Files
zeekayandHanzo Dev f43ad74104 kai: a turn on a model served on this machine never reaches Kai; decisions wait 1s, not 2.5s
Kai answers at api.hanzo.ai and a decision carries the request, the workspace's
files and the tools in play. A thread whose model provider is on loopback (Hanzo
Engine, Ollama, LM Studio) now gets no Kai at all, whatever kai.toml says: Dev
decides as it does without Kai. The check is the provider's base_url host
(loopback, unspecified, localhost); a name that contains localhost or a URL with
the address in its user info is not private.

A decision is advice, so its bound is short: fast 1s, step 1.5s, search 2s. An
answer from api.hanzo.ai takes 0.3s to 1s on an idle machine; a loaded one made
every turn wait out 2.5s per decision.

Co-authored-by: Hanzo Dev <[email protected]>
2026-11-21 06:22:29 -08:00
..