upkeep

One file describes what an app's cloud footprint should be. upkeep reads what it actually is, prints the difference, and — only if you ask — closes it.

$ upkeep plan
MANUAL  render-env  FEE_ACCOUNT_INFO      real bank details, shown to a customer verbatim
MANUAL  render-env  R2_SECRET_ACCESS_KEY  set $R2_SECRET_ACCESS_KEY locally, then re-run
CREATE  render-env  R2_BUCKET             not set on the service → my-bucket
UPDATE  r2-cors     my-bucket             no CORS rule at all, so a browser cannot upload

2 of these cannot be automated and are yours to do.

Why it exists

The failure it was written after: an app shipped with none of its object-storage variables set. Nothing errored. The service booted, served every page, and answered 404 to one endpoint — photo upload — for days. Every layer reported success, because from each layer's point of view nothing had gone wrong.

So upkeep has a fourth kind of action, MANUAL, for changes no tool can make: an R2 API token cannot be minted from any CLI, and real bank details are nobody's to invent. A tool that could only automate would have reported that same clean success. Naming the gap is the feature.

For the same reason apply re-reads afterwards. A provider returning 200 is not evidence a change took.

What it manages

Surface	Reconciles	Reports only
Render	service environment, deploys
Cloudflare R2	bucket, public access, CORS
Cloudflare Pages	production branch	domains it does not know about
Neon		branch exists, DATABASE_URL agreement

It never deletes by omission: a variable the config does not name is left exactly as it is. Deleting a database or repointing a live service at another one are not things a config file should do quietly, so it does not.

Install

go install github.com/zoolcoder/upkeep/cmd/upkeep@latest

Credentials

Every provider reuses the session its own CLI already stored, so on a machine where render, wrangler and neonctl work, upkeep works too.

Variable	Falls back to
RENDER_API_KEY	~/.render/cli.yaml after render login
CLOUDFLARE_API_TOKEN	wrangler's config after wrangler login
NEON_API_KEY	~/.config/neonctl/credentials.json after neonctl auth

A surface whose credential is missing is reported, never skipped silently. An expired borrowed session says which command to re-run, rather than failing as a 401 that reads like a permissions problem.

Getting a config

Do not write one by hand — that is how a config ends up describing an app that does not exist. Read it off the account:

$ upkeep import -name my-app -render srv-… -bucket my-bucket -pages my-app

import and plan compose, so you can check a config you have not saved:

$ upkeep import -name my-app -bucket my-bucket \
  | { echo "version: 1"; echo "apps:"; cat; } \
  | upkeep plan -config -

See upkeep.example.yaml for the full shape.

Secrets

The config file is meant to be committed. Nothing secret goes in it.

A variable either carries a literal value, or names the local environment variable its value comes from:

- key: R2_BUCKET             # not secret
  value: my-bucket
- key: R2_SECRET_ACCESS_KEY  # secret: resolved at apply time, never written
  valueEnv: R2_SECRET_ACCESS_KEY
- key: FEE_ACCOUNT_INFO      # nobody's to invent
  manual: true
  why: real bank details, set by hand on the service

Three things enforce that, rather than trusting you to remember:

1. A literal under a credential-shaped name is refused at parse time. Write value: under a key containing SECRET, TOKEN, PASSWORD, API_KEY, DATABASE_URL and friends and the config does not load, with an error naming the field to use instead. The check is deliberately eager — a public value wrongly refused costs one line of config; a secret wrongly committed costs a rotation.
2. import never copies a live secret out. A variable whose name looks like a credential comes back as a valueEnv reference with no value, using the same rule as the validator, so an import always parses back in.
3. No output ever prints a value. A plan says differs from the config (value from $R2_SECRET, not shown). -json follows the identical rule — a machine-readable view that leaked what the text view hides would be a way around the guarantee, not a second view of it.

Threat model

upkeep holds credentials in memory for the length of one run and writes none to disk. Every API client redacts its token in String(), so no accidental %v can print one.

It does not protect against a hostile config file: a config can point at any service id or account you have access to, and valueEnv can name any variable in your environment. Review a config you did not write, the same as any script.

The heuristic that classifies a name as secret is a heuristic. It will not catch a credential in a variable called SETTINGS_BLOB. Use valueEnv for anything sensitive regardless of what it is called.

Exit codes

0	nothing to do, or the apply converged and was verified
1	something failed, or an applied change is still not in place
2	plan -exit-code only: the live state differs from the config

So CI can gate on drift:

$ upkeep plan -exit-code -json > plan.json

Documentation

The full manual is at https://zoolcoder.github.io/upkeep — getting started, a page per surface, the threat model, and how it differs from Terraform.

Build it locally:

npx antora antora-playbook.yml && open build/site/index.html

Developing

make test     # go test ./...
make lint     # go vet + gofmt check
make build    # ./upkeep

No network is needed to run the tests. Every provider is behind an interface and every planner is tested against a fake, including one that accepts every write and changes nothing — the failure apply's verification exists to catch, and one no real API will perform on request.

Where it is going

See CHANGELOG.md for what has landed, and ROADMAP.md — including what has been considered and declined, which is the more useful half.

Contributing

See CONTRIBUTING.md. Everything runs offline; CI has no credentials configured on purpose, so a fork can be tested without being trusted.

Security

See SECURITY.md — including what is deliberately not in scope, which is the more useful half.

Licence

MIT. See LICENSE.