15 · The Kitty Pool

What you'll do

Watch — and, if you like, tune — the Kitty Pool: the set of AI helpers TofuFactory keeps warmed up and waiting, so a new job starts working straight away instead of pausing to boot one up.

Why it exists

Starting an AI from cold takes a few seconds — it has to launch, load its settings, and sign in. The Kitty Pool keeps a few helpers already started and idling, so when a job arrives:

1. It grabs a warm helper that's sitting ready.
2. The helper begins work immediately — no startup wait.
3. When the job's done, the helper goes back to idling for the next one.

Note: only some AIs can stay warm this way (Qwen and Gemini can). Others, like Claude, are started fresh for each job — that's normal, they just don't idle in the pool.

The Pool screen

Go to Pool from the navigation menu. It shows:

• How many warm helpers you have, by type.
• The standby list — each helper, marked idle (waiting), active (working), or recycling (being refreshed).
• A live log of helpers starting up, getting picked up for a job, and retiring.

Why helpers get refreshed

A warm helper can't run forever — over time it uses more memory and its memory of past jobs piles up, which makes it slower and pricier. So TofuFactory quietly recycles them — retires the old one and starts a fresh one — on sensible rules:

Trigger	Default	Why
Jobs run	15 (Gemini), 3 (Qwen)	Refresh after a helper has done a fair share of work.
Sitting idle	30 minutes	Free up your computer's memory when nothing's happening.
Memory of past jobs too full	60% full	Reset before a bloated context slows it down.
Hit an error	Always	A helper that stumbled is replaced rather than reused.

You can also refresh by hand on the Pool screen — Recycle next to one helper, or Recycle all idle at the top — if one seems to be misbehaving or hogging memory.

Power user: the pool's size and limits are set with environment variables read when the server starts — handy if you run TofuFactory yourself and want to tune it. The main ones: TOFU_POOL_ENABLED (true; set false to turn the pool off entirely), TOFU_POOL_MAX_PROCESSES (8, the ceiling on helpers at once), TOFU_POOL_MEM_THRESHOLD (85 — shrink the pool when system memory passes this %), and TOFU_POOL_TARGET_GEMINI / TOFU_POOL_TARGET_QWEN (1 each — how many to keep warm). You never need these to use the pool; it's on and sensible by default.

You should now see

• A live list of warm helpers on the Pool screen.
• Helpers starting and recycling in the event log.
• Jobs for Gemini and Qwen starting noticeably faster.

If something's not right

Problem	What to do
The screen says "Pool disabled"	The pool was turned off (TOFU_POOL_ENABLED set to false). Whoever runs your server can turn it back on.
Helpers recycle constantly	Usually fine — it just means they're finishing jobs quickly. Only a concern if it's clearly thrashing.
My computer's memory is maxed	Lower the maximum number of helpers, or have the pool shrink sooner under memory pressure (the power-user settings above).
The log shows "pool exhausted"	Every warm helper was busy, so a job started one from cold instead — nothing was lost. If it happens a lot and you run many jobs at once, keep more helpers warm.

Next

→ 16 · Settings & personalization — make the app look and read the way you want.