Lineup See the example run
OpenCode · Python

Role agents for OpenCode that pick their own models.

Every model you have answers the same fixed questions on your own machine. The best one for each role is written into the agents, with a backup and a put-back.

OpenCode · Python
lineup · example runexample run
Lineup screenshot
2platforms: OpenCode, Python
6role agents
18fixed questions per model with --hard
40tests, no network needed
Why it exists

A list of models is not a list of working models.

In the first example run, two of the nine listed models could not be used at all.

Lineup keeps a backup, checks that every changed agent still answers, and puts everything back if one does not.

Run it yourself

Install it, then tune.

From the README’s “Install and tune” section. Full options, requirements and troubleshooting are there too.

Open the README
shell
$ git clone https://github.com/juw12/lineup
$ cd lineup
$ python lineup.py install
# restart OpenCode, then:
$ python lineup.py tune --dry-run     # score only, change nothing
$ python lineup.py tune               # score, pick and apply
$ python lineup.py tune --hard        # also ask the harder third task per role
$ python lineup.py report             # open the score page
How it works

What tune does

  1. 1

    Roll call

    Every model gets one short question, and a model that does not answer in 60 seconds is left out.

  2. 2

    Questions and score

    Each role has two fixed tasks with pass-or-fail checks, or three with --hard, and the score is the share of checks passed, from 0 to 10.

  3. 3

    Pick

    The best score wins a role, the faster model wins on equal scores, and the model you already use stays.

  4. 4

    Apply

    The agent files are copied to a backup, the new model lines are written, and all files are put back if a changed agent does not answer.

What's inside

Six agents and one command.

Six role agents

Files, docs, numbers, writer, office and briefing. Pick one in OpenCode or call it as a subagent.

Each agent has its own permissions

Files, numbers and briefing may not edit files or run commands. Docs and writer ask before an edit. Office asks before an edit or a command.

One command chooses the models

tune asks every model the same fixed questions and writes the best one for each role into the agent files.

Pass-or-fail checks

For example "exactly 3 bullets" or "the sum is 4,815.50". The score is the share of checks passed, from 0 to 10. No model judges another model.

A backup and a put-back

The agent files are copied before any change. Each changed agent is asked one question, and if one does not answer, all files are put back.

It stops at a usage limit

If a provider refuses more questions, Lineup stops asking, changes no agent and keeps every answer it has. --resume continues from there, for up to 7 days. If it refuses one question only, that one counts as no answer and the run goes on.

Your own tasks

The built-in tasks are a floor, and --hard adds a harder third task per role. Add questions from your own work to tasks.json with four check types: regex, absent, bullets and max_words.

No packages to install

You need OpenCode with at least one working model, and Python 3.9 or newer on Windows, macOS or Linux. No Python packages are installed. On Windows without Python, lineup.exe does the same.

Your data

Results stay on your machine

  • Results stay on your machine, in a folder named .lineup in your home folder.
  • The questions go to the models you already use in OpenCode, through OpenCode itself.
  • The uninstall command removes the six agents again and touches nothing else.
Good to know

Honest limits.

Straight from the README, so you know what you're getting before you try it.

  • Twelve small tasks, or eighteen with --hard, and one or two runs are a small sample, so read a score of 10 as "passed these checks", not as "the best model".
  • The first example run did not finish, because OpenCode's free limit stopped it after 46 of 216 answers, so the example shown is a later run of two free models on a GitHub test machine.
  • On Ubuntu and macOS the full tuning run was done on GitHub's test machines only, and the provider refused one question there on both models, so 34 of 36 questions were answered.
  • Pass-or-fail checks measure whether an answer is correct and in the asked format, not style or depth.
  • Every question starts OpenCode once, so a question takes seconds even on a fast model.
Under the hood

How it's built.

Built with

Language
HTML/CSSPython
Testing
Automated tests

30 files in the repo · last updated 19 Nov 2026.

Browse the code

Recently shipped

  1. Limits: the tuning run tried on Ubuntu and macOS answered 17 of 18 questions, one refused by the free plan

  2. Hard tasks (tune --hard), macOS and Linux paths, lineup.exe release workflow, tests on three systems, comparison

  3. Six role agents for OpenCode that pick their own models

Full history on GitHub
Questions

Good questions.

What does it run on?

OpenCode, Python.

How do I get it running?

The commands in the “Install and tune” box above come straight from the README. Paste them into a terminal; the README covers options and troubleshooting.

Where does my data go?

Results stay on your machine, in a folder named .lineup in your home folder. The questions go to the models you already use in OpenCode, through OpenCode itself.

Why did the first example run stop early?

It asked OpenCode's nine free models 216 questions with no account, and OpenCode's free usage limit stopped it after 46 answers.

Does a score of 10 mean the best model?

No, it means the model passed these checks.

How do I undo it?

The uninstall command removes the six agents and touches nothing else, and every tune keeps the agent files from before in a backup folder.

What doesn't it do yet?

Twelve small tasks, or eighteen with --hard, and one or two runs are a small sample, so read a score of 10 as "passed these checks", not as "the best model". The first example run did not finish, because OpenCode's free limit stopped it after 46 of 216 answers, so the example shown is a later run of two free models on a…

Is it free?

The source is public on GitHub under the MIT licence.

Keep exploring

More from the same workbench.

Let every role pick its model.

Role agents for OpenCode that pick their own models.