Where the model is, which one to use, and how long a run is kept. Everything here saves itself as you change it.
Forge is another Mac on this network. It runs the model. Ollama there listens on its own loopback only, so this site reaches it by running a command over ssh rather than by opening a socket — the fleet's keys forbid port forwarding. These three settings are shared with every tool on this site.
/Volumes/WWW/WWW/conf/fleet/ssh_config
A Host block in the file above. The fleet writes one per device.
Forge's own loopback. A host and a port, no path.
Seconds for one model call. A 70B model writing a long section can take several minutes, so this is a ceiling on a stuck call rather than a deadline anybody waits out.
A number and s, m or h. Loading a large model takes most of a minute, and a run uses it four times — so unloading between passes costs more than it saves.
A model marked “reasons first” thinks before it answers: slower, usually more careful. The 70B model writes the warmest prose and takes several times as long. This can be changed per run.
How much the model may be given to read at once, in tokens. About 115,000 characters at 32,768 — a twenty thousand word account in one pass. A longer account is read in sections instead, which works and leaves visible joins. Raising this asks Forge for more memory and can make every pass slower.
Tokens the model may write in one answer. The rewrite needs the most; 4,096 is roughly three thousand words. Too low and a long answer is cut off mid-sentence.
0 writes correct, flat sentences and reaches for the same construction every paragraph. 0.35 reads better and changes no fact — the rules forbid it. Only the rewrite; the other three passes are always at 0 so they are the same twice.
Days. Every run holds somebody's account of their own life. 0 keeps them forever, which is a decision rather than a default. See what is on disk.
One per line. They appear as buttons beside the theme box and fill it in, editable. They are examples of an angle, not a list of subjects.
