The model picker starts with the machine. Every server you have is a tab — its model count, whether it is answering — and the list under it is that one server's catalogue: what you are running, starred and reached for lately in one short section, then the models on its own hardware, then the families. Under the tab, the providers are a filter: one chip per subscription or key the server reaches its models through — Ollama Cloud, OpenCode Go, OpenRouter, Anthropic, xAI, DeepSeek, local ollama — so you can read what each one will run as one list, and a pick made under a filter goes through that door. A search still finds a model on the other servers, under its own heading. The badges are gone; a row is a name and the one fact that would stop a send, and a heading's chevron always points the way the list is actually folded.
A model wears its own name. Only Claude's models answer to a family word — Fable, Opus, Sonnet, Haiku — and everything else is worn as itself: gpt-oss-120b, glm-5.3-flash, gemini-3-flash-preview, never a bare "GPT" that hides which model it is.
A queued message is still yours, and now it survives everything. A message written while a turn runs belongs to that conversation, not to the screen showing it: it is kept on disk, it comes back after the app is closed or killed, it goes the moment its turn yields whichever chat you are looking at, and it never lands in another conversation — leaving a chat with a message waiting used to carry it into the next one on the desktop.
Oh My Pi conversations name themselves after the first thing you say, and every Oh My Pi session — including one started in a terminal — has a real ledger: tokens and price per turn, in the chat and in Usage, read from the agent's own transcript.
A sent prompt rises to the top of the screen and the answer streams onto an empty page — the rest of the conversation is one scroll up, and the page stays put when the answer ends. The message you sent is drawn exactly like every sent message from the moment you press Send; no fading in, no "sending" and "sent" under it — only a message that did not go changes its ink.
The answer is the page: an assistant reply runs edge to edge with the same margin on both sides and no bubble behind it, so a long answer reads like a document rather than a chat balloon.
Tailscode now lives on iPad. The whole app is universal: every screen keeps to a readable column on the big display, a chat opens in its own window beside another, and a hardware keyboard gets a real menu bar with ⌘-shortcuts. Slide Over, Split View and Stage Manager all work, in every orientation.
Oh My Pi joins Claude Code and opencode as a first-class agent — in setup, on the discovery radar (which now scans this machine too and no longer shows duplicate rows for one server), and on every server screen.
Quick Ask aims like a sentence: a server chip and a model chip with real menus, so the question goes to the machine and model you meant.
The model directory learned stars — favourites float to the top of one fleet-wide list — and Usage grew an Ollama Cloud meter that speaks in every state, while free, local and OpenRouter models no longer wear quota walls that were never theirs.
A turn's clock now reads the transcript rather than a client stopwatch, and it counts from the prompt that started the running turn — never from the first message in the chat — so the elapsed time is how long you have actually been waiting.
The jump-to-bottom button no longer counts new messages — it is simply the way back when you have scrolled up — and the transcript only resumes following once you are actually at the bottom, so lifting your finger a few lines short of it no longer drags you down.