Blog · release

Latch 0.21.1

2026-10-18

A warm Handler, and a grammar that cannot write an empty action. No migration.

• Keep the model loaded (ai.keep_alive_minutes). Every model call now carries a keep_alive of your own: blank means your cycle plus 10 minutes, so the next cycle finds the model resident and its prefix cached — on the maintainer's 27B that is 1–2 s a cycle instead of 8–10 s, and a chat turn you can feel the difference in; a number is yours (up to 8 days); 0 frees the GPU right after each call. Before this a fixed 30 minutes under a 180-minute cycle made every cycle a cold start (58 of 62 production calls, measured 2026-10-18). Shown on Settings → AI Handler. Cycles, chat, the kennel and the profile's posts all send it.
• The reply schema says what the game checks. Ollama's structured output used to be told only that an action has a type; a model could close an action after {"type": "prompt"} — an empty shell the game then logged as invalid (9 of 11 replay actions, 7 of 23 production actions on the 27B) — or write an assign with add_time's keys. The schema now carries one branch per action type with its required keys and no other, the integer ranges (minutes 1–1440, due_hours 1–168, an order's minutes 1–600, maintenance 10–60), the text lengths and the 280-character line, and at most 4 actions. The Python check is unchanged and still runs on everything.