Doc chats now save — pick up where you left off

You know that AI chat that lives next to a doc — the one where you ask "what's a better way to phrase this?" or "summarize the key points" or just brainstorm next steps with the doc as context? Until today it was in-memory only. Have a great exchange, refresh by accident, lose it. We've all been there.

Doc chats now persist. Open a doc → Chat tab → your last conversation is right there, ready to continue. Refresh, close the browser, come back tomorrow — same chat, same context.

How it works

One conversation per doc, automatically loaded when you open the chat tab. No "new chat" buttons to think about, no history sidebar to manage. The chat just remembers what you were talking about.

Under the hood, doc chats use the same storage as your SP3K conversations — but they don't clutter the SP3K list (you don't want every "fix this typo" thread in your main thinking inbox). They're per-user private, so your chats about a shared doc are yours alone. And when you delete a doc, the chat goes with it.

Two small input polishes

While we were in there, two things that had been quietly bugging us:

• Cursor stays in the input when you hit Enter to send. Used to lose focus the moment a response started streaming, which meant clicking back in every time you wanted to follow up. Modern chat tools don't do that. Now neither do we.
• The input grows with what you type. Was a thin one-line strip even when your text wrapped to four lines. Now expands up to about 5-6 lines of comfortable typing room.

Two tiny things you only notice when they're broken — and now they're not.

Trade-offs worth naming

There's one chat per doc today, not many. We considered multiple ongoing threads per doc (gives you a "new chat" reset for fresh context) but landed on simpler. The mental model is just "open doc, see your chat." No new surface to manage. If you want to wipe a chat and start over, that's a button we can add when somebody actually asks for it.

Each chat also has a quiet 15-message cap when sent to the model — the full conversation stays in your UI, but the model only sees the recent slice. Keeps cost and latency from growing unbounded as chats get long. A smarter version (using Haiku to summarize older turns so the model gets to keep that context cheaply) is its own future piece of work — already designed, just not built yet. Let us know how you're using doc chat.