Changeset 0.12.0 (#63)

server/providers/openai.go

@@ -37,8 +37,14 @@ func (p *OpenAIProvider) ChatCompletion(ctx context.Context, cfg ProviderConfig,
 		return nil, fmt.Errorf("no choices in response")
 	}
 
+	// Extract content string from response (responses are always string, not array)
+	var contentStr string
+	if s, ok := resp.Choices[0].Message.Content.(string); ok {
+		contentStr = s
+	}
+
 	result := &CompletionResponse{
-		Content:      resp.Choices[0].Message.Content,
+		Content:      contentStr,
 		Model:        resp.Model,
 		FinishReason: resp.Choices[0].FinishReason,
 		InputTokens:  resp.Usage.PromptTokens,
@@ -360,11 +366,42 @@ func (p *OpenAIProvider) doRequest(ctx context.Context, cfg ProviderConfig, req
 		})
 	}
 
-	// Convert messages — handle all roles including tool
+	// Convert messages — handle all roles including tool and multimodal
 	for _, m := range req.Messages {
 		oaiMsg := openaiMessage{
-			Role:    m.Role,
-			Content: m.Content,
+			Role: m.Role,
+		}
+
+		// Build content: multimodal parts or plain string
+		if len(m.ContentParts) > 0 && m.Role == "user" {
+			// Multimodal: convert to OpenAI content array format
+			parts := make([]openaiContentPart, 0, len(m.ContentParts))
+			for _, p := range m.ContentParts {
+				switch p.Type {
+				case "image_url":
+					if p.ImageURL != nil {
+						detail := p.ImageURL.Detail
+						if detail == "" {
+							detail = "auto"
+						}
+						parts = append(parts, openaiContentPart{
+							Type: "image_url",
+							ImageURL: &openaiImageURL{
+								URL:    p.ImageURL.URL,
+								Detail: detail,
+							},
+						})
+					}
+				default: // "text", "document" → text
+					parts = append(parts, openaiContentPart{
+						Type: "text",
+						Text: p.Text,
+					})
+				}
+			}
+			oaiMsg.Content = parts
+		} else {
+			oaiMsg.Content = m.Content
 		}
 
 		// Assistant messages with tool calls
@@ -425,11 +462,24 @@ func (p *OpenAIProvider) doRequest(ctx context.Context, cfg ProviderConfig, req
 // ── OpenAI Wire Types ───────────────────────
 
 type openaiMessage struct {
-	Role       string              `json:"role"`
-	Content    string              `json:"content,omitempty"`
-	ToolCalls  []openaiToolCall    `json:"tool_calls,omitempty"`   // assistant requesting tools
-	ToolCallID string              `json:"tool_call_id,omitempty"` // role=tool result
-	Name       string              `json:"name,omitempty"`         // tool name for role=tool
+	Role         string              `json:"role"`
+	Content      interface{}         `json:"content,omitempty"`      // string or []openaiContentPart
+	ToolCalls    []openaiToolCall    `json:"tool_calls,omitempty"`   // assistant requesting tools
+	ToolCallID   string              `json:"tool_call_id,omitempty"` // role=tool result
+	Name         string              `json:"name,omitempty"`         // tool name for role=tool
+}
+
+// openaiContentPart represents a single element in a multimodal content array.
+// OpenAI format: [{"type":"text","text":"..."}, {"type":"image_url","image_url":{"url":"..."}}]
+type openaiContentPart struct {
+	Type     string          `json:"type"`
+	Text     string          `json:"text,omitempty"`
+	ImageURL *openaiImageURL `json:"image_url,omitempty"`
+}
+
+type openaiImageURL struct {
+	URL    string `json:"url"`
+	Detail string `json:"detail,omitempty"`
 }
 
 type openaiToolCallFunction struct {