mindfork 0.16.0 — Ollama, with nothing to keep in step

mindfork 0.16.0 is out. It is a release about Ollama: most people who run models at home already have it, and mindfork talks to it as to any OpenAI-compatible server. This time the way was measured end to end, and what it found is fixed.

Two lines

With Ollama running and its window set to 16k — the context slider in the Ollama app's settings, or OLLAMA_CONTEXT_LENGTH=16384 for ollama serve:

ollama pull gemma4:e4b
mindfork setup --set engine.mode=external --set engine.external.url=http://localhost:11434/v1 --set engine.external.model_name=gemma4:e4b

The replies stream, the model's thoughts fold above them, its tools run, and Enter on the settings' model row lists what Ollama has pulled.

The window, read from Ollama

A conversation is measured against the model's context window: near its end, mindfork folds the older part into a summary and carries on. Against Ollama it could not, because it had no way to ask how large the window is — and the window matters more there than anywhere:

• Ollama cuts a prompt that does not fit, and says nothing. Measured: a prompt a dozen tokens over the window was cut to half of it from the front, the system message first, and answered as if nothing had happened. The model answered without its instructions.
• Its default window is small where most GPUs are: 4096 tokens under 24 GB of memory. mindfork's first turn — its instructions, the memory, the tools — is about 3500.

mindfork now asks Ollama which window it loaded the model with (its /api/ps) and measures against that. Ollama loads a model on its first request, so the answer comes after the first turn; from then on a long conversation is folded before Ollama would cut it. Nothing to type, and nothing that can drift from Ollama's own setting. 0.15.0's recipe typed the window into mindfork as well; that still works, and a typed number still wins.

When the window is too small

Knowing the window does not make 4096 enough. When a turn ends near the end of the window with nothing older left to fold, mindfork now says so, once: how large the window is, how much the conversation already takes, and where to raise it — OLLAMA_CONTEXT_LENGTH, or the slider. Measured at Ollama's default, the note came after the second turn, before anything had been cut.

Fixed

• Advice meant for llama-server. Ollama now runs llama.cpp inside and reports the same timings, and its first request includes loading the model. So the first reply through Ollama read a slow server and ended with Launch it with -b 256 -ub 256 — a flag Ollama does not take, for a figure that was the load. The advice now goes only to a server that is a llama-server.

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.