Skip to content

Coding agents

dryfit never asks questions and every command answers in JSON, so an agent can drive it the way you do. What an agent lacks is knowing that dryfit exists and how to use it safely. dryfit ships that as a skill.

Terminal window
dryfit agent install

This writes the skill to ~/.claude/skills/dryfit/SKILL.md, where Claude Code finds it in every project. Start a new session and ask for what you want, such as “try my changes to the ui library in the shop app”. Claude Code looks at dryfit status, links the app, syncs while you work, and restores it when you are done.

To share the skill with everyone working on a project, install it into the project instead and commit it:

Terminal window
dryfit agent install --project

dryfit agent uninstall removes it, with the same options.

--dir installs the skill into any other folder of skills. For an agent that reads its instructions from a file, such as AGENTS.md, print the skill and copy what you need:

Terminal window
dryfit agent show
  • Pass --json to every command, read the result from stdout and the exit code.
  • Look at dryfit status before changing anything, and set dryfit up only when it never was.
  • Link only the apps you asked for, never commit a linked app, and restore the apps it linked when it is done.
  • When restoring fails because the app’s package.json changed, ask you before restore --force overwrites those changes.

With --json, an error is one line of JSON on stderr:

{
"error": {
"code": "library-linked",
"message": "@acme/ui is still linked into shop.",
"hint": "Run `dryfit restore` for those consumers first."
},
"exitCode": 1
}

The exit code is 0 when everything succeeded, 1 when something failed, 2 when the command was used wrongly, and 130 when it was interrupted. Commands that act on several apps, such as link and restore, report each app in their result and exit 1 if any of them failed.