fix(media): drop response_format for gpt-image-* on /v1/images/edits (#5727) (#5760)

OpenAI rejects response_format on /v1/images/edits when the model is a
gpt-image-* model with HTTP 400 'Unknown parameter: response_format'.
The generations branch already avoids setting response_format for
gpt-image-*, but the edit (image-to-image) branch set it unconditionally
whenever a reference image was supplied.

Set response_format only for non-gpt-image-* models in the edit branch
to match the generations behaviour.

Author: xxiaoxiong

apps/daemon/src/media/index.ts

@@ -1246,7 +1246,13 @@ async function renderCustomOpenAIImage(ctx: MediaContext, credentials: ProviderC
   };
   let url = buildOpenAIImageUrl(baseUrl, false);
   if (ctx.imageRef?.dataUrl) {
-    body.response_format = 'b64_json';
+    // gpt-image-* does NOT accept response_format on /v1/images/edits
+    // (HTTP 400 "Unknown parameter: 'response_format'"). dall-e-2
+    // accepts it; the base b64 path is what callers expect either way,
+    // so we only set response_format for non-gpt-image models.
+    if (!wireModel.startsWith('gpt-image-')) {
+      body.response_format = 'b64_json';
+    }
     body.images = [{ image_url: ctx.imageRef.dataUrl }];
     url = buildOpenAIImageEditUrl(baseUrl);
   }