# Timezone-aware datetime migration — handoff

> Follow-up artifact spawned by the timeago-shared-display work
> (`docs/superpowers/specs/2026-05-16-timeago-shared-display-design.md`).
> This is **not an implementation plan** — it is a brief for the next
> session to brainstorm the actual spec from evidence. Handoffs that
> pre-bake a plan tend to bury false premises; treat the "open
> questions" below as load-bearing.

## Context

`left4me`'s `l4d2web` package uses naive datetimes throughout the
runtime because `models.py` declares `DateTime` columns without
`timezone=True`. SQLAlchemy strips tzinfo on store (SQLite has no
native TZ type) and returns naive datetimes on read. The
`now_utc()` factory at `models.py:22-23` writes aware UTC, but the DB
layer collapses it back to naive on round-trip.

The pattern's cost is a defensive shell of `.replace(tzinfo=None)`
calls sprinkled across routes and services, plus a permanent mental
tax: *all naive datetimes in this codebase are UTC by convention, but
nothing in the type system says so.* Aware-everywhere would let the
defensive shell come down, the convention become explicit, and the
`_ensure_utc` defensive line in
`l4d2web/l4d2web/services/timeago.py` become a no-op safety net rather
than load-bearing.

This was discussed during the timeago-shared-display brainstorm and
deliberately scoped out to keep the diff focused. The user explicitly
asked for this handoff so the work isn't forgotten.

## What is currently known (evidence)

### The convention is documented in code

`l4d2web/l4d2web/routes/profile_routes.py:86-88` carries the canonical
explanation:

```python
# Strip tz so the marker matches what a subsequent DB read returns
# (SQLite DateTime columns don't preserve tzinfo).
user.password_changed_at = now_utc().replace(tzinfo=None)
new_marker = user.password_changed_at.isoformat()
```

The `pw_changed_at` session marker (`auth.py:52`, set in many routes)
is compared as an ISO string, so the on-DB and on-session
representations must match. Today they're both naive; an aware
migration must keep both sides in sync.

### Known workaround sites (definitive enumeration not done)

From a grep at the time of the timeago work:

- `l4d2web/l4d2web/routes/profile_routes.py:88` — password-changed-at write
- `l4d2web/l4d2web/routes/server_routes.py:210, 234` — `cutoff` / `recent_cutoff` for live-state staleness
- `l4d2web/l4d2web/routes/page_routes.py:174` — `cutoff` for the same
- `l4d2web/l4d2web/services/live_state_poller.py:37` — `_now()` helper returns naive UTC

The session that picks this up **must re-enumerate fresh** — the
list above is a starting point, not an exhaustive inventory. New
sites may have appeared after this handoff was written.

### Column declarations to flip

`l4d2web/l4d2web/models.py` declares ~20 `DateTime` columns without
`timezone=True`. Concrete count and list to be produced fresh
during the next session — column lists drift.

### What's *already* aware

- `now_utc()` (`models.py:22-23`) returns `datetime.now(UTC)`
- Most route-local time variables built via `datetime.now(UTC)` *before*
  the defensive `.replace(tzinfo=None)` is applied
- `timeago.py::_ensure_utc` (added in this session) handles both
  inputs defensively

### Tests touching naive datetimes

`l4d2web/tests/*.py` contain ~30+ instances of
`datetime.now(UTC).isoformat()` for the `pw_changed_at` session
marker, and at least a few `datetime.now(UTC).replace(tzinfo=None)` /
`now - timedelta(...)` patterns for fixture construction (e.g.
`test_servers.py:491, 535, 540`). These will need to be reviewed
once the production-side convention flips.

## Open questions the next session must answer

These are load-bearing. Do not draft an implementation plan that
assumes any particular answer.

1. **SQLite + `DateTime(timezone=True)` round-trip behaviour in
   SQLAlchemy 2.x:** does storing an aware UTC datetime and reading it
   back yield an *aware* UTC datetime (Python 3.13 + SQLAlchemy 2.x +
   pysqlite)? Or does it still strip and require a custom
   `TypeDecorator`? Spike with a throwaway test before assuming.

2. **Existing data:** the production SQLite DB at left4.me has rows
   stored as text-without-timezone (current behaviour). If the new
   column type changes how SQLAlchemy parses those strings on read,
   we may need a one-shot data migration (rewrite values to include
   `+00:00`). Verify behaviour on a copy of the production DB.

3. **Alembic migrations** (`l4d2web/migrations/` if it exists, or
   wherever migrations live): do any existing migration files
   reference `DateTime` columns? If yes, do those need to be updated
   too? Any backfills that build datetime literals?

4. **The session `pw_changed_at` marker round-trip** is the highest-risk
   site. It's compared as an ISO string in `auth.py:52`. After the
   migration, both sides must produce the same ISO form (either both
   naive UTC or both aware UTC). Plan must spell out the marker
   format and where the conversion happens.

5. **Comparisons crossing the boundary** (DB-side aware vs Python-side
   naive constants, or vice versa) will raise `TypeError` post-migration.
   `server_routes.py:210` (`latest.last_seen_at >= cutoff`) is the
   canonical example. Enumerate every such comparison; ensure both
   sides get fixed in lockstep.

6. **Scope cut for tests:** do the test fixtures need to flip in the
   same PR, or can the production code flip first and tests trail?
   (Likely needs to flip together — comparisons in test assertions
   will break if either side is mixed.)

## Suggested first steps for the next session

1. Read this handoff and the parent spec
   (`docs/superpowers/specs/2026-05-16-timeago-shared-display-design.md`).
2. Re-enumerate `.replace(tzinfo=None)` sites:
   `grep -rn 'replace(tzinfo=None)' l4d2web/`.
3. Re-enumerate `DateTime` column declarations:
   `grep -n 'mapped_column(DateTime' l4d2web/l4d2web/models.py`.
4. Run the round-trip spike (open question #1) in a throwaway script
   or test. Document the actual behaviour observed.
5. Take a copy of the production SQLite DB; read existing
   datetime rows back through the candidate column definition.
   Document whether they parse as aware (and to what timezone) or
   require backfill.
6. Brainstorm scope and ordering with the user — only then write
   the spec.

## Pointers — files to read first

- `docs/superpowers/specs/2026-05-16-timeago-shared-display-design.md`
  — parent spec, "Out of scope" + "Follow-up spec" sections.
- `l4d2web/l4d2web/models.py:1-30, 22-23` — `now_utc()` factory, first
  few column declarations.
- `l4d2web/l4d2web/routes/profile_routes.py:75-93` — the documented
  workaround comment that gave the convention its name.
- `l4d2web/l4d2web/auth.py:40-60` — session marker comparison.
- `l4d2web/l4d2web/services/timeago.py` — `_ensure_utc` defensive
  line that becomes a no-op after the migration.

## Anti-goals

- **Don't bundle this with unrelated cleanup.** It's a mechanical
  refactor when sized correctly. Adding "while I'm in there"
  scope grows the blast radius.
- **Don't skip the round-trip spike.** Believing SQLAlchemy's
  `timezone=True` "just works" on SQLite without checking is exactly
  the kind of false premise this handoff is shaped to prevent.
- **Don't trust this handoff's enumerations as exhaustive.** They
  are starting points from a point-in-time grep. Re-run them.