Architect: start
Architect
The board running the infrastructure your repositories run on, and you still deciding what changes. What Architect is, how a change goes, what stays true, its words, and where to read on.
Architect is the part of the board that runs what your repositories run on: their Workers, databases, namespaces, buckets, queues, containers, routes, and custom domains. It came in 2.0.0, it runs on Cloudflare, and it’s off until you connect a provider. Until then the board makes no call for it.
It works the way the rest of the board does. What should exist is a file in the repository. Agents change it by pull request. The board works out the exact change, a plan, and the plan waits for you. You press Approve, and a workflow in the repository applies it, with a write token the board never holds. Then the board checks health, and rolls back by itself if the check fails.
#How a change goes
the file: what should exist Cloudflare: what runs
staging.json (the read-only token)
\ /
the board compares them
|
v
a plan: what changes, what it costs,
what else it touches, can it be undone
|
the policy: it waits for you, or it
fits bounds you approved once
|
you press Approve
|
v
the repository's apply workflow, with
staging's write token, from a GitHub
environment only the default branch uses
|
v
the board checks health, and rolls
back by itself if the check fails
Every step is on the board: the plan as it forms, your approval, the run, and the health check, on the environment’s page and in its audit trail. A plan that waits for you sends one push, so you can approve from your phone.
#Two lanes: code, and what it runs on
Architect doesn’t replace the deploy flow. They’re two lanes, and a repository can use either or both.
| The deploy flow | Architect | |
|---|---|---|
| What it changes | The code: a new version of a Worker | What the code runs on: a Worker’s settings and bindings, and the resources around it |
| Where it’s written | The code and its wrangler config | .github/breakaway-infra/<environment>.json |
| How it goes | Merging deploys staging; you press Promote and Roll back | Merging plans; you press Approve, the board applies |
| Its buttons | Promote, Roll back, Release | Approve, Reject, Freeze |
On a repository with both, staging and production are one pair of environments: their deploys, Promotes, and Roll backs show on the same pages as their plans. Architect and the deploy flow says how to set up an app and its infrastructure together.
#What stays true
- Agents propose and never apply. They read everything and change infrastructure only by pull request. They never hold a write token, start the apply workflow, or press Approve.
- Nothing changes without your approval, or inside bounds you approved once on an environment: an envelope, or a policy you loosened. Until then, every plan waits for you, in every environment.
- The board holds no write credentials. It keeps one read-only token per provider. Each environment’s write token lives in a GitHub environment in its repository, and only the apply workflow reads it, for one plan you approved.
- The board only watches its own install. The environment the board runs on is always observe only: no plans, no envelope, no apply. Its way back is Recover without the board.
- Your data stays in your install. The board keeps names, kinds, settings, health, and cost, never a secret’s value, a stored value, or code. The read-only token goes only to its provider.
#The words
Architect uses the same few words everywhere: on the board, in pushes, in the CLI, and here.
| Word | What it means |
|---|---|
| environment | A named place a repository runs: production, staging, or short-lived (one task’s own). |
| provider | A platform you connect so the board can see what runs there. Cloudflare, in 2.0. |
| desired state | What should exist in one environment: its file in .github/breakaway-infra/. |
| change | What you edit on an environment’s page before it becomes a pull request and a plan. |
| plan | The exact change the board would make to one environment: what changes, what it costs, what else it touches, and whether it can be undone. |
| approve, reject | Your answer to a plan that waits for you. Only you can, and only on the board. |
| apply | What the board does with a plan you approved. It’s a status, never a button. |
| policy | The rules that decide which plans wait for you. |
| envelope | Bounds you approve once on one environment, like “2 to 10 instances” or “3 restarts a day”. |
| signal | One thing the board heard about an environment: its health, a platform’s alert, or its cost. |
| incident | A task tagged +incident, opened when a signal crosses a rule. |
| drift | What runs no longer matches its file. |
| break-glass | A change you made by hand outside a plan, on purpose. The board records it and never undoes it. |
| freeze | Stop every plan on one environment until you unfreeze it. |
| observe only | An environment the board watches and never changes. |
#Where it shows
- Infrastructure, in the sidebar (
gthenn): each repository’s environments, with health, drift, estimated cost, freeze, and the plan waiting. Add an environment and Policy are here. - An environment’s page, its console: what runs there as a map, a live stream of what happens, its plans, incidents, drift, deploys, cost, and audit trail. You change it here too.
- A plan’s page, the link a waiting plan’s push carries, with Approve and Reject.
- The inbox: incidents lead it, production’s first.
- A repository’s settings page, under Infrastructure: the policy in force, each environment’s freeze, and its envelope.
- Connections: the provider’s token, and each repository’s Infrastructure tokens checklist.
- Settings, under The board: the currency costs show in.
- The terminal:
npx breakaway infraand its reads, for you and your agents (Agents and infrastructure).
#Read on
| You want to | Read |
|---|---|
| Go from nothing to your first approved plan | Get started with Architect |
| Know what it can run, change, and scale | What’s supported |
| Add an environment, or build one from nothing | Environments |
Write or read the files in .github/breakaway-infra/ | Describe it as code |
| Change what runs, from the board or by pull request | Change an environment |
| Read a plan, approve it, and follow it to applied | Plans and approvals |
| Decide which plans wait for you | Policy |
| Let the board scale and restart within bounds | Envelopes and scaling rules |
| Hear when something breaks, and work it | Signals, incidents, and runbooks |
| See what it costs, and set budgets | Cost and budgets |
| Copy a setup that works | Patterns |
| Set up the tokens, GitHub environments, and the apply workflow | Tokens, GitHub, and the apply workflow |
| Run an app’s deploys and its infrastructure together | Architect and the deploy flow |
| Know what agents may do | Agents and infrastructure |
| Stop every change, or see who holds an environment | Freeze, gates, and locks |
| Change something by hand, and put it back in code | Drift, break-glass, and clean up |
| Fix what didn’t go through, or recover without the board | When it goes wrong |