🐛 fix: use JSON object for video image reference (#14900)

.github/workflows/claude-pr-assign.yml

@@ -21,6 +21,46 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v6
 
+      # Remind contributors when a non-release PR targets `main`.
+      # Day-to-day PRs should target `canary`; `main` is reserved for releases
+      # (see .agents/skills/version-release/SKILL.md). Allowed exceptions:
+      #   - PR title matches `🚀 release: v{x.y.z}` (minor release)
+      #   - head branch matches `hotfix/*` or `release/*` (patch release)
+      - name: Remind contributor if base branch is not canary
+        if: github.event.action == 'opened' && github.event.pull_request.base.ref == 'main'
+        env:
+          HEAD_REF: ${{ github.event.pull_request.head.ref }}
+          PR_TITLE: ${{ github.event.pull_request.title }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          GH_TOKEN: ${{ secrets.GH_TOKEN }}
+        run: |
+          if [[ "$HEAD_REF" == hotfix/* ]] || [[ "$HEAD_REF" == release/* ]]; then
+            echo "✅ Release/hotfix branch ($HEAD_REF) -> main is allowed"
+            exit 0
+          fi
+          if [[ "$PR_TITLE" =~ ^🚀[[:space:]]+release: ]]; then
+            echo "✅ Release-titled PR -> main is allowed"
+            exit 0
+          fi
+          echo "⚠️ Non-release PR targets main; posting reminder comment."
+          gh pr comment "$PR_NUMBER" --body "$(cat <<'EOF'
+          👋 Thanks for your contribution!
+
+          This PR currently targets the **`main`** branch, but `main` is reserved for release PRs only. Day-to-day development (features, fixes, refactors, docs, etc.) should target the **`canary`** branch.
+
+          ### How to fix
+
+          On the PR page, click **Edit** next to the title, then change the base branch from `main` to `canary`.
+
+          ### When targeting `main` is allowed
+
+          - PR title starts with `🚀 release: v{x.y.z}` (minor release)
+          - Head branch matches `hotfix/*` or `release/*` (patch release)
+
+          If your PR fits one of these cases, please ignore this message.
+          EOF
+          )"
+
       - name: Check if author is a team member
         id: check-team
         run: |

packages/model-runtime/src/core/openaiCompatibleFactory/createVideo.test.ts

@@ -19,6 +19,12 @@ const mockOptions: CreateVideoOptions = {
   provider: 'openai',
 };
 
+const mockVllmOptions: CreateVideoOptions = {
+  apiKey: 'test-api-key',
+  baseURL: 'http://localhost:8000/v1',
+  provider: 'vllm',
+};
+
 beforeEach(() => {
   vi.clearAllMocks();
 });
@@ -95,7 +101,7 @@ describe('createOpenAICompatibleVideo', () => {
       expect(body.size).toBe('1920x1080');
     });
 
-    it('should include imageUrl as input_reference', async () => {
+    it('should include imageUrl as JSON input_reference object', async () => {
       global.fetch = vi.fn().mockResolvedValueOnce({
         ok: true,
         json: async () => ({ id: 'video-task-img' }),
@@ -111,6 +117,26 @@ describe('createOpenAICompatibleVideo', () => {
 
       await createOpenAICompatibleVideo(payload, mockOptions);
 
+      const body = JSON.parse((global.fetch as any).mock.calls[0][1].body);
+      expect(body.input_reference).toEqual({ image_url: 'https://example.com/image.jpg' });
+    });
+
+    it('should preserve string input_reference for non-OpenAI compatible providers', async () => {
+      global.fetch = vi.fn().mockResolvedValueOnce({
+        ok: true,
+        json: async () => ({ id: 'video-task-vllm-img' }),
+      });
+
+      const payload: CreateVideoPayload = {
+        model: 'vllm-omni',
+        params: {
+          prompt: 'Continue this scene',
+          imageUrl: 'https://example.com/image.jpg',
+        },
+      };
+
+      await createOpenAICompatibleVideo(payload, mockVllmOptions);
+
       const body = JSON.parse((global.fetch as any).mock.calls[0][1].body);
       expect(body.input_reference).toBe('https://example.com/image.jpg');
     });

packages/model-runtime/src/core/openaiCompatibleFactory/createVideo.ts

@@ -1,4 +1,5 @@
 import createDebug from 'debug';
+import { ModelProvider } from 'model-bank';
 
 import type {
   CreateVideoPayload,
@@ -115,7 +116,7 @@ export async function pollOpenAICompatibleVideoStatus(
  *   model: string,
  *   prompt: string,
  *   seconds?: string,      // OpenAI Sora format (string type)
- *   input_reference?: string | object,  // For image-to-video
+ *   input_reference?: string | { image_url: string } | { file_id: string },  // For image-to-video
  * }
  *
  * Creates a video generation task and returns immediately with inferenceId.
@@ -150,7 +151,10 @@ export async function createOpenAICompatibleVideo(
 
   // Image-to-video support
   if (imageUrl) {
-    body['input_reference'] = imageUrl;
+    // OpenAI JSON requests reject bare strings, for example:
+    // `input_reference: "https://example.com/image.jpg"`.
+    body['input_reference'] =
+      options.provider === ModelProvider.OpenAI ? { image_url: imageUrl } : imageUrl;
   }
 
   log('OpenAI-compatible video API request body: %O', body);