Railway preview environment is missing the agenta_oss_supertokens database, so deploy/deploy fails

[mmabrouk]

Found during the 0.106.2 release. Not a release blocker, but it makes the deploy / deploy check red and useless as a signal.

What fails

The deploy / deploy job, at the step "Deploy preview environment". Supertokens crashes on startup, repeatedly:

What caused the crash: io.supertokens.pluginInterface.exceptions.DbInitException:
java.sql.SQLException: com.zaxxer.hikari.pool.HikariPool$PoolInitializationException:
Failed to initialize pool: FATAL: database "agenta_oss_supertokens" does not exist

The job then exits 1.

Why it matters

The environment was bootstrapped without that database, so every deploy into it fails deterministically. This is the same shape of problem #5588 addresses for missing services: a preview environment that was created incompletely and then fails on every later deploy.

It is not caused by this release. #5567 showed the same red deploy / deploy before the train was merged.

Scope

This affects the Railway preview environments used for pull requests only. The release stages, staging.preview.agenta.dev and oss.preview.agenta.dev, deploy through the agenta_cloud workflows onto their own databases and are unaffected.

Suggested fix

Either create the missing database during bootstrap the same way #5588 now verifies services exist, or make the bootstrap fail loudly when a required database is absent so the environment can be repaired by a re-run instead of failing silently forever.

[linear-code]

AGE-4040

[mmabrouk]

Root cause pinned, from the full deploy log of the PR #5586 environment

Trigger: a fresh environment. No prior deployment for 'Postgres'; skipping redeploy (fresh env) plus a full initdb in the Postgres log. Fresh envs get an empty database server; reused envs carry a volume where the databases already exist, which is why only new PR environments fail.

The missing step. The stock Postgres image creates only POSTGRES_DB. The log shows the init-scripts directory was empty:

/usr/local/bin/docker-entrypoint.sh: ignoring /docker-entrypoint-initdb.d/*

docker-compose mounts api/oss/databases/postgres/init-db-oss.sql there to create agenta_oss_core, agenta_oss_tracing, agenta_oss_supertokens. Railway mounts nothing.

The tool exists but nothing automated calls it. hosting/railway/oss/scripts/init-databases.sh creates exactly those three databases, idempotently. It is called from upgrade.sh and documented in the README as a manual step. Neither bootstrap.sh nor deploy-from-images.sh (the CI path that failed at line 269 → smoke.sh) invokes it.

Failure chain. Supertokens crash-loops (FATAL: database "agenta_oss_supertokens" does not exist every ~2.5s) → the API, which needs supertokens plus the two app databases, never comes up → gateway answers 502 → smoke.sh fails /api/health after its 300s budget. /w passes because the web container needs none of it.

Fix

Call init-databases.sh from the deploy path after Postgres is healthy and before dependent services start (in deploy-from-images.sh, or at the end of bootstrap). The script already skips databases that exist, so running it on every deploy is safe. Same class of gap #5588 closed for missing services: the bootstrap step exists but is not wired into the automated path.