fix(provider) prevent audio as image url

[afjcjsbx]

📝 Description

This PR fixes a critical bug where transcribed voice messages were causing text-only LLMs (like Zhipu glm-4.7-flash) to crash with a 404 error.

Previously, after the transcriber processed an audio file, the audio reference remained in the msg.Media array. The openai_compat provider then blindly converted all media files into image_url blocks. When sending an image_url payload to a text-only model, the API rejected the request entirely.

Changes made:

• Updated pkg/providers/openai_compat/provider.go (serializeMessages) to strictly filter media types. It now only appends image_url blocks for actual images (strings.HasPrefix(mediaURL, "data:image/")). Transcribed audio is safely ignored at the serialization layer since its content is already injected as text.

🗣️ Type of Change

• 🐞 Bug fix (non-breaking change which fixes an issue)
• ✨ New feature (non-breaking change which adds functionality)
• 📖 Documentation update
• ⚡ Code refactoring (no functional changes, no api changes)

🤖 AI Code Generation

• 🤖 Fully AI-generated (100% AI, 0% Human)
• 🛠️ Mostly AI-generated (AI draft, Human verified/modified)
• 👨‍💻 Mostly Human-written (Human lead, AI assisted or none)

🔗 Related Issue

📚 Technical Context (Skip for Docs)

• Reference URL: N/A
• Reasoning:

🧪 Test Environment

• Hardware: PC
• OS: macOS
• Model/Provider: Zhipu / glm-4.7-flash
• Channels: Telegram

📸 Evidence (Optional)

Click to view Logs/Screenshots

Before the fix (Error Log):

2026/03/04 19:03:49 [DEBUG] voice: Sending transcription request to Groq API
2026/03/04 19:03:50 [INFO] voice: Transcription completed successfully {text_length=4, transcription_preview= Hi.}
...
2026/03/04 19:03:50 [ERROR] agent: LLM call failed {iteration=1, error=API request failed:
  Status: 404
  Body:   {"error":{"message":"No endpoints found that support image input","code":404}}, agent_id=main}