Pick something small for the first run, like a new field or a simple page. You want to see the whole loop once before you hand it anything big.
Start it
In Claude Code, in the project:
/droost:workflow:start add a "Subtitle" text field to the Article content typeThe agent opens a run and starts planning. You can watch where it is at any time:
/droost:workflow:statusor, in a terminal:
tools/droost/vendor/bin/droost workflow:statusWhat you will see
The four steps from How it works: the agent looks at the brain, writes a spec, builds, and the checks run; it writes tests and they run; it writes down what it built. If it is unsure about something, it asks you. Answer it the way you would answer a colleague.
Finishing
A fresh install lets the agent finish the run itself (lead is autonomous).
If you set lead: human-led in droost/droost.yml, the agent ends with a
question instead:
This is the last hold: finishing ends the run and leaves the record in place for review. Ready?
Pick "Finish the run", or "Not yet" if something still needs work. Then only your answer finishes it.
Read the report
tools/droost/vendor/bin/droost workflow:reportCheck each gate's word (what they mean), that each requirement names the test that proves it, and the review's findings. The agent does not commit for you unless you ask: review the diff and commit it as you normally would.
If something goes wrong
- A gate keeps failing. After the retries are used up, the phase stops
and the run says why. Fix the cause, or, if the gate is wrong for this
change, waive it from your own terminal with a reason:
droost workflow:gate-waive phpstan "why". - A tool is missing. The gate says "could not run" and names the tool.
Install it (see Set up by hand), then
/droost:workflow:continue. - You want to start over.
droost workflow:reset --forcecloses the run and keeps its record in history. If the run left code in your working folder, stash or commit it first: a new run treats code it finds already changed as not its own.
More in Troubleshooting.