Workflows.
A workflow is a team of agents working on one large task in planned stages, for example: review every module, then fix what the review found, then check the fixes. It suits jobs too big for one conversation, such as an audit of the whole codebase, a migration across many files or research on many topics at once.
Nothing starts until you have seen the plan and agreed to it.
Start a workflow
Begin your message with /workflow and describe the job:
bash❯/workflow review every API route for missing input validation, then fix what you find
Nilux writes a plan for the run and shows you a summary before anything starts:
- how many agents it will use, and your limit;
- the models;
- what the agents may do: only read, or also change files in listed folders and run listed commands;
- the stages, in order.
Choose Start workflow to run it, or Cancel. Deleting files is never allowed in a workflow, and an agent that wants to change a file outside the listed folders has to ask you.
You can also discuss the plan with Nilux first in an ordinary conversation, and start it with /workflow once you are happy with it.
Follow the run
The run works in the background, so you can keep talking to Nilux. Type /workflow to open the board:
- each stage with how many of its agents are done;
- each agent: its model, what it is doing now, and any error;
Enteron an agent opens its conversation,Escgoes back.
On the board, f shows only failures, s stops the run, and r resumes a paused run or runs an ended one again.
When something fails
An agent that fails gets up to three more tries. If it still fails, the run finishes the work already in progress and pauses. Nilux tells you what failed. Then:
/workflow resumecontinues the run;/workflow stopends it.
When it ends
Nilux writes a full report to a file and tells you in the conversation where it is. /workflow list shows this session's past runs. A run that did not finish can be run again from there: the agents that already finished are reused, and only the rest run.
Settings
/workflow config sets:
- the most agents one run may use (up to 50);
- the model agents use when a stage does not name one.
A workflow can use many agents at once, and each agent uses its own tokens. The summary shows how many agents a run will use before you start it.