Workspaces

The isolation mechanism. Every agent works in its own jj workspace, so parallel work on one repository genuinely cannot collide.

What a workspace is

A jj workspace is a second working copy of the same repository, with its own working-copy commit. Edits in one are invisible to the others until they are merged. Rudder gives every agent one.

This is the reason several agents can work on the same files at the same time. They are not coordinating, taking turns, or locking anything. They cannot reach each other.

Where they live

your-repo/
  .rudder-workspaces/
    fix-login-redirect/
    port-settings-form/
    drop-sessions-table/

They sit beside your checkout, never nested inside it, so your editor, your build and your git status are unaffected by what the agents are doing.

How merging works

Merging a row runs jj new against your branch and that agent's change. It is a real merge commit, not a squash of the working copy, so the agent's history stays intact and the merge is a normal thing to inspect or undo.

m merges the selected row. M merges everything that is ready. u undoes a merge through jj's operation log.

Conflicts

jj records conflicts in the change rather than refusing the merge. A conflicted merge still completes; the conflict is then a thing you can see, resolve, or undo, instead of a wall you hit mid-operation.

This is what lets Rudder integrate several overlapping agents without stopping the whole session on the first collision.

Rudder settles the conflicts that need no judgment on its own: its own coordination files union their entries, package.json deep-merges its dependency maps, lock files take the most complete side. Source code is never merged mechanically - guessing at a real disagreement is worse than stopping. What is left starts a resolver agent in the row that already owns the work, rather than a prompt waiting for a keypress. An unanswered prompt used to park finished nodes in review and leave everything downstream of them unable to launch.

The main checkout

Some jobs are genuinely about your working tree: tagging a release, running a deploy, poking at something by hand. Those are /main, and they run in your real checkout with nothing to merge.

Several /main agents can run at once. Rudder allows it because it is sometimes what you want; managing what they do to each other is yours.