mindfork mindfork

mindfork 0.16.1 — when Ollama cuts the prompt, you are told

· 3 min read

mindfork 0.16.1 is out. It finishes what 0.16.0 started for Ollama: that release read Ollama's context window, so a long conversation is folded into a summary before Ollama would cut it. This one is about the cut that happens anyway.

Two cuts, and one that matters

A server that cannot hold a prompt should refuse it, and llama-server does. Ollama answers, and cuts. Measured on Ollama 0.35.1, it cuts in one of two ways:

  • The oldest messages go. A conversation over the window loses its first messages whole until the rest fits. The instructions and your last message stay, and what Ollama reports is just under the window — which is where mindfork folds the conversation anyway.
  • The start of the prompt goes. When the instructions and your last message alone do not fit — a long paste, a large tool result — the prompt is cut to half the window, from the start. The instructions go first. In the test, a model asked for a code word from its instructions answered with fragments of the pasted text instead.

Either way the reply arrives as if nothing happened, and what Ollama reports is the cut size. Before this release, mindfork took that size at its word.

What mindfork does now

It knows the exact size of the chat's previous request, and a request that extends it cannot hold less. When Ollama reports less, the request was cut:

  • The feed says so, once a session: how much of the request the server processed, and how to give Ollama a larger window — OLLAMA_CONTEXT_LENGTH, or the context slider in its app's settings.
  • The conversation is folded into a summary when there is anything to fold, even before Ollama has said what its window is. Then /regen asks again with the shorter conversation.

Measured at a window of 4096: a page pasted after a first turn of 3446 tokens was cut by Ollama from 5210 tokens to 2051, and the feed said of at least 4357 tokens it processed 2051. At 16384, and at 4096 without the paste, nothing was told, because nothing was cut.

Only an external server is judged. llama-server and the cloud providers refuse a prompt that does not fit, and mindfork already says what to do then.

Fixed

  • The window-too-small note described Ollama's cut wrongly. It said Ollama cuts from the start, the instructions first. That happens only when the last message alone does not fit; otherwise the oldest messages go. The note and the install guide say both now.

Your data

Nothing in the data's format changes, and no step runs at the start.

The full list is in the changelog; downloads are on the releases page, and cargo install mindfork builds the same version from crates.io.