Message Transforms

When a prompt is longer than the model's context window, the request fails. The context compression plugin makes it fit instead: it removes or shortens messages from the middle of the prompt until the prompt fits, keeping the beginning and the end. Models pay the least attention to the middle of a long sequence, so it is the part to lose.

Use it where perfect recall is not required — a long chat, an agent's running log.

Turn it on#

{
  "model": "anthropic/claude-haiku-4.5",
  "messages": [],
  "plugins": [{ "id": "context-compression" }]
}

The plugin also handles a count of messages above what a model accepts, not only a length.

When it runs without being asked#

On a model endpoint with a context of 8 192 tokens or less, compression is on by default. Turn it off with:

{ "plugins": [{ "id": "context-compression", "enabled": false }] }

It is skipped for picture models that answer only with pictures, so the reference pictures of an edit are kept whole.

Billing#

A plugin runs once on every request that names it. The request is billed in DA on the model's own bill, as its usage.cost shows.