Useless compression and buggy contextWindowSize

[AliakseiLasevich]

What happened?

I'm not sure if the problem is with qwen code or my model, but when I limit the context, compression seems useless.
I use llama.cpp locally with unsloth Qwen3-Coder-Next-UD-Q4_K_XL.gguf.
First compression was 82k -> 25k which looks pretty normal. The second is already useless 81651 to 81273
Log from qwen-code:

ℹIMPORTANT: This conversation approached the input token limit for unsloth/Qwen3-Coder-Next. A compressed context will be sent for future
    messages (compressed from: 81651 to 81273 tokens).

Then qwen code continued its execution and ended with an error

  ✕ [API Error: 400 request (100582 tokens) exceeds the available context size (100096 tokens), try increasing it]

settings.json:

{
  "$version": 3,
  "general": {
    "language": "ru"
  },
  "env": {
    "LOCAL_LLM_API_KEY": "local-llm"
  },
  "modelProviders": {
    "openai": [
      {
        "id": "unsloth/Qwen3-Coder-Next",
        "name": "unsloth/Qwen3-Coder-Next",
        "description": "Local Qwen model via OpenAI-compatible API",
        "baseUrl": "http://192.168.0.33:8001/v1",
        "envKey": "LOCAL_LLM_API_KEY",
        "generationConfig": {
          "contextWindowSize": 95000
        }
      }
    ]
  },
  "security": {
    "auth": {
      "selectedType": "openai"
    }
  },
  "model": {
    "name": "unsloth/Qwen3-Coder-Next",
    "chatCompression": {
      "contextPercentageThreshold": 0.85
    },
    "generationConfig": {
      "timeout": 1200000,
      "maxRetries": 3
    }
  },
  "tools": {
    "approvalMode": "default"
  }
}

 │ Status                                                                                           │
  │                                                                                                  │
  │ Qwen Code                         0.10.5 (135b47db)                                              │
  │ Runtime                           Node.js v24.11.0 / npm 11.6.1                                  │
  │ OS                                darwin arm64 (24.5.0)                                          │
  │                                                                                                  │
  │ Auth                              openai (http://192.168.0.33:8001/v1)                           │
  │ Model                             unsloth/Qwen3-Coder-Next                                       │
  │ Session ID                        178d7455-45b3-4be0-9d2c-716be523109f                           │
  │ Sandbox                           no sandbox                                                     │
  │ Proxy                             no proxy                                                       │
  │ Memory Usage                      361.9 MB

[Mingholy]

First context compression happens at the length of 82k which approximately just exceeds 95000 * 0.85 = 80750. So I suppose that the contextWindowSize and contextPercentageThreshold are just fine.

Could you describe the specifics of the two compressions? Were these two compressions performed very close together?

[AliakseiLasevich]

Not quite close. I just started the planning process using the qwen code. 100k of context wasn't enough, so it kept running. The first iteration was successful, but the second one couldn't compress the context and for some reason kept running as if nothing had happened, causing the context to overflow.

[tanzhenxin]

Need a better compression fallback logic, when model fails to compress context automatically.
Will work on it as a feature.

[HeKeHenryZhang]

Not quite close. I just started the planning process using the qwen code. 100k of context wasn't enough, so it kept running. The first iteration was successful, but the second one couldn't compress the context and for some reason kept running as if nothing had happened, causing the context to overflow.

The auto compression just triggered once. Release v0.11.0