fix: support multimodal image content in OpenAI provider

src/agent.ts

@@ -229,7 +229,7 @@ export class Agent {
    * Run a query with streaming events.
    */
   async *query(
-    prompt: string,
+    prompt: string | any[],
     overrides?: Partial<AgentOptions>,
   ): AsyncGenerator<SDKMessage, void> {
     await this.setupDone
@@ -504,7 +504,7 @@ export function createAgent(options: AgentOptions = {}): Agent {
  * The agent is created, used, and cleaned up automatically.
  */
 export async function* query(params: {
-  prompt: string
+  prompt: string | any[]
   options?: AgentOptions
 }): AsyncGenerator<SDKMessage, void> {
   const ephemeral = createAgent(params.options)

src/providers/openai.ts

@@ -23,11 +23,15 @@ import type {
 
 interface OpenAIChatMessage {
   role: 'system' | 'user' | 'assistant' | 'tool'
-  content?: string | null
+  content?: string | OpenAIContentPart[] | null
   tool_calls?: OpenAIToolCall[]
   tool_call_id?: string
 }
 
+type OpenAIContentPart =
+  | { type: 'text'; text: string }
+  | { type: 'image_url'; image_url: { url: string; detail?: 'auto' | 'low' | 'high' } }
+
 interface OpenAIToolCall {
   id: string
   type: 'function'
@@ -153,13 +157,30 @@ export class OpenAIProvider implements LLMProvider {
       return
     }
 
-    // Content blocks may contain text and/or tool_result blocks
-    const textParts: string[] = []
+    // Content blocks may contain text, image, and/or tool_result blocks
+    const contentParts: OpenAIContentPart[] = []
     const toolResults: Array<{ tool_use_id: string; content: string }> = []
 
     for (const block of msg.content) {
       if (block.type === 'text') {
-        textParts.push(block.text)
+        contentParts.push({ type: 'text', text: block.text })
+      } else if (block.type === 'image') {
+        const source = block.source
+        let imageUrl: string
+        if (source?.type === 'base64' && source.data) {
+          const mediaType = source.media_type || 'image/png'
+          imageUrl = `data:${mediaType};base64,${source.data}`
+        } else if (source?.type === 'url' && source.url) {
+          imageUrl = source.url
+        } else if (typeof source === 'string') {
+          imageUrl = source
+        } else {
+          continue
+        }
+        contentParts.push({
+          type: 'image_url',
+          image_url: { url: imageUrl, detail: 'high' },
+        })
       } else if (block.type === 'tool_result') {
         toolResults.push({
           tool_use_id: block.tool_use_id,
@@ -177,9 +198,18 @@ export class OpenAIProvider implements LLMProvider {
       })
     }
 
-    // Text parts become a user message
-    if (textParts.length > 0) {
-      result.push({ role: 'user', content: textParts.join('\n') })
+    // Content parts become a user message
+    if (contentParts.length > 0) {
+      const hasImages = contentParts.some(p => p.type === 'image_url')
+      if (!hasImages) {
+        const text = contentParts
+          .filter((p): p is { type: 'text'; text: string } => p.type === 'text')
+          .map(p => p.text)
+          .join('\n')
+        result.push({ role: 'user', content: text })
+      } else {
+        result.push({ role: 'user', content: contentParts })
+      }
     }
   }