[Bug] Generated media history passes image metadata dicts as image URLs

[zhonggang666]

Problem Description

The generated media nodes store image_list entries as dictionaries, but later replay the whole dictionary as the image_url.url string in history messages.

For example, image generation stores:

self.context['image_list'] = [{'file_id': path.split('/')[-1], 'url': path} for path in file_urls]

But history replay uses:

*[{'type': 'image_url', 'image_url': {'url': f'{file_url}'}} for file_url in image_list]

Since each file_url is a dict, f'{file_url}' becomes a string like "{'file_id': '...', 'url': './oss/file/...'}", not an actual image URL.

Code Locations

• apps/application/flow/step_node/image_generate_step_node/impl/base_image_generate_node.py:74
• apps/application/flow/step_node/image_generate_step_node/impl/base_image_generate_node.py:87
• Same pattern also appears in:
  • apps/application/flow/step_node/image_to_video_step_node/impl/base_image_to_video_node.py:153
  • apps/application/flow/step_node/text_to_video_step_node/impl/base_text_to_video_node.py:130

Steps to Reproduce

1. Create a workflow using an image generation node with dialogue_type not equal to WORKFLOW and dialogue_number > 0.
2. Run the node once so it stores image_list in execution details.
3. Run a follow-up turn that includes history.

Actual Result

The previous image is replayed to the model as an invalid URL string representation of a Python dict, for example:

{'file_id': '...', 'url': './oss/file/...'}

The model cannot fetch or interpret the previous image correctly.

Expected Result

History replay should pass the actual URL field, for example file_url.get('url'), or convert file_id to a valid accessible file URL/base64 payload as other multimodal nodes do.

Impact

Multi-turn image generation / image-to-video / text-to-video workflows can lose visual context from prior turns, causing follow-up prompts to behave as if the previous generated image was unavailable.

MaxKB Version

Current v2 branch.