feat: backend core - auth, user, role, permission, device, webhook, monitoring, cache, repository, service, middleware, API handlers

internal/pkg/apicompat/anthropic_to_responses.go

@@ -0,0 +1,451 @@
+package apicompat
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+)
+
+// AnthropicToResponses converts an Anthropic Messages request directly into
+// a Responses API request. This preserves fields that would be lost in a
+// Chat Completions intermediary round-trip (e.g. thinking, cache_control,
+// structured system prompts).
+func AnthropicToResponses(req *AnthropicRequest) (*ResponsesRequest, error) {
+	input, err := convertAnthropicToResponsesInput(req.System, req.Messages)
+	if err != nil {
+		return nil, err
+	}
+
+	inputJSON, err := json.Marshal(input)
+	if err != nil {
+		return nil, err
+	}
+
+	out := &ResponsesRequest{
+		Model:       req.Model,
+		Input:       inputJSON,
+		Temperature: req.Temperature,
+		TopP:        req.TopP,
+		Stream:      req.Stream,
+		Include:     []string{"reasoning.encrypted_content"},
+	}
+
+	storeFalse := false
+	out.Store = &storeFalse
+
+	if req.MaxTokens > 0 {
+		v := req.MaxTokens
+		if v < minMaxOutputTokens {
+			v = minMaxOutputTokens
+		}
+		out.MaxOutputTokens = &v
+	}
+
+	if len(req.Tools) > 0 {
+		out.Tools = convertAnthropicToolsToResponses(req.Tools)
+	}
+
+	// Determine reasoning effort: only output_config.effort controls the
+	// level; thinking.type is ignored. Default is high when unset (both
+	// Anthropic and OpenAI default to high).
+	// Anthropic levels map 1:1 to OpenAI: low→low, medium→medium, high→high, max→xhigh.
+	effort := "high" // default → both sides' default
+	if req.OutputConfig != nil && req.OutputConfig.Effort != "" {
+		effort = req.OutputConfig.Effort
+	}
+	out.Reasoning = &ResponsesReasoning{
+		Effort:  mapAnthropicEffortToResponses(effort),
+		Summary: "auto",
+	}
+
+	// Convert tool_choice
+	if len(req.ToolChoice) > 0 {
+		tc, err := convertAnthropicToolChoiceToResponses(req.ToolChoice)
+		if err != nil {
+			return nil, fmt.Errorf("convert tool_choice: %w", err)
+		}
+		out.ToolChoice = tc
+	}
+
+	return out, nil
+}
+
+// convertAnthropicToolChoiceToResponses maps Anthropic tool_choice to Responses format.
+//
+//	{"type":"auto"}            → "auto"
+//	{"type":"any"}             → "required"
+//	{"type":"none"}            → "none"
+//	{"type":"tool","name":"X"} → {"type":"function","function":{"name":"X"}}
+func convertAnthropicToolChoiceToResponses(raw json.RawMessage) (json.RawMessage, error) {
+	var tc struct {
+		Type string `json:"type"`
+		Name string `json:"name"`
+	}
+	if err := json.Unmarshal(raw, &tc); err != nil {
+		return nil, err
+	}
+
+	switch tc.Type {
+	case "auto":
+		return json.Marshal("auto")
+	case "any":
+		return json.Marshal("required")
+	case "none":
+		return json.Marshal("none")
+	case "tool":
+		return json.Marshal(map[string]any{
+			"type":     "function",
+			"function": map[string]string{"name": tc.Name},
+		})
+	default:
+		// Pass through unknown types as-is
+		return raw, nil
+	}
+}
+
+// convertAnthropicToResponsesInput builds the Responses API input items array
+// from the Anthropic system field and message list.
+func convertAnthropicToResponsesInput(system json.RawMessage, msgs []AnthropicMessage) ([]ResponsesInputItem, error) {
+	var out []ResponsesInputItem
+
+	// System prompt → system role input item.
+	if len(system) > 0 {
+		sysText, err := parseAnthropicSystemPrompt(system)
+		if err != nil {
+			return nil, err
+		}
+		if sysText != "" {
+			content, _ := json.Marshal(sysText)
+			out = append(out, ResponsesInputItem{
+				Role:    "system",
+				Content: content,
+			})
+		}
+	}
+
+	for _, m := range msgs {
+		items, err := anthropicMsgToResponsesItems(m)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, items...)
+	}
+	return out, nil
+}
+
+// parseAnthropicSystemPrompt handles the Anthropic system field which can be
+// a plain string or an array of text blocks.
+func parseAnthropicSystemPrompt(raw json.RawMessage) (string, error) {
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return s, nil
+	}
+	var blocks []AnthropicContentBlock
+	if err := json.Unmarshal(raw, &blocks); err != nil {
+		return "", err
+	}
+	var parts []string
+	for _, b := range blocks {
+		if b.Type == "text" && b.Text != "" {
+			parts = append(parts, b.Text)
+		}
+	}
+	return strings.Join(parts, "\n\n"), nil
+}
+
+// anthropicMsgToResponsesItems converts a single Anthropic message into one
+// or more Responses API input items.
+func anthropicMsgToResponsesItems(m AnthropicMessage) ([]ResponsesInputItem, error) {
+	switch m.Role {
+	case "user":
+		return anthropicUserToResponses(m.Content)
+	case "assistant":
+		return anthropicAssistantToResponses(m.Content)
+	default:
+		return anthropicUserToResponses(m.Content)
+	}
+}
+
+// anthropicUserToResponses handles an Anthropic user message. Content can be a
+// plain string or an array of blocks. tool_result blocks are extracted into
+// function_call_output items. Image blocks are converted to input_image parts.
+func anthropicUserToResponses(raw json.RawMessage) ([]ResponsesInputItem, error) {
+	// Try plain string.
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		content, _ := json.Marshal(s)
+		return []ResponsesInputItem{{Role: "user", Content: content}}, nil
+	}
+
+	var blocks []AnthropicContentBlock
+	if err := json.Unmarshal(raw, &blocks); err != nil {
+		return nil, err
+	}
+
+	var out []ResponsesInputItem
+	var toolResultImageParts []ResponsesContentPart
+
+	// Extract tool_result blocks → function_call_output items.
+	// Images inside tool_results are extracted separately because the
+	// Responses API function_call_output.output only accepts strings.
+	for _, b := range blocks {
+		if b.Type != "tool_result" {
+			continue
+		}
+		outputText, imageParts := convertToolResultOutput(b)
+		out = append(out, ResponsesInputItem{
+			Type:   "function_call_output",
+			CallID: toResponsesCallID(b.ToolUseID),
+			Output: outputText,
+		})
+		toolResultImageParts = append(toolResultImageParts, imageParts...)
+	}
+
+	// Remaining text + image blocks → user message with content parts.
+	// Also include images extracted from tool_results so the model can see them.
+	var parts []ResponsesContentPart
+	for _, b := range blocks {
+		switch b.Type {
+		case "text":
+			if b.Text != "" {
+				parts = append(parts, ResponsesContentPart{Type: "input_text", Text: b.Text})
+			}
+		case "image":
+			if uri := anthropicImageToDataURI(b.Source); uri != "" {
+				parts = append(parts, ResponsesContentPart{Type: "input_image", ImageURL: uri})
+			}
+		}
+	}
+	parts = append(parts, toolResultImageParts...)
+
+	if len(parts) > 0 {
+		content, err := json.Marshal(parts)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, ResponsesInputItem{Role: "user", Content: content})
+	}
+
+	return out, nil
+}
+
+// anthropicAssistantToResponses handles an Anthropic assistant message.
+// Text content → assistant message with output_text parts.
+// tool_use blocks → function_call items.
+// thinking blocks → ignored (OpenAI doesn't accept them as input).
+func anthropicAssistantToResponses(raw json.RawMessage) ([]ResponsesInputItem, error) {
+	// Try plain string.
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		parts := []ResponsesContentPart{{Type: "output_text", Text: s}}
+		partsJSON, err := json.Marshal(parts)
+		if err != nil {
+			return nil, err
+		}
+		return []ResponsesInputItem{{Role: "assistant", Content: partsJSON}}, nil
+	}
+
+	var blocks []AnthropicContentBlock
+	if err := json.Unmarshal(raw, &blocks); err != nil {
+		return nil, err
+	}
+
+	var items []ResponsesInputItem
+
+	// Text content → assistant message with output_text content parts.
+	text := extractAnthropicTextFromBlocks(blocks)
+	if text != "" {
+		parts := []ResponsesContentPart{{Type: "output_text", Text: text}}
+		partsJSON, err := json.Marshal(parts)
+		if err != nil {
+			return nil, err
+		}
+		items = append(items, ResponsesInputItem{Role: "assistant", Content: partsJSON})
+	}
+
+	// tool_use → function_call items.
+	for _, b := range blocks {
+		if b.Type != "tool_use" {
+			continue
+		}
+		args := "{}"
+		if len(b.Input) > 0 {
+			args = string(b.Input)
+		}
+		fcID := toResponsesCallID(b.ID)
+		items = append(items, ResponsesInputItem{
+			Type:      "function_call",
+			CallID:    fcID,
+			Name:      b.Name,
+			Arguments: args,
+		})
+	}
+
+	return items, nil
+}
+
+// toResponsesCallID converts an Anthropic tool ID (toolu_xxx / call_xxx) to a
+// Responses API function_call ID that starts with "fc_".
+func toResponsesCallID(id string) string {
+	if strings.HasPrefix(id, "fc_") {
+		return id
+	}
+	return "fc_" + id
+}
+
+// fromResponsesCallID reverses toResponsesCallID, stripping the "fc_" prefix
+// that was added during request conversion.
+func fromResponsesCallID(id string) string {
+	if after, ok := strings.CutPrefix(id, "fc_"); ok {
+		// Only strip if the remainder doesn't look like it was already "fc_" prefixed.
+		// E.g. "fc_toolu_xxx" → "toolu_xxx", "fc_call_xxx" → "call_xxx"
+		if strings.HasPrefix(after, "toolu_") || strings.HasPrefix(after, "call_") {
+			return after
+		}
+	}
+	return id
+}
+
+// anthropicImageToDataURI converts an AnthropicImageSource to a data URI string.
+// Returns "" if the source is nil or has no data.
+func anthropicImageToDataURI(src *AnthropicImageSource) string {
+	if src == nil || src.Data == "" {
+		return ""
+	}
+	mediaType := src.MediaType
+	if mediaType == "" {
+		mediaType = "image/png"
+	}
+	return "data:" + mediaType + ";base64," + src.Data
+}
+
+// convertToolResultOutput extracts text and image content from a tool_result
+// block. Returns the text as a string for the function_call_output Output
+// field, plus any image parts that must be sent in a separate user message
+// (the Responses API output field only accepts strings).
+func convertToolResultOutput(b AnthropicContentBlock) (string, []ResponsesContentPart) {
+	if len(b.Content) == 0 {
+		return "(empty)", nil
+	}
+
+	// Try plain string content.
+	var s string
+	if err := json.Unmarshal(b.Content, &s); err == nil {
+		if s == "" {
+			s = "(empty)"
+		}
+		return s, nil
+	}
+
+	// Array of content blocks — may contain text and/or images.
+	var inner []AnthropicContentBlock
+	if err := json.Unmarshal(b.Content, &inner); err != nil {
+		return "(empty)", nil
+	}
+
+	// Separate text (for function_call_output) from images (for user message).
+	var textParts []string
+	var imageParts []ResponsesContentPart
+	for _, ib := range inner {
+		switch ib.Type {
+		case "text":
+			if ib.Text != "" {
+				textParts = append(textParts, ib.Text)
+			}
+		case "image":
+			if uri := anthropicImageToDataURI(ib.Source); uri != "" {
+				imageParts = append(imageParts, ResponsesContentPart{Type: "input_image", ImageURL: uri})
+			}
+		}
+	}
+
+	text := strings.Join(textParts, "\n\n")
+	if text == "" {
+		text = "(empty)"
+	}
+	return text, imageParts
+}
+
+// extractAnthropicTextFromBlocks joins all text blocks, ignoring thinking/
+// tool_use/tool_result blocks.
+func extractAnthropicTextFromBlocks(blocks []AnthropicContentBlock) string {
+	var parts []string
+	for _, b := range blocks {
+		if b.Type == "text" && b.Text != "" {
+			parts = append(parts, b.Text)
+		}
+	}
+	return strings.Join(parts, "\n\n")
+}
+
+// mapAnthropicEffortToResponses converts Anthropic reasoning effort levels to
+// OpenAI Responses API effort levels.
+//
+// Both APIs default to "high". The mapping is 1:1 for shared levels;
+// only Anthropic's "max" (Opus 4.6 exclusive) maps to OpenAI's "xhigh"
+// (GPT-5.2+ exclusive) as both represent the highest reasoning tier.
+//
+//	low    → low
+//	medium → medium
+//	high   → high
+//	max    → xhigh
+func mapAnthropicEffortToResponses(effort string) string {
+	if effort == "max" {
+		return "xhigh"
+	}
+	return effort // low→low, medium→medium, high→high, unknown→passthrough
+}
+
+// convertAnthropicToolsToResponses maps Anthropic tool definitions to
+// Responses API tools. Server-side tools like web_search are mapped to their
+// OpenAI equivalents; regular tools become function tools.
+func convertAnthropicToolsToResponses(tools []AnthropicTool) []ResponsesTool {
+	var out []ResponsesTool
+	for _, t := range tools {
+		// Anthropic server tools like "web_search_20250305" → OpenAI {"type":"web_search"}
+		if strings.HasPrefix(t.Type, "web_search") {
+			out = append(out, ResponsesTool{Type: "web_search"})
+			continue
+		}
+		out = append(out, ResponsesTool{
+			Type:        "function",
+			Name:        t.Name,
+			Description: t.Description,
+			Parameters:  normalizeToolParameters(t.InputSchema),
+		})
+	}
+	return out
+}
+
+// normalizeToolParameters ensures the tool parameter schema is valid for
+// OpenAI's Responses API, which requires "properties" on object schemas.
+//
+//   - nil/empty → {"type":"object","properties":{}}
+//   - type=object without properties → adds "properties": {}
+//   - otherwise → returned unchanged
+func normalizeToolParameters(schema json.RawMessage) json.RawMessage {
+	if len(schema) == 0 || string(schema) == "null" {
+		return json.RawMessage(`{"type":"object","properties":{}}`)
+	}
+
+	var m map[string]json.RawMessage
+	if err := json.Unmarshal(schema, &m); err != nil {
+		return schema
+	}
+
+	typ := m["type"]
+	if string(typ) != `"object"` {
+		return schema
+	}
+
+	if _, ok := m["properties"]; ok {
+		return schema
+	}
+
+	m["properties"] = json.RawMessage(`{}`)
+	out, err := json.Marshal(m)
+	if err != nil {
+		return schema
+	}
+	return out
+}

internal/pkg/apicompat/anthropic_to_responses_response.go

@@ -0,0 +1,521 @@
+package apicompat
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"time"
+)
+
+// ---------------------------------------------------------------------------
+// Non-streaming: AnthropicResponse → ResponsesResponse
+// ---------------------------------------------------------------------------
+
+// AnthropicToResponsesResponse converts an Anthropic Messages response into a
+// Responses API response. This is the reverse of ResponsesToAnthropic and
+// enables Anthropic upstream responses to be returned in OpenAI Responses format.
+func AnthropicToResponsesResponse(resp *AnthropicResponse) *ResponsesResponse {
+	id := resp.ID
+	if id == "" {
+		id = generateResponsesID()
+	}
+
+	out := &ResponsesResponse{
+		ID:     id,
+		Object: "response",
+		Model:  resp.Model,
+	}
+
+	var outputs []ResponsesOutput
+	var msgParts []ResponsesContentPart
+
+	for _, block := range resp.Content {
+		switch block.Type {
+		case "thinking":
+			if block.Thinking != "" {
+				outputs = append(outputs, ResponsesOutput{
+					Type: "reasoning",
+					ID:   generateItemID(),
+					Summary: []ResponsesSummary{{
+						Type: "summary_text",
+						Text: block.Thinking,
+					}},
+				})
+			}
+		case "text":
+			if block.Text != "" {
+				msgParts = append(msgParts, ResponsesContentPart{
+					Type: "output_text",
+					Text: block.Text,
+				})
+			}
+		case "tool_use":
+			args := "{}"
+			if len(block.Input) > 0 {
+				args = string(block.Input)
+			}
+			outputs = append(outputs, ResponsesOutput{
+				Type:      "function_call",
+				ID:        generateItemID(),
+				CallID:    toResponsesCallID(block.ID),
+				Name:      block.Name,
+				Arguments: args,
+				Status:    "completed",
+			})
+		}
+	}
+
+	// Assemble message output item from text parts
+	if len(msgParts) > 0 {
+		outputs = append(outputs, ResponsesOutput{
+			Type:    "message",
+			ID:      generateItemID(),
+			Role:    "assistant",
+			Content: msgParts,
+			Status:  "completed",
+		})
+	}
+
+	if len(outputs) == 0 {
+		outputs = append(outputs, ResponsesOutput{
+			Type:    "message",
+			ID:      generateItemID(),
+			Role:    "assistant",
+			Content: []ResponsesContentPart{{Type: "output_text", Text: ""}},
+			Status:  "completed",
+		})
+	}
+	out.Output = outputs
+
+	// Map stop_reason → status
+	out.Status = anthropicStopReasonToResponsesStatus(resp.StopReason, resp.Content)
+	if out.Status == "incomplete" {
+		out.IncompleteDetails = &ResponsesIncompleteDetails{Reason: "max_output_tokens"}
+	}
+
+	// Usage
+	out.Usage = &ResponsesUsage{
+		InputTokens:  resp.Usage.InputTokens,
+		OutputTokens: resp.Usage.OutputTokens,
+		TotalTokens:  resp.Usage.InputTokens + resp.Usage.OutputTokens,
+	}
+	if resp.Usage.CacheReadInputTokens > 0 {
+		out.Usage.InputTokensDetails = &ResponsesInputTokensDetails{
+			CachedTokens: resp.Usage.CacheReadInputTokens,
+		}
+	}
+
+	return out
+}
+
+// anthropicStopReasonToResponsesStatus maps Anthropic stop_reason to Responses status.
+func anthropicStopReasonToResponsesStatus(stopReason string, blocks []AnthropicContentBlock) string {
+	switch stopReason {
+	case "max_tokens":
+		return "incomplete"
+	case "end_turn", "tool_use", "stop_sequence":
+		return "completed"
+	default:
+		return "completed"
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Streaming: AnthropicStreamEvent → []ResponsesStreamEvent (stateful converter)
+// ---------------------------------------------------------------------------
+
+// AnthropicEventToResponsesState tracks state for converting a sequence of
+// Anthropic SSE events into Responses SSE events.
+type AnthropicEventToResponsesState struct {
+	ResponseID     string
+	Model          string
+	Created        int64
+	SequenceNumber int
+
+	// CreatedSent tracks whether response.created has been emitted.
+	CreatedSent bool
+	// CompletedSent tracks whether the terminal event has been emitted.
+	CompletedSent bool
+
+	// Current output tracking
+	OutputIndex     int
+	CurrentItemID   string
+	CurrentItemType string // "message" | "function_call" | "reasoning"
+
+	// For message output: accumulate text parts
+	ContentIndex int
+
+	// For function_call: track per-output info
+	CurrentCallID string
+	CurrentName   string
+
+	// Usage from message_delta
+	InputTokens          int
+	OutputTokens         int
+	CacheReadInputTokens int
+}
+
+// NewAnthropicEventToResponsesState returns an initialised stream state.
+func NewAnthropicEventToResponsesState() *AnthropicEventToResponsesState {
+	return &AnthropicEventToResponsesState{
+		Created: time.Now().Unix(),
+	}
+}
+
+// AnthropicEventToResponsesEvents converts a single Anthropic SSE event into
+// zero or more Responses SSE events, updating state as it goes.
+func AnthropicEventToResponsesEvents(
+	evt *AnthropicStreamEvent,
+	state *AnthropicEventToResponsesState,
+) []ResponsesStreamEvent {
+	switch evt.Type {
+	case "message_start":
+		return anthToResHandleMessageStart(evt, state)
+	case "content_block_start":
+		return anthToResHandleContentBlockStart(evt, state)
+	case "content_block_delta":
+		return anthToResHandleContentBlockDelta(evt, state)
+	case "content_block_stop":
+		return anthToResHandleContentBlockStop(evt, state)
+	case "message_delta":
+		return anthToResHandleMessageDelta(evt, state)
+	case "message_stop":
+		return anthToResHandleMessageStop(state)
+	default:
+		return nil
+	}
+}
+
+// FinalizeAnthropicResponsesStream emits synthetic termination events if the
+// stream ended without a proper message_stop.
+func FinalizeAnthropicResponsesStream(state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	if !state.CreatedSent || state.CompletedSent {
+		return nil
+	}
+
+	var events []ResponsesStreamEvent
+
+	// Close any open item
+	events = append(events, closeCurrentResponsesItem(state)...)
+
+	// Emit response.completed
+	events = append(events, makeResponsesCompletedEvent(state, "completed", nil))
+	state.CompletedSent = true
+	return events
+}
+
+// ResponsesEventToSSE formats a ResponsesStreamEvent as an SSE data line.
+func ResponsesEventToSSE(evt ResponsesStreamEvent) (string, error) {
+	data, err := json.Marshal(evt)
+	if err != nil {
+		return "", err
+	}
+	return fmt.Sprintf("event: %s\ndata: %s\n\n", evt.Type, data), nil
+}
+
+// --- internal handlers ---
+
+func anthToResHandleMessageStart(evt *AnthropicStreamEvent, state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	if evt.Message != nil {
+		state.ResponseID = evt.Message.ID
+		if state.Model == "" {
+			state.Model = evt.Message.Model
+		}
+		if evt.Message.Usage.InputTokens > 0 {
+			state.InputTokens = evt.Message.Usage.InputTokens
+		}
+	}
+
+	if state.CreatedSent {
+		return nil
+	}
+	state.CreatedSent = true
+
+	// Emit response.created
+	return []ResponsesStreamEvent{makeResponsesCreatedEvent(state)}
+}
+
+func anthToResHandleContentBlockStart(evt *AnthropicStreamEvent, state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	if evt.ContentBlock == nil {
+		return nil
+	}
+
+	var events []ResponsesStreamEvent
+
+	switch evt.ContentBlock.Type {
+	case "thinking":
+		state.CurrentItemID = generateItemID()
+		state.CurrentItemType = "reasoning"
+		state.ContentIndex = 0
+
+		events = append(events, makeResponsesEvent(state, "response.output_item.added", &ResponsesStreamEvent{
+			OutputIndex: state.OutputIndex,
+			Item: &ResponsesOutput{
+				Type: "reasoning",
+				ID:   state.CurrentItemID,
+			},
+		}))
+
+	case "text":
+		// If we don't have an open message item, open one
+		if state.CurrentItemType != "message" {
+			state.CurrentItemID = generateItemID()
+			state.CurrentItemType = "message"
+			state.ContentIndex = 0
+
+			events = append(events, makeResponsesEvent(state, "response.output_item.added", &ResponsesStreamEvent{
+				OutputIndex: state.OutputIndex,
+				Item: &ResponsesOutput{
+					Type:   "message",
+					ID:     state.CurrentItemID,
+					Role:   "assistant",
+					Status: "in_progress",
+				},
+			}))
+		}
+
+	case "tool_use":
+		// Close previous item if any
+		events = append(events, closeCurrentResponsesItem(state)...)
+
+		state.CurrentItemID = generateItemID()
+		state.CurrentItemType = "function_call"
+		state.CurrentCallID = toResponsesCallID(evt.ContentBlock.ID)
+		state.CurrentName = evt.ContentBlock.Name
+
+		events = append(events, makeResponsesEvent(state, "response.output_item.added", &ResponsesStreamEvent{
+			OutputIndex: state.OutputIndex,
+			Item: &ResponsesOutput{
+				Type:   "function_call",
+				ID:     state.CurrentItemID,
+				CallID: state.CurrentCallID,
+				Name:   state.CurrentName,
+				Status: "in_progress",
+			},
+		}))
+	}
+
+	return events
+}
+
+func anthToResHandleContentBlockDelta(evt *AnthropicStreamEvent, state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	if evt.Delta == nil {
+		return nil
+	}
+
+	switch evt.Delta.Type {
+	case "text_delta":
+		if evt.Delta.Text == "" {
+			return nil
+		}
+		return []ResponsesStreamEvent{makeResponsesEvent(state, "response.output_text.delta", &ResponsesStreamEvent{
+			OutputIndex:  state.OutputIndex,
+			ContentIndex: state.ContentIndex,
+			Delta:        evt.Delta.Text,
+			ItemID:       state.CurrentItemID,
+		})}
+
+	case "thinking_delta":
+		if evt.Delta.Thinking == "" {
+			return nil
+		}
+		return []ResponsesStreamEvent{makeResponsesEvent(state, "response.reasoning_summary_text.delta", &ResponsesStreamEvent{
+			OutputIndex:  state.OutputIndex,
+			SummaryIndex: 0,
+			Delta:        evt.Delta.Thinking,
+			ItemID:       state.CurrentItemID,
+		})}
+
+	case "input_json_delta":
+		if evt.Delta.PartialJSON == "" {
+			return nil
+		}
+		return []ResponsesStreamEvent{makeResponsesEvent(state, "response.function_call_arguments.delta", &ResponsesStreamEvent{
+			OutputIndex: state.OutputIndex,
+			Delta:       evt.Delta.PartialJSON,
+			ItemID:      state.CurrentItemID,
+			CallID:      state.CurrentCallID,
+			Name:        state.CurrentName,
+		})}
+
+	case "signature_delta":
+		// Anthropic signature deltas have no Responses equivalent; skip
+		return nil
+	}
+
+	return nil
+}
+
+func anthToResHandleContentBlockStop(evt *AnthropicStreamEvent, state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	switch state.CurrentItemType {
+	case "reasoning":
+		// Emit reasoning summary done + output item done
+		events := []ResponsesStreamEvent{
+			makeResponsesEvent(state, "response.reasoning_summary_text.done", &ResponsesStreamEvent{
+				OutputIndex:  state.OutputIndex,
+				SummaryIndex: 0,
+				ItemID:       state.CurrentItemID,
+			}),
+		}
+		events = append(events, closeCurrentResponsesItem(state)...)
+		return events
+
+	case "function_call":
+		// Emit function_call_arguments.done + output item done
+		events := []ResponsesStreamEvent{
+			makeResponsesEvent(state, "response.function_call_arguments.done", &ResponsesStreamEvent{
+				OutputIndex: state.OutputIndex,
+				ItemID:      state.CurrentItemID,
+				CallID:      state.CurrentCallID,
+				Name:        state.CurrentName,
+			}),
+		}
+		events = append(events, closeCurrentResponsesItem(state)...)
+		return events
+
+	case "message":
+		// Emit output_text.done (text block is done, but message item stays open for potential more blocks)
+		return []ResponsesStreamEvent{
+			makeResponsesEvent(state, "response.output_text.done", &ResponsesStreamEvent{
+				OutputIndex:  state.OutputIndex,
+				ContentIndex: state.ContentIndex,
+				ItemID:       state.CurrentItemID,
+			}),
+		}
+	}
+
+	return nil
+}
+
+func anthToResHandleMessageDelta(evt *AnthropicStreamEvent, state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	// Update usage
+	if evt.Usage != nil {
+		state.OutputTokens = evt.Usage.OutputTokens
+		if evt.Usage.CacheReadInputTokens > 0 {
+			state.CacheReadInputTokens = evt.Usage.CacheReadInputTokens
+		}
+	}
+
+	return nil
+}
+
+func anthToResHandleMessageStop(state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	if state.CompletedSent {
+		return nil
+	}
+
+	var events []ResponsesStreamEvent
+
+	// Close any open item
+	events = append(events, closeCurrentResponsesItem(state)...)
+
+	// Determine status
+	status := "completed"
+	var incompleteDetails *ResponsesIncompleteDetails
+
+	// Emit response.completed
+	events = append(events, makeResponsesCompletedEvent(state, status, incompleteDetails))
+	state.CompletedSent = true
+	return events
+}
+
+// --- helper functions ---
+
+func closeCurrentResponsesItem(state *AnthropicEventToResponsesState) []ResponsesStreamEvent {
+	if state.CurrentItemType == "" {
+		return nil
+	}
+
+	itemType := state.CurrentItemType
+	itemID := state.CurrentItemID
+
+	// Reset
+	state.CurrentItemType = ""
+	state.CurrentItemID = ""
+	state.CurrentCallID = ""
+	state.CurrentName = ""
+	state.OutputIndex++
+	state.ContentIndex = 0
+
+	return []ResponsesStreamEvent{makeResponsesEvent(state, "response.output_item.done", &ResponsesStreamEvent{
+		OutputIndex: state.OutputIndex - 1, // Use the index before increment
+		Item: &ResponsesOutput{
+			Type:   itemType,
+			ID:     itemID,
+			Status: "completed",
+		},
+	})}
+}
+
+func makeResponsesCreatedEvent(state *AnthropicEventToResponsesState) ResponsesStreamEvent {
+	seq := state.SequenceNumber
+	state.SequenceNumber++
+	return ResponsesStreamEvent{
+		Type:           "response.created",
+		SequenceNumber: seq,
+		Response: &ResponsesResponse{
+			ID:     state.ResponseID,
+			Object: "response",
+			Model:  state.Model,
+			Status: "in_progress",
+			Output: []ResponsesOutput{},
+		},
+	}
+}
+
+func makeResponsesCompletedEvent(
+	state *AnthropicEventToResponsesState,
+	status string,
+	incompleteDetails *ResponsesIncompleteDetails,
+) ResponsesStreamEvent {
+	seq := state.SequenceNumber
+	state.SequenceNumber++
+
+	usage := &ResponsesUsage{
+		InputTokens:  state.InputTokens,
+		OutputTokens: state.OutputTokens,
+		TotalTokens:  state.InputTokens + state.OutputTokens,
+	}
+	if state.CacheReadInputTokens > 0 {
+		usage.InputTokensDetails = &ResponsesInputTokensDetails{
+			CachedTokens: state.CacheReadInputTokens,
+		}
+	}
+
+	return ResponsesStreamEvent{
+		Type:           "response.completed",
+		SequenceNumber: seq,
+		Response: &ResponsesResponse{
+			ID:                state.ResponseID,
+			Object:            "response",
+			Model:             state.Model,
+			Status:            status,
+			Output:            []ResponsesOutput{}, // Simplified; full output tracking would add complexity
+			Usage:             usage,
+			IncompleteDetails: incompleteDetails,
+		},
+	}
+}
+
+func makeResponsesEvent(state *AnthropicEventToResponsesState, eventType string, template *ResponsesStreamEvent) ResponsesStreamEvent {
+	seq := state.SequenceNumber
+	state.SequenceNumber++
+
+	evt := *template
+	evt.Type = eventType
+	evt.SequenceNumber = seq
+	return evt
+}
+
+func generateResponsesID() string {
+	b := make([]byte, 12)
+	_, _ = rand.Read(b)
+	return "resp_" + hex.EncodeToString(b)
+}
+
+func generateItemID() string {
+	b := make([]byte, 12)
+	_, _ = rand.Read(b)
+	return "item_" + hex.EncodeToString(b)
+}

internal/pkg/apicompat/chatcompletions_responses_test.go

@@ -0,0 +1,878 @@
+package apicompat
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// ---------------------------------------------------------------------------
+// ChatCompletionsToResponses tests
+// ---------------------------------------------------------------------------
+
+func TestChatCompletionsToResponses_BasicText(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Hello"`)},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+	assert.Equal(t, "gpt-4o", resp.Model)
+	assert.True(t, resp.Stream) // always forced true
+	assert.False(t, *resp.Store)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 1)
+	assert.Equal(t, "user", items[0].Role)
+}
+
+func TestChatCompletionsToResponses_SystemMessage(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "system", Content: json.RawMessage(`"You are helpful."`)},
+			{Role: "user", Content: json.RawMessage(`"Hi"`)},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 2)
+	assert.Equal(t, "system", items[0].Role)
+	assert.Equal(t, "user", items[1].Role)
+}
+
+func TestChatCompletionsToResponses_ToolCalls(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Call the function"`)},
+			{
+				Role: "assistant",
+				ToolCalls: []ChatToolCall{
+					{
+						ID:   "call_1",
+						Type: "function",
+						Function: ChatFunctionCall{
+							Name:      "ping",
+							Arguments: `{"host":"example.com"}`,
+						},
+					},
+				},
+			},
+			{
+				Role:       "tool",
+				ToolCallID: "call_1",
+				Content:    json.RawMessage(`"pong"`),
+			},
+		},
+		Tools: []ChatTool{
+			{
+				Type: "function",
+				Function: &ChatFunction{
+					Name:        "ping",
+					Description: "Ping a host",
+					Parameters:  json.RawMessage(`{"type":"object"}`),
+				},
+			},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	// user + function_call + function_call_output = 3
+	// (assistant message with empty content + tool_calls → only function_call items emitted)
+	require.Len(t, items, 3)
+
+	// Check function_call item
+	assert.Equal(t, "function_call", items[1].Type)
+	assert.Equal(t, "call_1", items[1].CallID)
+	assert.Empty(t, items[1].ID)
+	assert.Equal(t, "ping", items[1].Name)
+
+	// Check function_call_output item
+	assert.Equal(t, "function_call_output", items[2].Type)
+	assert.Equal(t, "call_1", items[2].CallID)
+	assert.Equal(t, "pong", items[2].Output)
+
+	// Check tools
+	require.Len(t, resp.Tools, 1)
+	assert.Equal(t, "function", resp.Tools[0].Type)
+	assert.Equal(t, "ping", resp.Tools[0].Name)
+}
+
+func TestChatCompletionsToResponses_MaxTokens(t *testing.T) {
+	t.Run("max_tokens", func(t *testing.T) {
+		maxTokens := 100
+		req := &ChatCompletionsRequest{
+			Model:     "gpt-4o",
+			MaxTokens: &maxTokens,
+			Messages:  []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+		}
+		resp, err := ChatCompletionsToResponses(req)
+		require.NoError(t, err)
+		require.NotNil(t, resp.MaxOutputTokens)
+		// Below minMaxOutputTokens (128), should be clamped
+		assert.Equal(t, minMaxOutputTokens, *resp.MaxOutputTokens)
+	})
+
+	t.Run("max_completion_tokens_preferred", func(t *testing.T) {
+		maxTokens := 100
+		maxCompletion := 500
+		req := &ChatCompletionsRequest{
+			Model:               "gpt-4o",
+			MaxTokens:           &maxTokens,
+			MaxCompletionTokens: &maxCompletion,
+			Messages:            []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+		}
+		resp, err := ChatCompletionsToResponses(req)
+		require.NoError(t, err)
+		require.NotNil(t, resp.MaxOutputTokens)
+		assert.Equal(t, 500, *resp.MaxOutputTokens)
+	})
+}
+
+func TestChatCompletionsToResponses_ReasoningEffort(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model:           "gpt-4o",
+		ReasoningEffort: "high",
+		Messages:        []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+	}
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+	require.NotNil(t, resp.Reasoning)
+	assert.Equal(t, "high", resp.Reasoning.Effort)
+	assert.Equal(t, "auto", resp.Reasoning.Summary)
+}
+
+func TestChatCompletionsToResponses_ImageURL(t *testing.T) {
+	content := `[{"type":"text","text":"Describe this"},{"type":"image_url","image_url":{"url":"data:image/png;base64,abc123"}}]`
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(content)},
+		},
+	}
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 1)
+
+	var parts []ResponsesContentPart
+	require.NoError(t, json.Unmarshal(items[0].Content, &parts))
+	require.Len(t, parts, 2)
+	assert.Equal(t, "input_text", parts[0].Type)
+	assert.Equal(t, "Describe this", parts[0].Text)
+	assert.Equal(t, "input_image", parts[1].Type)
+	assert.Equal(t, "data:image/png;base64,abc123", parts[1].ImageURL)
+}
+
+func TestChatCompletionsToResponses_SystemArrayContent(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "system", Content: json.RawMessage(`[{"type":"text","text":"You are a careful visual assistant."}]`)},
+			{Role: "user", Content: json.RawMessage(`[{"type":"text","text":"Describe this image"},{"type":"image_url","image_url":{"url":"data:image/png;base64,abc123"}}]`)},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 2)
+
+	var systemParts []ResponsesContentPart
+	require.NoError(t, json.Unmarshal(items[0].Content, &systemParts))
+	require.Len(t, systemParts, 1)
+	assert.Equal(t, "input_text", systemParts[0].Type)
+	assert.Equal(t, "You are a careful visual assistant.", systemParts[0].Text)
+
+	var userParts []ResponsesContentPart
+	require.NoError(t, json.Unmarshal(items[1].Content, &userParts))
+	require.Len(t, userParts, 2)
+	assert.Equal(t, "input_image", userParts[1].Type)
+	assert.Equal(t, "data:image/png;base64,abc123", userParts[1].ImageURL)
+}
+
+func TestChatCompletionsToResponses_LegacyFunctions(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Hi"`)},
+		},
+		Functions: []ChatFunction{
+			{
+				Name:        "get_weather",
+				Description: "Get weather",
+				Parameters:  json.RawMessage(`{"type":"object"}`),
+			},
+		},
+		FunctionCall: json.RawMessage(`{"name":"get_weather"}`),
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+	require.Len(t, resp.Tools, 1)
+	assert.Equal(t, "function", resp.Tools[0].Type)
+	assert.Equal(t, "get_weather", resp.Tools[0].Name)
+
+	// tool_choice should be converted
+	require.NotNil(t, resp.ToolChoice)
+	var tc map[string]any
+	require.NoError(t, json.Unmarshal(resp.ToolChoice, &tc))
+	assert.Equal(t, "function", tc["type"])
+}
+
+func TestChatCompletionsToResponses_ServiceTier(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model:       "gpt-4o",
+		ServiceTier: "flex",
+		Messages:    []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+	}
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+	assert.Equal(t, "flex", resp.ServiceTier)
+}
+
+func TestChatCompletionsToResponses_AssistantWithTextAndToolCalls(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Do something"`)},
+			{
+				Role:    "assistant",
+				Content: json.RawMessage(`"Let me call a function."`),
+				ToolCalls: []ChatToolCall{
+					{
+						ID:   "call_abc",
+						Type: "function",
+						Function: ChatFunctionCall{
+							Name:      "do_thing",
+							Arguments: `{}`,
+						},
+					},
+				},
+			},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	// user + assistant message (with text) + function_call
+	require.Len(t, items, 3)
+	assert.Equal(t, "user", items[0].Role)
+	assert.Equal(t, "assistant", items[1].Role)
+	assert.Equal(t, "function_call", items[2].Type)
+	assert.Empty(t, items[2].ID)
+}
+
+func TestChatCompletionsToResponses_AssistantArrayContentPreserved(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Hi"`)},
+			{Role: "assistant", Content: json.RawMessage(`[{"type":"text","text":"A"},{"type":"text","text":"B"}]`)},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 2)
+	assert.Equal(t, "assistant", items[1].Role)
+
+	var parts []ResponsesContentPart
+	require.NoError(t, json.Unmarshal(items[1].Content, &parts))
+	require.Len(t, parts, 1)
+	assert.Equal(t, "output_text", parts[0].Type)
+	assert.Equal(t, "AB", parts[0].Text)
+}
+
+func TestChatCompletionsToResponses_AssistantThinkingTagPreserved(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Hi"`)},
+			{Role: "assistant", Content: json.RawMessage(`[{"type":"thinking","thinking":"internal plan"},{"type":"text","text":"final answer"}]`)},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 2)
+
+	var parts []ResponsesContentPart
+	require.NoError(t, json.Unmarshal(items[1].Content, &parts))
+	require.Len(t, parts, 1)
+	assert.Equal(t, "output_text", parts[0].Type)
+	assert.Contains(t, parts[0].Text, "<thinking>internal plan</thinking>")
+	assert.Contains(t, parts[0].Text, "final answer")
+}
+
+// ---------------------------------------------------------------------------
+// ResponsesToChatCompletions tests
+// ---------------------------------------------------------------------------
+
+func TestResponsesToChatCompletions_BasicText(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:     "resp_123",
+		Status: "completed",
+		Output: []ResponsesOutput{
+			{
+				Type: "message",
+				Content: []ResponsesContentPart{
+					{Type: "output_text", Text: "Hello, world!"},
+				},
+			},
+		},
+		Usage: &ResponsesUsage{
+			InputTokens:  10,
+			OutputTokens: 5,
+			TotalTokens:  15,
+		},
+	}
+
+	chat := ResponsesToChatCompletions(resp, "gpt-4o")
+	assert.Equal(t, "chat.completion", chat.Object)
+	assert.Equal(t, "gpt-4o", chat.Model)
+	require.Len(t, chat.Choices, 1)
+	assert.Equal(t, "stop", chat.Choices[0].FinishReason)
+
+	var content string
+	require.NoError(t, json.Unmarshal(chat.Choices[0].Message.Content, &content))
+	assert.Equal(t, "Hello, world!", content)
+
+	require.NotNil(t, chat.Usage)
+	assert.Equal(t, 10, chat.Usage.PromptTokens)
+	assert.Equal(t, 5, chat.Usage.CompletionTokens)
+	assert.Equal(t, 15, chat.Usage.TotalTokens)
+}
+
+func TestResponsesToChatCompletions_ToolCalls(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:     "resp_456",
+		Status: "completed",
+		Output: []ResponsesOutput{
+			{
+				Type:      "function_call",
+				CallID:    "call_xyz",
+				Name:      "get_weather",
+				Arguments: `{"city":"NYC"}`,
+			},
+		},
+	}
+
+	chat := ResponsesToChatCompletions(resp, "gpt-4o")
+	require.Len(t, chat.Choices, 1)
+	assert.Equal(t, "tool_calls", chat.Choices[0].FinishReason)
+
+	msg := chat.Choices[0].Message
+	require.Len(t, msg.ToolCalls, 1)
+	assert.Equal(t, "call_xyz", msg.ToolCalls[0].ID)
+	assert.Equal(t, "function", msg.ToolCalls[0].Type)
+	assert.Equal(t, "get_weather", msg.ToolCalls[0].Function.Name)
+	assert.Equal(t, `{"city":"NYC"}`, msg.ToolCalls[0].Function.Arguments)
+}
+
+func TestResponsesToChatCompletions_Reasoning(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:     "resp_789",
+		Status: "completed",
+		Output: []ResponsesOutput{
+			{
+				Type: "reasoning",
+				Summary: []ResponsesSummary{
+					{Type: "summary_text", Text: "I thought about it."},
+				},
+			},
+			{
+				Type: "message",
+				Content: []ResponsesContentPart{
+					{Type: "output_text", Text: "The answer is 42."},
+				},
+			},
+		},
+	}
+
+	chat := ResponsesToChatCompletions(resp, "gpt-4o")
+	require.Len(t, chat.Choices, 1)
+
+	var content string
+	require.NoError(t, json.Unmarshal(chat.Choices[0].Message.Content, &content))
+	assert.Equal(t, "The answer is 42.", content)
+	assert.Equal(t, "I thought about it.", chat.Choices[0].Message.ReasoningContent)
+}
+
+func TestChatCompletionsToResponses_ToolArrayContent(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Use the tool"`)},
+			{
+				Role: "assistant",
+				ToolCalls: []ChatToolCall{
+					{
+						ID:   "call_1",
+						Type: "function",
+						Function: ChatFunctionCall{
+							Name:      "inspect_image",
+							Arguments: `{}`,
+						},
+					},
+				},
+			},
+			{
+				Role:       "tool",
+				ToolCallID: "call_1",
+				Content: json.RawMessage(
+					`[{"type":"text","text":"image width: 100"},{"type":"image_url","image_url":{"url":"data:image/png;base64,ignored"}},{"type":"text","text":"; image height: 200"}]`,
+				),
+			},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 3)
+	assert.Equal(t, "function_call_output", items[2].Type)
+	assert.Equal(t, "call_1", items[2].CallID)
+	assert.Equal(t, "image width: 100; image height: 200", items[2].Output)
+}
+
+func TestResponsesToChatCompletions_Incomplete(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:                "resp_inc",
+		Status:            "incomplete",
+		IncompleteDetails: &ResponsesIncompleteDetails{Reason: "max_output_tokens"},
+		Output: []ResponsesOutput{
+			{
+				Type: "message",
+				Content: []ResponsesContentPart{
+					{Type: "output_text", Text: "partial..."},
+				},
+			},
+		},
+	}
+
+	chat := ResponsesToChatCompletions(resp, "gpt-4o")
+	require.Len(t, chat.Choices, 1)
+	assert.Equal(t, "length", chat.Choices[0].FinishReason)
+}
+
+func TestResponsesToChatCompletions_CachedTokens(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:     "resp_cache",
+		Status: "completed",
+		Output: []ResponsesOutput{
+			{
+				Type:    "message",
+				Content: []ResponsesContentPart{{Type: "output_text", Text: "cached"}},
+			},
+		},
+		Usage: &ResponsesUsage{
+			InputTokens:  100,
+			OutputTokens: 10,
+			TotalTokens:  110,
+			InputTokensDetails: &ResponsesInputTokensDetails{
+				CachedTokens: 80,
+			},
+		},
+	}
+
+	chat := ResponsesToChatCompletions(resp, "gpt-4o")
+	require.NotNil(t, chat.Usage)
+	require.NotNil(t, chat.Usage.PromptTokensDetails)
+	assert.Equal(t, 80, chat.Usage.PromptTokensDetails.CachedTokens)
+}
+
+func TestResponsesToChatCompletions_WebSearch(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:     "resp_ws",
+		Status: "completed",
+		Output: []ResponsesOutput{
+			{
+				Type:   "web_search_call",
+				Action: &WebSearchAction{Type: "search", Query: "test"},
+			},
+			{
+				Type:    "message",
+				Content: []ResponsesContentPart{{Type: "output_text", Text: "search results"}},
+			},
+		},
+	}
+
+	chat := ResponsesToChatCompletions(resp, "gpt-4o")
+	require.Len(t, chat.Choices, 1)
+	assert.Equal(t, "stop", chat.Choices[0].FinishReason)
+
+	var content string
+	require.NoError(t, json.Unmarshal(chat.Choices[0].Message.Content, &content))
+	assert.Equal(t, "search results", content)
+}
+
+// ---------------------------------------------------------------------------
+// Streaming: ResponsesEventToChatChunks tests
+// ---------------------------------------------------------------------------
+
+func TestResponsesEventToChatChunks_TextDelta(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+
+	// response.created → role chunk
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.created",
+		Response: &ResponsesResponse{
+			ID: "resp_stream",
+		},
+	}, state)
+	require.Len(t, chunks, 1)
+	assert.Equal(t, "assistant", chunks[0].Choices[0].Delta.Role)
+	assert.True(t, state.SentRole)
+
+	// response.output_text.delta → content chunk
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:  "response.output_text.delta",
+		Delta: "Hello",
+	}, state)
+	require.Len(t, chunks, 1)
+	require.NotNil(t, chunks[0].Choices[0].Delta.Content)
+	assert.Equal(t, "Hello", *chunks[0].Choices[0].Delta.Content)
+}
+
+func TestResponsesEventToChatChunks_ToolCallDelta(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.SentRole = true
+
+	// response.output_item.added (function_call) — output_index=1 (e.g. after a message item at 0)
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:        "response.output_item.added",
+		OutputIndex: 1,
+		Item: &ResponsesOutput{
+			Type:   "function_call",
+			CallID: "call_1",
+			Name:   "get_weather",
+		},
+	}, state)
+	require.Len(t, chunks, 1)
+	require.Len(t, chunks[0].Choices[0].Delta.ToolCalls, 1)
+	tc := chunks[0].Choices[0].Delta.ToolCalls[0]
+	assert.Equal(t, "call_1", tc.ID)
+	assert.Equal(t, "get_weather", tc.Function.Name)
+	require.NotNil(t, tc.Index)
+	assert.Equal(t, 0, *tc.Index)
+
+	// response.function_call_arguments.delta — uses output_index (NOT call_id) to find tool
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:        "response.function_call_arguments.delta",
+		OutputIndex: 1, // matches the output_index from output_item.added above
+		Delta:       `{"city":`,
+	}, state)
+	require.Len(t, chunks, 1)
+	tc = chunks[0].Choices[0].Delta.ToolCalls[0]
+	require.NotNil(t, tc.Index)
+	assert.Equal(t, 0, *tc.Index, "argument delta must use same index as the tool call")
+	assert.Equal(t, `{"city":`, tc.Function.Arguments)
+
+	// Add a second function call at output_index=2
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:        "response.output_item.added",
+		OutputIndex: 2,
+		Item: &ResponsesOutput{
+			Type:   "function_call",
+			CallID: "call_2",
+			Name:   "get_time",
+		},
+	}, state)
+	require.Len(t, chunks, 1)
+	tc = chunks[0].Choices[0].Delta.ToolCalls[0]
+	require.NotNil(t, tc.Index)
+	assert.Equal(t, 1, *tc.Index, "second tool call should get index 1")
+
+	// Argument delta for second tool call
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:        "response.function_call_arguments.delta",
+		OutputIndex: 2,
+		Delta:       `{"tz":"UTC"}`,
+	}, state)
+	require.Len(t, chunks, 1)
+	tc = chunks[0].Choices[0].Delta.ToolCalls[0]
+	require.NotNil(t, tc.Index)
+	assert.Equal(t, 1, *tc.Index, "second tool arg delta must use index 1")
+
+	// Argument delta for first tool call (interleaved)
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:        "response.function_call_arguments.delta",
+		OutputIndex: 1,
+		Delta:       `"Tokyo"}`,
+	}, state)
+	require.Len(t, chunks, 1)
+	tc = chunks[0].Choices[0].Delta.ToolCalls[0]
+	require.NotNil(t, tc.Index)
+	assert.Equal(t, 0, *tc.Index, "first tool arg delta must still use index 0")
+}
+
+func TestResponsesEventToChatChunks_Completed(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.IncludeUsage = true
+
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.completed",
+		Response: &ResponsesResponse{
+			Status: "completed",
+			Usage: &ResponsesUsage{
+				InputTokens:  50,
+				OutputTokens: 20,
+				TotalTokens:  70,
+				InputTokensDetails: &ResponsesInputTokensDetails{
+					CachedTokens: 30,
+				},
+			},
+		},
+	}, state)
+	// finish chunk + usage chunk
+	require.Len(t, chunks, 2)
+
+	// First chunk: finish_reason
+	require.NotNil(t, chunks[0].Choices[0].FinishReason)
+	assert.Equal(t, "stop", *chunks[0].Choices[0].FinishReason)
+
+	// Second chunk: usage
+	require.NotNil(t, chunks[1].Usage)
+	assert.Equal(t, 50, chunks[1].Usage.PromptTokens)
+	assert.Equal(t, 20, chunks[1].Usage.CompletionTokens)
+	assert.Equal(t, 70, chunks[1].Usage.TotalTokens)
+	require.NotNil(t, chunks[1].Usage.PromptTokensDetails)
+	assert.Equal(t, 30, chunks[1].Usage.PromptTokensDetails.CachedTokens)
+}
+
+func TestResponsesEventToChatChunks_CompletedWithToolCalls(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.SawToolCall = true
+
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.completed",
+		Response: &ResponsesResponse{
+			Status: "completed",
+		},
+	}, state)
+	require.Len(t, chunks, 1)
+	require.NotNil(t, chunks[0].Choices[0].FinishReason)
+	assert.Equal(t, "tool_calls", *chunks[0].Choices[0].FinishReason)
+}
+
+func TestResponsesEventToChatChunks_ReasoningDelta(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.SentRole = true
+
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:  "response.reasoning_summary_text.delta",
+		Delta: "Thinking...",
+	}, state)
+	require.Len(t, chunks, 1)
+	require.NotNil(t, chunks[0].Choices[0].Delta.ReasoningContent)
+	assert.Equal(t, "Thinking...", *chunks[0].Choices[0].Delta.ReasoningContent)
+
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.reasoning_summary_text.done",
+	}, state)
+	require.Len(t, chunks, 0)
+}
+
+func TestResponsesEventToChatChunks_ReasoningThenTextAutoCloseTag(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.SentRole = true
+
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:  "response.reasoning_summary_text.delta",
+		Delta: "plan",
+	}, state)
+	require.Len(t, chunks, 1)
+	require.NotNil(t, chunks[0].Choices[0].Delta.ReasoningContent)
+	assert.Equal(t, "plan", *chunks[0].Choices[0].Delta.ReasoningContent)
+
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:  "response.output_text.delta",
+		Delta: "answer",
+	}, state)
+	require.Len(t, chunks, 1)
+	require.NotNil(t, chunks[0].Choices[0].Delta.Content)
+	assert.Equal(t, "answer", *chunks[0].Choices[0].Delta.Content)
+}
+
+func TestFinalizeResponsesChatStream(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.IncludeUsage = true
+	state.Usage = &ChatUsage{
+		PromptTokens:     100,
+		CompletionTokens: 50,
+		TotalTokens:      150,
+	}
+
+	chunks := FinalizeResponsesChatStream(state)
+	require.Len(t, chunks, 2)
+
+	// Finish chunk
+	require.NotNil(t, chunks[0].Choices[0].FinishReason)
+	assert.Equal(t, "stop", *chunks[0].Choices[0].FinishReason)
+
+	// Usage chunk
+	require.NotNil(t, chunks[1].Usage)
+	assert.Equal(t, 100, chunks[1].Usage.PromptTokens)
+
+	// Idempotent: second call returns nil
+	assert.Nil(t, FinalizeResponsesChatStream(state))
+}
+
+func TestFinalizeResponsesChatStream_AfterCompleted(t *testing.T) {
+	// If response.completed already emitted the finish chunk, FinalizeResponsesChatStream
+	// must be a no-op (prevents double finish_reason being sent to the client).
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.IncludeUsage = true
+
+	// Simulate response.completed
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.completed",
+		Response: &ResponsesResponse{
+			Status: "completed",
+			Usage: &ResponsesUsage{
+				InputTokens:  10,
+				OutputTokens: 5,
+				TotalTokens:  15,
+			},
+		},
+	}, state)
+	require.NotEmpty(t, chunks) // finish + usage chunks
+
+	// Now FinalizeResponsesChatStream should return nil — already finalized.
+	assert.Nil(t, FinalizeResponsesChatStream(state))
+}
+
+func TestChatChunkToSSE(t *testing.T) {
+	chunk := ChatCompletionsChunk{
+		ID:      "chatcmpl-test",
+		Object:  "chat.completion.chunk",
+		Created: 1700000000,
+		Model:   "gpt-4o",
+		Choices: []ChatChunkChoice{
+			{
+				Index:        0,
+				Delta:        ChatDelta{Role: "assistant"},
+				FinishReason: nil,
+			},
+		},
+	}
+
+	sse, err := ChatChunkToSSE(chunk)
+	require.NoError(t, err)
+	assert.Contains(t, sse, "data: ")
+	assert.Contains(t, sse, "chatcmpl-test")
+	assert.Contains(t, sse, "assistant")
+	assert.True(t, len(sse) > 10)
+}
+
+// ---------------------------------------------------------------------------
+// Stream round-trip test
+// ---------------------------------------------------------------------------
+
+func TestChatCompletionsStreamRoundTrip(t *testing.T) {
+	// Simulate: client sends chat completions request, upstream returns Responses SSE events.
+	// Verify that the streaming state machine produces correct chat completions chunks.
+
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.IncludeUsage = true
+
+	var allChunks []ChatCompletionsChunk
+
+	// 1. response.created
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:     "response.created",
+		Response: &ResponsesResponse{ID: "resp_rt"},
+	}, state)
+	allChunks = append(allChunks, chunks...)
+
+	// 2. text deltas
+	for _, text := range []string{"Hello", ", ", "world", "!"} {
+		chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+			Type:  "response.output_text.delta",
+			Delta: text,
+		}, state)
+		allChunks = append(allChunks, chunks...)
+	}
+
+	// 3. response.completed
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.completed",
+		Response: &ResponsesResponse{
+			Status: "completed",
+			Usage: &ResponsesUsage{
+				InputTokens:  10,
+				OutputTokens: 4,
+				TotalTokens:  14,
+			},
+		},
+	}, state)
+	allChunks = append(allChunks, chunks...)
+
+	// Verify: role chunk + 4 text chunks + finish chunk + usage chunk = 7
+	require.Len(t, allChunks, 7)
+
+	// First chunk has role
+	assert.Equal(t, "assistant", allChunks[0].Choices[0].Delta.Role)
+
+	// Text chunks
+	var fullText string
+	for i := 1; i <= 4; i++ {
+		require.NotNil(t, allChunks[i].Choices[0].Delta.Content)
+		fullText += *allChunks[i].Choices[0].Delta.Content
+	}
+	assert.Equal(t, "Hello, world!", fullText)
+
+	// Finish chunk
+	require.NotNil(t, allChunks[5].Choices[0].FinishReason)
+	assert.Equal(t, "stop", *allChunks[5].Choices[0].FinishReason)
+
+	// Usage chunk
+	require.NotNil(t, allChunks[6].Usage)
+	assert.Equal(t, 10, allChunks[6].Usage.PromptTokens)
+	assert.Equal(t, 4, allChunks[6].Usage.CompletionTokens)
+
+	// All chunks share the same ID
+	for _, c := range allChunks {
+		assert.Equal(t, "resp_rt", c.ID)
+	}
+}

internal/pkg/apicompat/chatcompletions_to_responses.go

@@ -0,0 +1,425 @@
+package apicompat
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+)
+
+type chatMessageContent struct {
+	Text  *string
+	Parts []ChatContentPart
+}
+
+// ChatCompletionsToResponses converts a Chat Completions request into a
+// Responses API request. The upstream always streams, so Stream is forced to
+// true. store is always false and reasoning.encrypted_content is always
+// included so that the response translator has full context.
+func ChatCompletionsToResponses(req *ChatCompletionsRequest) (*ResponsesRequest, error) {
+	input, err := convertChatMessagesToResponsesInput(req.Messages)
+	if err != nil {
+		return nil, err
+	}
+
+	inputJSON, err := json.Marshal(input)
+	if err != nil {
+		return nil, err
+	}
+
+	out := &ResponsesRequest{
+		Model:       req.Model,
+		Input:       inputJSON,
+		Temperature: req.Temperature,
+		TopP:        req.TopP,
+		Stream:      true, // upstream always streams
+		Include:     []string{"reasoning.encrypted_content"},
+		ServiceTier: req.ServiceTier,
+	}
+
+	storeFalse := false
+	out.Store = &storeFalse
+
+	// max_tokens / max_completion_tokens → max_output_tokens, prefer max_completion_tokens
+	maxTokens := 0
+	if req.MaxTokens != nil {
+		maxTokens = *req.MaxTokens
+	}
+	if req.MaxCompletionTokens != nil {
+		maxTokens = *req.MaxCompletionTokens
+	}
+	if maxTokens > 0 {
+		v := maxTokens
+		if v < minMaxOutputTokens {
+			v = minMaxOutputTokens
+		}
+		out.MaxOutputTokens = &v
+	}
+
+	// reasoning_effort → reasoning.effort + reasoning.summary="auto"
+	if req.ReasoningEffort != "" {
+		out.Reasoning = &ResponsesReasoning{
+			Effort:  req.ReasoningEffort,
+			Summary: "auto",
+		}
+	}
+
+	// tools[] and legacy functions[] → ResponsesTool[]
+	if len(req.Tools) > 0 || len(req.Functions) > 0 {
+		out.Tools = convertChatToolsToResponses(req.Tools, req.Functions)
+	}
+
+	// tool_choice: already compatible format — pass through directly.
+	// Legacy function_call needs mapping.
+	if len(req.ToolChoice) > 0 {
+		out.ToolChoice = req.ToolChoice
+	} else if len(req.FunctionCall) > 0 {
+		tc, err := convertChatFunctionCallToToolChoice(req.FunctionCall)
+		if err != nil {
+			return nil, fmt.Errorf("convert function_call: %w", err)
+		}
+		out.ToolChoice = tc
+	}
+
+	return out, nil
+}
+
+// convertChatMessagesToResponsesInput converts the Chat Completions messages
+// array into a Responses API input items array.
+func convertChatMessagesToResponsesInput(msgs []ChatMessage) ([]ResponsesInputItem, error) {
+	var out []ResponsesInputItem
+	for _, m := range msgs {
+		items, err := chatMessageToResponsesItems(m)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, items...)
+	}
+	return out, nil
+}
+
+// chatMessageToResponsesItems converts a single ChatMessage into one or more
+// ResponsesInputItem values.
+func chatMessageToResponsesItems(m ChatMessage) ([]ResponsesInputItem, error) {
+	switch m.Role {
+	case "system":
+		return chatSystemToResponses(m)
+	case "user":
+		return chatUserToResponses(m)
+	case "assistant":
+		return chatAssistantToResponses(m)
+	case "tool":
+		return chatToolToResponses(m)
+	case "function":
+		return chatFunctionToResponses(m)
+	default:
+		return chatUserToResponses(m)
+	}
+}
+
+// chatSystemToResponses converts a system message.
+func chatSystemToResponses(m ChatMessage) ([]ResponsesInputItem, error) {
+	parsed, err := parseChatMessageContent(m.Content)
+	if err != nil {
+		return nil, err
+	}
+	content, err := marshalChatInputContent(parsed)
+	if err != nil {
+		return nil, err
+	}
+	return []ResponsesInputItem{{Role: "system", Content: content}}, nil
+}
+
+// chatUserToResponses converts a user message, handling both plain strings and
+// multi-modal content arrays.
+func chatUserToResponses(m ChatMessage) ([]ResponsesInputItem, error) {
+	parsed, err := parseChatMessageContent(m.Content)
+	if err != nil {
+		return nil, fmt.Errorf("parse user content: %w", err)
+	}
+	content, err := marshalChatInputContent(parsed)
+	if err != nil {
+		return nil, err
+	}
+	return []ResponsesInputItem{{Role: "user", Content: content}}, nil
+}
+
+// chatAssistantToResponses converts an assistant message. If there is both
+// text content and tool_calls, the text is emitted as an assistant message
+// first, then each tool_call becomes a function_call item. If the content is
+// empty/nil and there are tool_calls, only function_call items are emitted.
+func chatAssistantToResponses(m ChatMessage) ([]ResponsesInputItem, error) {
+	var items []ResponsesInputItem
+
+	// Emit assistant message with output_text if content is non-empty.
+	if len(m.Content) > 0 {
+		s, err := parseAssistantContent(m.Content)
+		if err != nil {
+			return nil, err
+		}
+		if s != "" {
+			parts := []ResponsesContentPart{{Type: "output_text", Text: s}}
+			partsJSON, err := json.Marshal(parts)
+			if err != nil {
+				return nil, err
+			}
+			items = append(items, ResponsesInputItem{Role: "assistant", Content: partsJSON})
+		}
+	}
+
+	// Emit one function_call item per tool_call.
+	for _, tc := range m.ToolCalls {
+		args := tc.Function.Arguments
+		if args == "" {
+			args = "{}"
+		}
+		items = append(items, ResponsesInputItem{
+			Type:      "function_call",
+			CallID:    tc.ID,
+			Name:      tc.Function.Name,
+			Arguments: args,
+		})
+	}
+
+	return items, nil
+}
+
+// parseAssistantContent returns assistant content as plain text.
+//
+// Supported formats:
+// - JSON string
+// - JSON array of typed parts (e.g. [{"type":"text","text":"..."}])
+//
+// For structured thinking/reasoning parts, it preserves semantics by wrapping
+// the text in explicit tags so downstream can still distinguish it from normal text.
+func parseAssistantContent(raw json.RawMessage) (string, error) {
+	if len(raw) == 0 {
+		return "", nil
+	}
+
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return s, nil
+	}
+
+	var parts []map[string]any
+	if err := json.Unmarshal(raw, &parts); err != nil {
+		// Keep compatibility with prior behavior: unsupported assistant content
+		// formats are ignored instead of failing the whole request conversion.
+		return "", nil
+	}
+
+	var b strings.Builder
+	write := func(v string) error {
+		_, err := b.WriteString(v)
+		return err
+	}
+	for _, p := range parts {
+		typ, _ := p["type"].(string)
+		text, _ := p["text"].(string)
+		thinking, _ := p["thinking"].(string)
+
+		switch typ {
+		case "thinking", "reasoning":
+			if thinking != "" {
+				if err := write("<thinking>"); err != nil {
+					return "", err
+				}
+				if err := write(thinking); err != nil {
+					return "", err
+				}
+				if err := write("</thinking>"); err != nil {
+					return "", err
+				}
+			} else if text != "" {
+				if err := write("<thinking>"); err != nil {
+					return "", err
+				}
+				if err := write(text); err != nil {
+					return "", err
+				}
+				if err := write("</thinking>"); err != nil {
+					return "", err
+				}
+			}
+		default:
+			if text != "" {
+				if err := write(text); err != nil {
+					return "", err
+				}
+			}
+		}
+	}
+
+	return b.String(), nil
+}
+
+// chatToolToResponses converts a tool result message (role=tool) into a
+// function_call_output item.
+func chatToolToResponses(m ChatMessage) ([]ResponsesInputItem, error) {
+	output, err := parseChatContent(m.Content)
+	if err != nil {
+		return nil, err
+	}
+	if output == "" {
+		output = "(empty)"
+	}
+	return []ResponsesInputItem{{
+		Type:   "function_call_output",
+		CallID: m.ToolCallID,
+		Output: output,
+	}}, nil
+}
+
+// chatFunctionToResponses converts a legacy function result message
+// (role=function) into a function_call_output item. The Name field is used as
+// call_id since legacy function calls do not carry a separate call_id.
+func chatFunctionToResponses(m ChatMessage) ([]ResponsesInputItem, error) {
+	output, err := parseChatContent(m.Content)
+	if err != nil {
+		return nil, err
+	}
+	if output == "" {
+		output = "(empty)"
+	}
+	return []ResponsesInputItem{{
+		Type:   "function_call_output",
+		CallID: m.Name,
+		Output: output,
+	}}, nil
+}
+
+// parseChatContent returns the string value of a ChatMessage Content field.
+// Content can be a JSON string or an array of typed parts. Array content is
+// flattened to text by concatenating text parts and ignoring non-text parts.
+func parseChatContent(raw json.RawMessage) (string, error) {
+	parsed, err := parseChatMessageContent(raw)
+	if err != nil {
+		return "", err
+	}
+	if parsed.Text != nil {
+		return *parsed.Text, nil
+	}
+	return flattenChatContentParts(parsed.Parts), nil
+}
+
+func parseChatMessageContent(raw json.RawMessage) (chatMessageContent, error) {
+	if len(raw) == 0 {
+		return chatMessageContent{Text: stringPtr("")}, nil
+	}
+
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return chatMessageContent{Text: &s}, nil
+	}
+
+	var parts []ChatContentPart
+	if err := json.Unmarshal(raw, &parts); err == nil {
+		return chatMessageContent{Parts: parts}, nil
+	}
+
+	return chatMessageContent{}, fmt.Errorf("parse content as string or parts array")
+}
+
+func marshalChatInputContent(content chatMessageContent) (json.RawMessage, error) {
+	if content.Text != nil {
+		return json.Marshal(*content.Text)
+	}
+	return json.Marshal(convertChatContentPartsToResponses(content.Parts))
+}
+
+func convertChatContentPartsToResponses(parts []ChatContentPart) []ResponsesContentPart {
+	var responseParts []ResponsesContentPart
+	for _, p := range parts {
+		switch p.Type {
+		case "text":
+			if p.Text != "" {
+				responseParts = append(responseParts, ResponsesContentPart{
+					Type: "input_text",
+					Text: p.Text,
+				})
+			}
+		case "image_url":
+			if p.ImageURL != nil && p.ImageURL.URL != "" {
+				responseParts = append(responseParts, ResponsesContentPart{
+					Type:     "input_image",
+					ImageURL: p.ImageURL.URL,
+				})
+			}
+		}
+	}
+	return responseParts
+}
+
+func flattenChatContentParts(parts []ChatContentPart) string {
+	var textParts []string
+	for _, p := range parts {
+		if p.Type == "text" && p.Text != "" {
+			textParts = append(textParts, p.Text)
+		}
+	}
+	return strings.Join(textParts, "")
+}
+
+func stringPtr(s string) *string {
+	return &s
+}
+
+// convertChatToolsToResponses maps Chat Completions tool definitions and legacy
+// function definitions to Responses API tool definitions.
+func convertChatToolsToResponses(tools []ChatTool, functions []ChatFunction) []ResponsesTool {
+	var out []ResponsesTool
+
+	for _, t := range tools {
+		if t.Type != "function" || t.Function == nil {
+			continue
+		}
+		rt := ResponsesTool{
+			Type:        "function",
+			Name:        t.Function.Name,
+			Description: t.Function.Description,
+			Parameters:  t.Function.Parameters,
+			Strict:      t.Function.Strict,
+		}
+		out = append(out, rt)
+	}
+
+	// Legacy functions[] are treated as function-type tools.
+	for _, f := range functions {
+		rt := ResponsesTool{
+			Type:        "function",
+			Name:        f.Name,
+			Description: f.Description,
+			Parameters:  f.Parameters,
+			Strict:      f.Strict,
+		}
+		out = append(out, rt)
+	}
+
+	return out
+}
+
+// convertChatFunctionCallToToolChoice maps the legacy function_call field to a
+// Responses API tool_choice value.
+//
+//	"auto" → "auto"
+//	"none" → "none"
+//	{"name":"X"} → {"type":"function","function":{"name":"X"}}
+func convertChatFunctionCallToToolChoice(raw json.RawMessage) (json.RawMessage, error) {
+	// Try string first ("auto", "none", etc.) — pass through as-is.
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return json.Marshal(s)
+	}
+
+	// Object form: {"name":"X"}
+	var obj struct {
+		Name string `json:"name"`
+	}
+	if err := json.Unmarshal(raw, &obj); err != nil {
+		return nil, err
+	}
+	return json.Marshal(map[string]any{
+		"type":     "function",
+		"function": map[string]string{"name": obj.Name},
+	})
+}

internal/pkg/apicompat/responses_to_anthropic.go

@@ -0,0 +1,516 @@
+package apicompat
+
+import (
+	"encoding/json"
+	"fmt"
+	"time"
+)
+
+// ---------------------------------------------------------------------------
+// Non-streaming: ResponsesResponse → AnthropicResponse
+// ---------------------------------------------------------------------------
+
+// ResponsesToAnthropic converts a Responses API response directly into an
+// Anthropic Messages response. Reasoning output items are mapped to thinking
+// blocks; function_call items become tool_use blocks.
+func ResponsesToAnthropic(resp *ResponsesResponse, model string) *AnthropicResponse {
+	out := &AnthropicResponse{
+		ID:    resp.ID,
+		Type:  "message",
+		Role:  "assistant",
+		Model: model,
+	}
+
+	var blocks []AnthropicContentBlock
+
+	for _, item := range resp.Output {
+		switch item.Type {
+		case "reasoning":
+			summaryText := ""
+			for _, s := range item.Summary {
+				if s.Type == "summary_text" && s.Text != "" {
+					summaryText += s.Text
+				}
+			}
+			if summaryText != "" {
+				blocks = append(blocks, AnthropicContentBlock{
+					Type:     "thinking",
+					Thinking: summaryText,
+				})
+			}
+		case "message":
+			for _, part := range item.Content {
+				if part.Type == "output_text" && part.Text != "" {
+					blocks = append(blocks, AnthropicContentBlock{
+						Type: "text",
+						Text: part.Text,
+					})
+				}
+			}
+		case "function_call":
+			blocks = append(blocks, AnthropicContentBlock{
+				Type:  "tool_use",
+				ID:    fromResponsesCallID(item.CallID),
+				Name:  item.Name,
+				Input: json.RawMessage(item.Arguments),
+			})
+		case "web_search_call":
+			toolUseID := "srvtoolu_" + item.ID
+			query := ""
+			if item.Action != nil {
+				query = item.Action.Query
+			}
+			inputJSON, _ := json.Marshal(map[string]string{"query": query})
+			blocks = append(blocks, AnthropicContentBlock{
+				Type:  "server_tool_use",
+				ID:    toolUseID,
+				Name:  "web_search",
+				Input: inputJSON,
+			})
+			emptyResults, _ := json.Marshal([]struct{}{})
+			blocks = append(blocks, AnthropicContentBlock{
+				Type:      "web_search_tool_result",
+				ToolUseID: toolUseID,
+				Content:   emptyResults,
+			})
+		}
+	}
+
+	if len(blocks) == 0 {
+		blocks = append(blocks, AnthropicContentBlock{Type: "text", Text: ""})
+	}
+	out.Content = blocks
+
+	out.StopReason = responsesStatusToAnthropicStopReason(resp.Status, resp.IncompleteDetails, blocks)
+
+	if resp.Usage != nil {
+		out.Usage = AnthropicUsage{
+			InputTokens:  resp.Usage.InputTokens,
+			OutputTokens: resp.Usage.OutputTokens,
+		}
+		if resp.Usage.InputTokensDetails != nil {
+			out.Usage.CacheReadInputTokens = resp.Usage.InputTokensDetails.CachedTokens
+		}
+	}
+
+	return out
+}
+
+func responsesStatusToAnthropicStopReason(status string, details *ResponsesIncompleteDetails, blocks []AnthropicContentBlock) string {
+	switch status {
+	case "incomplete":
+		if details != nil && details.Reason == "max_output_tokens" {
+			return "max_tokens"
+		}
+		return "end_turn"
+	case "completed":
+		if len(blocks) > 0 && blocks[len(blocks)-1].Type == "tool_use" {
+			return "tool_use"
+		}
+		return "end_turn"
+	default:
+		return "end_turn"
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Streaming: ResponsesStreamEvent → []AnthropicStreamEvent (stateful converter)
+// ---------------------------------------------------------------------------
+
+// ResponsesEventToAnthropicState tracks state for converting a sequence of
+// Responses SSE events directly into Anthropic SSE events.
+type ResponsesEventToAnthropicState struct {
+	MessageStartSent bool
+	MessageStopSent  bool
+
+	ContentBlockIndex int
+	ContentBlockOpen  bool
+	CurrentBlockType  string // "text" | "thinking" | "tool_use"
+
+	// OutputIndexToBlockIdx maps Responses output_index → Anthropic content block index.
+	OutputIndexToBlockIdx map[int]int
+
+	InputTokens          int
+	OutputTokens         int
+	CacheReadInputTokens int
+
+	ResponseID string
+	Model      string
+	Created    int64
+}
+
+// NewResponsesEventToAnthropicState returns an initialised stream state.
+func NewResponsesEventToAnthropicState() *ResponsesEventToAnthropicState {
+	return &ResponsesEventToAnthropicState{
+		OutputIndexToBlockIdx: make(map[int]int),
+		Created:               time.Now().Unix(),
+	}
+}
+
+// ResponsesEventToAnthropicEvents converts a single Responses SSE event into
+// zero or more Anthropic SSE events, updating state as it goes.
+func ResponsesEventToAnthropicEvents(
+	evt *ResponsesStreamEvent,
+	state *ResponsesEventToAnthropicState,
+) []AnthropicStreamEvent {
+	switch evt.Type {
+	case "response.created":
+		return resToAnthHandleCreated(evt, state)
+	case "response.output_item.added":
+		return resToAnthHandleOutputItemAdded(evt, state)
+	case "response.output_text.delta":
+		return resToAnthHandleTextDelta(evt, state)
+	case "response.output_text.done":
+		return resToAnthHandleBlockDone(state)
+	case "response.function_call_arguments.delta":
+		return resToAnthHandleFuncArgsDelta(evt, state)
+	case "response.function_call_arguments.done":
+		return resToAnthHandleBlockDone(state)
+	case "response.output_item.done":
+		return resToAnthHandleOutputItemDone(evt, state)
+	case "response.reasoning_summary_text.delta":
+		return resToAnthHandleReasoningDelta(evt, state)
+	case "response.reasoning_summary_text.done":
+		return resToAnthHandleBlockDone(state)
+	case "response.completed", "response.incomplete", "response.failed":
+		return resToAnthHandleCompleted(evt, state)
+	default:
+		return nil
+	}
+}
+
+// FinalizeResponsesAnthropicStream emits synthetic termination events if the
+// stream ended without a proper completion event.
+func FinalizeResponsesAnthropicStream(state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if !state.MessageStartSent || state.MessageStopSent {
+		return nil
+	}
+
+	var events []AnthropicStreamEvent
+	events = append(events, closeCurrentBlock(state)...)
+
+	events = append(events,
+		AnthropicStreamEvent{
+			Type: "message_delta",
+			Delta: &AnthropicDelta{
+				StopReason: "end_turn",
+			},
+			Usage: &AnthropicUsage{
+				InputTokens:          state.InputTokens,
+				OutputTokens:         state.OutputTokens,
+				CacheReadInputTokens: state.CacheReadInputTokens,
+			},
+		},
+		AnthropicStreamEvent{Type: "message_stop"},
+	)
+	state.MessageStopSent = true
+	return events
+}
+
+// ResponsesAnthropicEventToSSE formats an AnthropicStreamEvent as an SSE line pair.
+func ResponsesAnthropicEventToSSE(evt AnthropicStreamEvent) (string, error) {
+	data, err := json.Marshal(evt)
+	if err != nil {
+		return "", err
+	}
+	return fmt.Sprintf("event: %s\ndata: %s\n\n", evt.Type, data), nil
+}
+
+// --- internal handlers ---
+
+func resToAnthHandleCreated(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if evt.Response != nil {
+		state.ResponseID = evt.Response.ID
+		// Only use upstream model if no override was set (e.g. originalModel)
+		if state.Model == "" {
+			state.Model = evt.Response.Model
+		}
+	}
+
+	if state.MessageStartSent {
+		return nil
+	}
+	state.MessageStartSent = true
+
+	return []AnthropicStreamEvent{{
+		Type: "message_start",
+		Message: &AnthropicResponse{
+			ID:      state.ResponseID,
+			Type:    "message",
+			Role:    "assistant",
+			Content: []AnthropicContentBlock{},
+			Model:   state.Model,
+			Usage: AnthropicUsage{
+				InputTokens:  0,
+				OutputTokens: 0,
+			},
+		},
+	}}
+}
+
+func resToAnthHandleOutputItemAdded(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if evt.Item == nil {
+		return nil
+	}
+
+	switch evt.Item.Type {
+	case "function_call":
+		var events []AnthropicStreamEvent
+		events = append(events, closeCurrentBlock(state)...)
+
+		idx := state.ContentBlockIndex
+		state.OutputIndexToBlockIdx[evt.OutputIndex] = idx
+		state.ContentBlockOpen = true
+		state.CurrentBlockType = "tool_use"
+
+		events = append(events, AnthropicStreamEvent{
+			Type:  "content_block_start",
+			Index: &idx,
+			ContentBlock: &AnthropicContentBlock{
+				Type:  "tool_use",
+				ID:    fromResponsesCallID(evt.Item.CallID),
+				Name:  evt.Item.Name,
+				Input: json.RawMessage("{}"),
+			},
+		})
+		return events
+
+	case "reasoning":
+		var events []AnthropicStreamEvent
+		events = append(events, closeCurrentBlock(state)...)
+
+		idx := state.ContentBlockIndex
+		state.OutputIndexToBlockIdx[evt.OutputIndex] = idx
+		state.ContentBlockOpen = true
+		state.CurrentBlockType = "thinking"
+
+		events = append(events, AnthropicStreamEvent{
+			Type:  "content_block_start",
+			Index: &idx,
+			ContentBlock: &AnthropicContentBlock{
+				Type:     "thinking",
+				Thinking: "",
+			},
+		})
+		return events
+
+	case "message":
+		return nil
+	}
+
+	return nil
+}
+
+func resToAnthHandleTextDelta(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if evt.Delta == "" {
+		return nil
+	}
+
+	var events []AnthropicStreamEvent
+
+	if !state.ContentBlockOpen || state.CurrentBlockType != "text" {
+		events = append(events, closeCurrentBlock(state)...)
+
+		idx := state.ContentBlockIndex
+		state.ContentBlockOpen = true
+		state.CurrentBlockType = "text"
+
+		events = append(events, AnthropicStreamEvent{
+			Type:  "content_block_start",
+			Index: &idx,
+			ContentBlock: &AnthropicContentBlock{
+				Type: "text",
+				Text: "",
+			},
+		})
+	}
+
+	idx := state.ContentBlockIndex
+	events = append(events, AnthropicStreamEvent{
+		Type:  "content_block_delta",
+		Index: &idx,
+		Delta: &AnthropicDelta{
+			Type: "text_delta",
+			Text: evt.Delta,
+		},
+	})
+	return events
+}
+
+func resToAnthHandleFuncArgsDelta(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if evt.Delta == "" {
+		return nil
+	}
+
+	blockIdx, ok := state.OutputIndexToBlockIdx[evt.OutputIndex]
+	if !ok {
+		return nil
+	}
+
+	return []AnthropicStreamEvent{{
+		Type:  "content_block_delta",
+		Index: &blockIdx,
+		Delta: &AnthropicDelta{
+			Type:        "input_json_delta",
+			PartialJSON: evt.Delta,
+		},
+	}}
+}
+
+func resToAnthHandleReasoningDelta(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if evt.Delta == "" {
+		return nil
+	}
+
+	blockIdx, ok := state.OutputIndexToBlockIdx[evt.OutputIndex]
+	if !ok {
+		return nil
+	}
+
+	return []AnthropicStreamEvent{{
+		Type:  "content_block_delta",
+		Index: &blockIdx,
+		Delta: &AnthropicDelta{
+			Type:     "thinking_delta",
+			Thinking: evt.Delta,
+		},
+	}}
+}
+
+func resToAnthHandleBlockDone(state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if !state.ContentBlockOpen {
+		return nil
+	}
+	return closeCurrentBlock(state)
+}
+
+func resToAnthHandleOutputItemDone(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if evt.Item == nil {
+		return nil
+	}
+
+	// Handle web_search_call → synthesize server_tool_use + web_search_tool_result blocks.
+	if evt.Item.Type == "web_search_call" && evt.Item.Status == "completed" {
+		return resToAnthHandleWebSearchDone(evt, state)
+	}
+
+	if state.ContentBlockOpen {
+		return closeCurrentBlock(state)
+	}
+	return nil
+}
+
+// resToAnthHandleWebSearchDone converts an OpenAI web_search_call output item
+// into Anthropic server_tool_use + web_search_tool_result content block pairs.
+// This allows Claude Code to count the searches performed.
+func resToAnthHandleWebSearchDone(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	var events []AnthropicStreamEvent
+	events = append(events, closeCurrentBlock(state)...)
+
+	toolUseID := "srvtoolu_" + evt.Item.ID
+	query := ""
+	if evt.Item.Action != nil {
+		query = evt.Item.Action.Query
+	}
+	inputJSON, _ := json.Marshal(map[string]string{"query": query})
+
+	// Emit server_tool_use block (start + stop).
+	idx1 := state.ContentBlockIndex
+	events = append(events, AnthropicStreamEvent{
+		Type:  "content_block_start",
+		Index: &idx1,
+		ContentBlock: &AnthropicContentBlock{
+			Type:  "server_tool_use",
+			ID:    toolUseID,
+			Name:  "web_search",
+			Input: inputJSON,
+		},
+	})
+	events = append(events, AnthropicStreamEvent{
+		Type:  "content_block_stop",
+		Index: &idx1,
+	})
+	state.ContentBlockIndex++
+
+	// Emit web_search_tool_result block (start + stop).
+	// Content is empty because OpenAI does not expose individual search results;
+	// the model consumes them internally and produces text output.
+	emptyResults, _ := json.Marshal([]struct{}{})
+	idx2 := state.ContentBlockIndex
+	events = append(events, AnthropicStreamEvent{
+		Type:  "content_block_start",
+		Index: &idx2,
+		ContentBlock: &AnthropicContentBlock{
+			Type:      "web_search_tool_result",
+			ToolUseID: toolUseID,
+			Content:   emptyResults,
+		},
+	})
+	events = append(events, AnthropicStreamEvent{
+		Type:  "content_block_stop",
+		Index: &idx2,
+	})
+	state.ContentBlockIndex++
+
+	return events
+}
+
+func resToAnthHandleCompleted(evt *ResponsesStreamEvent, state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if state.MessageStopSent {
+		return nil
+	}
+
+	var events []AnthropicStreamEvent
+	events = append(events, closeCurrentBlock(state)...)
+
+	stopReason := "end_turn"
+	if evt.Response != nil {
+		if evt.Response.Usage != nil {
+			state.InputTokens = evt.Response.Usage.InputTokens
+			state.OutputTokens = evt.Response.Usage.OutputTokens
+			if evt.Response.Usage.InputTokensDetails != nil {
+				state.CacheReadInputTokens = evt.Response.Usage.InputTokensDetails.CachedTokens
+			}
+		}
+		switch evt.Response.Status {
+		case "incomplete":
+			if evt.Response.IncompleteDetails != nil && evt.Response.IncompleteDetails.Reason == "max_output_tokens" {
+				stopReason = "max_tokens"
+			}
+		case "completed":
+			if state.ContentBlockIndex > 0 && state.CurrentBlockType == "tool_use" {
+				stopReason = "tool_use"
+			}
+		}
+	}
+
+	events = append(events,
+		AnthropicStreamEvent{
+			Type: "message_delta",
+			Delta: &AnthropicDelta{
+				StopReason: stopReason,
+			},
+			Usage: &AnthropicUsage{
+				InputTokens:          state.InputTokens,
+				OutputTokens:         state.OutputTokens,
+				CacheReadInputTokens: state.CacheReadInputTokens,
+			},
+		},
+		AnthropicStreamEvent{Type: "message_stop"},
+	)
+	state.MessageStopSent = true
+	return events
+}
+
+func closeCurrentBlock(state *ResponsesEventToAnthropicState) []AnthropicStreamEvent {
+	if !state.ContentBlockOpen {
+		return nil
+	}
+	idx := state.ContentBlockIndex
+	state.ContentBlockOpen = false
+	state.ContentBlockIndex++
+	return []AnthropicStreamEvent{{
+		Type:  "content_block_stop",
+		Index: &idx,
+	}}
+}

internal/pkg/apicompat/responses_to_anthropic_request.go

@@ -0,0 +1,464 @@
+package apicompat
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+)
+
+// ResponsesToAnthropicRequest converts a Responses API request into an
+// Anthropic Messages request. This is the reverse of AnthropicToResponses and
+// enables Anthropic platform groups to accept OpenAI Responses API requests
+// by converting them to the native /v1/messages format before forwarding upstream.
+func ResponsesToAnthropicRequest(req *ResponsesRequest) (*AnthropicRequest, error) {
+	system, messages, err := convertResponsesInputToAnthropic(req.Input)
+	if err != nil {
+		return nil, err
+	}
+
+	out := &AnthropicRequest{
+		Model:       req.Model,
+		Messages:    messages,
+		Temperature: req.Temperature,
+		TopP:        req.TopP,
+		Stream:      req.Stream,
+	}
+
+	if len(system) > 0 {
+		out.System = system
+	}
+
+	// max_output_tokens → max_tokens
+	if req.MaxOutputTokens != nil && *req.MaxOutputTokens > 0 {
+		out.MaxTokens = *req.MaxOutputTokens
+	}
+	if out.MaxTokens == 0 {
+		// Anthropic requires max_tokens; default to a sensible value.
+		out.MaxTokens = 8192
+	}
+
+	// Convert tools
+	if len(req.Tools) > 0 {
+		out.Tools = convertResponsesToAnthropicTools(req.Tools)
+	}
+
+	// Convert tool_choice (reverse of convertAnthropicToolChoiceToResponses)
+	if len(req.ToolChoice) > 0 {
+		tc, err := convertResponsesToAnthropicToolChoice(req.ToolChoice)
+		if err != nil {
+			return nil, fmt.Errorf("convert tool_choice: %w", err)
+		}
+		out.ToolChoice = tc
+	}
+
+	// reasoning.effort → output_config.effort + thinking
+	if req.Reasoning != nil && req.Reasoning.Effort != "" {
+		effort := mapResponsesEffortToAnthropic(req.Reasoning.Effort)
+		out.OutputConfig = &AnthropicOutputConfig{Effort: effort}
+		// Enable thinking for non-low efforts
+		if effort != "low" {
+			out.Thinking = &AnthropicThinking{
+				Type:         "enabled",
+				BudgetTokens: defaultThinkingBudget(effort),
+			}
+		}
+	}
+
+	return out, nil
+}
+
+// defaultThinkingBudget returns a sensible thinking budget based on effort level.
+func defaultThinkingBudget(effort string) int {
+	switch effort {
+	case "low":
+		return 1024
+	case "medium":
+		return 4096
+	case "high":
+		return 10240
+	case "max":
+		return 32768
+	default:
+		return 10240
+	}
+}
+
+// mapResponsesEffortToAnthropic converts OpenAI Responses reasoning effort to
+// Anthropic effort levels. Reverse of mapAnthropicEffortToResponses.
+//
+//	low    → low
+//	medium → medium
+//	high   → high
+//	xhigh  → max
+func mapResponsesEffortToAnthropic(effort string) string {
+	if effort == "xhigh" {
+		return "max"
+	}
+	return effort // low→low, medium→medium, high→high, unknown→passthrough
+}
+
+// convertResponsesInputToAnthropic extracts system prompt and messages from
+// a Responses API input array. Returns the system as raw JSON (for Anthropic's
+// polymorphic system field) and a list of Anthropic messages.
+func convertResponsesInputToAnthropic(inputRaw json.RawMessage) (json.RawMessage, []AnthropicMessage, error) {
+	// Try as plain string input.
+	var inputStr string
+	if err := json.Unmarshal(inputRaw, &inputStr); err == nil {
+		content, _ := json.Marshal(inputStr)
+		return nil, []AnthropicMessage{{Role: "user", Content: content}}, nil
+	}
+
+	var items []ResponsesInputItem
+	if err := json.Unmarshal(inputRaw, &items); err != nil {
+		return nil, nil, fmt.Errorf("parse responses input: %w", err)
+	}
+
+	var system json.RawMessage
+	var messages []AnthropicMessage
+
+	for _, item := range items {
+		switch {
+		case item.Role == "system":
+			// System prompt → Anthropic system field
+			text := extractTextFromContent(item.Content)
+			if text != "" {
+				system, _ = json.Marshal(text)
+			}
+
+		case item.Type == "function_call":
+			// function_call → assistant message with tool_use block
+			input := json.RawMessage("{}")
+			if item.Arguments != "" {
+				input = json.RawMessage(item.Arguments)
+			}
+			block := AnthropicContentBlock{
+				Type:  "tool_use",
+				ID:    fromResponsesCallIDToAnthropic(item.CallID),
+				Name:  item.Name,
+				Input: input,
+			}
+			blockJSON, _ := json.Marshal([]AnthropicContentBlock{block})
+			messages = append(messages, AnthropicMessage{
+				Role:    "assistant",
+				Content: blockJSON,
+			})
+
+		case item.Type == "function_call_output":
+			// function_call_output → user message with tool_result block
+			outputContent := item.Output
+			if outputContent == "" {
+				outputContent = "(empty)"
+			}
+			contentJSON, _ := json.Marshal(outputContent)
+			block := AnthropicContentBlock{
+				Type:      "tool_result",
+				ToolUseID: fromResponsesCallIDToAnthropic(item.CallID),
+				Content:   contentJSON,
+			}
+			blockJSON, _ := json.Marshal([]AnthropicContentBlock{block})
+			messages = append(messages, AnthropicMessage{
+				Role:    "user",
+				Content: blockJSON,
+			})
+
+		case item.Role == "user":
+			content, err := convertResponsesUserToAnthropicContent(item.Content)
+			if err != nil {
+				return nil, nil, err
+			}
+			messages = append(messages, AnthropicMessage{
+				Role:    "user",
+				Content: content,
+			})
+
+		case item.Role == "assistant":
+			content, err := convertResponsesAssistantToAnthropicContent(item.Content)
+			if err != nil {
+				return nil, nil, err
+			}
+			messages = append(messages, AnthropicMessage{
+				Role:    "assistant",
+				Content: content,
+			})
+
+		default:
+			// Unknown role/type — attempt as user message
+			if item.Content != nil {
+				messages = append(messages, AnthropicMessage{
+					Role:    "user",
+					Content: item.Content,
+				})
+			}
+		}
+	}
+
+	// Merge consecutive same-role messages (Anthropic requires alternating roles)
+	messages = mergeConsecutiveMessages(messages)
+
+	return system, messages, nil
+}
+
+// extractTextFromContent extracts text from a content field that may be a
+// plain string or an array of content parts.
+func extractTextFromContent(raw json.RawMessage) string {
+	if len(raw) == 0 {
+		return ""
+	}
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return s
+	}
+	var parts []ResponsesContentPart
+	if err := json.Unmarshal(raw, &parts); err == nil {
+		var texts []string
+		for _, p := range parts {
+			if (p.Type == "input_text" || p.Type == "output_text" || p.Type == "text") && p.Text != "" {
+				texts = append(texts, p.Text)
+			}
+		}
+		return strings.Join(texts, "\n\n")
+	}
+	return ""
+}
+
+// convertResponsesUserToAnthropicContent converts a Responses user message
+// content field into Anthropic content blocks JSON.
+func convertResponsesUserToAnthropicContent(raw json.RawMessage) (json.RawMessage, error) {
+	if len(raw) == 0 {
+		return json.Marshal("") // empty string content
+	}
+
+	// Try plain string.
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return json.Marshal(s)
+	}
+
+	// Array of content parts → Anthropic content blocks.
+	var parts []ResponsesContentPart
+	if err := json.Unmarshal(raw, &parts); err != nil {
+		// Pass through as-is if we can't parse
+		return raw, nil
+	}
+
+	var blocks []AnthropicContentBlock
+	for _, p := range parts {
+		switch p.Type {
+		case "input_text", "text":
+			if p.Text != "" {
+				blocks = append(blocks, AnthropicContentBlock{
+					Type: "text",
+					Text: p.Text,
+				})
+			}
+		case "input_image":
+			src := dataURIToAnthropicImageSource(p.ImageURL)
+			if src != nil {
+				blocks = append(blocks, AnthropicContentBlock{
+					Type:   "image",
+					Source: src,
+				})
+			}
+		}
+	}
+
+	if len(blocks) == 0 {
+		return json.Marshal("")
+	}
+	return json.Marshal(blocks)
+}
+
+// convertResponsesAssistantToAnthropicContent converts a Responses assistant
+// message content field into Anthropic content blocks JSON.
+func convertResponsesAssistantToAnthropicContent(raw json.RawMessage) (json.RawMessage, error) {
+	if len(raw) == 0 {
+		return json.Marshal([]AnthropicContentBlock{{Type: "text", Text: ""}})
+	}
+
+	// Try plain string.
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return json.Marshal([]AnthropicContentBlock{{Type: "text", Text: s}})
+	}
+
+	// Array of content parts → Anthropic content blocks.
+	var parts []ResponsesContentPart
+	if err := json.Unmarshal(raw, &parts); err != nil {
+		return raw, nil
+	}
+
+	var blocks []AnthropicContentBlock
+	for _, p := range parts {
+		switch p.Type {
+		case "output_text", "text":
+			if p.Text != "" {
+				blocks = append(blocks, AnthropicContentBlock{
+					Type: "text",
+					Text: p.Text,
+				})
+			}
+		}
+	}
+
+	if len(blocks) == 0 {
+		blocks = append(blocks, AnthropicContentBlock{Type: "text", Text: ""})
+	}
+	return json.Marshal(blocks)
+}
+
+// fromResponsesCallIDToAnthropic converts an OpenAI function call ID back to
+// Anthropic format. Reverses toResponsesCallID.
+func fromResponsesCallIDToAnthropic(id string) string {
+	// If it has our "fc_" prefix wrapping a known Anthropic prefix, strip it
+	if after, ok := strings.CutPrefix(id, "fc_"); ok {
+		if strings.HasPrefix(after, "toolu_") || strings.HasPrefix(after, "call_") {
+			return after
+		}
+	}
+	// Generate a synthetic Anthropic tool ID
+	if !strings.HasPrefix(id, "toolu_") && !strings.HasPrefix(id, "call_") {
+		return "toolu_" + id
+	}
+	return id
+}
+
+// dataURIToAnthropicImageSource parses a data URI into an AnthropicImageSource.
+func dataURIToAnthropicImageSource(dataURI string) *AnthropicImageSource {
+	if !strings.HasPrefix(dataURI, "data:") {
+		return nil
+	}
+	// Format: data:<media_type>;base64,<data>
+	rest := strings.TrimPrefix(dataURI, "data:")
+	semicolonIdx := strings.Index(rest, ";")
+	if semicolonIdx < 0 {
+		return nil
+	}
+	mediaType := rest[:semicolonIdx]
+	rest = rest[semicolonIdx+1:]
+	if !strings.HasPrefix(rest, "base64,") {
+		return nil
+	}
+	data := strings.TrimPrefix(rest, "base64,")
+	return &AnthropicImageSource{
+		Type:      "base64",
+		MediaType: mediaType,
+		Data:      data,
+	}
+}
+
+// mergeConsecutiveMessages merges consecutive messages with the same role
+// because Anthropic requires alternating user/assistant turns.
+func mergeConsecutiveMessages(messages []AnthropicMessage) []AnthropicMessage {
+	if len(messages) <= 1 {
+		return messages
+	}
+
+	var merged []AnthropicMessage
+	for _, msg := range messages {
+		if len(merged) == 0 || merged[len(merged)-1].Role != msg.Role {
+			merged = append(merged, msg)
+			continue
+		}
+
+		// Same role — merge content arrays
+		last := &merged[len(merged)-1]
+		lastBlocks := parseContentBlocks(last.Content)
+		newBlocks := parseContentBlocks(msg.Content)
+		combined := append(lastBlocks, newBlocks...)
+		last.Content, _ = json.Marshal(combined)
+	}
+	return merged
+}
+
+// parseContentBlocks attempts to parse content as []AnthropicContentBlock.
+// If it's a string, wraps it in a text block.
+func parseContentBlocks(raw json.RawMessage) []AnthropicContentBlock {
+	var blocks []AnthropicContentBlock
+	if err := json.Unmarshal(raw, &blocks); err == nil {
+		return blocks
+	}
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return []AnthropicContentBlock{{Type: "text", Text: s}}
+	}
+	return nil
+}
+
+// convertResponsesToAnthropicTools maps Responses API tools to Anthropic format.
+// Reverse of convertAnthropicToolsToResponses.
+func convertResponsesToAnthropicTools(tools []ResponsesTool) []AnthropicTool {
+	var out []AnthropicTool
+	for _, t := range tools {
+		switch t.Type {
+		case "web_search":
+			out = append(out, AnthropicTool{
+				Type: "web_search_20250305",
+				Name: "web_search",
+			})
+		case "function":
+			out = append(out, AnthropicTool{
+				Name:        t.Name,
+				Description: t.Description,
+				InputSchema: normalizeAnthropicInputSchema(t.Parameters),
+			})
+		default:
+			// Pass through unknown tool types
+			out = append(out, AnthropicTool{
+				Type:        t.Type,
+				Name:        t.Name,
+				Description: t.Description,
+				InputSchema: t.Parameters,
+			})
+		}
+	}
+	return out
+}
+
+// normalizeAnthropicInputSchema ensures the input_schema has a "type" field.
+func normalizeAnthropicInputSchema(schema json.RawMessage) json.RawMessage {
+	if len(schema) == 0 || string(schema) == "null" {
+		return json.RawMessage(`{"type":"object","properties":{}}`)
+	}
+	return schema
+}
+
+// convertResponsesToAnthropicToolChoice maps Responses tool_choice to Anthropic format.
+// Reverse of convertAnthropicToolChoiceToResponses.
+//
+//	"auto"                                     → {"type":"auto"}
+//	"required"                                 → {"type":"any"}
+//	"none"                                     → {"type":"none"}
+//	{"type":"function","function":{"name":"X"}} → {"type":"tool","name":"X"}
+func convertResponsesToAnthropicToolChoice(raw json.RawMessage) (json.RawMessage, error) {
+	// Try as string first
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		switch s {
+		case "auto":
+			return json.Marshal(map[string]string{"type": "auto"})
+		case "required":
+			return json.Marshal(map[string]string{"type": "any"})
+		case "none":
+			return json.Marshal(map[string]string{"type": "none"})
+		default:
+			return raw, nil
+		}
+	}
+
+	// Try as object with type=function
+	var tc struct {
+		Type     string `json:"type"`
+		Function struct {
+			Name string `json:"name"`
+		} `json:"function"`
+	}
+	if err := json.Unmarshal(raw, &tc); err == nil && tc.Type == "function" && tc.Function.Name != "" {
+		return json.Marshal(map[string]string{
+			"type": "tool",
+			"name": tc.Function.Name,
+		})
+	}
+
+	// Pass through unknown
+	return raw, nil
+}

internal/pkg/apicompat/responses_to_chatcompletions.go

@@ -0,0 +1,374 @@
+package apicompat
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"time"
+)
+
+// ---------------------------------------------------------------------------
+// Non-streaming: ResponsesResponse → ChatCompletionsResponse
+// ---------------------------------------------------------------------------
+
+// ResponsesToChatCompletions converts a Responses API response into a Chat
+// Completions response. Text output items are concatenated into
+// choices[0].message.content; function_call items become tool_calls.
+func ResponsesToChatCompletions(resp *ResponsesResponse, model string) *ChatCompletionsResponse {
+	id := resp.ID
+	if id == "" {
+		id = generateChatCmplID()
+	}
+
+	out := &ChatCompletionsResponse{
+		ID:      id,
+		Object:  "chat.completion",
+		Created: time.Now().Unix(),
+		Model:   model,
+	}
+
+	var contentText string
+	var reasoningText string
+	var toolCalls []ChatToolCall
+
+	for _, item := range resp.Output {
+		switch item.Type {
+		case "message":
+			for _, part := range item.Content {
+				if part.Type == "output_text" && part.Text != "" {
+					contentText += part.Text
+				}
+			}
+		case "function_call":
+			toolCalls = append(toolCalls, ChatToolCall{
+				ID:   item.CallID,
+				Type: "function",
+				Function: ChatFunctionCall{
+					Name:      item.Name,
+					Arguments: item.Arguments,
+				},
+			})
+		case "reasoning":
+			for _, s := range item.Summary {
+				if s.Type == "summary_text" && s.Text != "" {
+					reasoningText += s.Text
+				}
+			}
+		case "web_search_call":
+			// silently consumed — results already incorporated into text output
+		}
+	}
+
+	msg := ChatMessage{Role: "assistant"}
+	if len(toolCalls) > 0 {
+		msg.ToolCalls = toolCalls
+	}
+	if contentText != "" {
+		raw, _ := json.Marshal(contentText)
+		msg.Content = raw
+	}
+	if reasoningText != "" {
+		msg.ReasoningContent = reasoningText
+	}
+
+	finishReason := responsesStatusToChatFinishReason(resp.Status, resp.IncompleteDetails, toolCalls)
+
+	out.Choices = []ChatChoice{{
+		Index:        0,
+		Message:      msg,
+		FinishReason: finishReason,
+	}}
+
+	if resp.Usage != nil {
+		usage := &ChatUsage{
+			PromptTokens:     resp.Usage.InputTokens,
+			CompletionTokens: resp.Usage.OutputTokens,
+			TotalTokens:      resp.Usage.InputTokens + resp.Usage.OutputTokens,
+		}
+		if resp.Usage.InputTokensDetails != nil && resp.Usage.InputTokensDetails.CachedTokens > 0 {
+			usage.PromptTokensDetails = &ChatTokenDetails{
+				CachedTokens: resp.Usage.InputTokensDetails.CachedTokens,
+			}
+		}
+		out.Usage = usage
+	}
+
+	return out
+}
+
+func responsesStatusToChatFinishReason(status string, details *ResponsesIncompleteDetails, toolCalls []ChatToolCall) string {
+	switch status {
+	case "incomplete":
+		if details != nil && details.Reason == "max_output_tokens" {
+			return "length"
+		}
+		return "stop"
+	case "completed":
+		if len(toolCalls) > 0 {
+			return "tool_calls"
+		}
+		return "stop"
+	default:
+		return "stop"
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Streaming: ResponsesStreamEvent → []ChatCompletionsChunk (stateful converter)
+// ---------------------------------------------------------------------------
+
+// ResponsesEventToChatState tracks state for converting a sequence of Responses
+// SSE events into Chat Completions SSE chunks.
+type ResponsesEventToChatState struct {
+	ID                     string
+	Model                  string
+	Created                int64
+	SentRole               bool
+	SawToolCall            bool
+	SawText                bool
+	Finalized              bool        // true after finish chunk has been emitted
+	NextToolCallIndex      int         // next sequential tool_call index to assign
+	OutputIndexToToolIndex map[int]int // Responses output_index → Chat tool_calls index
+	IncludeUsage           bool
+	Usage                  *ChatUsage
+}
+
+// NewResponsesEventToChatState returns an initialised stream state.
+func NewResponsesEventToChatState() *ResponsesEventToChatState {
+	return &ResponsesEventToChatState{
+		ID:                     generateChatCmplID(),
+		Created:                time.Now().Unix(),
+		OutputIndexToToolIndex: make(map[int]int),
+	}
+}
+
+// ResponsesEventToChatChunks converts a single Responses SSE event into zero
+// or more Chat Completions chunks, updating state as it goes.
+func ResponsesEventToChatChunks(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	switch evt.Type {
+	case "response.created":
+		return resToChatHandleCreated(evt, state)
+	case "response.output_text.delta":
+		return resToChatHandleTextDelta(evt, state)
+	case "response.output_item.added":
+		return resToChatHandleOutputItemAdded(evt, state)
+	case "response.function_call_arguments.delta":
+		return resToChatHandleFuncArgsDelta(evt, state)
+	case "response.reasoning_summary_text.delta":
+		return resToChatHandleReasoningDelta(evt, state)
+	case "response.reasoning_summary_text.done":
+		return nil
+	case "response.completed", "response.incomplete", "response.failed":
+		return resToChatHandleCompleted(evt, state)
+	default:
+		return nil
+	}
+}
+
+// FinalizeResponsesChatStream emits a final chunk with finish_reason if the
+// stream ended without a proper completion event (e.g. upstream disconnect).
+// It is idempotent: if a completion event already emitted the finish chunk,
+// this returns nil.
+func FinalizeResponsesChatStream(state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	if state.Finalized {
+		return nil
+	}
+	state.Finalized = true
+
+	finishReason := "stop"
+	if state.SawToolCall {
+		finishReason = "tool_calls"
+	}
+
+	chunks := []ChatCompletionsChunk{makeChatFinishChunk(state, finishReason)}
+
+	if state.IncludeUsage && state.Usage != nil {
+		chunks = append(chunks, ChatCompletionsChunk{
+			ID:      state.ID,
+			Object:  "chat.completion.chunk",
+			Created: state.Created,
+			Model:   state.Model,
+			Choices: []ChatChunkChoice{},
+			Usage:   state.Usage,
+		})
+	}
+
+	return chunks
+}
+
+// ChatChunkToSSE formats a ChatCompletionsChunk as an SSE data line.
+func ChatChunkToSSE(chunk ChatCompletionsChunk) (string, error) {
+	data, err := json.Marshal(chunk)
+	if err != nil {
+		return "", err
+	}
+	return fmt.Sprintf("data: %s\n\n", data), nil
+}
+
+// --- internal handlers ---
+
+func resToChatHandleCreated(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	if evt.Response != nil {
+		if evt.Response.ID != "" {
+			state.ID = evt.Response.ID
+		}
+		if state.Model == "" && evt.Response.Model != "" {
+			state.Model = evt.Response.Model
+		}
+	}
+	// Emit the role chunk.
+	if state.SentRole {
+		return nil
+	}
+	state.SentRole = true
+
+	role := "assistant"
+	return []ChatCompletionsChunk{makeChatDeltaChunk(state, ChatDelta{Role: role})}
+}
+
+func resToChatHandleTextDelta(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	if evt.Delta == "" {
+		return nil
+	}
+	state.SawText = true
+	content := evt.Delta
+	return []ChatCompletionsChunk{makeChatDeltaChunk(state, ChatDelta{Content: &content})}
+}
+
+func resToChatHandleOutputItemAdded(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	if evt.Item == nil || evt.Item.Type != "function_call" {
+		return nil
+	}
+
+	state.SawToolCall = true
+	idx := state.NextToolCallIndex
+	state.OutputIndexToToolIndex[evt.OutputIndex] = idx
+	state.NextToolCallIndex++
+
+	return []ChatCompletionsChunk{makeChatDeltaChunk(state, ChatDelta{
+		ToolCalls: []ChatToolCall{{
+			Index: &idx,
+			ID:    evt.Item.CallID,
+			Type:  "function",
+			Function: ChatFunctionCall{
+				Name: evt.Item.Name,
+			},
+		}},
+	})}
+}
+
+func resToChatHandleFuncArgsDelta(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	if evt.Delta == "" {
+		return nil
+	}
+
+	idx, ok := state.OutputIndexToToolIndex[evt.OutputIndex]
+	if !ok {
+		return nil
+	}
+
+	return []ChatCompletionsChunk{makeChatDeltaChunk(state, ChatDelta{
+		ToolCalls: []ChatToolCall{{
+			Index: &idx,
+			Function: ChatFunctionCall{
+				Arguments: evt.Delta,
+			},
+		}},
+	})}
+}
+
+func resToChatHandleReasoningDelta(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	if evt.Delta == "" {
+		return nil
+	}
+	reasoning := evt.Delta
+	return []ChatCompletionsChunk{makeChatDeltaChunk(state, ChatDelta{ReasoningContent: &reasoning})}
+}
+
+func resToChatHandleCompleted(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {
+	state.Finalized = true
+	finishReason := "stop"
+
+	if evt.Response != nil {
+		if evt.Response.Usage != nil {
+			u := evt.Response.Usage
+			usage := &ChatUsage{
+				PromptTokens:     u.InputTokens,
+				CompletionTokens: u.OutputTokens,
+				TotalTokens:      u.InputTokens + u.OutputTokens,
+			}
+			if u.InputTokensDetails != nil && u.InputTokensDetails.CachedTokens > 0 {
+				usage.PromptTokensDetails = &ChatTokenDetails{
+					CachedTokens: u.InputTokensDetails.CachedTokens,
+				}
+			}
+			state.Usage = usage
+		}
+
+		switch evt.Response.Status {
+		case "incomplete":
+			if evt.Response.IncompleteDetails != nil && evt.Response.IncompleteDetails.Reason == "max_output_tokens" {
+				finishReason = "length"
+			}
+		case "completed":
+			if state.SawToolCall {
+				finishReason = "tool_calls"
+			}
+		}
+	} else if state.SawToolCall {
+		finishReason = "tool_calls"
+	}
+
+	var chunks []ChatCompletionsChunk
+	chunks = append(chunks, makeChatFinishChunk(state, finishReason))
+
+	if state.IncludeUsage && state.Usage != nil {
+		chunks = append(chunks, ChatCompletionsChunk{
+			ID:      state.ID,
+			Object:  "chat.completion.chunk",
+			Created: state.Created,
+			Model:   state.Model,
+			Choices: []ChatChunkChoice{},
+			Usage:   state.Usage,
+		})
+	}
+
+	return chunks
+}
+
+func makeChatDeltaChunk(state *ResponsesEventToChatState, delta ChatDelta) ChatCompletionsChunk {
+	return ChatCompletionsChunk{
+		ID:      state.ID,
+		Object:  "chat.completion.chunk",
+		Created: state.Created,
+		Model:   state.Model,
+		Choices: []ChatChunkChoice{{
+			Index:        0,
+			Delta:        delta,
+			FinishReason: nil,
+		}},
+	}
+}
+
+func makeChatFinishChunk(state *ResponsesEventToChatState, finishReason string) ChatCompletionsChunk {
+	empty := ""
+	return ChatCompletionsChunk{
+		ID:      state.ID,
+		Object:  "chat.completion.chunk",
+		Created: state.Created,
+		Model:   state.Model,
+		Choices: []ChatChunkChoice{{
+			Index:        0,
+			Delta:        ChatDelta{Content: &empty},
+			FinishReason: &finishReason,
+		}},
+	}
+}
+
+// generateChatCmplID returns a "chatcmpl-" prefixed random hex ID.
+func generateChatCmplID() string {
+	b := make([]byte, 12)
+	_, _ = rand.Read(b)
+	return "chatcmpl-" + hex.EncodeToString(b)
+}

internal/pkg/apicompat/types.go

@@ -0,0 +1,482 @@
+// Package apicompat provides type definitions and conversion utilities for
+// translating between Anthropic Messages and OpenAI Responses API formats.
+// It enables multi-protocol support so that clients using different API
+// formats can be served through a unified gateway.
+package apicompat
+
+import "encoding/json"
+
+// ---------------------------------------------------------------------------
+// Anthropic Messages API types
+// ---------------------------------------------------------------------------
+
+// AnthropicRequest is the request body for POST /v1/messages.
+type AnthropicRequest struct {
+	Model        string                 `json:"model"`
+	MaxTokens    int                    `json:"max_tokens"`
+	System       json.RawMessage        `json:"system,omitempty"` // string or []AnthropicContentBlock
+	Messages     []AnthropicMessage     `json:"messages"`
+	Tools        []AnthropicTool        `json:"tools,omitempty"`
+	Stream       bool                   `json:"stream,omitempty"`
+	Temperature  *float64               `json:"temperature,omitempty"`
+	TopP         *float64               `json:"top_p,omitempty"`
+	StopSeqs     []string               `json:"stop_sequences,omitempty"`
+	Thinking     *AnthropicThinking     `json:"thinking,omitempty"`
+	ToolChoice   json.RawMessage        `json:"tool_choice,omitempty"`
+	OutputConfig *AnthropicOutputConfig `json:"output_config,omitempty"`
+}
+
+// AnthropicOutputConfig controls output generation parameters.
+type AnthropicOutputConfig struct {
+	Effort string `json:"effort,omitempty"` // "low" | "medium" | "high"
+}
+
+// AnthropicThinking configures extended thinking in the Anthropic API.
+type AnthropicThinking struct {
+	Type         string `json:"type"`                    // "enabled" | "adaptive" | "disabled"
+	BudgetTokens int    `json:"budget_tokens,omitempty"` // max thinking tokens
+}
+
+// AnthropicMessage is a single message in the Anthropic conversation.
+type AnthropicMessage struct {
+	Role    string          `json:"role"` // "user" | "assistant"
+	Content json.RawMessage `json:"content"`
+}
+
+// AnthropicContentBlock is one block inside a message's content array.
+type AnthropicContentBlock struct {
+	Type string `json:"type"`
+
+	// type=text
+	Text string `json:"text,omitempty"`
+
+	// type=thinking
+	Thinking string `json:"thinking,omitempty"`
+
+	// type=image
+	Source *AnthropicImageSource `json:"source,omitempty"`
+
+	// type=tool_use
+	ID    string          `json:"id,omitempty"`
+	Name  string          `json:"name,omitempty"`
+	Input json.RawMessage `json:"input,omitempty"`
+
+	// type=tool_result
+	ToolUseID string          `json:"tool_use_id,omitempty"`
+	Content   json.RawMessage `json:"content,omitempty"` // string or []AnthropicContentBlock
+	IsError   bool            `json:"is_error,omitempty"`
+}
+
+// AnthropicImageSource describes the source data for an image content block.
+type AnthropicImageSource struct {
+	Type      string `json:"type"` // "base64"
+	MediaType string `json:"media_type"`
+	Data      string `json:"data"`
+}
+
+// AnthropicTool describes a tool available to the model.
+type AnthropicTool struct {
+	Type        string          `json:"type,omitempty"` // e.g. "web_search_20250305" for server tools
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	InputSchema json.RawMessage `json:"input_schema"` // JSON Schema object
+}
+
+// AnthropicResponse is the non-streaming response from POST /v1/messages.
+type AnthropicResponse struct {
+	ID           string                  `json:"id"`
+	Type         string                  `json:"type"` // "message"
+	Role         string                  `json:"role"` // "assistant"
+	Content      []AnthropicContentBlock `json:"content"`
+	Model        string                  `json:"model"`
+	StopReason   string                  `json:"stop_reason"`
+	StopSequence *string                 `json:"stop_sequence,omitempty"`
+	Usage        AnthropicUsage          `json:"usage"`
+}
+
+// AnthropicUsage holds token counts in Anthropic format.
+type AnthropicUsage struct {
+	InputTokens              int `json:"input_tokens"`
+	OutputTokens             int `json:"output_tokens"`
+	CacheCreationInputTokens int `json:"cache_creation_input_tokens"`
+	CacheReadInputTokens     int `json:"cache_read_input_tokens"`
+}
+
+// ---------------------------------------------------------------------------
+// Anthropic SSE event types
+// ---------------------------------------------------------------------------
+
+// AnthropicStreamEvent is a single SSE event in the Anthropic streaming protocol.
+type AnthropicStreamEvent struct {
+	Type string `json:"type"`
+
+	// message_start
+	Message *AnthropicResponse `json:"message,omitempty"`
+
+	// content_block_start
+	Index        *int                   `json:"index,omitempty"`
+	ContentBlock *AnthropicContentBlock `json:"content_block,omitempty"`
+
+	// content_block_delta
+	Delta *AnthropicDelta `json:"delta,omitempty"`
+
+	// message_delta
+	Usage *AnthropicUsage `json:"usage,omitempty"`
+}
+
+// AnthropicDelta carries incremental content in streaming events.
+type AnthropicDelta struct {
+	Type string `json:"type,omitempty"` // "text_delta" | "input_json_delta" | "thinking_delta" | "signature_delta"
+
+	// text_delta
+	Text string `json:"text,omitempty"`
+
+	// input_json_delta
+	PartialJSON string `json:"partial_json,omitempty"`
+
+	// thinking_delta
+	Thinking string `json:"thinking,omitempty"`
+
+	// signature_delta
+	Signature string `json:"signature,omitempty"`
+
+	// message_delta fields
+	StopReason   string  `json:"stop_reason,omitempty"`
+	StopSequence *string `json:"stop_sequence,omitempty"`
+}
+
+// ---------------------------------------------------------------------------
+// OpenAI Responses API types
+// ---------------------------------------------------------------------------
+
+// ResponsesRequest is the request body for POST /v1/responses.
+type ResponsesRequest struct {
+	Model           string              `json:"model"`
+	Input           json.RawMessage     `json:"input"` // string or []ResponsesInputItem
+	MaxOutputTokens *int                `json:"max_output_tokens,omitempty"`
+	Temperature     *float64            `json:"temperature,omitempty"`
+	TopP            *float64            `json:"top_p,omitempty"`
+	Stream          bool                `json:"stream,omitempty"`
+	Tools           []ResponsesTool     `json:"tools,omitempty"`
+	Include         []string            `json:"include,omitempty"`
+	Store           *bool               `json:"store,omitempty"`
+	Reasoning       *ResponsesReasoning `json:"reasoning,omitempty"`
+	ToolChoice      json.RawMessage     `json:"tool_choice,omitempty"`
+	ServiceTier     string              `json:"service_tier,omitempty"`
+}
+
+// ResponsesReasoning configures reasoning effort in the Responses API.
+type ResponsesReasoning struct {
+	Effort  string `json:"effort"`            // "low" | "medium" | "high"
+	Summary string `json:"summary,omitempty"` // "auto" | "concise" | "detailed"
+}
+
+// ResponsesInputItem is one item in the Responses API input array.
+// The Type field determines which other fields are populated.
+type ResponsesInputItem struct {
+	// Common
+	Type string `json:"type,omitempty"` // "" for role-based messages
+
+	// Role-based messages (system/user/assistant)
+	Role    string          `json:"role,omitempty"`
+	Content json.RawMessage `json:"content,omitempty"` // string or []ResponsesContentPart
+
+	// type=function_call
+	CallID    string `json:"call_id,omitempty"`
+	Name      string `json:"name,omitempty"`
+	Arguments string `json:"arguments,omitempty"`
+	ID        string `json:"id,omitempty"`
+
+	// type=function_call_output
+	Output string `json:"output,omitempty"`
+}
+
+// ResponsesContentPart is a typed content part in a Responses message.
+type ResponsesContentPart struct {
+	Type     string `json:"type"` // "input_text" | "output_text" | "input_image"
+	Text     string `json:"text,omitempty"`
+	ImageURL string `json:"image_url,omitempty"` // data URI for input_image
+}
+
+// ResponsesTool describes a tool in the Responses API.
+type ResponsesTool struct {
+	Type        string          `json:"type"` // "function" | "web_search" | "local_shell" etc.
+	Name        string          `json:"name,omitempty"`
+	Description string          `json:"description,omitempty"`
+	Parameters  json.RawMessage `json:"parameters,omitempty"`
+	Strict      *bool           `json:"strict,omitempty"`
+}
+
+// ResponsesResponse is the non-streaming response from POST /v1/responses.
+type ResponsesResponse struct {
+	ID     string            `json:"id"`
+	Object string            `json:"object"` // "response"
+	Model  string            `json:"model"`
+	Status string            `json:"status"` // "completed" | "incomplete" | "failed"
+	Output []ResponsesOutput `json:"output"`
+	Usage  *ResponsesUsage   `json:"usage,omitempty"`
+
+	// incomplete_details is present when status="incomplete"
+	IncompleteDetails *ResponsesIncompleteDetails `json:"incomplete_details,omitempty"`
+
+	// Error is present when status="failed"
+	Error *ResponsesError `json:"error,omitempty"`
+}
+
+// ResponsesError describes an error in a failed response.
+type ResponsesError struct {
+	Code    string `json:"code"`
+	Message string `json:"message"`
+}
+
+// ResponsesIncompleteDetails explains why a response is incomplete.
+type ResponsesIncompleteDetails struct {
+	Reason string `json:"reason"` // "max_output_tokens" | "content_filter"
+}
+
+// ResponsesOutput is one output item in a Responses API response.
+type ResponsesOutput struct {
+	Type string `json:"type"` // "message" | "reasoning" | "function_call" | "web_search_call"
+
+	// type=message
+	ID      string                 `json:"id,omitempty"`
+	Role    string                 `json:"role,omitempty"`
+	Content []ResponsesContentPart `json:"content,omitempty"`
+	Status  string                 `json:"status,omitempty"`
+
+	// type=reasoning
+	EncryptedContent string             `json:"encrypted_content,omitempty"`
+	Summary          []ResponsesSummary `json:"summary,omitempty"`
+
+	// type=function_call
+	CallID    string `json:"call_id,omitempty"`
+	Name      string `json:"name,omitempty"`
+	Arguments string `json:"arguments,omitempty"`
+
+	// type=web_search_call
+	Action *WebSearchAction `json:"action,omitempty"`
+}
+
+// WebSearchAction describes the search action in a web_search_call output item.
+type WebSearchAction struct {
+	Type  string `json:"type,omitempty"`  // "search"
+	Query string `json:"query,omitempty"` // primary search query
+}
+
+// ResponsesSummary is a summary text block inside a reasoning output.
+type ResponsesSummary struct {
+	Type string `json:"type"` // "summary_text"
+	Text string `json:"text"`
+}
+
+// ResponsesUsage holds token counts in Responses API format.
+type ResponsesUsage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+	TotalTokens  int `json:"total_tokens"`
+
+	// Optional detailed breakdown
+	InputTokensDetails  *ResponsesInputTokensDetails  `json:"input_tokens_details,omitempty"`
+	OutputTokensDetails *ResponsesOutputTokensDetails `json:"output_tokens_details,omitempty"`
+}
+
+// ResponsesInputTokensDetails breaks down input token usage.
+type ResponsesInputTokensDetails struct {
+	CachedTokens int `json:"cached_tokens,omitempty"`
+}
+
+// ResponsesOutputTokensDetails breaks down output token usage.
+type ResponsesOutputTokensDetails struct {
+	ReasoningTokens int `json:"reasoning_tokens,omitempty"`
+}
+
+// ---------------------------------------------------------------------------
+// Responses SSE event types
+// ---------------------------------------------------------------------------
+
+// ResponsesStreamEvent is a single SSE event in the Responses streaming protocol.
+// The Type field corresponds to the "type" in the JSON payload.
+type ResponsesStreamEvent struct {
+	Type string `json:"type"`
+
+	// response.created / response.completed / response.failed / response.incomplete
+	Response *ResponsesResponse `json:"response,omitempty"`
+
+	// response.output_item.added / response.output_item.done
+	Item *ResponsesOutput `json:"item,omitempty"`
+
+	// response.output_text.delta / response.output_text.done
+	OutputIndex  int    `json:"output_index,omitempty"`
+	ContentIndex int    `json:"content_index,omitempty"`
+	Delta        string `json:"delta,omitempty"`
+	Text         string `json:"text,omitempty"`
+	ItemID       string `json:"item_id,omitempty"`
+
+	// response.function_call_arguments.delta / done
+	CallID    string `json:"call_id,omitempty"`
+	Name      string `json:"name,omitempty"`
+	Arguments string `json:"arguments,omitempty"`
+
+	// response.reasoning_summary_text.delta / done
+	// Reuses Text/Delta fields above, SummaryIndex identifies which summary part
+	SummaryIndex int `json:"summary_index,omitempty"`
+
+	// error event fields
+	Code  string `json:"code,omitempty"`
+	Param string `json:"param,omitempty"`
+
+	// Sequence number for ordering events
+	SequenceNumber int `json:"sequence_number,omitempty"`
+}
+
+// ---------------------------------------------------------------------------
+// OpenAI Chat Completions API types
+// ---------------------------------------------------------------------------
+
+// ChatCompletionsRequest is the request body for POST /v1/chat/completions.
+type ChatCompletionsRequest struct {
+	Model               string             `json:"model"`
+	Messages            []ChatMessage      `json:"messages"`
+	MaxTokens           *int               `json:"max_tokens,omitempty"`
+	MaxCompletionTokens *int               `json:"max_completion_tokens,omitempty"`
+	Temperature         *float64           `json:"temperature,omitempty"`
+	TopP                *float64           `json:"top_p,omitempty"`
+	Stream              bool               `json:"stream,omitempty"`
+	StreamOptions       *ChatStreamOptions `json:"stream_options,omitempty"`
+	Tools               []ChatTool         `json:"tools,omitempty"`
+	ToolChoice          json.RawMessage    `json:"tool_choice,omitempty"`
+	ReasoningEffort     string             `json:"reasoning_effort,omitempty"` // "low" | "medium" | "high"
+	ServiceTier         string             `json:"service_tier,omitempty"`
+	Stop                json.RawMessage    `json:"stop,omitempty"` // string or []string
+
+	// Legacy function calling (deprecated but still supported)
+	Functions    []ChatFunction  `json:"functions,omitempty"`
+	FunctionCall json.RawMessage `json:"function_call,omitempty"`
+}
+
+// ChatStreamOptions configures streaming behavior.
+type ChatStreamOptions struct {
+	IncludeUsage bool `json:"include_usage,omitempty"`
+}
+
+// ChatMessage is a single message in the Chat Completions conversation.
+type ChatMessage struct {
+	Role             string          `json:"role"` // "system" | "user" | "assistant" | "tool" | "function"
+	Content          json.RawMessage `json:"content,omitempty"`
+	ReasoningContent string          `json:"reasoning_content,omitempty"`
+	Name             string          `json:"name,omitempty"`
+	ToolCalls        []ChatToolCall  `json:"tool_calls,omitempty"`
+	ToolCallID       string          `json:"tool_call_id,omitempty"`
+
+	// Legacy function calling
+	FunctionCall *ChatFunctionCall `json:"function_call,omitempty"`
+}
+
+// ChatContentPart is a typed content part in a multi-modal message.
+type ChatContentPart struct {
+	Type     string        `json:"type"` // "text" | "image_url"
+	Text     string        `json:"text,omitempty"`
+	ImageURL *ChatImageURL `json:"image_url,omitempty"`
+}
+
+// ChatImageURL contains the URL for an image content part.
+type ChatImageURL struct {
+	URL    string `json:"url"`
+	Detail string `json:"detail,omitempty"` // "auto" | "low" | "high"
+}
+
+// ChatTool describes a tool available to the model.
+type ChatTool struct {
+	Type     string        `json:"type"` // "function"
+	Function *ChatFunction `json:"function,omitempty"`
+}
+
+// ChatFunction describes a function tool definition.
+type ChatFunction struct {
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	Parameters  json.RawMessage `json:"parameters,omitempty"`
+	Strict      *bool           `json:"strict,omitempty"`
+}
+
+// ChatToolCall represents a tool call made by the assistant.
+// Index is only populated in streaming chunks (omitted in non-streaming responses).
+type ChatToolCall struct {
+	Index    *int             `json:"index,omitempty"`
+	ID       string           `json:"id,omitempty"`
+	Type     string           `json:"type,omitempty"` // "function"
+	Function ChatFunctionCall `json:"function"`
+}
+
+// ChatFunctionCall contains the function name and arguments.
+type ChatFunctionCall struct {
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+}
+
+// ChatCompletionsResponse is the non-streaming response from POST /v1/chat/completions.
+type ChatCompletionsResponse struct {
+	ID                string       `json:"id"`
+	Object            string       `json:"object"` // "chat.completion"
+	Created           int64        `json:"created"`
+	Model             string       `json:"model"`
+	Choices           []ChatChoice `json:"choices"`
+	Usage             *ChatUsage   `json:"usage,omitempty"`
+	SystemFingerprint string       `json:"system_fingerprint,omitempty"`
+	ServiceTier       string       `json:"service_tier,omitempty"`
+}
+
+// ChatChoice is a single completion choice.
+type ChatChoice struct {
+	Index        int         `json:"index"`
+	Message      ChatMessage `json:"message"`
+	FinishReason string      `json:"finish_reason"` // "stop" | "length" | "tool_calls" | "content_filter"
+}
+
+// ChatUsage holds token counts in Chat Completions format.
+type ChatUsage struct {
+	PromptTokens        int               `json:"prompt_tokens"`
+	CompletionTokens    int               `json:"completion_tokens"`
+	TotalTokens         int               `json:"total_tokens"`
+	PromptTokensDetails *ChatTokenDetails `json:"prompt_tokens_details,omitempty"`
+}
+
+// ChatTokenDetails provides a breakdown of token usage.
+type ChatTokenDetails struct {
+	CachedTokens int `json:"cached_tokens,omitempty"`
+}
+
+// ChatCompletionsChunk is a single streaming chunk from POST /v1/chat/completions.
+type ChatCompletionsChunk struct {
+	ID                string            `json:"id"`
+	Object            string            `json:"object"` // "chat.completion.chunk"
+	Created           int64             `json:"created"`
+	Model             string            `json:"model"`
+	Choices           []ChatChunkChoice `json:"choices"`
+	Usage             *ChatUsage        `json:"usage,omitempty"`
+	SystemFingerprint string            `json:"system_fingerprint,omitempty"`
+	ServiceTier       string            `json:"service_tier,omitempty"`
+}
+
+// ChatChunkChoice is a single choice in a streaming chunk.
+type ChatChunkChoice struct {
+	Index        int       `json:"index"`
+	Delta        ChatDelta `json:"delta"`
+	FinishReason *string   `json:"finish_reason"` // pointer: null when not final
+}
+
+// ChatDelta carries incremental content in a streaming chunk.
+type ChatDelta struct {
+	Role             string         `json:"role,omitempty"`
+	Content          *string        `json:"content,omitempty"` // pointer: omit when not present, null vs "" matters
+	ReasoningContent *string        `json:"reasoning_content,omitempty"`
+	ToolCalls        []ChatToolCall `json:"tool_calls,omitempty"`
+}
+
+// ---------------------------------------------------------------------------
+// Shared constants
+// ---------------------------------------------------------------------------
+
+// minMaxOutputTokens is the floor for max_output_tokens in a Responses request.
+// Very small values may cause upstream API errors, so we enforce a minimum.
+const minMaxOutputTokens = 128