Only the schema (migrations on pgvector) job declares a postgres service
(.github/workflows/ci.yml:146-148). The middleware test job does not, so every
Postgres-backed suite silently skips in CI — the migration backfill tests, the embedding-gate
suites, devJobTaskStore.pg.test.ts, and the rest.
Silent skipping is the dangerous part: the job reports green and the suites look covered.
Surfaced during PR #550, where a restored migration assertion could only be verified locally, on a
privately-run pgvector/pg16.
Evidence that switching it on is probably cheap: a full local run with postgres enabled on node
22.22.3 gave 5982 tests / 5972 pass / 6 fail / 4 skip in 169 s, and none of the six failures were
Postgres-related (devplatform/goldenFixture.e2e, builder preview routes, runner-router
credential tests — all local-environment ENOENT/worker-provisioning artefacts).
Deliberately not fixed inside #550: it changes CI infrastructure for the whole repo, would be
invisible inside a 116-file MCP change, and if it does red-line on pre-existing failures it blocks a
PR that is otherwise ready.
Filed as a follow-up from PR #550 (MCP 2026-07-28 cluster, wave 4–6).
Only the
schema (migrations on pgvector)job declares a postgres service(
.github/workflows/ci.yml:146-148). Themiddlewaretest job does not, so everyPostgres-backed suite silently skips in CI — the migration backfill tests, the embedding-gate
suites,
devJobTaskStore.pg.test.ts, and the rest.Silent skipping is the dangerous part: the job reports green and the suites look covered.
Surfaced during PR #550, where a restored migration assertion could only be verified locally, on a
privately-run
pgvector/pg16.Evidence that switching it on is probably cheap: a full local run with postgres enabled on node
22.22.3 gave 5982 tests / 5972 pass / 6 fail / 4 skip in 169 s, and none of the six failures were
Postgres-related (
devplatform/goldenFixture.e2e,builder preview routes, runner-routercredential tests — all local-environment
ENOENT/worker-provisioning artefacts).Deliberately not fixed inside #550: it changes CI infrastructure for the whole repo, would be
invisible inside a 116-file MCP change, and if it does red-line on pre-existing failures it blocks a
PR that is otherwise ready.
Filed as a follow-up from PR #550 (MCP 2026-07-28 cluster, wave 4–6).