diff --git a/lib/multimodal.test.ts b/lib/multimodal.test.ts
index d6e2f6e..46ece4b 100644
--- a/lib/multimodal.test.ts
+++ b/lib/multimodal.test.ts
@@ -165,26 +165,23 @@ describe('buildApiContent — files', () => {
 // --- audio attachments ---
 
 describe('buildApiContent — audio', () => {
-  it('skips audio binary, keeps text content (transcript from Whisper)', () => {
+  it('returns plain string for audio-only message (transcript from Whisper)', () => {
     const result = buildApiContent(msg({
       content: 'Hello this is my transcribed message',
       media: [audioAttachment()],
     }))
-    const parts = result as Array<{ type: string; text?: string }>
-    // Should only have the text part, no audio binary
-    expect(parts).toHaveLength(1)
-    expect(parts[0]).toEqual({ type: 'text', text: 'Hello this is my transcribed message' })
+    // Audio is skipped and no other non-text parts exist, so return plain string
+    // to avoid wrapping in ContentPart[] which the gateway may not handle
+    expect(result).toBe('Hello this is my transcribed message')
   })
 
-  it('returns text content string when audio is the only attachment and content exists', () => {
+  it('returns plain string when audio is the only attachment', () => {
     const result = buildApiContent(msg({
       content: 'transcribed words',
       media: [audioAttachment()],
     }))
-    // With audio-only + text content, should return a parts array with just text
-    const parts = result as Array<{ type: string }>
-    expect(parts).toHaveLength(1)
-    expect(parts[0].type).toBe('text')
+    expect(typeof result).toBe('string')
+    expect(result).toBe('transcribed words')
   })
 })
 
diff --git a/lib/multimodal.ts b/lib/multimodal.ts
index 0bc202e..b3b6ab7 100644
--- a/lib/multimodal.ts
+++ b/lib/multimodal.ts
@@ -14,6 +14,7 @@ export function buildApiContent(msg: Message): MessageContent {
   if (!media || media.length === 0) return msg.content
 
   const parts: ContentPart[] = []
+  let attachmentAdded = false
 
   if (msg.content) {
     parts.push({ type: 'text', text: msg.content })
@@ -22,6 +23,7 @@ export function buildApiContent(msg: Message): MessageContent {
   for (const attachment of media) {
     if (attachment.type === 'image') {
       parts.push({ type: 'image_url', image_url: { url: attachment.url } })
+      attachmentAdded = true
     } else if (attachment.type === 'file') {
       const label = attachment.name || 'unknown'
       const sizeNote = attachment.size ? ` (${Math.round(attachment.size / 1024)} KB)` : ''
@@ -32,10 +34,16 @@ export function buildApiContent(msg: Message): MessageContent {
       } else {
         parts.push({ type: 'text', text: `[Attached file: ${label}${sizeNote}]` })
       }
+      attachmentAdded = true
     }
     // Audio: transcript already in msg.content via Whisper — skip binary
   }
 
+  // If no attachment actually contributed to parts (e.g., audio-only message
+  // where the transcript is already in msg.content), return a plain string
+  // so the gateway doesn't receive an unnecessary ContentPart[] wrapper.
+  if (!attachmentAdded) return msg.content
+
   return parts.length > 0 ? parts : msg.content
 }