Factory

When you ask Lux to build a feature, screen, fix, or redesign, Lux hands the goal to the factory. The Manager breaks it into tasks, runs agents in parallel, and lands the result while you watch.

#1. Frame the goal

Describe what you want in plain language, either to Lux in the chat sidebar or to the Manager in the Factory section. Describe the outcome rather than the implementation: “add a profile screen with an editable avatar and bio” beats a list of files to touch. The Manager turns that into a spec and a set of tasks.

It asks you the genuine product decisions, like brand, colors, or a real tradeoff, through a structured question, and keeps unblocked work moving while it waits. Answer those and leave the rest to the agents.

#2. Watch it build

Two places show you the run:

• The Plan board is a kanban of every task moving through backlog, in progress, review, and done. Click a card for its detail, and use the chat button on an active card to jump to the agent working it.
• The Agents pane lists the agents and their threads side by side. Turn on Follow active agent from its header and the chat tracks whoever is most active. Picking an agent yourself, or typing into a chat, turns it back off so the pane stays where you put it. Its Floor view, opened from the same header, shows the agents themselves, each at a station for what it is doing.

Every agent has its own thread. Open one to read exactly what it ran and why, rather than waiting for the final diff. The previews update live as work lands, so the feature appears in the same surfaces you would edit by hand.

#3. Steer in flight

You are not locked out once a run starts. Message Lux or the Manager to redirect scope, correct a wrong assumption, or refine the goal, and the Manager adjusts the remaining work. If an agent gets stuck, the factory’s own checks catch it, and you can always step in.

#4. Review and land

Finished work does not land unchecked. A reviewer runs a fresh pass on each task and either approves it or sends it back. On the board, each card carries its own signals, including a review-difficulty indicator and pass or fail badges for the checks that ran, so you can see at a glance what was verified. Once everything is done, QA exercises the running app like a user and files anything that is broken.

For a reviewed task on its own branch, approval moves the card to done and the factory auto-merges the branch. Landing is not a separate manual step there.

#Inspect it yourself

Whether work came from an agent or from your own branch, you can review it the same way before it merges. Open the branch menu from the top-bar branch indicator and Diff the branch against main. The diff is a full file-level review of everything the branch touched, so read it the way you would read a pull request.

To see it running, Open the branch and let the previews render it. A branch is fully isolated, so this is safe: main is untouched until you merge.

#Merge it

When the diff and the previews look right, Merge the branch into main. Contrast saves the branch’s unsaved files, writes them into main, marks the branch merged, and returns you to main, and the editor and previews refresh to reflect it. If it is not right, leave the branch unmerged and send the work back. For an agent task, ask the Manager to revise it.

Manual merge from the branch menu is for your own branches and for recovery, such as when an auto-merge hits a conflict and stops. Either way, nothing reaches main without passing through review.

#GitHub-linked projects

If the project is synced to GitHub, the branch menu also shows ahead and behind counts and lets you push a branch, pull updates, or import an open PR as a review branch, so the same review-and-land loop spans Contrast and your repo.