Add OpenAI Responses API (/v1/responses) endpoint

Translate Responses API requests (input→messages, instructions→system, max_output_tokens→max_tokens, text.format→response_format, flat tools→nested {function:{…}}) to upstream chat/completions, then convert responses back to Responses format (streaming SSE event lifecycle + non-streaming JSON). Verified against OpenAI migration guide and Python SDK Response model: - Echo back required fields parallel_tool_calls/tool_choice/tools - Include content:[] in reasoning items, logprobs:[] in output_text parts - Support function_call/function_call_output multi-turn input items - Map usage fields prompt_tokens→input_tokens, completion_tokens→output_tokens
2026-08-20 10:02:11 +08:00
parent c8938cb514
commit 2517b4f730
6 changed files with 1576 additions and 4 deletions
@@ -0,0 +1,347 @@
+package openai
+
+import (
+	"encoding/json"
+	"errors"
+)
+
+// --- Responses API request parsing & conversion ---
+
+// ParseResponsesRequest converts a raw OpenAI Responses API (POST /v1/responses)
+// request body into a ChatCompletionRequest suitable for the upstream gateway.
+// It translates input→messages, instructions→system message, max_output_tokens→
+// max_tokens, and reshapes tools from the Responses flat format to the Chat
+// Completions nested {function:{…}} format.
+func ParseResponsesRequest(body []byte) (ChatCompletionRequest, error) {
+	var raw map[string]json.RawMessage
+	if err := json.Unmarshal(body, &raw); err != nil {
+		return ChatCompletionRequest{}, err
+	}
+
+	var req ChatCompletionRequest
+	if v, ok := raw["model"]; ok {
+		_ = json.Unmarshal(v, &req.Model)
+	}
+	if v, ok := raw["stream"]; ok {
+		_ = json.Unmarshal(v, &req.Stream)
+	}
+
+	// Build messages from instructions + input.
+	messages := make([]map[string]any, 0, 4)
+	if v, ok := raw["instructions"]; ok {
+		var s string
+		if json.Unmarshal(v, &s) == nil && s != "" {
+			messages = append(messages, map[string]any{"role": "system", "content": s})
+		}
+	}
+	if v, ok := raw["input"]; ok {
+		msgs, err := responsesInputToMessages(v)
+		if err != nil {
+			return req, err
+		}
+		messages = append(messages, msgs...)
+	}
+	req.Messages = messages
+
+	// Extra: renamed + pass-through fields sent to upstream as-is.
+	extra := map[string]json.RawMessage{}
+	if v, ok := raw["max_output_tokens"]; ok {
+		extra["max_tokens"] = v
+	}
+	for _, key := range []string{
+		"temperature", "top_p", "top_k", "frequency_penalty",
+		"presence_penalty", "stop", "seed", "user",
+	} {
+		if v, ok := raw[key]; ok {
+			extra[key] = v
+		}
+	}
+	if v, ok := raw["tools"]; ok {
+		if translated, err := translateResponsesTools(v); err == nil {
+			extra["tools"] = translated
+		} else {
+			extra["tools"] = v // fall back to pass-through
+		}
+	}
+	if v, ok := raw["tool_choice"]; ok {
+		extra["tool_choice"] = v
+	}
+	// Structured Outputs: Responses API uses text.format instead of
+	// response_format. Translate text.format → response_format for the
+	// upstream Chat Completions endpoint.
+	if v, ok := raw["text"]; ok {
+		if rf, err := translateTextFormat(v); err == nil {
+			extra["response_format"] = rf
+		}
+	}
+	req.Extra = extra
+	return req, nil
+}
+
+// translateTextFormat converts the Responses API "text" field (containing a
+// "format" sub-object) into a Chat Completions response_format value.
+//   - {text:{format:{type:"json_object"}}}        → {type:"json_object"}
+//   - {text:{format:{type:"json_schema",name,schema,strict}}} →
+//     {type:"json_schema",json_schema:{name,schema,strict}}
+func translateTextFormat(textRaw json.RawMessage) (json.RawMessage, error) {
+	var text struct {
+		Format map[string]any `json:"format"`
+	}
+	if err := json.Unmarshal(textRaw, &text); err != nil {
+		return nil, err
+	}
+	if len(text.Format) == 0 {
+		return nil, errors.New("empty text.format")
+	}
+	t, _ := text.Format["type"].(string)
+	switch t {
+	case "json_object":
+		return json.Marshal(map[string]any{"type": "json_object"})
+	case "json_schema":
+		js := map[string]any{}
+		for _, k := range []string{"name", "schema", "strict"} {
+			if v, ok := text.Format[k]; ok {
+				js[k] = v
+			}
+		}
+		return json.Marshal(map[string]any{"type": "json_schema", "json_schema": js})
+	default:
+		return json.Marshal(text.Format) // pass through unknown types
+	}
+}
+
+// ResponsesMeta holds fields from the Responses API request that should be
+// echoed back in the response object (the OpenAI SDK requires them).
+type ResponsesMeta struct {
+	ParallelToolCalls any
+	ToolChoice        any
+	Tools             any
+	Temperature       any
+	TopP              any
+	MaxOutputTokens   any
+}
+
+// ParseResponsesMeta extracts echo-back fields from the raw request body.
+func ParseResponsesMeta(body []byte) ResponsesMeta {
+	var m map[string]any
+	_ = json.Unmarshal(body, &m)
+	meta := ResponsesMeta{
+		ParallelToolCalls: false,
+		ToolChoice:        "auto",
+		Tools:             []any{},
+	}
+	if v, ok := m["parallel_tool_calls"]; ok {
+		meta.ParallelToolCalls = v
+	}
+	if v, ok := m["tool_choice"]; ok {
+		meta.ToolChoice = v
+	}
+	// Tools are echoed back in the Responses API flat format (not the
+	// translated Chat Completions format).
+	if v, ok := m["tools"]; ok {
+		meta.Tools = v
+	} else {
+		meta.Tools = []any{}
+	}
+	if v, ok := m["temperature"]; ok {
+		meta.Temperature = v
+	}
+	if v, ok := m["top_p"]; ok {
+		meta.TopP = v
+	}
+	if v, ok := m["max_output_tokens"]; ok {
+		meta.MaxOutputTokens = v
+	}
+	return meta
+}
+
+// responsesInputToMessages converts the Responses API "input" field (which may
+// be a plain string or an array of input items) into Chat Completions messages.
+func responsesInputToMessages(input json.RawMessage) ([]map[string]any, error) {
+	// Case 1: input is a plain string.
+	var s string
+	if json.Unmarshal(input, &s) == nil {
+		return []map[string]any{{"role": "user", "content": s}}, nil
+	}
+
+	// Case 2: input is an array of items.
+	var items []map[string]json.RawMessage
+	if err := json.Unmarshal(input, &items); err != nil {
+		return nil, err
+	}
+
+	messages := make([]map[string]any, 0, len(items))
+	for _, item := range items {
+		// Determine the type — most items are message-like with a role.
+		var role string
+		if v, ok := item["role"]; ok {
+			_ = json.Unmarshal(v, &role)
+		}
+		var itemType string
+		if v, ok := item["type"]; ok {
+			_ = json.Unmarshal(v, &itemType)
+		}
+
+		switch {
+		case itemType == "function_call":
+			// An assistant tool call from a previous turn.
+			msg := map[string]any{"role": "assistant"}
+			tc := map[string]any{"type": "function"}
+			inner := map[string]any{}
+			if v, ok := item["name"]; ok {
+				var name string
+				_ = json.Unmarshal(v, &name)
+				inner["name"] = name
+			}
+			if v, ok := item["arguments"]; ok {
+				var args string
+				_ = json.Unmarshal(v, &args)
+				inner["arguments"] = args
+			}
+			if v, ok := item["call_id"]; ok {
+				var id string
+				_ = json.Unmarshal(v, &id)
+				tc["id"] = id
+			}
+			tc["function"] = inner
+			msg["tool_calls"] = []any{tc}
+			messages = append(messages, msg)
+		case itemType == "function_call_output":
+			// A tool result from a previous turn.
+			msg := map[string]any{"role": "tool"}
+			if v, ok := item["call_id"]; ok {
+				var id string
+				_ = json.Unmarshal(v, &id)
+				msg["tool_call_id"] = id
+			}
+			if v, ok := item["output"]; ok {
+				var out string
+				_ = json.Unmarshal(v, &out)
+				msg["content"] = out
+			}
+			messages = append(messages, msg)
+		default:
+			// Standard message item with role + content.
+			if role == "" {
+				role = "user"
+			}
+			// "developer" maps to "system" for broad upstream compatibility.
+			if role == "developer" {
+				role = "system"
+			}
+			msg := map[string]any{"role": role}
+			if v, ok := item["content"]; ok {
+				msg["content"] = convertContentParts(v)
+			} else {
+				msg["content"] = ""
+			}
+			messages = append(messages, msg)
+		}
+	}
+	return messages, nil
+}
+
+// convertContentParts converts a Responses API content field (string or array
+// of content parts) into the Chat Completions content format.
+func convertContentParts(raw json.RawMessage) any {
+	// Content is a plain string.
+	var s string
+	if json.Unmarshal(raw, &s) == nil {
+		return s
+	}
+	// Content is an array of parts.
+	var parts []map[string]any
+	if json.Unmarshal(raw, &parts) != nil {
+		return string(raw) // fallback
+	}
+	result := make([]map[string]any, 0, len(parts))
+	for _, p := range parts {
+		pt, _ := p["type"].(string)
+		switch pt {
+		case "input_text", "output_text", "text":
+			result = append(result, map[string]any{"type": "text", "text": p["text"]})
+		case "input_image":
+			img := map[string]any{"type": "image_url"}
+			if url, ok := p["image_url"]; ok {
+				img["image_url"] = map[string]any{"url": url}
+			} else if d, ok := p["image"]; ok {
+				img["image_url"] = map[string]any{"url": d}
+			}
+			result = append(result, img)
+		default:
+			// Pass through unknown part types as-is.
+			result = append(result, p)
+		}
+	}
+	if len(result) == 0 {
+		return ""
+	}
+	return result
+}
+
+// translateResponsesTools converts tools from the Responses API flat format
+// to the Chat Completions nested {type:"function",function:{…}} format.
+func translateResponsesTools(raw json.RawMessage) (json.RawMessage, error) {
+	var tools []map[string]any
+	if err := json.Unmarshal(raw, &tools); err != nil {
+		return nil, err
+	}
+	out := make([]map[string]any, 0, len(tools))
+	for _, t := range tools {
+		tt, _ := t["type"].(string)
+		if tt != "function" && tt != "" {
+			// Non-function tool types (web_search, file_search, etc.) —
+			// pass through as-is; upstream may or may not support them.
+			out = append(out, t)
+			continue
+		}
+		fn := map[string]any{}
+		for _, key := range []string{"name", "description", "parameters", "strict"} {
+			if v, ok := t[key]; ok {
+				fn[key] = v
+			}
+		}
+		out = append(out, map[string]any{"type": "function", "function": fn})
+	}
+	return json.Marshal(out)
+}
+
+// --- Responses API response building ---
+
+// UsageToResponses converts a Chat Completions usage value (as decoded by
+// json.Unmarshal into any) to the Responses API usage field names
+// (input_tokens / output_tokens instead of prompt_tokens / completion_tokens).
+func UsageToResponses(v any) any {
+	if v == nil {
+		return nil
+	}
+	b, err := json.Marshal(v)
+	if err != nil {
+		return v
+	}
+	var m map[string]any
+	if json.Unmarshal(b, &m) != nil {
+		return v
+	}
+	out := map[string]any{}
+	if pt, ok := m["prompt_tokens"]; ok {
+		out["input_tokens"] = pt
+	}
+	if ct, ok := m["completion_tokens"]; ok {
+		out["output_tokens"] = ct
+	}
+	if tt, ok := m["total_tokens"]; ok {
+		out["total_tokens"] = tt
+	}
+	if d, ok := m["prompt_tokens_details"]; ok {
+		out["input_tokens_details"] = d
+	} else if d, ok := m["input_tokens_details"]; ok {
+		out["input_tokens_details"] = d
+	}
+	if d, ok := m["completion_tokens_details"]; ok {
+		out["output_tokens_details"] = d
+	} else if d, ok := m["output_tokens_details"]; ok {
+		out["output_tokens_details"] = d
+	}
+	return out
+}