ci: run backend tests against Postgres, not only in-memory SQLite

[lucastononro]

Summary

• backend/tests/conftest.py unconditionally forced DATABASE_URL=sqlite+aiosqlite://, while production runs postgresql+asyncpg. That hid real divergences: models.py uses Column(JSON) in ~9 places (SQLite vs Postgres JSON semantics differ), the FK ON DELETE CASCADE behavior, and the hand-rolled db._run_migrations were never exercised against the engine we actually ship.
• backend-test now runs a real postgres:16-alpine service container and points the suite at it.

Changes

• .github/workflows/ci.yml (backend-test): add a postgres:16-alpine services: container (health-checked via pg_isready) and a job-level TEST_DATABASE_URL pointing at it.
• backend/tests/conftest.py: respect TEST_DATABASE_URL (highest priority — this is what CI sets), then DATABASE_URL if already set, else fall back to the existing in-memory SQLite default for local dev. No behavior change for anyone not setting either var.
• backend/pytest.ini (new): pins asyncio_default_fixture_loop_scope / asyncio_default_test_loop_scope to session. This was a necessary correctness fix, not just plumbing — without it, pytest-asyncio 1.x's default per-test event loop silently works with aiosqlite but breaks the module-level asyncpg engine with RuntimeError: ... got Future <Future ...> attached to a different loop as soon as a second test needs the DB (asyncpg connections are strictly bound to the loop that created them; the engine is created once at import time). Session-scoping the loop keeps every test on the same loop as the engine, for both SQLite and Postgres.

Test plan

Existing

• Full suite against the existing SQLite default (unchanged): 296 passed, 8 skipped.

New

• Full suite against a real dockerized postgres:16-alpine via TEST_DATABASE_URL: 296 passed, 8 skipped — identical pass count to SQLite. The 8 skips are the pre-existing RUN_LLM_E2E-gated live-provider tests in test_providers_e2e.py, unrelated to the database backend; nothing had to be newly skipped or marked to keep Postgres green.
• Confirmed the coverage gate from Add coverage measurement and a coverage gate to CI #116 still passes unchanged on top of this (65.81% ≥ 60% floor).

Caveats

• No test needed skipping/marking for Postgres-only reasons — the suite was already engine-agnostic once the event-loop scoping bug was fixed.
• CI now always runs against Postgres (matching prod); the SQLite fallback in conftest.py remains for fast local iteration when TEST_DATABASE_URL/DATABASE_URL aren't set.

Closes #117

🤖 Generated with Claude Code

Greptile Summary

This PR wires the backend-test CI job to a real postgres:16-alpine service container so the test suite exercises the same engine that ships in production, and fixes a previously hidden correctness bug where pytest-asyncio's per-function event loop silently worked with aiosqlite but would raise RuntimeError: got Future attached to a different loop the moment asyncpg reused its pooled connections across tests.

• CI (ci.yml): adds a health-checked postgres:16-alpine service and sets TEST_DATABASE_URL so the job points at it automatically.
• backend/pytest.ini (new): pins asyncio_mode = strict and both loop-scope options to session, which is the correct pytest-asyncio 1.x replacement for the old event_loop fixture override.
• backend/tests/conftest.py: replaces the hard-coded SQLite assignment with an explicit TEST_DATABASE_URL-wins / SQLite-fallback guard, closing the previous concern about ambient DATABASE_URL env vars pointing at developer databases; adds a regression test to lock this behaviour.

Confidence Score: 5/5

Safe to merge — the only finding is a dead pre-existing fixture that pytest-asyncio 1.x silently ignores, causing no runtime harm.

The DATABASE_URL guard in conftest.py is correct and directly addresses the previous review concern; the pytest.ini loop-scope settings are the proper pytest-asyncio 1.x mechanism; the CI Postgres service is health-checked and the credentials are appropriate for an ephemeral test container. The one cleanup item (dead event_loop fixture) is harmless.

backend/tests/conftest.py — the pre-existing event_loop fixture at lines 71-75 is now dead code and can be removed.

Important Files Changed

Filename	Overview
.github/workflows/ci.yml	Adds a postgres:16-alpine service container with health checks and sets TEST_DATABASE_URL for the backend-test job; setup is correct and well-structured.
backend/pytest.ini	New file pinning asyncio_mode=strict and both loop-scope options to "session"; valid pytest-asyncio 1.x configuration that replaces the old event_loop fixture approach.
backend/tests/conftest.py	DATABASE_URL logic now correctly honours only TEST_DATABASE_URL (falling back to SQLite), addressing the previous concern about ambient env vars; custom event_loop fixture (pre-existing, lines 71-75) is now dead code in pytest-asyncio 1.x.
backend/tests/test_conftest_db_url.py	Adds a regression test verifying DATABASE_URL is always either TEST_DATABASE_URL or the SQLite default after conftest import.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[pytest starts] --> B{TEST_DATABASE_URL set?}
    B -- Yes CI path --> C[os.environ DATABASE_URL = TEST_DATABASE_URL]
    B -- No local dev path --> D[os.environ DATABASE_URL = sqlite+aiosqlite://]
    C --> E[SQLAlchemy async engine against Postgres]
    D --> F[SQLAlchemy async engine against in-memory SQLite]
    E --> G[pytest.ini: asyncio_default_test_loop_scope = session]
    F --> G
    G --> H[setup_db fixture: create_all / drop_all per test]
    H --> I[Tests run on single session-scoped event loop]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[pytest starts] --> B{TEST_DATABASE_URL set?}
    B -- Yes CI path --> C[os.environ DATABASE_URL = TEST_DATABASE_URL]
    B -- No local dev path --> D[os.environ DATABASE_URL = sqlite+aiosqlite://]
    C --> E[SQLAlchemy async engine against Postgres]
    D --> F[SQLAlchemy async engine against in-memory SQLite]
    E --> G[pytest.ini: asyncio_default_test_loop_scope = session]
    F --> G
    G --> H[setup_db fixture: create_all / drop_all per test]
    H --> I[Tests run on single session-scoped event loop]

Loading

_{Reviews (2): Last reviewed commit: "fix(review): address Greptile findings o..." | Re-trigger Greptile}

[lucastononro]

Greptile review findings addressed (de197b8)

Fixed: 1 / 1

P1 — backend/tests/conftest.py:34 (ambient DATABASE_URL could be dropped by tests) — FIXED

Valid finding. The middle elif branch silently adopted whatever DATABASE_URL was already in the shell, and the setup_db fixture runs Base.metadata.drop_all after every test — a developer with DATABASE_URL pointing at a real dev database would have it wiped by pytest. Applied the suggested fix: only the explicit TEST_DATABASE_URL opt-in is honored; everything else falls back to in-memory SQLite (the pre-PR behavior). CI is unaffected since it sets TEST_DATABASE_URL.

Tests

• Added backend/tests/test_conftest_db_url.py pinning the guarantee: after conftest import, DATABASE_URL is either TEST_DATABASE_URL or the SQLite default — never an ambient shell value. Verified it passes under all three scenarios (clean env, bogus ambient DATABASE_URL, explicit TEST_DATABASE_URL).
• Behavioral check: ran suite subsets with DATABASE_URL=postgresql+asyncpg://fake-dev-host:5432/production_db in the environment — tests run on SQLite and pass instead of connecting anywhere.

Validation

• Full backend suite (py3.12, SQLite default): 297 passed, 8 skipped
• ruff check + ruff format --check: clean
• .github/workflows/ci.yml: valid YAML