v0.4.0

CCProxy Agent v0.4 Release Notes

Summary

v0.4 adds multimodal image support for text-only downstream models.

Claude Desktop can send image blocks, but your downstream model may be GLM-5.2 or another text-only model. In that case, CCProxy Agent now intercepts the image, calls a dedicated vision model, injects a text summary, and strips the original image before forwarding downstream.

Default Vision Models

{
  "vision": {
    "enabled": true,
    "models": [
      "qwen3-vl-30b-a3b-instruct",
      "Qwen3.6-35B-A3B"
    ],
    "compareModels": true
  }
}

Both models are called by default. Their summaries are combined into one [VISION SUMMARY] block.

Secret Handling

The API key is not committed to config. Set it as a user environment variable:

[Environment]::SetEnvironmentVariable("CCPROXY_VISION_API_KEY", "your-token", "User")

Restart the proxy after setting the variable.

Flow

Claude Desktop image input
-> CCProxy Agent extracts image blocks
-> calls qwen3-vl-30b-a3b-instruct
-> calls Qwen3.6-35B-A3B
-> injects [VISION SUMMARY]
-> strips image blocks
-> forwards text-only request to ccswitch / GLM-5.2

Limits

Default limits:

{
  "maxImagesPerRequest": 5,
  "maxImageBytes": 5000000,
  "summaryMaxTokens": 1500
}

These guardrails prevent image payloads from exploding context size.

Validation

• TypeScript check passes.
• Unit tests pass.
• Added regression coverage for dual vision model calls.
• Live test confirmed qwen3-vl-30b-a3b-instruct can OCR the provided screenshot.
• Live test confirmed Qwen3.6-35B-A3B can return normal content when enable_thinking=false.