run-llama / llama-agents / 39873834246

85%

push

github

Serve pushed deployment UIs from a fixed link in the build artifact (#824)

## What changed

Push-created deployments never get `spec.staticAssetsPath`, so nginx
proxied their whole UI to the Python app. The build now records where
its UI landed inside the artifact itself: after `npm run build` it
writes a relative symlink `/opt/app/.llama-ui` pointing at the output
directory (`build_ui` returns that directory instead of a bool). When
`spec.staticAssetsPath` is empty, nginx aliases `/opt/app/.llama-ui/`
for the UI location, with the same `try_files` fallback the
git-deployment path uses.

Because the link travels with the tarball, every artifact describes its
own UI. Cached artifacts, rollbacks to an earlier build, and config-only
updates that reuse a build all serve the UI that was actually built,
with no extra state in the operator, the Build API, or S3.

## Fallback

The `@python_upstream` named location proxied to `127.0.0.1:8081`, which
is nginx's own port. It now proxies to the app on `8080`. That is what
makes the alias safe to use by default: an artifact without the link (no
UI configured, or built before this change) misses every `try_files`
candidate and falls through to Python, which is what those deployments
get today.

14 of 18 new or added lines in 2 files covered. (77.78%)

18953 of 22263 relevant lines covered (85.13%)

1.15 hits per line