> ## Documentation Index
> Fetch the complete documentation index at: https://docs.arouter.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Message Transforms

> Automatically compress prompts that exceed a model's context window using the context-compression plugin.

When a prompt exceeds a model's context length, ARouter can automatically compress it using the `context-compression` plugin — rather than failing the request.

## Context Compression

Enable context compression per-request by passing the plugin in the request body:

```json theme={null}
{
  "model": "anthropic/claude-sonnet-4.6",
  "messages": [...],
  "plugins": [{"id": "context-compression"}]
}
```

<Tabs>
  <Tab title="TypeScript">
    ```typescript theme={null}
    import OpenAI from "openai";

    const client = new OpenAI({
      baseURL: "https://api.arouter.ai/v1",
      apiKey: "lr_live_xxxx",
    });

    const response = await client.chat.completions.create({
      model: "anthropic/claude-sonnet-4.6",
      messages: veryLongConversation,
      // @ts-ignore
      plugins: [{ id: "context-compression" }],
    });
    ```
  </Tab>

  <Tab title="Python">
    ```python theme={null}
    from openai import OpenAI

    client = OpenAI(
        base_url="https://api.arouter.ai/v1",
        api_key="lr_live_xxxx",
    )

    response = client.chat.completions.create(
        model="anthropic/claude-sonnet-4.6",
        messages=very_long_conversation,
        extra_body={"plugins": [{"id": "context-compression"}]},
    )
    ```
  </Tab>

  <Tab title="cURL">
    ```bash theme={null}
    curl https://api.arouter.ai/v1/chat/completions \
      -H "Authorization: Bearer lr_live_xxxx" \
      -H "Content-Type: application/json" \
      -d '{
        "model": "anthropic/claude-sonnet-4.6",
        "messages": [...],
        "plugins": [{"id": "context-compression"}]
      }'
    ```
  </Tab>
</Tabs>

***

## How It Works

The plugin removes or truncates messages from the **middle** of the conversation until the prompt fits within the model's context window.

This strategy is based on research showing that [LLMs pay less attention to the middle of long sequences](https://arxiv.org/abs/2307.03172). Preserving the beginning (system instructions, initial context) and end (most recent messages) of a conversation generally produces better results than truncating from either end.

**Compression steps:**

1. Check if total tokens (prompt + estimated completion) exceed the model's context length
2. If over limit: remove or truncate messages from the middle of `messages[]`
3. Repeat until the prompt fits
4. Forward the compressed prompt to the model

### Message Count Limits

Some models enforce a maximum number of messages regardless of token count. For example, Anthropic Claude models have a maximum message count. When this limit is exceeded with context compression enabled, the plugin keeps half of the messages from the start and half from the end of the conversation.

***

## Default Behavior for Small Context Models

All models with **8,192 tokens or fewer context length** have context compression **enabled by default**.

To explicitly disable compression for these models:

```json theme={null}
{
  "model": "some-small-context-model",
  "messages": [...],
  "plugins": [{"id": "context-compression", "enabled": false}]
}
```

Without compression enabled, if your total tokens exceed the model's context length, the request fails with an error suggesting you reduce input length or enable compression.

***

## Model Selection with Compression

When context compression is active, ARouter first tries to find models whose context length is at least **half** of your total required tokens (input + estimated completion).

For example, if your prompt requires 10,000 tokens total:

* Models with at least 5,000 context length are considered
* If no models meet this threshold, ARouter uses the model with the highest available context length

***

## When to Use

Context compression is useful when:

* You have long multi-turn conversations that grow over time
* You're processing documents that may occasionally exceed the context window
* You want resilient behavior without manually managing context length

Context compression is **not** ideal when:

* Perfect recall of all conversation history is required (e.g. document Q\&A where any message may contain the answer)
* You need deterministic behavior (compression is non-deterministic in which messages are removed)

For use cases requiring full context retention, consider models with larger context windows (see [Model Variants `:extended`](/en/guides/features/model-variants#extended--extended-context)).

***

## Combining with Other Plugins

Context compression can be combined with other plugins:

```json theme={null}
{
  "model": "openai/gpt-5.4:online",
  "messages": [...],
  "plugins": [
    {"id": "context-compression"},
    {"id": "web"}
  ]
}
```

See [Plugins Overview](/en/guides/features/plugins-overview) for the complete list of available plugins.