feat: add gemini3 support

.gitignore

@@ -19,4 +19,6 @@
 .idea
 
 # Generated by tests
-test.mp3
+test.mp3
+
+go.sum

chat.go

@@ -19,6 +19,13 @@ const (
 	ChatMessageRoleDeveloper = "developer"
 )
 
+const (
+	ReasoningEffortHigh    = "high"
+	ReasoningEffortMedium  = "medium"
+	ReasoningEffortLow     = "low"
+	ReasoningEffortMinimal = "minimal"
+)
+
 const chatCompletionsSuffix = "/chat/completions"
 
 var (
@@ -92,13 +99,23 @@ type ChatMessagePart struct {
 	Type     ChatMessagePartType  `json:"type,omitempty"`
 	Text     string               `json:"text,omitempty"`
 	ImageURL *ChatMessageImageURL `json:"image_url,omitempty"`
+	// ExtraPart is a vendor-specific extension container. For example, Vertex AI's
+	// OpenAI-compatible endpoint returns thought signatures inside
+	// content parts under extra_part.google.thought_signature.
+	ExtraPart map[string]any `json:"extra_part,omitempty"`
+	// ExtraContent carries vendor-specific data that would otherwise be stripped by
+	// the OpenAI-compatible schema. Vertex AI allows per-part extra_content.
+	ExtraContent map[string]any `json:"extra_content,omitempty"`
 }
 
 type ChatCompletionMessage struct {
 	Role         string `json:"role"`
 	Content      string `json:"content,omitempty"`
 	Refusal      string `json:"refusal,omitempty"`
 	MultiContent []ChatMessagePart
+	// ExtraContent is a vendor-specific extension container. For example, Vertex AI
+	// uses extra_content to return additional metadata alongside content arrays.
+	ExtraContent map[string]any `json:"extra_content,omitempty"`
 
 	// This property isn't in the official documentation, but it's in
 	// the documentation for the official library for python:
@@ -128,15 +145,25 @@ func (m ChatCompletionMessage) MarshalJSON() ([]byte, error) {
 	if len(m.MultiContent) > 0 {
 		msg := struct {
 			Role             string            `json:"role"`
-			Content          string            `json:"-"`
 			Refusal          string            `json:"refusal,omitempty"`
 			MultiContent     []ChatMessagePart `json:"content,omitempty"`
 			Name             string            `json:"name,omitempty"`
 			ReasoningContent string            `json:"reasoning_content,omitempty"`
 			FunctionCall     *FunctionCall     `json:"function_call,omitempty"`
 			ToolCalls        []ToolCall        `json:"tool_calls,omitempty"`
 			ToolCallID       string            `json:"tool_call_id,omitempty"`
-		}(m)
+			ExtraContent     map[string]any    `json:"extra_content,omitempty"`
+		}{
+			Role:             m.Role,
+			Refusal:          m.Refusal,
+			MultiContent:     m.MultiContent,
+			Name:             m.Name,
+			ReasoningContent: m.ReasoningContent,
+			FunctionCall:     m.FunctionCall,
+			ToolCalls:        m.ToolCalls,
+			ToolCallID:       m.ToolCallID,
+			ExtraContent:     m.ExtraContent,
+		}
 		return json.Marshal(msg)
 	}
 
@@ -150,7 +177,18 @@ func (m ChatCompletionMessage) MarshalJSON() ([]byte, error) {
 		FunctionCall     *FunctionCall     `json:"function_call,omitempty"`
 		ToolCalls        []ToolCall        `json:"tool_calls,omitempty"`
 		ToolCallID       string            `json:"tool_call_id,omitempty"`
-	}(m)
+		ExtraContent     map[string]any    `json:"extra_content,omitempty"`
+	}{
+		Role:             m.Role,
+		Content:          m.Content,
+		Refusal:          m.Refusal,
+		Name:             m.Name,
+		ReasoningContent: m.ReasoningContent,
+		FunctionCall:     m.FunctionCall,
+		ToolCalls:        m.ToolCalls,
+		ToolCallID:       m.ToolCallID,
+		ExtraContent:     m.ExtraContent,
+	}
 	return json.Marshal(msg)
 }
 
@@ -160,15 +198,25 @@ func (m *ChatCompletionMessage) UnmarshalJSON(bs []byte) error {
 		Content          string `json:"content"`
 		Refusal          string `json:"refusal,omitempty"`
 		MultiContent     []ChatMessagePart
-		Name             string        `json:"name,omitempty"`
-		ReasoningContent string        `json:"reasoning_content,omitempty"`
-		FunctionCall     *FunctionCall `json:"function_call,omitempty"`
-		ToolCalls        []ToolCall    `json:"tool_calls,omitempty"`
-		ToolCallID       string        `json:"tool_call_id,omitempty"`
+		Name             string         `json:"name,omitempty"`
+		ReasoningContent string         `json:"reasoning_content,omitempty"`
+		FunctionCall     *FunctionCall  `json:"function_call,omitempty"`
+		ToolCalls        []ToolCall     `json:"tool_calls,omitempty"`
+		ToolCallID       string         `json:"tool_call_id,omitempty"`
+		ExtraContent     map[string]any `json:"extra_content,omitempty"`
 	}{}
 
 	if err := json.Unmarshal(bs, &msg); err == nil {
-		*m = ChatCompletionMessage(msg)
+		m.Role = msg.Role
+		m.Content = msg.Content
+		m.Refusal = msg.Refusal
+		m.MultiContent = msg.MultiContent
+		m.Name = msg.Name
+		m.ReasoningContent = msg.ReasoningContent
+		m.FunctionCall = msg.FunctionCall
+		m.ToolCalls = msg.ToolCalls
+		m.ToolCallID = msg.ToolCallID
+		m.ExtraContent = msg.ExtraContent
 		return nil
 	}
 	multiMsg := struct {
@@ -181,11 +229,21 @@ func (m *ChatCompletionMessage) UnmarshalJSON(bs []byte) error {
 		FunctionCall     *FunctionCall     `json:"function_call,omitempty"`
 		ToolCalls        []ToolCall        `json:"tool_calls,omitempty"`
 		ToolCallID       string            `json:"tool_call_id,omitempty"`
+		ExtraContent     map[string]any    `json:"extra_content,omitempty"`
 	}{}
 	if err := json.Unmarshal(bs, &multiMsg); err != nil {
 		return err
 	}
-	*m = ChatCompletionMessage(multiMsg)
+	m.Role = multiMsg.Role
+	m.Content = multiMsg.Content
+	m.Refusal = multiMsg.Refusal
+	m.MultiContent = multiMsg.MultiContent
+	m.Name = multiMsg.Name
+	m.ReasoningContent = multiMsg.ReasoningContent
+	m.FunctionCall = multiMsg.FunctionCall
+	m.ToolCalls = multiMsg.ToolCalls
+	m.ToolCallID = multiMsg.ToolCallID
+	m.ExtraContent = multiMsg.ExtraContent
 	return nil
 }
 
@@ -195,6 +253,9 @@ type ToolCall struct {
 	ID       string       `json:"id,omitempty"`
 	Type     ToolType     `json:"type"`
 	Function FunctionCall `json:"function"`
+	// ExtraContent preserves provider-specific metadata (e.g. thought signatures)
+	// attached to tool calls, as used by Vertex AI.
+	ExtraContent map[string]any `json:"extra_content,omitempty"`
 }
 
 type FunctionCall struct {
@@ -307,7 +368,7 @@ type ChatCompletionRequest struct {
 	// Store can be set to true to store the output of this completion request for use in distillations and evals.
 	// https://platform.openai.com/docs/api-reference/chat/create#chat-create-store
 	Store bool `json:"store,omitempty"`
-	// Controls effort on reasoning for reasoning models. It can be set to "low", "medium", or "high".
+	// Controls effort on reasoning for reasoning models. It can be set to "low", "medium", "high" or "minimal".
 	ReasoningEffort string `json:"reasoning_effort,omitempty"`
 	// Metadata to store with the completion.
 	Metadata map[string]string `json:"metadata,omitempty"`
@@ -333,6 +394,9 @@ type ChatCompletionRequest struct {
 	SafetyIdentifier string `json:"safety_identifier,omitempty"`
 	// Embedded struct for non-OpenAI extensions
 	ChatCompletionRequestExtensions
+	// ExtraBody lets provider-specific fields pass through. Vertex AI uses
+	// extra_body.google.* for features like thinking_config and safety_settings.
+	ExtraBody map[string]any `json:"extra_body,omitempty"`
 }
 
 type StreamOptions struct {

chat_stream.go

@@ -2,15 +2,17 @@ package openai
 
 import (
 	"context"
+	"encoding/json"
 	"net/http"
 )
 
 type ChatCompletionStreamChoiceDelta struct {
-	Content      string        `json:"content,omitempty"`
-	Role         string        `json:"role,omitempty"`
-	FunctionCall *FunctionCall `json:"function_call,omitempty"`
-	ToolCalls    []ToolCall    `json:"tool_calls,omitempty"`
-	Refusal      string        `json:"refusal,omitempty"`
+	Content      string            `json:"content,omitempty"`
+	MultiContent []ChatMessagePart `json:"-"`
+	Role         string            `json:"role,omitempty"`
+	FunctionCall *FunctionCall     `json:"function_call,omitempty"`
+	ToolCalls    []ToolCall        `json:"tool_calls,omitempty"`
+	Refusal      string            `json:"refusal,omitempty"`
 
 	// This property is used for the "reasoning" feature supported by deepseek-reasoner
 	// which is not in the official documentation.
@@ -19,6 +21,81 @@ type ChatCompletionStreamChoiceDelta struct {
 	ReasoningContent string `json:"reasoning_content,omitempty"`
 }
 
+func (d *ChatCompletionStreamChoiceDelta) UnmarshalJSON(bs []byte) error {
+	// Probe the content type using json.RawMessage to support both string and array payloads.
+	type probe struct {
+		Content          json.RawMessage `json:"content,omitempty"`
+		Role             string          `json:"role,omitempty"`
+		FunctionCall     *FunctionCall   `json:"function_call,omitempty"`
+		ToolCalls        []ToolCall      `json:"tool_calls,omitempty"`
+		Refusal          string          `json:"refusal,omitempty"`
+		ReasoningContent string          `json:"reasoning_content,omitempty"`
+	}
+
+	var p probe
+	if err := json.Unmarshal(bs, &p); err != nil {
+		return err
+	}
+
+	// No content — simply copy scalar fields.
+	if len(p.Content) == 0 {
+		*d = ChatCompletionStreamChoiceDelta{
+			Role:             p.Role,
+			FunctionCall:     p.FunctionCall,
+			ToolCalls:        p.ToolCalls,
+			Refusal:          p.Refusal,
+			ReasoningContent: p.ReasoningContent,
+		}
+		return nil
+	}
+
+	// content is a JSON string.
+	if len(p.Content) > 0 && p.Content[0] == '"' {
+		var r struct {
+			Content          string        `json:"content,omitempty"`
+			Role             string        `json:"role,omitempty"`
+			FunctionCall     *FunctionCall `json:"function_call,omitempty"`
+			ToolCalls        []ToolCall    `json:"tool_calls,omitempty"`
+			Refusal          string        `json:"refusal,omitempty"`
+			ReasoningContent string        `json:"reasoning_content,omitempty"`
+		}
+		if err := json.Unmarshal(bs, &r); err != nil {
+			return err
+		}
+		*d = ChatCompletionStreamChoiceDelta{
+			Content:          r.Content,
+			Role:             r.Role,
+			FunctionCall:     r.FunctionCall,
+			ToolCalls:        r.ToolCalls,
+			Refusal:          r.Refusal,
+			ReasoningContent: r.ReasoningContent,
+		}
+		return nil
+	}
+
+	// Otherwise treat content as an array of parts.
+	var a struct {
+		MultiContent     []ChatMessagePart `json:"content,omitempty"`
+		Role             string            `json:"role,omitempty"`
+		FunctionCall     *FunctionCall     `json:"function_call,omitempty"`
+		ToolCalls        []ToolCall        `json:"tool_calls,omitempty"`
+		Refusal          string            `json:"refusal,omitempty"`
+		ReasoningContent string            `json:"reasoning_content,omitempty"`
+	}
+	if err := json.Unmarshal(bs, &a); err != nil {
+		return err
+	}
+	*d = ChatCompletionStreamChoiceDelta{
+		MultiContent:     a.MultiContent,
+		Role:             a.Role,
+		FunctionCall:     a.FunctionCall,
+		ToolCalls:        a.ToolCalls,
+		Refusal:          a.Refusal,
+		ReasoningContent: a.ReasoningContent,
+	}
+	return nil
+}
+
 type ChatCompletionStreamChoiceLogprobs struct {
 	Content []ChatCompletionTokenLogprob `json:"content,omitempty"`
 	Refusal []ChatCompletionTokenLogprob `json:"refusal,omitempty"`

chat_test.go

@@ -1205,3 +1205,72 @@ func TestChatCompletionRequest_UnmarshalJSON(t *testing.T) {
 		})
 	}
 }
+
+func TestChatCompletionMessageExtraFields(t *testing.T) {
+	raw := []byte(`{
+		"role":"assistant",
+		"content":[{"type":"text","text":"Hi","extra_part":{"google":{"thought_signature":"sig"}},"extra_content":{"google":{"foo":"bar"}}}],
+		"extra_content":{"google":{"safety_rationale":"ok"}},
+		"tool_calls":[{"id":"call_1","type":"function","function":{"name":"f","arguments":"{}"},"extra_content":{"google":{"thought_signature":"sig2"}}}]
+	}`)
+
+	var msg openai.ChatCompletionMessage
+	if err := json.Unmarshal(raw, &msg); err != nil {
+		t.Fatalf("unmarshal failed: %v", err)
+	}
+
+	if len(msg.MultiContent) != 1 {
+		t.Fatalf("expected 1 content part, got %d", len(msg.MultiContent))
+	}
+	if msg.MultiContent[0].ExtraPart == nil || msg.MultiContent[0].ExtraPart["google"] == nil {
+		t.Fatalf("expected extra_part to round-trip")
+	}
+	if msg.ExtraContent == nil || msg.ExtraContent["google"] == nil {
+		t.Fatalf("expected message extra_content to round-trip")
+	}
+	if len(msg.ToolCalls) != 1 || msg.ToolCalls[0].ExtraContent == nil {
+		t.Fatalf("expected tool call extra_content to round-trip")
+	}
+
+	out, err := json.Marshal(msg)
+	if err != nil {
+		t.Fatalf("marshal failed: %v", err)
+	}
+	if !strings.Contains(string(out), "extra_part") || !strings.Contains(string(out), "extra_content") {
+		t.Fatalf("marshal did not preserve extra fields: %s", string(out))
+	}
+}
+
+func TestChatCompletionRequestExtraBodyMarshal(t *testing.T) {
+	req := openai.ChatCompletionRequest{
+		Model: openai.Gemini3FlashPreview,
+		Messages: []openai.ChatCompletionMessage{{
+			Role:    openai.ChatMessageRoleUser,
+			Content: "hello",
+		}},
+		ExtraBody: map[string]any{
+			"google": map[string]any{
+				"include_thoughts": true,
+			},
+		},
+	}
+
+	out, err := json.Marshal(req)
+	if err != nil {
+		t.Fatalf("marshal failed: %v", err)
+	}
+	if !strings.Contains(string(out), `"extra_body"`) {
+		t.Fatalf("expected extra_body in marshaled JSON, got: %s", string(out))
+	}
+}
+
+func TestChatCompletionStreamChoiceDeltaMultiContent(t *testing.T) {
+	raw := []byte(`{"content":[{"type":"text","text":"Hello","extra_part":{"google":{"thought_signature":"sig"}}}]}`)
+	var d openai.ChatCompletionStreamChoiceDelta
+	if err := json.Unmarshal(raw, &d); err != nil {
+		t.Fatalf("unmarshal failed: %v", err)
+	}
+	if len(d.MultiContent) != 1 || d.MultiContent[0].ExtraPart == nil {
+		t.Fatalf("expected multi content with extra_part, got %#v", d)
+	}
+}

completion.go

@@ -53,14 +53,18 @@ const (
 	GPT5Mini                = "gpt-5-mini"
 	GPT5Nano                = "gpt-5-nano"
 	GPT5ChatLatest          = "gpt-5-chat-latest"
-	GPT3Dot5Turbo0125       = "gpt-3.5-turbo-0125"
-	GPT3Dot5Turbo1106       = "gpt-3.5-turbo-1106"
-	GPT3Dot5Turbo0613       = "gpt-3.5-turbo-0613"
-	GPT3Dot5Turbo0301       = "gpt-3.5-turbo-0301"
-	GPT3Dot5Turbo16K        = "gpt-3.5-turbo-16k"
-	GPT3Dot5Turbo16K0613    = "gpt-3.5-turbo-16k-0613"
-	GPT3Dot5Turbo           = "gpt-3.5-turbo"
-	GPT3Dot5TurboInstruct   = "gpt-3.5-turbo-instruct"
+	// Google Vertex AI (OpenAI-compatible) Gemini 3 preview model.
+	// Vertex OpenAI endpoint expects publisher/model format.
+	Gemini3FlashPreview   = "google/gemini-3-flash-preview"
+	Gemini3ProPreview     = "google/gemini-3-pro-preview"
+	GPT3Dot5Turbo0125     = "gpt-3.5-turbo-0125"
+	GPT3Dot5Turbo1106     = "gpt-3.5-turbo-1106"
+	GPT3Dot5Turbo0613     = "gpt-3.5-turbo-0613"
+	GPT3Dot5Turbo0301     = "gpt-3.5-turbo-0301"
+	GPT3Dot5Turbo16K      = "gpt-3.5-turbo-16k"
+	GPT3Dot5Turbo16K0613  = "gpt-3.5-turbo-16k-0613"
+	GPT3Dot5Turbo         = "gpt-3.5-turbo"
+	GPT3Dot5TurboInstruct = "gpt-3.5-turbo-instruct"
 	// Deprecated: Model is shutdown. Use gpt-3.5-turbo-instruct instead.
 	GPT3TextDavinci003 = "text-davinci-003"
 	// Deprecated: Model is shutdown. Use gpt-3.5-turbo-instruct instead.
@@ -101,6 +105,8 @@ const (
 
 var disabledModelsForEndpoints = map[string]map[string]bool{
 	"/completions": {
+		Gemini3FlashPreview:     true,
+		Gemini3ProPreview:       true,
 		O1Mini:                  true,
 		O1Mini20240912:          true,
 		O1Preview:               true,