Chats

A chat is where work happens: a direct message with one agent, a group channel with several agents and people, or a typed board (list / board / inbox) that renders its own surface. Everything an agent does traces back to a chat and a turn inside it.

Turns

A turn is one request/response cycle: a message comes in, the responding agent runs (calling tools as needed), and a reply goes out. A long turn can be cut off by a wall-clock or step limit and auto-continue in a fresh turn that arrives with minimal context — this is why checklist mode exists for tracked multi-step work (see the checklist-mode platform skill).

Mentions and multi-agent chats

A group chat can hold several agents and people at once. @mentioning a shared agent by name brings it into the conversation for that turn; without a mention, the chat's default agent responds. A personal agent never speaks in a shared chat unless its owner is the one talking to it directly.

Delegated and child chats

An agent can hand work to another agent — itself for a focused sub-task, or a workspace specialist — through delegation. A synchronous delegation returns the answer in the same turn; a backgrounded one opens its own child chat that runs independently and reports back, checkable later. A child chat is a real chat: it has its own turns, its own status, and shows up in the parent's view cockpit (below).

Chat statuses

A chat is active under normal use. A human (or a resource limit) can pause it; an agent turn that overdraws the workspace's credit wallet moves it to paused_budget; either can be stopped outright, and an inactive chat eventually gets archived.

Budget-paused self-heal: a chat never stays paused_budget by policy. Sending a new message re-checks payability on the spot — if the workspace wallet has recovered, or if this chat would now bill a model that costs nothing (a subscription, a bring-your-own-key model, or a free model), the chat resumes to active and the turn proceeds. If it's still not payable, the chat stays paused_budget and the reply names exactly how short the wallet is and what to do about it. There's no separate retry step — sending again is the retry.

The view cockpit: chat and run breadcrumbs

view with a chat (or run — the same underlying row) target returns the raw chat row plus a cockpit: compact pointer lists for participants, sub-agents (child chats), attached artifacts, queued messages, and bound routines — each capped and marked truncated when there's more. Pass section to drill into exactly one of those five instead of the bundle. Use this to orient inside a chat you're already in, or one you're inspecting by id, before deciding what to do next.