Chat

Chat is one conversational feed to your buildd agent. Ask what your fleet has been doing, work out an approach to something you haven't filed yet, then say "make this a mission" or "run it daily" — asking and filing work happen in the same place, without translating the answer into a form.

Scope: your workspaces, routed per turn

A new conversation covers every workspace you can reach. On each turn, buildd works out which workspace you mean — from what you said, or a workspace you've pinned — and uses that as the default for tools, knowledge, and any tier override for that turn. When it isn't confident enough to guess, it lists the candidate workspaces and asks instead.

Pin a workspace explicitly from the workspace switcher in the composer; the same switcher appears wherever else you filter by workspace (the phone header, the desktop bar, Home).

Tools and approval

Chat reads fleet state freely — missions, tasks, PRs, budgets — with no confirmation needed.

Anything chat proposes to write — creating a mission, scheduling a task, arming a held mission, resuming stalled work — shows as a before/after card first. Nothing changes until you approve it. Starting recurring or unattended work always asks, even for a tool group you've otherwise allowed.

Your own tool permissions live in the composer's tool controls: flip a tool group from Ask first to Allow and a write in that group can run for you without a card — but only once per turn, only when there's no earlier tool result already in the conversation's context, and only after the same preview buildd would have shown you succeeds underneath. Admin-level writes, and any tool nobody's opened up to chat, always ask regardless of your settings.

Tier

The chip showing what chat is "running on" is a tier switch: Auto routes each turn on its own — simple turns go to budget, complex ones to premium, and anything in between falls back to standard — or you can pin a tier for the whole conversation. Opening it shows each tier's model, its expected price per 1,000 tokens, and what the conversation has cost so far.

A thumbs up or down on any reply feeds the same model-quality signal buildd uses across the fleet; a thumbs-down asks for a one-line reason.

Standing Rules

Say "always" or "never" in a message — "Always round per line," "Never force-push to dev" — and chat offers a small card: Remember this? Nothing is saved until you tap it.

• Where it applies. A rule saves as everywhere by default — every workspace, every conversation. If the rule names something workspace-specific (a repo, a path, a mission), the card instead pre-selects only this workspace. Either way, you choose the scope before saving, and can change it later.
• Yours only. Standing rules belong to you, not the team. Nobody else's chat loads them, and nobody else can see or edit them.
• Settings → Profile. Every rule you've saved is listed there, newest first, each with Edit and Remove. You can add one directly from that page too, without waiting for a card to offer it.
• Every reply. Chat loads your applicable rules — everywhere, plus the current workspace's — into every turn, and follows them ahead of anything else it's inferred, up to a small budget. The oldest rules drop first if you have a lot; the Settings page always shows the complete list even when a turn doesn't load all of them.
• Reaching filed work. When a chat turn files a task or mission for you, the rules that applied to that turn are copied into its description under a "Standing rules (from chat)" heading — only your own rules, never another user's — so the agent that picks up the work follows them, and anyone reading the task afterward can see why it behaved the way it did.