OpenKanban

The deciding trait is that everything lands on one board. Tickets from every repository you work in appear together, and each one carries its own running session, so the question of what is in flight has an answer you can see rather than a row of terminal tabs you have to identify by squinting at the titles.

That is a real problem solved simply, and there is not much else here. Pick it if you juggle projects. Pick something else if the juggling was never the hard part.

Cleanup as a setting is cleanup that does not happen. Each ticket spawns a branch and a separate copy of the working tree, and after a busy month a developer's disk holds dozens of half-finished checkouts, each with its own untracked files and stale dependencies. Nothing in the row records a retention policy, a size warning, or a way to see what is orphaned.

What it does right is putting the running agent inside the board rather than beside it, so a session and its ticket cannot drift apart.

1. This is the correct primitive and it is borrowed rather than invented, which is a point in its favour. Concurrent agents editing one checkout is a data race with extra steps; separate trees make the isolation real at the filesystem level instead of asking a scheduler to be careful. 2. Merging remains entirely the user's problem, and the design does not pretend otherwise.

2. No evaluation is offered of whether parallel agents actually produce more finished work, which is the claim the whole arrangement rests on.

Coordination surfaces are the first thing a well-funded agent vendor adds once its own tool works, because the parallel-session problem is created by their product and solving it keeps users inside. A solo project holding that position has neither the distribution to defend it nor revenue to fund a fight.

Moat: none. Likely acquirer: none; the feature gets shipped rather than bought. Likely path: genuinely useful for a year, then unnecessary. Position: use it now, do not plan around it.

A board nobody else can see is a personal to-do list with columns. Across sixty engineers this creates sixty private views of what is in progress, none of which reach the tracker my delivery managers actually read, and reconciling those two pictures becomes somebody's weekly chore.

There is no identity layer, no directory sync, no central record and nothing that runs unattended, so it never becomes a stage I can measure. The licence costs nothing and the model spend sits on keys we issue. Not yet.

Copyleft here is a feature. Anyone who builds a hosted product on this has to publish their changes, which is a stronger guarantee for me than a permissive licence that quietly turns into somebody's closed service. And the whole configuration, including which agent runs and how branches are named, sits in one file I keep in version control.

It takes whatever command-line agent I already trust, which is the right posture. What is absent is a protocol port and any route to weights on my own machine.