Hangar

SDLC · Stage 03 of 07

Build and test

Build without special privileges, change shared things carefully, and test against the real thing.

Principles: The platform is a product, Event-driven by default

Nobody should have to be a CI expert to ship software. The goal here is a pipeline that feels boring in the best way: it runs, it's fast, and when it fails it tells you why in plain words.

The pipeline is a fixed superset DAG whose stages are switched on and parameterised by cicd.yaml, not an arbitrary graph compiled from user config. That is a deliberate limit: it keeps every tenant's pipeline recognisable and supportable, and a compiler is a heavier pattern than the problem has needed so far.

Images are built with rootless BuildKit, one daemon per build: no privileged container, no host access, and Pod Security baseline. It replaced kaniko in October 2026, which was archived upstream and emulated every cross-architecture build; the earlier claim that builds met restricted turned out never to have been true, and the docs now say so. The shared Tekton catalog is a Helm release pinned by git tag, so upgrading a tenant is a reviewed change to a pin and a catalog change can be tried on a canary tenant first. Stages are separate PipelineRuns chained by events, which keeps each one small and independently re-runnable.

Tests run in Testkube, and that decision is a good example of the habit behind this site: the documented multi-namespace mode turned out, in the source code, to be gated to a paid edition. The design that shipped runs in one shared namespace and closes the resulting secret-isolation gap with an admission policy.

Where Hangar does this

Flowchart · 02 of 07 · What you change What you change, and what the platform owns Open full page ↗
Sequence · 05 of 07 · Chaining How test gets started when build finishes Open full page ↗

And for AI agents

CI is the loop an agent retries against

For an agent, the pipeline is not just a gate; it is the feedback signal. Fast, deterministic checks with readable errors are what let an agent fix its own change before a human ever looks at it.

Flowchart · 04 of 10 · CI/CD validation loop CI/CD as a validation loop the agent can retry against Open full page ↗

The full Autopilot walkthrough →