Type: "input_text", Text: *req.Instructions}},
+		})
+	}
+	providerMsgs = append(providerMsgs, storeMsgs...)
 
 	provider, err := s.resolveProvider(&req)
 	if err != nil {
@@ -91,64 +113,44 @@ func (s *GatewayServer) handleResponses(w http.ResponseWriter, r *http.Request)
 		return
 	}
 
-	// Resolve provider_model_id (e.g., Azure deployment name) before sending to provider
-	fullReq.Model = s.registry.ResolveModelID(req.Model)
+	// Resolve provider_model_id (e.g., Azure deployment name)
+	resolvedReq := req
+	resolvedReq.Model = s.registry.ResolveModelID(req.Model)
 
-	// Handle streaming vs non-streaming
 	if req.Stream {
-		s.handleStreamingResponse(w, r, provider, &fullReq, &req)
+		s.handleStreamingResponse(w, r, provider, providerMsgs, &resolvedReq, &req, storeMsgs)
 	} else {
-		s.handleSyncResponse(w, r, provider, &fullReq, &req)
+		s.handleSyncResponse(w, r, provider, providerMsgs, &resolvedReq, &req, storeMsgs)
 	}
 }
 
-func (s *GatewayServer) buildMessageHistory(req *api.ResponseRequest) []api.Message {
-	// If no previous_response_id, use input as-is
-	if req.PreviousResponseID == "" {
-		return req.Input
-	}
-
-	// Load previous conversation
-	conv, ok := s.convs.Get(req.PreviousResponseID)
-	if !ok {
-		return nil
-	}
-
-	// Append new input to conversation history
-	messages := make([]api.Message, len(conv.Messages))
-	copy(messages, conv.Messages)
-	messages = append(messages, req.Input...)
-
-	return messages
-}
-
-func (s *GatewayServer) handleSyncResponse(w http.ResponseWriter, r *http.Request, provider providers.Provider, fullReq *api.ResponseRequest, origReq *api.ResponseRequest) {
-	resp, err := provider.Generate(r.Context(), fullReq)
+func (s *GatewayServer) handleSyncResponse(w http.ResponseWriter, r *http.Request, provider providers.Provider, providerMsgs []api.Message, resolvedReq *api.ResponseRequest, origReq *api.ResponseRequest, storeMsgs []api.Message) {
+	result, err := provider.Generate(r.Context(), providerMsgs, resolvedReq)
 	if err != nil {
 		s.logger.Printf("provider %s error: %v", provider.Name(), err)
 		http.Error(w, "provider error", http.StatusBadGateway)
 		return
 	}
 
-	// Store conversation - use previous_response_id if continuing, otherwise use new ID
-	conversationID := origReq.PreviousResponseID
-	if conversationID == "" {
-		conversationID = resp.ID
+	responseID := generateID("resp_")
+
+	// Build assistant message for conversation store
+	assistantMsg := api.Message{
+		Role:    "assistant",
+		Content: []api.ContentBlock{{Type: "output_text", Text: result.Text}},
 	}
+	allMsgs := append(storeMsgs, assistantMsg)
+	s.convs.Create(responseID, result.Model, allMsgs)
 
-	messages := append(fullReq.Input, resp.Output...)
-	s.convs.Create(conversationID, resp.Model, messages)
-	
-	// Return the conversation ID (not the provider's response ID)
-	resp.ID = conversationID
+	// Build spec-compliant response
+	resp := s.buildResponse(origReq, result, provider.Name(), responseID)
 
 	w.Header().Set("Content-Type", "application/json")
 	w.WriteHeader(http.StatusOK)
 	_ = json.NewEncoder(w).Encode(resp)
 }
 
-func (s *GatewayServer) handleStreamingResponse(w http.ResponseWriter, r *http.Request, provider providers.Provider, fullReq *api.ResponseRequest, origReq *api.ResponseRequest) {
-	// Set headers for SSE
+func (s *GatewayServer) handleStreamingResponse(w http.ResponseWriter, r *http.Request, provider providers.Provider, providerMsgs []api.Message, resolvedReq *api.ResponseRequest, origReq *api.ResponseRequest, storeMsgs []api.Message) {
 	w.Header().Set("Content-Type", "text/event-stream")
 	w.Header().Set("Cache-Control", "no-cache")
 	w.Header().Set("Connection", "keep-alive")
@@ -160,89 +162,322 @@ func (s *GatewayServer) handleStreamingResponse(w http.ResponseWriter, r *http.R
 		return
 	}
 
-	chunkChan, errChan := provider.GenerateStream(r.Context(), fullReq)
+	responseID := generateID("resp_")
+	itemID := generateID("msg_")
+	seq := 0
+	outputIdx := 0
+	contentIdx := 0
+
+	// Build initial response snapshot (in_progress, no output yet)
+	initialResp := s.buildResponse(origReq, &api.ProviderResult{
+		Model: origReq.Model,
+	}, provider.Name(), responseID)
+	initialResp.Status = "in_progress"
+	initialResp.CompletedAt = nil
+	initialResp.Output = []api.OutputItem{}
+	initialResp.Usage = nil
+
+	// response.created
+	s.sendSSE(w, flusher, &seq, "response.created", &api.StreamEvent{
+		Type:     "response.created",
+		Response: initialResp,
+	})
+
+	// response.in_progress
+	s.sendSSE(w, flusher, &seq, "response.in_progress", &api.StreamEvent{
+		Type:     "response.in_progress",
+		Response: initialResp,
+	})
+
+	// response.output_item.added
+	inProgressItem := &api.OutputItem{
+		ID:      itemID,
+		Type:    "message",
+		Status:  "in_progress",
+		Role:    "assistant",
+		Content: []api.ContentPart{},
+	}
+	s.sendSSE(w, flusher, &seq, "response.output_item.added", &api.StreamEvent{
+		Type:        "response.output_item.added",
+		OutputIndex: &outputIdx,
+		Item:        inProgressItem,
+	})
+
+	// response.content_part.added
+	emptyPart := &api.ContentPart{
+		Type:        "output_text",
+		Text:        "",
+		Annotations: []api.Annotation{},
+	}
+	s.sendSSE(w, flusher, &seq, "response.content_part.added", &api.StreamEvent{
+		Type:         "response.content_part.added",
+		ItemID:       itemID,
+		OutputIndex:  &outputIdx,
+		ContentIndex: &contentIdx,
+		Part:         emptyPart,
+	})
+
+	// Start provider stream
+	deltaChan, errChan := provider.GenerateStream(r.Context(), providerMsgs, resolvedReq)
 
-	var responseID string
 	var fullText string
+	var streamErr error
+	var providerModel string
 
+loop:
 	for {
 		select {
-		case chunk, ok := <-chunkChan:
+		case delta, ok := <-deltaChan:
 			if !ok {
-				return
+				break loop
 			}
-			
-			// Capture response ID
-			if chunk.ID != "" && responseID == "" {
-				responseID = chunk.ID
+			if delta.Model != "" && providerModel == "" {
+				providerModel = delta.Model
 			}
-			
-			// Override chunk ID with conversation ID
-			if origReq.PreviousResponseID != "" {
-				chunk.ID = origReq.PreviousResponseID
-			} else if responseID != "" {
-				chunk.ID = responseID
+			if delta.Text != "" {
+				fullText += delta.Text
+				s.sendSSE(w, flusher, &seq, "response.output_text.delta", &api.StreamEvent{
+					Type:         "response.output_text.delta",
+					ItemID:       itemID,
+					OutputIndex:  &outputIdx,
+					ContentIndex: &contentIdx,
+					Delta:        delta.Text,
+				})
 			}
-			
-			// Accumulate text from deltas
-			if chunk.Delta != nil && len(chunk.Delta.Content) > 0 {
-				for _, block := range chunk.Delta.Content {
-					fullText += block.Text
-				}
+			if delta.Done {
+				break loop
 			}
-			
-			data, err := json.Marshal(chunk)
-			if err != nil {
-				s.logger.Printf("failed to marshal chunk: %v", err)
-				continue
-			}
-			
-			fmt.Fprintf(w, "data: %s\n\n", data)
-			flusher.Flush()
-			
-			if chunk.Done {
-				// Store conversation with a single consolidated assistant message
-				s.storeStreamConversation(fullReq, origReq, responseID, fullText)
-				return
-			}
-			
 		case err := <-errChan:
 			if err != nil {
-				s.logger.Printf("stream error: %v", err)
-				errData, _ := json.Marshal(map[string]string{"error": err.Error()})
-				fmt.Fprintf(w, "data: %s\n\n", errData)
-				flusher.Flush()
+				streamErr = err
 			}
-			// Store whatever we accumulated before the error
-			s.storeStreamConversation(fullReq, origReq, responseID, fullText)
-			return
-			
+			break loop
 		case <-r.Context().Done():
 			s.logger.Printf("client disconnected")
 			return
 		}
 	}
-}
 
-func (s *GatewayServer) storeStreamConversation(fullReq *api.ResponseRequest, origReq *api.ResponseRequest, responseID string, fullText string) {
-	if responseID == "" || fullText == "" {
+	if streamErr != nil {
+		s.logger.Printf("stream error: %v", streamErr)
+		failedResp := s.buildResponse(origReq, &api.ProviderResult{
+			Model: origReq.Model,
+		}, provider.Name(), responseID)
+		failedResp.Status = "failed"
+		failedResp.CompletedAt = nil
+		failedResp.Output = []api.OutputItem{}
+		failedResp.Error = &api.ResponseError{
+			Type:    "server_error",
+			Message: streamErr.Error(),
+		}
+		s.sendSSE(w, flusher, &seq, "response.failed", &api.StreamEvent{
+			Type:     "response.failed",
+			Response: failedResp,
+		})
 		return
 	}
 
-	assistantMsg := api.Message{
-		Role: "assistant",
-		Content: []api.ContentBlock{
-			{Type: "output_text", Text: fullText},
-		},
-	}
-	messages := append(fullReq.Input, assistantMsg)
+	// response.output_text.done
+	s.sendSSE(w, flusher, &seq, "response.output_text.done", &api.StreamEvent{
+		Type:         "response.output_text.done",
+		ItemID:       itemID,
+		OutputIndex:  &outputIdx,
+		ContentIndex: &contentIdx,
+		Text:         fullText,
+	})
 
-	conversationID := origReq.PreviousResponseID
-	if conversationID == "" {
-		conversationID = responseID
+	// response.content_part.done
+	completedPart := &api.ContentPart{
+		Type:        "output_text",
+		Text:        fullText,
+		Annotations: []api.Annotation{},
+	}
+	s.sendSSE(w, flusher, &seq, "response.content_part.done", &api.StreamEvent{
+		Type:         "response.content_part.done",
+		ItemID:       itemID,
+		OutputIndex:  &outputIdx,
+		ContentIndex: &contentIdx,
+		Part:         completedPart,
+	})
+
+	// response.output_item.done
+	completedItem := &api.OutputItem{
+		ID:     itemID,
+		Type:   "message",
+		Status: "completed",
+		Role:   "assistant",
+		Content: []api.ContentPart{*completedPart},
+	}
+	s.sendSSE(w, flusher, &seq, "response.output_item.done", &api.StreamEvent{
+		Type:        "response.output_item.done",
+		OutputIndex: &outputIdx,
+		Item:        completedItem,
+	})
+
+	// Build final completed response
+	model := origReq.Model
+	if providerModel != "" {
+		model = providerModel
+	}
+	finalResult := &api.ProviderResult{
+		Model: model,
+		Text:  fullText,
+	}
+	completedResp := s.buildResponse(origReq, finalResult, provider.Name(), responseID)
+	completedResp.Output[0].ID = itemID
+
+	// response.completed
+	s.sendSSE(w, flusher, &seq, "response.completed", &api.StreamEvent{
+		Type:     "response.completed",
+		Response: completedResp,
+	})
+
+	// Store conversation
+	if fullText != "" {
+		assistantMsg := api.Message{
+			Role:    "assistant",
+			Content: []api.ContentBlock{{Type: "output_text", Text: fullText}},
+		}
+		allMsgs := append(storeMsgs, assistantMsg)
+		s.convs.Create(responseID, model, allMsgs)
+	}
+}
+
+func (s *GatewayServer) sendSSE(w http.ResponseWriter, flusher http.Flusher, seq *int, eventType string, event *api.StreamEvent) {
+	event.SequenceNumber = *seq
+	*seq++
+	data, err := json.Marshal(event)
+	if err != nil {
+		s.logger.Printf("failed to marshal SSE event: %v", err)
+		return
+	}
+	fmt.Fprintf(w, "event: %s\ndata: %s\n\n", eventType, data)
+	flusher.Flush()
+}
+
+func (s *GatewayServer) buildResponse(req *api.ResponseRequest, result *api.ProviderResult, providerName string, responseID string) *api.Response {
+	now := time.Now().Unix()
+
+	model := result.Model
+	if model == "" {
+		model = req.Model
 	}
 
-	s.convs.Create(conversationID, fullReq.Model, messages)
+	// Build output item
+	itemID := generateID("msg_")
+	outputItem := api.OutputItem{
+		ID:     itemID,
+		Type:   "message",
+		Status: "completed",
+		Role:   "assistant",
+		Content: []api.ContentPart{{
+			Type:        "output_text",
+			Text:        result.Text,
+			Annotations: []api.Annotation{},
+		}},
+	}
+
+	// Echo back request params with defaults
+	tools := req.Tools
+	if tools == nil {
+		tools = json.RawMessage(`[]`)
+	}
+	toolChoice := req.ToolChoice
+	if toolChoice == nil {
+		toolChoice = json.RawMessage(`"auto"`)
+	}
+	text := req.Text
+	if text == nil {
+		text = json.RawMessage(`{"format":{"type":"text"}}`)
+	}
+	truncation := "disabled"
+	if req.Truncation != nil {
+		truncation = *req.Truncation
+	}
+	temperature := 1.0
+	if req.Temperature != nil {
+		temperature = *req.Temperature
+	}
+	topP := 1.0
+	if req.TopP != nil {
+		topP = *req.TopP
+	}
+	presencePenalty := 0.0
+	if req.PresencePenalty != nil {
+		presencePenalty = *req.PresencePenalty
+	}
+	frequencyPenalty := 0.0
+	if req.FrequencyPenalty != nil {
+		frequencyPenalty = *req.FrequencyPenalty
+	}
+	topLogprobs := 0
+	if req.TopLogprobs != nil {
+		topLogprobs = *req.TopLogprobs
+	}
+	parallelToolCalls := true
+	if req.ParallelToolCalls != nil {
+		parallelToolCalls = *req.ParallelToolCalls
+	}
+	store := true
+	if req.Store != nil {
+		store = *req.Store
+	}
+	background := false
+	if req.Background != nil {
+		background = *req.Background
+	}
+	serviceTier := "default"
+	if req.ServiceTier != nil {
+		serviceTier = *req.ServiceTier
+	}
+	var reasoning json.RawMessage
+	if req.Reasoning != nil {
+		reasoning = req.Reasoning
+	}
+	metadata := req.Metadata
+	if metadata == nil {
+		metadata = map[string]string{}
+	}
+
+	var usage *api.Usage
+	if result.Text != "" {
+		usage = &result.Usage
+	}
+
+	return &api.Response{
+		ID:                 responseID,
+		Object:             "response",
+		CreatedAt:          now,
+		CompletedAt:        &now,
+		Status:             "completed",
+		IncompleteDetails:  nil,
+		Model:              model,
+		PreviousResponseID: req.PreviousResponseID,
+		Instructions:       req.Instructions,
+		Output:             []api.OutputItem{outputItem},
+		Error:              nil,
+		Tools:              tools,
+		ToolChoice:         toolChoice,
+		Truncation:         truncation,
+		ParallelToolCalls:  parallelToolCalls,
+		Text:               text,
+		TopP:               topP,
+		PresencePenalty:    presencePenalty,
+		FrequencyPenalty:   frequencyPenalty,
+		TopLogprobs:        topLogprobs,
+		Temperature:        temperature,
+		Reasoning:          reasoning,
+		Usage:              usage,
+		MaxOutputTokens:    req.MaxOutputTokens,
+		MaxToolCalls:       req.MaxToolCalls,
+		Store:              store,
+		Background:         background,
+		ServiceTier:        serviceTier,
+		Metadata:           metadata,
+		SafetyIdentifier:   nil,
+		PromptCacheKey:     nil,
+		Provider:           providerName,
+	}
 }
 
 func (s *GatewayServer) resolveProvider(req *api.ResponseRequest) (providers.Provider, error) {
@@ -254,3 +489,8 @@ func (s *GatewayServer) resolveProvider(req *api.ResponseRequest) (providers.Pro
 	}
 	return s.registry.Default(req.Model)
 }
+
+func generateID(prefix string) string {
+	id := strings.ReplaceAll(uuid.NewString(), "-", "")
+	return prefix + id[:24]
+}