Add structured output support to Responses API

Author: dmedovich

e2e/inference_test.go

@@ -106,6 +106,53 @@ func TestE2E_Inference(t *testing.T) {
 				t.Logf("responses API: %q", resp.OutputText)
 			})
 
+			t.Run("ResponsesAPIStructuredOutput", func(t *testing.T) {
+				status, body := doJSON(t, http.MethodPost, serverURL+"/responses",
+					map[string]any{
+						"model":             bc.model,
+						"input":             "Return a JSON object with answer set to yes.",
+						"max_output_tokens": 64,
+						"text": map[string]any{
+							"format": map[string]any{
+								"type":   "json_schema",
+								"name":   "answer_payload",
+								"strict": true,
+								"schema": map[string]any{
+									"type": "object",
+									"properties": map[string]any{
+										"answer": map[string]any{"type": "string"},
+									},
+									"required":             []string{"answer"},
+									"additionalProperties": false,
+								},
+							},
+						},
+					})
+				if status != http.StatusOK {
+					t.Fatalf("responses structured output: status=%d body=%s", status, body)
+				}
+				var resp struct {
+					Status     string `json:"status"`
+					OutputText string `json:"output_text"`
+				}
+				if err := json.Unmarshal(body, &resp); err != nil {
+					t.Fatalf("decode: %v (body=%s)", err, body)
+				}
+				if resp.Status != "completed" {
+					t.Errorf("expected status=completed, got %q", resp.Status)
+				}
+
+				var output struct {
+					Answer string `json:"answer"`
+				}
+				if err := json.Unmarshal([]byte(resp.OutputText), &output); err != nil {
+					t.Fatalf("output_text is not valid structured JSON: %v (output_text=%q)", err, resp.OutputText)
+				}
+				if output.Answer == "" {
+					t.Fatalf("answer field is empty in output_text=%q", resp.OutputText)
+				}
+			})
+
 			t.Run("AnthropicMessages", func(t *testing.T) {
 				status, body := doJSON(t, http.MethodPost, serverURL+"/anthropic/v1/messages",
 					map[string]any{

pkg/responses/api.go

@@ -91,9 +91,42 @@ type CreateRequest struct {
 	// MaxOutputTokens limits the response length.
 	MaxOutputTokens *int `json:"max_output_tokens,omitempty"`
 
+	// Store controls whether the response is stored for later retrieval.
+	Store *bool `json:"store,omitempty"`
+
+	// Text configures text output formatting.
+	Text *ResponseTextConfig `json:"text,omitempty"`
+
+	// Include requests additional hosted-tool output fields. Unsupported locally.
+	Include []string `json:"include,omitempty"`
+
 	// Stream enables streaming responses.
 	Stream bool `json:"stream,omitempty"`
 
+	// StreamOptions configures response streaming. Unsupported locally.
+	StreamOptions json.RawMessage `json:"stream_options,omitempty"`
+
+	// TopLogprobs requests token log probabilities. Unsupported locally.
+	TopLogprobs *int `json:"top_logprobs,omitempty"`
+
+	// Truncation configures OpenAI context truncation behavior.
+	Truncation string `json:"truncation,omitempty"`
+
+	// Background requests asynchronous background response execution.
+	Background *bool `json:"background,omitempty"`
+
+	// Conversation requests OpenAI hosted conversation state. Unsupported locally.
+	Conversation json.RawMessage `json:"conversation,omitempty"`
+
+	// Prompt requests an OpenAI hosted prompt template. Unsupported locally.
+	Prompt json.RawMessage `json:"prompt,omitempty"`
+
+	// ServiceTier selects OpenAI service tier. Unsupported locally.
+	ServiceTier string `json:"service_tier,omitempty"`
+
+	// SafetyIdentifier is an OpenAI safety identifier. Unsupported locally.
+	SafetyIdentifier string `json:"safety_identifier,omitempty"`
+
 	// Metadata is user-defined metadata for the response.
 	Metadata map[string]string `json:"metadata,omitempty"`
 
@@ -104,6 +137,21 @@ type CreateRequest struct {
 	User string `json:"user,omitempty"`
 }
 
+// ResponseTextConfig configures text output for a Responses API request.
+type ResponseTextConfig struct {
+	Format    *ResponseTextFormat `json:"format,omitempty"`
+	Verbosity string              `json:"verbosity,omitempty"`
+}
+
+// ResponseTextFormat configures the output text format.
+type ResponseTextFormat struct {
+	Type        string          `json:"type"`
+	Name        string          `json:"name,omitempty"`
+	Description string          `json:"description,omitempty"`
+	Schema      json.RawMessage `json:"schema,omitempty"`
+	Strict      *bool           `json:"strict,omitempty"`
+}
+
 // Response represents a complete response from the API.
 type Response struct {
 	// ID is the unique identifier for this response.
@@ -160,6 +208,12 @@ type Response struct {
 	// MaxOutputTokens limit used.
 	MaxOutputTokens *int `json:"max_output_tokens"`
 
+	// Store indicates whether the response was stored.
+	Store *bool `json:"store,omitempty"`
+
+	// Text config used for the response.
+	Text *ResponseTextConfig `json:"text,omitempty"`
+
 	// PreviousResponseID is the ID of the previous response in the chain.
 	PreviousResponseID *string `json:"previous_response_id"`
 

pkg/responses/handler.go

@@ -43,6 +43,19 @@ func NewHTTPHandler(log logging.Logger, schedulerHTTP http.Handler, allowedOrigi
 	h.router.HandleFunc("GET /v1"+APIPrefix+"/{id}", h.handleGet)
 	h.router.HandleFunc("GET /v1"+APIPrefix+"/{id}/input_items", h.handleListInputItems)
 	h.router.HandleFunc("DELETE /v1"+APIPrefix+"/{id}", h.handleDelete)
+	// Also register /engines/responses and /engines/v1/responses routes.
+	h.router.HandleFunc("POST /engines"+APIPrefix, h.handleCreate)
+	h.router.HandleFunc("GET /engines"+APIPrefix+"/{id}", h.handleGet)
+	h.router.HandleFunc("GET /engines"+APIPrefix+"/{id}/input_items", h.handleListInputItems)
+	h.router.HandleFunc("DELETE /engines"+APIPrefix+"/{id}", h.handleDelete)
+	h.router.HandleFunc("POST /engines/v1"+APIPrefix, h.handleCreate)
+	h.router.HandleFunc("GET /engines/v1"+APIPrefix+"/{id}", h.handleGet)
+	h.router.HandleFunc("GET /engines/v1"+APIPrefix+"/{id}/input_items", h.handleListInputItems)
+	h.router.HandleFunc("DELETE /engines/v1"+APIPrefix+"/{id}", h.handleDelete)
+	h.router.HandleFunc("POST /engines/{backend}/v1"+APIPrefix, h.handleCreate)
+	h.router.HandleFunc("GET /engines/{backend}/v1"+APIPrefix+"/{id}", h.handleGet)
+	h.router.HandleFunc("GET /engines/{backend}/v1"+APIPrefix+"/{id}/input_items", h.handleListInputItems)
+	h.router.HandleFunc("DELETE /engines/{backend}/v1"+APIPrefix+"/{id}", h.handleDelete)
 
 	// Apply CORS middleware
 	h.httpHandler = middleware.CorsMiddleware(allowedOrigins, h.router)
@@ -97,14 +110,21 @@ func (h *HTTPHandler) handleCreate(w http.ResponseWriter, r *http.Request) {
 		h.sendError(w, http.StatusBadRequest, "invalid_request", "model is required")
 		return
 	}
+	if err := validateUnsupportedRequestFields(&req); err != nil {
+		h.sendError(w, http.StatusBadRequest, "invalid_request", err.Error())
+		return
+	}
 
 	// Create a new response
 	respID := GenerateResponseID()
 	resp := NewResponse(respID, req.Model)
+	store := shouldStore(&req)
 	resp.Instructions = nilIfEmpty(req.Instructions)
 	resp.Temperature = req.Temperature
 	resp.TopP = req.TopP
 	resp.MaxOutputTokens = req.MaxOutputTokens
+	resp.Store = &store
+	resp.Text = req.Text
 	resp.Tools = req.Tools
 	resp.ToolChoice = req.ToolChoice
 	resp.ParallelToolCalls = req.ParallelToolCalls
@@ -134,7 +154,7 @@ func (h *HTTPHandler) handleCreate(w http.ResponseWriter, r *http.Request) {
 	}
 
 	// Create upstream request
-	upstreamReq, err := http.NewRequestWithContext(r.Context(), http.MethodPost, "/engines/v1/chat/completions", bytes.NewReader(chatBody))
+	upstreamReq, err := http.NewRequestWithContext(r.Context(), http.MethodPost, chatCompletionPathForRequest(r), bytes.NewReader(chatBody))
 	if err != nil {
 		h.sendError(w, http.StatusInternalServerError, "internal_error", "Failed to create request")
 		return
@@ -147,24 +167,29 @@ func (h *HTTPHandler) handleCreate(w http.ResponseWriter, r *http.Request) {
 
 	if req.Stream {
 		// Handle streaming response
-		h.handleStreaming(w, upstreamReq, resp)
+		h.handleStreaming(w, upstreamReq, resp, store)
 	} else {
 		// Handle non-streaming response
-		h.handleNonStreaming(w, upstreamReq, resp)
+		h.handleNonStreaming(w, upstreamReq, resp, store)
 	}
 }
 
 // handleStreaming handles streaming responses.
-func (h *HTTPHandler) handleStreaming(w http.ResponseWriter, upstreamReq *http.Request, resp *Response) {
+func (h *HTTPHandler) handleStreaming(w http.ResponseWriter, upstreamReq *http.Request, resp *Response, store bool) {
+	responseStore := h.store
+	if !store {
+		responseStore = nil
+	}
+
 	// Create streaming writer
-	streamWriter := NewStreamingResponseWriter(w, resp, h.store)
+	streamWriter := NewStreamingResponseWriter(w, resp, responseStore)
 
 	// Forward to scheduler
 	h.schedulerHTTP.ServeHTTP(streamWriter, upstreamReq)
 }
 
 // handleNonStreaming handles non-streaming responses.
-func (h *HTTPHandler) handleNonStreaming(w http.ResponseWriter, upstreamReq *http.Request, resp *Response) {
+func (h *HTTPHandler) handleNonStreaming(w http.ResponseWriter, upstreamReq *http.Request, resp *Response, store bool) {
 	// Capture upstream response
 	capture := NewNonStreamingResponseCapture()
 
@@ -187,7 +212,9 @@ func (h *HTTPHandler) handleNonStreaming(w http.ResponseWriter, upstreamReq *htt
 				Code:    errResp.Error.Code,
 				Message: errResp.Error.Message,
 			}
-			h.store.Save(resp)
+			if store {
+				h.store.Save(resp)
+			}
 			h.sendJSON(w, capture.StatusCode, resp)
 			return
 		}
@@ -197,7 +224,9 @@ func (h *HTTPHandler) handleNonStreaming(w http.ResponseWriter, upstreamReq *htt
 			Code:    "upstream_error",
 			Message: capture.Body.String(),
 		}
-		h.store.Save(resp)
+		if store {
+			h.store.Save(resp)
+		}
 		h.sendJSON(w, capture.StatusCode, resp)
 		return
 	}
@@ -210,7 +239,9 @@ func (h *HTTPHandler) handleNonStreaming(w http.ResponseWriter, upstreamReq *htt
 			Code:    "parse_error",
 			Message: "Failed to parse upstream response",
 		}
-		h.store.Save(resp)
+		if store {
+			h.store.Save(resp)
+		}
 		h.sendJSON(w, http.StatusInternalServerError, resp)
 		return
 	}
@@ -222,6 +253,8 @@ func (h *HTTPHandler) handleNonStreaming(w http.ResponseWriter, upstreamReq *htt
 	finalResp.Temperature = resp.Temperature
 	finalResp.TopP = resp.TopP
 	finalResp.MaxOutputTokens = resp.MaxOutputTokens
+	finalResp.Store = resp.Store
+	finalResp.Text = resp.Text
 	finalResp.Tools = resp.Tools
 	finalResp.ToolChoice = resp.ToolChoice
 	finalResp.ParallelToolCalls = resp.ParallelToolCalls
@@ -232,7 +265,9 @@ func (h *HTTPHandler) handleNonStreaming(w http.ResponseWriter, upstreamReq *htt
 	finalResp.CreatedAt = resp.CreatedAt
 
 	// Store the response
-	h.store.Save(finalResp)
+	if store {
+		h.store.Save(finalResp)
+	}
 
 	// Send response
 	h.sendJSON(w, http.StatusOK, finalResp)
@@ -332,6 +367,54 @@ func nilIfEmpty(s string) *string {
 	return &s
 }
 
+func shouldStore(req *CreateRequest) bool {
+	return req.Store == nil || *req.Store
+}
+
+func chatCompletionPathForRequest(r *http.Request) string {
+	if backend := r.PathValue("backend"); backend != "" {
+		return "/engines/" + backend + "/v1/chat/completions"
+	}
+	return "/engines/v1/chat/completions"
+}
+
+func validateUnsupportedRequestFields(req *CreateRequest) error {
+	if len(req.Include) > 0 {
+		return fmt.Errorf("include is not supported by Docker Model Runner Responses compatibility layer")
+	}
+	if req.StreamOptions != nil && !isNullJSON(req.StreamOptions) {
+		return fmt.Errorf("stream_options is not supported by Docker Model Runner Responses compatibility layer")
+	}
+	if req.TopLogprobs != nil {
+		return fmt.Errorf("top_logprobs is not supported by Docker Model Runner Responses compatibility layer")
+	}
+	switch req.Truncation {
+	case "", "disabled":
+	default:
+		return fmt.Errorf("truncation value %q is not supported by Docker Model Runner Responses compatibility layer", req.Truncation)
+	}
+	if req.Background != nil && *req.Background {
+		return fmt.Errorf("background responses are not supported by Docker Model Runner Responses compatibility layer")
+	}
+	if req.Conversation != nil && !isNullJSON(req.Conversation) {
+		return fmt.Errorf("conversation is not supported by Docker Model Runner Responses compatibility layer; use previous_response_id instead")
+	}
+	if req.Prompt != nil && !isNullJSON(req.Prompt) {
+		return fmt.Errorf("prompt is not supported by Docker Model Runner Responses compatibility layer")
+	}
+	if req.ServiceTier != "" {
+		return fmt.Errorf("service_tier is not supported by Docker Model Runner Responses compatibility layer")
+	}
+	if req.SafetyIdentifier != "" {
+		return fmt.Errorf("safety_identifier is not supported by Docker Model Runner Responses compatibility layer")
+	}
+	return nil
+}
+
+func isNullJSON(raw json.RawMessage) bool {
+	return strings.TrimSpace(string(raw)) == "null"
+}
+
 // GetStore returns the response store (for testing).
 func (h *HTTPHandler) GetStore() *Store {
 	return h.store