The task record

Every task is a record
you can read

A task starts from a brief and a list of what done means. While it runs you see its plan, each step it takes, and at the end whether every check was met.

Say what done means

Add up to 20 "Done when" checks to the brief: a command that passes, a file that exists. The agent is asked to mark each one met or not met, with its evidence, before it finishes.

A plan you can follow

When it plans the work, a bar under the title shows the state of each step. Steps tick off as they finish, only the live steps stay open, and a step a later plan drops is marked "replaced".

Every step, in plain words

Each row says what happened: Read, Searched, Edited, Created, Ran. Open a row to see the files, the diff or the command output.

Steer it while it works

Type an instruction during a run. Enter queues it for when the run ends; Shift+Enter sends it into the live run now. Esc stops the run.

A receipt at the end

Each run ends with its time, how many checks were met, tokens, output speed, cache hits and cost. A cost marked "~" is estimated from published prices.

Ask before you act

In Ask mode the agent reads and answers with read-only tools. Ctrl+. switches between Agent and Ask, and the agent can switch on its own unless you turn that off.

Start from a brief and a finish line

The New task page takes the brief and its Done when checks in one box, and shows what the agent will start from: the branch, the rule files, its memory, the code index and its tools.

Marked, with evidence

Each check ends marked met or not met, with what the agent saw. If one is still unmarked when it tries to finish, it is reminded once, and the result says how many were left not checked.