Deploy and publish
pf deploy (pf d)
pf deploy # staging if configured, else production, else the repo's own deploy
pf d # same thing
pf d "fix login" # with a commit message, when the repo's deploy commits
pf d --dry-run # show the steps onlyWith no environment, pf picks the first target the repo has:
- Staging, when pf finds one (see below).
- Production, when that is the only target.
- A bare
deploy: adeployentry inpulpfixer.json, adeployscript inpackage.jsonorcomposer.json, or the repo’s own./a deploy. A bare deploy does not name an environment, so pf treats it like production: it always asks first and needs--yeswithout a terminal. pf never guesses it from platform configs. - The publish, in a repo that only publishes (see Deploy means publish).
pf says which one it picked. pf deploy staging and pf deploy production (or pf ds, pf dp) pick one yourself. The coding agent and the MCP tools never run a bare deploy.
A commit message
pf d "fix login", pf ds fix login and pf dp "release 3.4" pass the text to a deploy command that makes a commit. -m "text" is the explicit form. Words after staging or production are the message, and a mistyped flag such as --dryrun is an error, not a message.
pf cannot tell whether an arbitrary command commits, so only a target that says it takes a message gets one:
- a
pulpfixer.jsoncommand with"message": true, for example{ "run": "./tl d", "message": true }; - a repo CLI command (
./a deploy) whose# usage:header mentionsmessage.
The message goes on as one quoted argument (./tl d 'fix login'), on the last step of the plan. A target that takes none prints a warning and runs without it.
Repos with their own helper script
Older repos deploy with a hand-written helper such as ./tl d or ./al d. Run pf rp (or pf rp --deploy) once: the coding agent reads the repo’s docs and helper, asks which environment it is, and writes the pulpfixer.json entry. See Tools and repo setup.
Staging
pf deploy staging # detect the target and run it
pf deploy staging --dry-run # show the steps only
pf ds # same as pf deploy stagingpf finds the staging target by itself, in this order:
- A
deploy:stagingentry inpulpfixer.json - A
deploy:stagingscript in apackage.jsonorcomposer.json, or the repo’s own./a deploy:stagingtool - A GitHub workflow with
stagin its file name and aworkflow_dispatchtrigger, run withgh workflow run - Platform configs: Firebase, Cloudflare Workers (
wrangler, with a staging env), Fly.io (fly.staging.toml), Serverless, SST, Netlify and Vercel (a preview)
Steps print before they run. Anything pf skips is listed with the reason.
Production
pf deploy production
pf dpProduction is never guessed from platform configs. pf runs it only when the repo defines it:
- a
deploy:productionscript (package.jsonorcomposer.json), - a
deploy:productionentry inpulpfixer.json, or - the repo’s own
./a deploy:productiontool.
It always asks for confirmation. Without a terminal it refuses unless you pass --yes. --dry-run previews it.
The coding agent and the MCP tools deploy to staging only. Production is never targeted by an agent.
Publish to the registry
pf publish staging # same as pf psRuns the repo’s publish:cellars script (or publish:staging), or the matching pulpfixer.json entry, and always confirms (--yes without a terminal, --dry-run to preview). A repo with neither (and no staging deploy, see below) fails with exit code 1.
Deploy means publish, and back
Some repos ship by publishing packages (a publish:cellars script) and have no deploy at all. In such a repo pf d, pf ds and pf dp run the publish instead, and say so: Staging publish (cellars registry) (the repo has no deploy target, so pf deploy publishes). pf only does this when the repo shows no sign of deploying: no staging, production or bare deploy target, and no platform config it skipped (a firebase.json without a staging alias, for example). A repo with only a production deploy still fails pf ds instead of publishing.
The other way round, pf ps in a repo with no publish target runs the staging deploy. It never falls back to a production or bare deploy.
Either way the plan prints first and pf asks before it runs (--yes without a terminal). The coding agent and the MCP tools never publish.