旧式openai接口 /v1/chat/completions 是不是不支持 tool 中返回图片的base64多模态 #2313

Problem

In the legacy OpenAI-compatible endpoint /v1/chat/completions, image content works correctly when it is sent in a normal user message.

However, if a tool message returns image content (for example structured array content containing image_url or file-based multimodal parts), the Codex path fails and the client receives:

empty_stream: upstream stream closed before first payload

Current behavior

• user message with image input: works
• tool message with image content: fails

This suggests the issue is in the translator path for tool-role content, not in general multimodal handling.

Expected behavior

Tool-role content should preserve multimodal payloads in the same way user-role content does, as long as the downstream Responses payload remains valid.

Suspected area

• internal/translator/codex/openai/chat-completions/codex_openai_request.go

The translator likely needs to support structured tool content arrays and correctly map multimodal parts for function_call_output.output.

Required behavior

Please update the translator so that:

• structured tool content arrays can preserve supported multimodal parts such as text, image_url, and valid file parts
• input_image is emitted only when a real image source exists (url or file_id)
• input_file is emitted only when a real file source exists (file_id, file_data, or file_url)
• unsupported or incomplete parts are not silently dropped in a lossy way; they should fall back to raw text or serialized JSON
• function_call_output.output should remain present even when the original tool content is not a string/array (for example null or other JSON values)

Reproduction hints

Observed symptom:

empty_stream: upstream stream closed before first payload

Repro pattern:

1. send request to /v1/chat/completions
2. let the model call a tool
3. return tool content containing image data or multimodal structured content
4. observe upstream stream closes before first payload

Regression coverage

Please add tests covering:

• tool content with image_url
• tool content with valid file parts
• filename-only file parts should not emit invalid input_file
• image parts without source should not emit invalid input_image
• content: null
• unknown tool content part types should not be silently dropped

Related context

• Related PR attempt: #2349