You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
One last quick update tonight, roleplay mode fixes
A reply's world update shows when it arrives. The world update, the activity line and the images of a reply are sent after the reply itself has finished, sometimes after the next turn has started, and the chat page dropped every one of them. They showed only once the conversation was reopened. They now appear under the reply they belong to, and an update for an earlier reply never lands on a newer one. Game mode's queued assets arrive the same way.
A change made during play no longer erases the story record. The page kept the copy of the story it had when the conversation was opened, and a gear setting, a removed turn, a scene picked from the scene menu or a redrawn or painted image saved that copy back whole. The events, the cast's states and the timeline points of every reply since were erased, and the narrator worked from a world several turns old until the story state was cleared. Each of these changes now starts from the story as stored.
The model picked in the composer's gear or the model menu is the story's model from the next reply. The pick changed the pill and nothing else: the story kept answering from the model chosen at setup.
Reopening a story right after a regeneration can no longer write an older copy over what the new reply saved.
Network Sharing: a chat mode played on a paired PC's model sends its whole request. A roleplay played from a paired PC ran on the host with no reply cap (a reply that came apart generated until the host's context was full), without the story's samplers (DRY, min-p) or its thinking budget, and with every bookkeeping call at the model family's temperature in place of its own 0.2. All of these now reach the host.
"Thinking off" is respected on a host's model. The thinking setting was sent under names no server reads, so the host applied its own dial and a call that asked for no thinking reasoned anyway.
The host's settings decide what you did not choose. This PC's Setup default for thinking belongs to its own local server and is no longer sent for a host's model. With nothing chosen on the chat's thinking pill or in the model's Tune setting, the host's dial applies, and the pill's tooltip says so ("Nothing is chosen here, so the setting of the host applies.").
Where both sides name a setting: the reply cap is the smaller of the client's and the host's Tune cap, and the host's own Tune sampler and temperature for the model win over the client's, except for a bookkeeping call's fixed temperature. The context window and the launch settings are always the host's.
Tools in a chat on a host's model are called in the text format. The host's sharing endpoint takes no tool schemas, so the model was shown no tools and could call none. Measured on the wire; a full agent turn through a host has not been run yet. Not covered: a gpt-oss (Harmony) model on a host, which needs the schemas passed through.
A chat mode's reply cap is also sent to every other remote provider.
Roleplay Lab:
npm run rp:text -- --via-host plays the story as a paired Network Sharing client does, through the app's own sharing handler on a loopback port, and reports which request fields arrived and which never crossed. Every Lab run used the local path before, so nothing a paired client does differently was exercised.
npm run rp:state has four more probes: deleting only the last reply, editing the last message, carrying on after a regeneration, and paging back to the first version of a regenerated reply. All pass without a model and on Qwen3.8-27B.