Never in a migration (spec §13.7). make seed applies every .sql here in name order, against a database that has already been migrated. Each file is idempotent (fixed ids, ON CONFLICT DO NOTHING) and is numbered at merge, one number per file in merge order; a file's header names the seeds it relies on.

What is here:

• 01-dev.sql — the Admin, one student and one class with that student in it. Enough to sign in and click through the app.
• 02-dev-assignments.sql — one published test, one open assignment, one submitted-and-flagged attempt, so the dashboard has something to count.
• 03-dev-students.sql — a few more students, including a Google-only one.
• 04-dev-e2e.sql — the take-test fixtures the live E2E suite uses.
• 05-dev-sections.sql — a published paper with three parts.
• 06-dev-access.sql — two Teachers, each with a class and a draft test, and an Assistant (R2).
• 07-dev-classroom.sql — thirty students in no class, for the classroom join test (R2).
• 08-dev-teacher-workspace.sql — the Teacher deck's fixtures for [email protected] (R4, T-R4.51): seven tests, eight media files, four classes with members, assignments in every status, attempts, imports. It follows 06 and 07 and relies on 06's teacher. See below.
• 99-assert.sql — runs last and refuses the seed if the rows above break a rule publish would have enforced, or if a user, class or owned row lacks the column R2's access model reads. The seed writes frozen version rows directly, which is the one path that skips publish's validation.

Every account's password is quizzivy-dev.

Everything belongs to Trần Thị Lan ([email protected]), named in every owner_id, teacher_id and created_by. Ids are fixed: …00000080 + a kind + a number (pg_temp.sid), or derived from a label (pg_temp.did); the header lists the kinds.

A published version is written the way the publish command writes it (group_v1, the draft's rows copied, the test moved to published), and the file is one transaction.

The seed has no object-store access, so media rows are metadata only: the library, the tests and the question editor load and list them, but an image shows as broken and an audio file does not play, because the signed URL points at a key with no object. Nothing here commits a binary. To give the audio rows sound for a look at the player, copy the one small fixture over their keys:

keys=$(psql "$MIGRATE_DSN" -Atc "select storage_key from app.media_assets where kind = 'audio' and id::text like '01935000-0000-7000-8000-00000080%'")
docker compose run --rm --no-deps -v "$PWD/web/tests/e2e/fixtures:/fx:ro" --entrypoint /bin/sh minio-init -c \
  'mc alias set local http://minio:9000 "$S3_ACCESS_KEY_ID" "$S3_SECRET_ACCESS_KEY" >/dev/null &&
   for key in "$@"; do mc cp /fx/unit5-listening.mp3 "local/$S3_BUCKET/$key"; done' sh $keys

MIGRATE_DSN is the Makefile's. Images stay broken: the repository holds no image fixture.

Rows only again: the import bucket holds no source file, artifact set or evidence. The history, the detail and the review (its questions, findings and counts) read the database and load; the review's source pane reads evidence from storage and answers with an error for these four. The processing import's run is queued at the extraction stage and not due for ten years, so no worker claims it and it takes none of the worker's slots; cancel it from the page to clear it. A database seeded before 2026-10-12 holds the old row, a run "running" on a ten-year lease that takes the worker's only slot; cancel that import or seed again.

Still planned: a volume seed (~50 students, 10k attempts, 500k integrity events) for the query review, which measures p95 latency at realistic volume rather than asserting plan shapes.

_{Generated from seed/README.md. Edit that file in the repository; changes made on this page are overwritten.}