shipline

Deploy only the Firebase functions your change actually affects.

One functions codebase means firebase deploy --only functions redeploys all of them on any change — slow, and heavy projects trip the Cloud Functions write-ops quota. shipline builds each function's transitive require-closure from source and deploys only the functions whose closure includes a changed file. On the project this was extracted from: a typical domain edit went from 83 functions to 4–14.

npm install -g shipline

shipline affected --since HEAD~1          # what would deploy?
shipline deploy --project my-staging --staged
shipline watch                            # fail-closed staging auto-deploy

Zero dependencies. Requires a barrel-style functions/index.js (exports.name = require("./domains/x").name;) — the pattern you want anyway.

Never under-deploys — safe by design

A targeted deploy updates only the named functions; everything else keeps its last-deployed code. So the resolver must be provably complete, and it falls back to a FULL deploy whenever it can't prove safety:

• index.js / package.json / lockfile / .env changed → full
• a changed source file that no function's closure reaches → full
• a shared file reaching ≥60% of functions → full (cleaner than 50 --only flags)
• index.js unparsable → full

Inline exports in index.js (no require delegation) can't be statically scoped — they ride along on every targeted deploy and get reported, so you know to move them into a module.

The watcher (shipline watch)

A polling staging auto-deployer with every guard we learned the hard way:

• Fail-closed config — the project id is explicit (never a remappable alias), must exist in .firebaserc, must not be in forbidProjects (production stays promote-only), and the repo path must not match forbidPathPatterns (deploying from a OneDrive/Dropbox-synced copy is a corruption vector).
• Hooks that can't ship stale artifacts — a preDeploy hook (your bundler) that FAILS never lets its target deploy; only a successful hook earns the deploy. Hook outputs get their mtimes adopted so the watcher's own writes don't re-trigger it forever. We once shipped a stale bundle under a green log; this rule is why you won't.
• Failure backoff with a loud record — retries back off 30s → 5min instead of hammering on the save-debounce, and 3+ consecutive failures append a LOUD line to .deploy-state/deploy-log.txt. We once lost an hour to deploys silently not landing, one terse scrollback line per cycle. Scrollback is not a record.
• Fingerprint skip — functions source is hashed (lockfile included, node_modules excluded); an unchanged hash skips the multi-minute functions build on watcher startup. The marker is written only after a successful FULL deploy.
• The 429 fix — GOOGLE_CLOUD_QUOTA_PROJECT is set to your project. firebase-tools bills its cloudbilling pre-check to a shared internal project whose global quota is saturated (firebase-tools#9895) — the reason functions deploys 429 while hosting sails through. Enable the Cloud Billing API on your project once, and the 429s stop.

Config: see example/shipline.config.json.

CI usage

- run: npx shipline deploy --project my-staging --since ${{ github.event.before }}

shipline affected --since <ref> prints the JSON verdict + the exact firebase deploy command, so you can wire it into anything.

Honest scope

CommonJS require-closure analysis only (ESM functions codebases: PRs welcome, the resolver is one module). Polling watcher by design — editor-agnostic, works where fs-events don't. Single-codebase Firebase projects; if you use functions codebases already, you have a different solution. Maintenance is best-effort; issues with a failing test reproduction get attention first.

Where this comes from

Extracted from the deploy line behind Doric — every guard in the watcher is a production incident we only had to have once. The write-up: doric.build/blog/shipline.

Sibling packages: keepline (context integrity for LLM sessions) · wireline (find code that is built but never wired) · plotline (the context-integrity benchmark).

MIT © Gabriel Kerner