Message Transforms

Not yet available. Message transforms are on the roadmap.

Every model has a fixed context window. Once a conversation grows beyond that limit, the request would simply fail. Message transforms intercept the prompt before it reaches the model and intelligently compress it so the most important content fits — without your application needing to manage truncation logic.

How it works#

When a request's token count exceeds a model's context window, OpenDunes applies a middle-out compression strategy before forwarding to the model:

1. Preserve the system prompt in full — it defines behavior and is always kept.
2. Preserve the most recent messages — the tail of the conversation is what the model needs most.
3. Compress the middle — older messages are summarised, deduplicated, or pruned based on information density. Tool results that are no longer referenced are removed first.

The transformed prompt fits within the model's context window. The model never sees the overflowing original; your application receives a response as normal.

When transforms apply#

Transforms only activate when the raw token count exceeds the target model's context limit. For conversations that fit within context, the messages array is forwarded unchanged.

You can check whether transforms were applied on a given request in your dashboard — the activity log shows both the raw input token count and the effective token count sent to the model.

What is preserved#

Opting out#

If your application manages context manually and you never want transforms applied, pass:

{
  "transforms": []
}

With an empty transforms array, requests that exceed the model's context window will return a 400 invalid_request error with code context_length_exceeded rather than being silently compressed.

Best practices#

• Design prompts with recency in mind. Important instructions the model must always follow belong in the system prompt, not in old user messages that may be pruned.
• Use explicit summaries for long sessions. For very long agent loops, periodically inject a system or user message that summarises prior context. This gives the compressor a clean, dense anchor point.
• Check the activity log. If responses seem to lose context on long conversations, the activity view shows what was actually sent to the model after transforms.