Bundle Postgres, Valkey and NATS into the kit image

[07prajwal2000]

Why

The kit is the trial build, but getting it running still meant copying an environment file, editing it, and bringing up a four-container compose stack. Caddy was already inside the image, so the backing services were the only reason a compose file was needed at all.

They are now bundled. docker run -p 8080:8080 is a complete Fluxify instance with nothing else installed.

What

Bundled backing services. PostgreSQL, Valkey and NATS ship in the image and start automatically. Each one is skipped when its connection variable is already set (PG_URL / REDIS_HOST / NATS_URL), so docker/kit/docker-compose.yml drives the external topology from the same image with no changes. One image, both worlds, no second Dockerfile.

State lives in a /data volume: the Postgres cluster, the JetStream store, and the encryption/session keys generated on first boot when not supplied. Image is 338MB.

Admin credentials are enforced. The seeder skips silently when SEED_USER_EMAIL / SEED_USER_PASSWORD are missing or malformed, which leaves a running instance nobody can log into. A fresh database now refuses to boot instead, before initdb, with a message saying why.

The admin UI is apps/portal. The kit still built the legacy Next.js app, which no longer typechecks against the current server schemas (it imports test-suites/run/dto, renamed to start-run). Per AGENT.md that app is legacy and must not be edited, so the kit moved to the shipping UI. It is a static Vite bundle served straight by Caddy, so the UI process is gone entirely.

Bugs found on the way

The kit could not boot without these. All predate this branch:

• entrypoint.sh did not supervise. A bare wait blocks until every child exits, so one dead service left the container "up" and quietly broken — that is how the AI gateway had been crash-looping unnoticed. It now exits non-zero on the first death and lets the restart policy handle it. Fixing this is what exposed the three below.
• DOCS_INDEX_PATH used path.join on an absolute env value, producing /app/ai-gateway/app/ai-gateway/docs-index.bin. path.resolve honours absolute paths and still resolves relative ones against the module directory.
• The AI gateway has no schema-wait and died on a missing app_config. Its startup is now gated on the admin server answering.
• HEALTHCHECK pointed at /_/admin/api/healthchecks/startup, which does not exist — the container could never report healthy. Uses public-settings.

Also in here

• Portal: the editor page is gone; routes open on the canvas. Two flow editors existed for the same job; the canvas is the real one (block picker, playground, route switcher, settings modal) while the editor was a thinner duplicate with "coming soon" tabs. All four entry points moved: the routes-list Open button, the post-create redirect, the AI resource chip, and the applied-artifact link. That last one had a target prop to choose between the two — dead flexibility now, so removed rather than pinned to a constant. components/editor/ went with it; nothing else imported it.
• Portal: sidebar and AI-home logos were broken. Both hardcoded /_/admin/ui/public/icons/logo.svg; files under public/ serve from the bundle root, so the public/ segment matched nothing. Because the SPA fallback answers unknown paths with index.html, the request returned 200 with an HTML body rather than 404 — the browser got markup where an image was expected. Now uses BASE_URL, and webp.
• .gitattributes pins shell scripts to LF. A CRLF shebang makes the kernel look for /bin/sh\r; the container dies with exec /app/entrypoint.sh failed: No such file or directory.
• Root dev script launched @fluxify/web; points at @fluxify/portal now. Vite binds 0.0.0.0 so a containerised Caddy can reach it via host.docker.internal, and the bare /_/admin/ui path redirects to the trailing-slash form Vite's base actually serves.

Testing

Verified against a running container throughout, not just at the end:

• Bundled mode boots clean to healthy on a fresh volume — migrations, seed user, compiler, worker, AI gateway.
• Restart reuses the cluster and the generated secrets; no re-init, no regenerated keys.
• UI, admin API, static assets and SPA deep links all serve; /_/admin/ui → 308 → 200.
• A user route (GET /users) returns 200 through the proxy once WORKER_PROJECT_ID is set.
• External mode (PG_URL / REDIS_HOST / NATS_URL set) starts zero bundled services, so compose is unaffected.
• apps/portal typechecks and builds; the route tree regenerates with no editor entries.
• Full suite: 740 passing, 2 skipped.

Not in scope

docker/production/ and docker/admin/ are untouched by request. They still build apps/web, which is worth a follow-up.

Docs

docs/deployments/kit.md leads with the one-command run and keeps bring-your-own-services as the second option. Covers the parts that bite: mount /data or lose everything, back up the generated keys, admin credentials are mandatory, and a bundled major database upgrade is not applied in place. Also documents WORKER_PROJECT_ID=* as a way to skip the create-project-then-restart step, with the path-collision caveat stated plainly.

Fixes two dead anchors in the deployments index and folds the backing services into the kit diagram.

---

Related: #251 tracks project-scoped route resolution, which is what would let the kit default to serving every project without the collision caveat.

🤖 Generated with Claude Code