Steer work while it runs
Add a note to a running task, stop it when it goes wrong, and run it again on a stronger crew, without starting over by hand.
Starting work should not mean letting go of it. You correct a colleague mid-task; you should be able to correct an agent the same way.
- also document the new endpoint
- stop the one that went off track
- redo it on a stronger crew
A task you start keeps running in its own copy of the folder while you do other things. That does not make it a sealed box. You can add to it while it works, stop it, or ask for the same work again with stronger models, all from the conversation that started it.
Try it in CodeAF
- Where
In the conversation that started a running task.
- Press
- Alt+T, move to the task's row in the Tasks column, and press Enter. The box now reads
a note for this task… (esc: main). - Alt+T again to hand the keyboard back to the box.
- Alt+T, move to the task's row in the Tasks column, and press Enter. The box now reads
- Type
Also add a Health section at the top of docs/API.mdThen press Enter.
- You see
Your line appears on the task's page under its brief, followed by
· the worker reads a note at its next step.
A note, not a restart
Nothing is re-planned and nothing is lost: the worker finishes the step it is in, then reads your words with everything it already knows. In our run the worker was inside a one-minute sleep; it read the note as soon as the command returned, and its report said it had added the section. Press Esc to go back to the conversation.
Stop it
When a task is going the wrong way, select its row in the Tasks column and press X with the box empty. A question asks Stop this task? Its work halts; the branch it wrote on is kept., with keep going as the safe answer, so a stray key never costs you work. You can also say it in plain words, stop task 6, and the chat stops it the same way. The task's card then reads stopped with how long it ran, and what it spent so far is on its crew line.
Redo it stronger
/redo stronger runs the last task this conversation started again, one rung up: the seat whose next model buys the most moves up, and the crew line names the change, for example worker glm-5.3 → claude-opus-5.5, with the new estimate.
Things to build
- Add the docs while it codes: a task adding an endpoint is running; leave the note "also document it in docs/API.md" and the same worker does both.
- One part of twenty, corrected: in a twenty-part run, tell the chat "note for the Wrap part only: ..." and only that part changes course. Close twenty issues at once
- Cheap first, stronger on demand: let auto pick the crew and read the result; if it falls short,
/redo strongerruns it again one rung up. - A stop in plain words: "stop task 6" from the chat halts a task that went the wrong way and keeps its branch for a look.
Go further
Close twenty issues at once: Hand one brief that names twenty issues, and twenty workers take one each, side by side.
Coming soon: describe the org you want and CodeAF builds it: teams, managers, budgets, standing orders.