Glossary

The vocabulary the platform is built on — each term defined once, used consistently across the engineering blog.

#Branch-promotion deploy

Going to production by promoting a copy-on-write database branch instead of dumping and restoring data. Production is a branch with its own timeline; a release pins a git commit and applies pending migrations to that branch. No bytes are copied, dev stays a live fork, and rollback is a point-in-time branch. Deep dive →

#Scale-to-zero Postgres

A per-app PostgreSQL whose compute scales to zero replicas while idle, so a fleet of mostly-idle databases costs nothing. A proxy that speaks the Postgres wire protocol parks the next incoming connection, wakes the compute, and forwards the query once it's live — clients never see the sleep. Deep dive →

#Copy-on-write database branch

An instant fork of a PostgreSQL database that shares storage pages with its parent (self-hosted Neon timeline). Creating one moves no data, which is what makes per-project databases, dev/prod separation, and pre-deploy checkpoints affordable at agent speed.

#Objective verification

Gating an AI agent's work with deterministic checks — the TypeScript compiler, live contract probes against the running pod, and the app's own unit tests — rather than an LLM judging its own output. A claim of "done" must survive evidence that the agent cannot talk its way past.

#Contract probe

A live check that every API procedure the architect declared actually responds on the running app without server errors, and that every declared database table exists. Run before a subtask may be marked complete; a missing procedure fails objectively.

#Born-testable scaffold

Every generated project ships with a test runner (vitest) and a working `npm test` from its very first commit, so verification can execute the app's own tests in its pod — there is no "we'll add tests later" phase.

#Code + data time-travel

Rolling a project back to one coherent moment across git source, the live Kubernetes runtime, and the database: the code restores to a commit, the data restores to the matching point-in-time branch. Undo for the migrations and bulk edits that git alone can't revert. Deep dive →

#Pre-deploy checkpoint

A point-in-time marker (LSN) tagged on production's database timeline immediately before each release is applied — the mandatory rollback anchor. Rolling back a release restores its pinned commit and, optionally, data to the checkpoint of whatever superseded it. Deep dive →

#Project namespace

The per-project Kubernetes namespace every generated app runs in, with its own resource quotas, limits, services, and ingress. The platform's isolation boundary: one project can never starve or touch another.

#Primary app

A project can host several isolated apps in one namespace; the primary app is the one serving the project's root preview domain. Sibling apps get their own subdomains, schemas, and deployments — and the platform converges to the declared set instead of silently falling back to the primary. Deep dive →

#Idle freeze

Suspending an inactive app's containers with a cgroup freeze: zero CPU, near-zero memory, instant thaw on the next visit. Together with scale-to-zero Postgres it's why an idle project costs effectively nothing.

#Keep-warm

A production tariff that keeps a deployed app always awake instead of freezing on idle — for apps whose users won't tolerate a cold-start. Priced per awake-hour.

#Production data timeline

Production's database history, independent of dev from the first deploy onward. Re-deploys apply migrations to it and never re-copy dev data; replacing prod data with dev's is a separate, explicit, snapshotted operation. Deep dive →

#Run reconciliation

The recovery sweep that detects builds interrupted by a crash or eviction and restores every piece of state — job status, task chain, locks, chat flow — to a consistent, resumable picture. The job system's answer to "the worker died mid-build".