command not found: formwork? Every command here has a second form that needs nothing installed.formwork/troubleshooting.mdhas both.
No. You need to be able to run a command in a terminal. That is the only skill assumed here. You do not need to know how a project is supposed to be run: that is the part this brings.
Three things. It gives your agent a team to work as instead of one general assistant. It gives the work an order, so the same steps happen every time. And it blocks some things outright, so a few mistakes cannot happen at all.
No. It is a folder of text files and small programs. Nothing imports it, nothing runs it as a library, and you can delete any part you do not want.
Because an agent that commits will eventually commit something you did not
read. You commit, the agent works. If you disagree, it is one line in
.formwork.toml.
Yes, in one way: every piece of work stops and waits for you. On a busy day
that is the bottleneck, and it is also the point. Everything else is fast. The
gate takes about three seconds, measured in
formwork/COSTS.md.
Write a brief. Six short headings, in
formwork/templates/brief.md. Then ask your
agent to do it.
The rest of the loop is in
formwork/loop.md.
Two kinds. One director thread, which decides what happens next and reads the reports. One working session per piece of work, in the repository, which does that one thing and stops.
The director is a role, formwork/roles/method/director.md. Inside your coding
tool, ask for it by name. Anywhere else, paste the file.
You carry the text between them. The plan does not live in the chat window: it
lives in docs/standing.md, so closing the window costs you nothing.
formwork/threads.md has the whole shape.
On Claude Code, ask for it by name in plain words:
Use the challenger on this plan. Tell me what is wrong with it.
The roles are installed where your tool looks for them. On the other three tools this is documented but untested.
No. Most people will use four or five. The rest sit there until you hit a situation that needs one.
Yes, and it takes two steps. Delete the source file in formwork/roles/,
then delete its generated copies:
rm formwork/roles/packs/mobile.md
rm .claude/agents/mobile.md
formwork roles
If you forget the second one, the gate goes red and tells you: a generated file with nothing producing it. That is the check doing its job.
Yes, and you should. Copy
formwork/roles/TEMPLATE.md and read
formwork/roles/HOW-TO-ADD-A-ROLE.md.
The kit will refuse a role that does not say what it owns.
Several agents on the same question at once, arguing, with one document at the
end. It costs real money, so use one only when a decision is expensive to undo.
See formwork/round.md.
Nobody has measured it. formwork/COSTS.md says so
and says what would settle it. Set a spending limit with your provider; nothing
here can do that for you.
Warning
Partly. The guards stop the normal path. Somebody who wants around them
can get around them. formwork/limits.md lists every
route anybody has found. If you need real containment, use a sandbox.
No. There is no network code in the kit at all. Your agent talks to its provider as it always did, which is between you and them.
It adds and never removes. Install writes your settings, your hook wiring and
your role files, and one file outside your project in ~/.formwork/.
formwork/first-run.md says what that last one is for.
Because nothing makes it obey one. docs/style.md is a document, and documents
are advice. The only thing a check can prove is that every role points at the
page, which it does.
If it keeps happening, say so in the turn. A style file is a default, not a rule that refuses.
Because you changed a file that enforces something, and that is a decision a person makes. If it happens constantly you are editing the kit itself, which is fine.
It was pulled out of two real projects that were already being run this way. Nothing here is theoretical. Every rule is in the kit because something went wrong without it.
Its failures are written down in docs/dogfood.md. Read that
before you decide what you think of the rest.
The guards and checks have 486 tests and five audit rounds behind them. Every
check has been watched failing on purpose, and you can watch them yourself with
formwork demo.
Unknown. The guards assume a unix shell. Nobody has tried.
If you get it working, that is a real contribution.
Claude Code has been watched refusing a real command. Codex, Cursor and Gemini CLI are documented by their publishers and untried.
Trying one is the single most useful thing you can do for this project. Wire it up, ask your agent to commit something, and say what happened.
MIT licence. Yes.
It was built around one person and their agents. Teams may well find it useful, and a team version is not a goal yet, so expect to adapt it.
Open an issue. The most useful ones contain the exact message you saw and the command you ran.