Delivery and Environments
DevStride's environments, and the path every change takes from a branch to production: review, testing, release, approval and the checks after a deploy — including how AI coding agents take part.
On this page
Delivery and Environments#
Environments#
| Environment | Purpose |
|---|---|
| Production | What customers use. It runs in its own AWS account and is deployed only from the master branch, after approval. |
| Staging | A shared, long-lived environment deployed from the develop branch. Every finished change runs here before any customer has it, and everything on it ships with the next release. |
| Personal environments | Short-lived copies of the stack for individual developers, in the non-production account, removed when no longer needed. |
| Local | The whole stack can also run on a developer's machine in Docker, with no AWS account at all. |
Production and staging keep their data if a stack is removed, and both are monitored. See Deployment for how deploys run in detail and Operations and Monitoring for what is watched.
How a change reaches production#
Branch ─► local checks ─► review (two AI reviewers) ─► epic branch or release train
─► develop (staging deploy) ─► release branch ─► CI once ─► approval
─► master (production deploy) ─► post-deploy health check
- Planned in DevStride. Every change starts as a work item. Larger work is grouped under an epic.
- Built on its own branch. A change under an epic collects on the epic's own branch, where unfinished work is allowed to wait. Small standalone fixes collect on a shared release branch called the release train.
- Checked locally. Before a change can merge anywhere, it must pass a local check: every type check, lint on the changed files, and every test the change can affect.
- Reviewed. Every pull request is reviewed by two independent AI reviewers, an adversarial Claude review and OpenAI's Codex reviewer. Every finding is fixed or dismissed with a reason. High-risk areas, such as sign-in, billing, the public API, migrations and infrastructure, get extra scrutiny.
- Tested once, on the reviewed code. Pull requests open as drafts, and the cloud test suite does not run on drafts. It runs once, when review is finished, on exactly the code that was reviewed.
- Only complete work reaches staging. An epic reaches
developonly when all of it is finished, reviewed and safe to go live. Any new infrastructure it needs must exist, and any new cost must already be agreed. - Released. A release is cut from
developas a protected release branch, which accepts fixes only. Only one release can be in progress at a time across the whole team, and urgent fixes are never held behind it. - Approved. Production deploys wait in a protected GitHub environment for approval. A release carrying a database migration needs that migration approved by a person.
- Checked after deploy. A health check confirms the release, then the next release is allowed. See Operations and Monitoring.
Releases are small and frequent, often several a day.
Continuous integration#
CI runs on GitHub Actions, on machines that RunsOn starts inside DevStride's own AWS account.
| Check | What it covers |
|---|---|
| Backend tests | The full backend suite against real PostgreSQL and DynamoDB, split into 20 parallel shards |
| Web app tests | Unit and component tests in 4 shards, plus the customer portal's size budget |
| Type checks and lint | The backend, the web app and the MCP server |
| Security | Semgrep static analysis and gitleaks secret scanning |
| Licences | Every dependency's licence checked against an approved list |
| Merge rules | Only the right kinds of branch may merge into develop and master |
Merges to develop and master are restricted to approvers, and master cannot be force-pushed. Deploys sign in to AWS with short-lived GitHub identities that only the deploy workflow on the right branch can use.
Testing#
- Backend. About 1,800 test files, run with Vitest against real PostgreSQL (with pgvector) and DynamoDB in containers. Each test worker gets its own fresh database.
- Golden dataset. A complete, realistic demo organization is built from code through the product's own commands, then checked against a list of assertions. The same dataset serves as an end-to-end test, as sales demo and training data, and as test data for analytics. It is rebuilt from nothing every time, and is never fixed by hand.
- Web app. Unit and component tests with Vitest, plus type checks.
- Infrastructure. The infrastructure definitions have their own tests. For example, they check that every scheduled job has its alarms, and that every handler a stack names actually exists.
How AI agents take part#
Much of DevStride's code is written and reviewed by AI coding agents (Claude Code and Codex) working through a plan in DevStride one item at a time, often several at once.
- Machine-checked rules. The process above is enforced by checks the agents cannot bypass: required CI checks, merge rules, protected branches and protected environments.
- Claims. Agents claim work in DevStride before starting it, so two agents never pick up the same item or release.
- The record lives on the item. Each work item's description is updated to say what actually shipped.
- People keep the decisions. Releases that carry migrations, new costs or changes that cannot be undone need a person's approval.
The full process is described in How DevStride Is Built, and Why.