Talk it over, and let it do it
Not every answer is a job. Open a conversation to think out loud — and when it lands on a change, the chat makes it.
Updated September 4, 2026
When you want to think out loud — what's worth doing next, what a card actually means, whether two cards are the same card — open a conversation instead of starting a run.
akb chat "what should I build next, and why?"
akb chat 241 "what is still unclear about this card?"
akb chat 241 "put it in v0.8 and drop the last todo"
akb chat "start a build on 241"
You explain nothing. A chat opens knowing what a coding agent knows after you type /kanban:
your goal, your modules, every open card in one line, and — when you name one —
that card in full. In the app it is a rail down the right of the window (Chat in the
header), following what you are reading: the board's chat on the board, a card's on its page.
It is one conversation either way, so you can start it in a terminal and carry it on in the app.
| You want | What happens |
|---|---|
| to decide what to build, or why | it answers — a question, not a job |
| to understand a card before you approve it | it answers |
| to compare two cards, or spot a duplicate | it answers |
| a change you have already settled — reword it, put it in v1, archive it | it makes the change |
| the card worked out, or the code written | the skill does its printed flow here unless you ask for a background run |
Chat is the kanban skill in a persistent agent session, so it follows the same board-specific
flows and by default does the work here. A card a run is already working on is off limits, and
the refusal names the card and what that run is doing. The chat itself is not a run: it shows in
no runs panel and holds no card. Each conversation lives on your machine under
docs/kanban/.chats/, out of git, and --clear starts it fresh.
A model per conversation
The agent and the model are one conversation's own. Both start on the board's; what you pick sticks to that conversation until you change it, and nothing about the board or any other chat moves — so a cheap model can take the quick questions while the expensive one takes the thinking.
Changing the model carries the same session on, and the transcript marks where it changed.
Changing the agent cannot: a session belongs to the CLI that opened it, so the old conversation
goes. In the app both sit on the message box's bottom row, beside Send; in a terminal they are
--model and --agent, which akb chat --help spells out.
Only an agent whose command can take a second message into its own session can hold a conversation — chat says so and names the ones that can.