kolu
Docs

before the first agent starts

One form,
answered once.

Every piece of work the Conductor opens starts with one form. You say how it runs, what happens when its PR is green, and which agent CLI (the harness) and model run the planner and the implementor.

What it asks

QuestionWhat you choose
ModeFull: a planner, your approval, an implementor and review rounds. Trivial: one implementor.
Planner, ImplementorThe harness that runs each agent, and the model it runs on.
On greenWhat to do once the PR is green, meaning it is not a draft, GitHub would let it merge, and no check is running or failing: merge it, or hand it to you.

The host is not in the form: it is part of the work itself. When your request does not name the machine, the Conductor asks in its reply, because a path like ~/code/app can exist on several machines.

A question your request already answers is not asked. Each question takes one answer, and a form you fill in whole is settled with one press of Answer.

A question left blank stays open

Pressing Answer records the questions you filled in. One you left blank is not answered: it stays open in the same form, on the card and in the chat, until you answer it. A question whose options are still loading says so on it and stays open the same way. Answer stays disabled until at least one question is filled in, and I’d rather not say declines whatever is still open.

Nothing starts on a missing answer. An agent waits for a question that is open in front of you, and the Conductor asks again for one that is not.

Where the harnesses and models come from

kolu offers both, not the Conductor, and starts each agent on exactly what you chose.

  • The harnesses are the ones this kolu carries.
  • The models are the ones that harness offers on the work’s host, asked of the host as the form is drawn.
  • The host asks in the same environment a new terminal there gets, after your shell has read its startup files. A key your ~/.bashrc exports reaches the list, as it reaches an agent you start by hand.
  • kolu asks bash and zsh. A host whose login shell is neither says so in the question, rather than guessing at a list.
  • Claude Code and Codex cannot list their models. For those you get the models you have run before, and a box to type one.
  • That box takes one model id: one word, no spaces. Anything else is refused with a line saying why.

A long list

A list of more than eight options, such as a gateway’s models, is drawn short:

  • Only the options you have used before show, the last one you chose first and already selected. With none, no options show.
  • One line says how many more there are, such as 87 more · search, or show all. Click show all to unfold the whole list, and show fewer to fold it back.
  • The search box searches the whole list, not only what is shown.
  • An option you pick from a search stays shown once you clear the box.

When the host cannot list its models

A list can fail: the gateway is down, or a key is missing. The question then shows the host’s own words, and below them the models you have run before with that harness and a box to type one. You can still answer, knowing the list was not read.

kolu records how you chose each model:

Youkolu recordsWhen the agent starts
Pick from the host’s listA pickChecked against the list, if the host can list then
Type in the boxA typed modelStarted as typed; never checked against a list

A typed model goes to the harness as you wrote it, so a short name the harness matches itself, like opus for omp, works. Next time it comes back in the box, still typed. Starting an agent never waits on the list: how a harness takes a model is a fixed fact kolu reads separately.

One press repeats your last setup

Every question comes with your last answer already selected, read off your saved works. For a model that is the same role’s own last model with that harness, falling back to the other role’s model with it. A setup you are happy with is one press of Answer.

When kolu refuses an answer

  • A host whose coding agents are switched off, still arriving, failed to land, or older than this kolu needs is named, with what to do about it, before any harness question is asked.
  • An answer is checked where it is recorded. A harness this kolu does not carry, or a picked model the host did not list, is refused there, so a work never holds a setup it cannot start.
  • The pair is checked again when the agent starts. kolu refuses in words that say which answer and why, rather than starting the agent on something else.
  • A saved answer kolu no longer offers is asked again when the work opens. The question says what the old answer was, and why it is not on offer now.

Next