Conversations moved into the app itself: pick a model per conversation, dial its reasoning, decide when it searches the web, and watch it think — plus an about page and a tidier profile.
29 August 2026
The biggest change in the 1.51 line is one you can't see directly: the separate agents service that used to answer your conversations is gone, and the AI now runs inside WanderNotes itself on TanStack AI. What you can see is everything that became possible once it did — a real model picker, control over how hard the assistant thinks, web search you decide the terms of, and a reply that shows its reasoning and cites where it looked. This note covers 1.51.0 through 1.51.2.
Pick a model per conversation. The composer now carries a model menu with twenty models grouped by provider — Gemini, Gemma, DeepSeek, Qwen, and Kimi — each listed with a one-line description and icons for what it can take: reasoning, images, files, audio. Your choice is stored on the conversation, so switching models mid-thread sticks, and every later message goes to the model you last picked rather than resetting to a default.
Decide how hard the assistant thinks. Alongside the model menu, a reasoning control sets the effort for the next reply: Low answers with little deliberation, Medium thinks before answering, High works the problem through at length. It appears only for models that actually reason, and defaults to Medium.
Web search on your terms. A grounding menu sets whether a reply is allowed to search: Off answers from the model alone, Auto searches when it would help, Always grounds every answer in search results. Auto is the default.
Watch the assistant think. Replies from a reasoning model now show a Thinking panel that opens on its own and streams the reasoning as it arrives, pinned to the newest line. When the answer starts, it collapses and relabels itself Thought process — still there if you want to reread how the model got there, out of the way if you don't.
Sources, read from the reply itself. When a reply was grounded in a search, the citations behind it are listed under the answer, deduplicated by URL. They're read off the reply's own tool activity rather than stored separately, so what you see is what the model actually consulted.
Usage for every conversation. Conversation settings gained a Usage panel: total tokens and cost at a glance, and underneath, the run count, input and output tokens, reasoning tokens, and cache reads and writes when there are any. It updates live as the conversation runs.
Note enhancement, configured. The enhance dialog now lets you choose the model, its reasoning effort, and its grounding before you run it, rather than taking whatever the server picked. It defaults to Gemini 3.7 Flash at high effort with grounding on auto.
An about page. /about explains what WanderNotes is and why it exists, reachable from the landing page and from your profile.
A generation you can leave and come back to. Every reply is now a tracked run with its own status, so reloading the page or closing the tab mid-answer no longer loses it — the app rejoins the run in progress and replays what you missed. Conversation lists read the same run status, so a thread that's still generating says so.
A proper "conversation not found". Opening a conversation that was deleted or never existed now shows an explanation with a way out — Close in a panel, Back on a full page — instead of an empty screen.
Resumed answers no longer vanish. A reply that reasoned before it wrote could disappear entirely when you rejoined it after a reload — the replay would build the message out of reasoning, then delete exactly that message when the first text arrived. Rejoining now keeps the answer.
A stuck enhancement can be dismissed. A note enhancement is produced by the server process that accepted it, so a deploy or crash mid-run used to leave the note pending forever with a locked editor. Pending enhancements now record when they started and are treated as failed after ten minutes, so you can dismiss and retry instead of losing the note.
Conversation titles are named from the right message. Naming a new conversation read its opening message inconsistently after the migration; it now takes the first user turn as written.