@cortexkit/aft-opencode@latest causes `System message must be at the beginning` with Qwen via LiteLLM/vLLM

[hille721]

After adding @cortexkit/aft-opencode@latest to OpenCode plugins, requests to qwen3.6-27b through LiteLLM/vLLM fail with:

litellm.BadRequestError: Hosted_vllmException - {"error":{"message":"System message must be at the beginning.","type":"BadRequestError","param":null,"code":400}}

Environment

• OpenCode 1.17.13
• Plugins:
  • @cortexkit/aft-opencode@latest
  • @cortexkit/opencode-magic-context@latest
• Model: qwen3.6-27b
• Provider: LiteLLM
• Endpoint: /v1/chat/completions

Reproduction

1. Add @cortexkit/aft-opencode@latest to the plugin list
2. Start a fresh OpenCode session with qwen3.6-27b
3. Send the first prompt
4. Request fails with the error above

Observations

• This started immediately after adding @cortexkit/aft-opencode@latest
• A brand new session fails on the first turn
• Another model through the same LiteLLM instance works
• In a mixed session, using another model works, then switching to qwen3.6-27b fails again with the same error

Relevant code location

Installed package contains:

"experimental.chat.system.transform": async (_input, output) => {
  if (hintsBlock) {
    output.system.push(hintsBlock);
  }
},

From @cortexkit/aft-opencode dist:
dist/index.js around lines 36794-36797

[ramhaidar]

I dont know wether its related or not but DeepSeek showing this too in Pi

Error: OpenAI API error (400): {"message":"Error from provider (DeepSeek): Invalid schema for function 'read': schema must be a JSON Schema of 'type: \"object\"', got 'type: null'.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}