Round-1 review fixes on the E3 branch:
- prune partial-fetch (correctness): `_paginate` now reports whether it reached
the server's reported `total`; `_prune_stale` prunes only on a COMPLETE
non-empty listing. A truncated fetch (a mid-sequence empty/short page before
total — a transient Plex hiccup) no longer deletes the rows it merely failed
to list. New test: a listing that claims 5 but returns 1 prunes nothing.
- `_prune_stale` (robustness): `stats.get("pruned", 0) + 1` — no KeyError when a
caller's stats dict omits the key.
- service transitions (conventions): pause/resume/redownload use the JobStatus
StrEnum instead of loose "queued"/"paused"/"error" literals, matching cancel.
- ffmpeg progress writer (correctness): a stuck-percent heartbeat now runs on its
OWN 1/interval budget, so it can't consume the progress-write slot and delay a
real percent change by up to an interval.
- backend/app/README.md (conventions): reword the transaction-ownership rule to
quote and extend CLAUDE.md's canonical ("handler/job-wrapper commits; helpers
flush; only quota owns its session") instead of stating a contradicting
"helpers MAY commit" carve-out.
ruff clean, full DB-lane 364 pass.
The layering rule ("request-only logic in routes/; scheduler+worker-shared logic
in the plex|downloads|sync packages; cross-cutting primitives as flat core
modules; extract at the SECOND consumer") was real but unwritten — it lived only
in three module docstrings and a one-line README mention. Collect it, plus the
transaction-ownership rule (a mid-operation helper stages, the caller commits),
the atomic-counter / JobStatus-StrEnum conventions, and the one-home-per-config
rule, into backend/app/README.md. Add docs/development.md (run locally + the gate
commands + branching), so a new session/contributor has one place to look instead
of copying whichever pattern they happen to find first.
worker.py (~1090 L) mixed the PARENT supervisor (claim + supervise + the process
loop) with the CHILD pipeline (the yt-dlp/ffmpeg download+edit that runs in the
spawned subprocess). Extract them:
worker.py 1090 -> 475 parent: claim/supervise/loop
downloads/pipeline.py (new 592) the child pipeline (_download/_process_edit +
hooks/progress/ffmpeg helpers + _Aborted)
downloads/jobdb.py (new 73) the short-lived-session helpers both sides
share (_set_job/_job_status/_touch_job/
_requeue_if_running/_reset_asset_pending +
the staging-dir paths)
The graph is one-directional (runner -> pipeline -> jobdb; worker -> jobdb; no
worker<->pipeline edge) — the parent never calls the child directly, it spawns
runner.py as a subprocess. Consumers rewired: runner import, and two tests follow
the moved symbols (_media_info/_produced_file -> pipeline; the ffmpeg-callback
patches target pipeline where the names are looked up).
Pure code motion, verbatim bodies. ruff clean, full lane 363 pass (unchanged).
request-access / demo / register / verify / verify-resend / password-login /
password-reset(request|confirm) / set-password each took an untyped `payload:
dict` and hand-picked fields with `.get(...)`. Replace with five small BaseModels
(EmailIn, EmailPasswordIn, TokenIn, TokenPasswordIn, SetPasswordIn) so FastAPI
validates the body shape and publishes the schema. Fields are `str | None` to
preserve exact behaviour — a missing/null field still coerces via the existing
`(... or "")`, no stricter 422s. Green under the auth-session suite (7), ruff clean.
pause/resume/cancel/delete/redownload each hand-rolled their asset bookkeeping
inline in the route (release, failed-asset reset, the 20-line fresh-asset
rebuild), duplicating logic the worker/GC also touch. Move each transition into a
service.* method (service.pause/resume/cancel/delete/redownload) that owns the
mutation + its commit; the routes keep only auth, the HTTP 409 guards, and
serialization. Drops the now-dead _set_status route helper and two stale imports.
Pure behaviour-preserving code motion (routes -56 lines). Green under the E3 S2
transition net + the full downloads suite (363 pass), ruff clean.
Pins the observable HTTP behaviour of pause/resume/cancel/delete/redownload
(status changes, asset release, failed-asset reset, fresh-asset rebuild) before
the inline route logic moves into service.* methods. Green on the current routes.
The supervisor kills a job that writes nothing for download_stall_timeout_seconds
(150s) and labels it a "network" stall. But the ffmpeg progress writer only wrote
on an integer-percent CHANGE, so a re-encode that sits on one percent for minutes
(AV1/VP9 -> H.264 or a crop on the CPU-only prod host) produced no row write and
was killed mid-encode as "stalled" — and the retry stalled the same way.
Fix: the writer now heartbeats (`_touch_job`, bumping updated_at without touching
the progress bar) once per interval even when the whole percent is unchanged, so
a live-but-quiet re-encode keeps a fresh updated_at. Reorders the rate-limit gate
first so a burst still costs one DB op — either a progress write or a heartbeat.
Pure-logic test with an injected clock: three stuck-at-10% intervals write no new
progress but heartbeat each time. Full lane green, ruff clean.
Two jobs on the same shared asset (same video+format queued twice, or by two
users) starved the worker pool: worker A downloads the asset; worker B claims the
sibling job, finds the asset `downloading`, requeues it, sleep(2)s, and re-claims
it on the next tick — burning a whole worker slot behind one in-flight asset for
the entire download (at concurrency=2, the queue stalls).
Fix: the claim SQL now excludes any job whose asset is already `downloading`
(`NOT EXISTS … media_assets.status='downloading'`), so a competing sibling simply
waits until the asset settles and the workers pick up unrelated jobs meanwhile.
The `_process_job` busy/lost-race branches drop their `time.sleep(2)` — the
exclusion keeps a requeued job out of the claim window, so there's no spin.
TDD: the new test failed on the old claim (claimed the sibling), passes now.
Full lane green, ruff clean.
MediaAsset.ref_count (how many jobs hold a shared cache asset) was a Python
read-modify-write at four sites (enqueue, enqueue_edit, release_asset, the
redownload route). Two concurrent enqueues of the same (source,format) both read
the same value and write the same +1 → the count ends up one short → the first
release drives it to zero and deletes the file out from under a live holder.
Replace all four with atomic DB ops: incr_ref_count (`ref_count = ref_count + 1`)
and decr_ref_count (`greatest(ref_count - 1, 0)` RETURNING the new value, so the
release decides whether to free the asset from the post-decrement count while
holding the row's write lock — Postgres serialises concurrent releases on that
lock, closing the double-free). Pure DB-atomicity change, no semantic shift.
Green under the E3 S2 ref_count net (5) + full lane (356 pass), ruff clean.
Locks the observable single-threaded ref_count accounting before the atomic-SQL
conversion: two enqueues of one (source,format) share a single asset and stack
ref_count to 2; release keeps the asset until the last holder then frees the
ready row; a non-holding (canceled) job's release is a no-op; the decrement
clamps at 0; an errored asset is kept at zero (only ready is freed). Green on the
current RMW code (5 pass) — the fix must keep every assertion.
A sync was upsert-only: a movie/show/season/episode removed from Plex stayed
browsable forever and 404'd on playback. Each section now diffs the rating_keys
it just listed against the stored ones and deletes the stale rows (watch-state /
playlist links cascade via the existing FKs), mirroring the collections prune
that already shipped.
Safety: the prune is GUARDED on a non-empty listing (_prune_stale returns early
when `seen` is empty), so a transient Plex glitch that yields an empty page can
never mass-delete a library — the metadata is cheap to re-sync, an accidental
wipe is the only real risk. Shows prune leaf-first (episodes -> seasons -> shows)
so a stale row is deleted directly before its parent's cascade reaches it.
Out of scope (separate follow-up): library-level reconcile of a whole removed
section + retiring the dead `PlexLibrary.enabled` column.
Tests: movie prune + watch-state cascade, the empty-listing guard (mutation-
verified — removing the guard wipes everything), show/season/episode prune.
351 DB-lane pass, ruff clean.
routes/plex/browse.py was a 1693-line god-module (libraries + facets +
collections + playlists + show/item detail + people). Split it into focused
sibling routers under the same /api/plex prefix — pure code motion, no SQL /
route-path / response-shape changes:
browse.py 1693 -> 469 (list_libraries, unified_library, facets)
collections.py (new 253) browse + admin collection editor; owns _collection_source
playlists.py (new 405) the user's orderable playlists
detail.py (new 438) show/item detail + enrichment; _lib_key lives here
people.py (new 139) cast/crew chips + person filmography
Cross-module helpers _norm_lang + _content_libs move to _shared.py (their one
home); detail imports _collection_source from collections (one-way, no cycle).
__init__.py includes the four new routers. Three DB-lane tests follow the moved
symbols (import path for _content_libs; monkeypatch targets for PlexClient /
_show_enrichment now on detail).
Behaviour-equivalent: ruff clean, 348 DB-lane pass (unchanged), parity net green.
The `show` scope (all|watched|in_progress|unwatched) filter and the status CASE
were hand-duplicated between unified_library (the grid) and facets.wsq — the
exact drift risk the review flagged (a hidden movie kept out of the grid must
not inflate facets). Extract them into four shared query.py helpers
(movie_status_case, show_status_case, apply_movie_state_filter,
apply_show_state_filter), model-agnostic (callers pass the joined state
expressions), and point all three call sites at them.
Pure behaviour-equivalence refactor: browse.py -67 lines, no query changes.
Green under the E3 S1 parity net + query_conversion/facets/field-scoped/lang
tests (348 DB-lane pass, ruff clean).
Coverage-first safety net for the E3 S1 Plex god-module split. The movie/show
watch-state predicates are currently hand-duplicated in unified_library (the
grid) and facets.wsq (the facet aggregation); this pins the invariant the
facets docstring promises — with no metadata filter active, a facet's genre /
content-rating counts equal the same counts over the grid's visible items — for
every (scope × show) combination, deriving the expected counts from the grid's
own response. Locks the two branch-sets together before they merge into one
shared helper, so the merge is provably behaviour-equivalent.
Mutation-verified: dropping the facets hidden-exclusion branch makes both tests
fail (Drama 3≠2, R 2≠1). Green on the current code (348 DB-lane pass).
Bump VERSION 0.67.0 -> 0.68.0 and add the user-facing release-notes entry for
the E2 epic (feed sort index, keyset/NOT NULL, search GIN + state inner-joins,
single /feed/summary pass, no card-action refetch storm, job partial indexes).
The via_search / shorts_probed partials used `WHERE via_search` / `WHERE NOT
shorts_probed`, but the queries filter with `.is_(True)`/`.is_(False)` (SQL
`IS TRUE`/`IS FALSE`), which Postgres's partial-index prover does NOT match to a
bare-boolean predicate — so both fell back to a full seq scan (the shorts-probe
job kept full-scanning for its work queue, the exact thing PERF-3 meant to fix).
Align the predicates to the query form: `WHERE via_search IS TRUE` and
`WHERE shorts_probed IS FALSE`. Verified on the prod-sized localdev: both now
Index Only Scan (via_search 44,683 -> 3 buffers; shorts_probed 44,683 -> 1).
Gate green (346 pytest).
Two scheduler jobs full-seq-scanned the 257k-row videos table every run:
- PERF-17 enrich_pending (`WHERE enriched_at IS NULL LIMIT N`, every few minutes):
44,680 buffers even with 0 pending -> a partial ix_videos_enrich_pending makes
it a 1-buffer index scan (empty when nothing pending).
- PERF-18 maintenance revalidate (`WHERE unavailable_since IS NULL ORDER BY
last_checked_at NULLS FIRST LIMIT N`, ~300 pages/run): 44,754 buffers + top-N
sort -> a partial ix_videos_recheck_order in that exact order = a 42-buffer
index scan. The order_by is rewritten to `asc().nulls_first()` (equivalent to
the old `(IS NULL) DESC, ASC`) so it rides the index; the unused plain
ix_videos_last_checked_at is dropped.
PERF-3: drop the never-picked boolean btrees (is_short, via_explore) and slim the
useful ones to partials over the queried subset (via_search=true, NOT
shorts_probed, unavailable_since IS NOT NULL) — less write amplification on the
biggest table, same lookups. Model: those 6 columns lose their inline index=True
(the partials are migration-defined, like the search_vector GIN); live_status
left alone.
Measured on the prod-sized localdev reseed: enrich 44,680 -> 1 buffer (25.9 ->
0.02 ms), maintenance 44,754 -> 42 buffers (35.5 -> 0.17 ms); all 11 CONCURRENTLY
index ops on 257k rows in 0.75s. Gate green (346 pytest incl. drift + downgrade
round-trip).
The tags_by_category extraction dropped feed_summary's `if selected:` guard, so
a /feed/summary request with no tag filter (the common case) ran a pointless
`SELECT ... WHERE id IN ()` round-trip. Early-return {} for empty tag_ids —
restores the no-query hot path and covers filtered_query + future callers too.
Gate green (346 pytest).
Review round on the E2 stack (S1-S5). No correctness bugs; the surviving
low/medium findings addressed:
- altitude (#1): extract the shared `tags_by_category` helper used by BOTH
filtered_query (SQL tag WHERE) and feed_summary (Python facet/count), removing
the category-grouping duplication, and add bidirectional "keep in lockstep —
equivalence-tested by test_feed_summary_matches_count_plus_facets" comments so
the SQL<->Python tag rule can't silently drift.
- test-coverage (#2): the /feed/summary equivalence test seeded only system tags,
leaving the `ChannelTag.user_id == user.id` visibility branch untested. Add a
user-owned Tag + link + two combos; the suite is now 346 (was 344) and the new
cases pass (feed_summary agrees with /feed/count + /facets on user-visible tags).
- dead-code (#3): annotate get_feed_count, get_facets, and the FE api.feedCount/
facets as "SUPERSEDED by /feed/summary — kept only as the equivalence oracle;
retire together" so the intentionally-retained surface reads as deliberate.
Not changed (confirmed no-op): the search channel-name resolve (#4) is
once-per-request on a 633-row table and a deliberate pre-resolve for PERF-12's
BitmapOr plan; the facet loop (#5) already computes passes() once per
(channel, category) with a distinct exclude each, so there is no recomputation.
Gate green: ruff + 346 pytest; frontend tsc/knip/prettier (FE change is a comment).
Marking a card watched/hidden/reset and closing the player each fired a full
`invalidateQueries(qk.feed())`, refetching EVERY loaded infinite-query page — 10
pages x a few quick marks = ~50 feed refetches, each a full scan (x PERF-1). The
optimistic `overrides` layer + the matchesView filter already render the correct
state (a hidden/watched card drops out instantly), so the refetch was pure waste.
Switch these paths to `refetchType: "none"`: the queries are marked stale but not
refetched now; the next filter change / remount pulls fresh server data. Applies
to the two Feed card-action paths (setState/clearState) and the PlayerModal
close (feed + playlists — progress is already persisted server-side, only the
on-screen resume bar waits for the next natural refetch).
Left as immediate refetch on purpose: channel/tag/playlist mutations (one-off, on
pages where the main feed is unmounted anyway) and the onboarding / notification-
undo paths (which must reveal the changed content right away). Gate green:
frontend tsc/eslint/knip/prettier/222 vitest (backend unchanged, 344 pytest).
The FE fired /feed/count (1 pass) + /facets (3 passes) on every filter change =
4 full scans of the subscribed set. New /api/feed/summary does ONE grouped query
(the base filter with tags stripped, GROUP BY channel_id -> per-channel video
counts) and derives the total count + every tag's distinct-channel facet count in
Python on the ~few-hundred-row channel map, faithfully mirroring filtered_query's
channel-based tag logic + the /facets drill-down (own category dropped, unless a
conjunctive topic-AND). Returns {count, facets} — the same numbers the two old
endpoints returned.
FE: a single shared query (Feed reads .count, Sidebar reads .facets); same key +
filters, so React Query dedupes to ONE request. Invalidations repointed to
qk.feedSummary. The old /feed/count + /facets endpoints (and their now-unused FE
api/qk) are kept as the equivalence-test oracle + safety net; a later cleanup can
drop them.
New equivalence test (test_feed_views, 8 filter combos incl. cross-category,
disjunctive/conjunctive topic modes, search): /feed/summary count == /feed/count
and facets == /facets, byte-identical. Structural win: 2 requests / 4-6 queries
-> 1 request / 2-3 queries (the review measured 530 -> 144 ms). Gate green:
344 pytest, frontend tsc/eslint/knip/prettier/222 vitest.
PERF-12: the search ORed `unaccent(channels.title) ILIKE '%q%'` straight into
the per-video predicate, which forced a full scan + detoast of every
search_vector and a per-candidate channel fetch. Pre-resolve the matching
channel ids (a few-hundred-row scan) and match `Video.channel_id.in_(ids)`
instead, so the planner keeps the search_vector GIN as a BitmapOr branch
(alongside ix_videos_channel_pub_desc). Empty match -> false() collapses the OR.
PERF-16: show=watched|hidden|in_progress used an outer join + coalesce(status,
'new'), so the planner couldn't start from the user's small video_states set.
Those views only return rows that HAVE a state, so inner-join video_states and
filter the raw column (status is NOT NULL on a present row, so it's identical).
'unwatched'/'all' keep the outer join. Dropped the now-dead position_expr.
New regression tests (test_feed_views): search returns both text- and channel-
name matches; the four watch-state views return the exact expected sets.
Measured on the prod-sized localdev reseed (prod planner, 0067/0068 applied):
- search /feed/count q=magyar: 663,463 -> 19,206 buffers (1,406 -> 31 ms; OLD
seq-scan+detoast -> NEW BitmapOr GIN+channel index)
- search feed page: 2,908 -> 1,479 buffers (32.6 -> 1.4 ms)
- show=watched (user with 64 watched): 40,661 -> 264 buffers (~154x)
Gate green (336 pytest, incl. the 2 new equivalence tests).
A nullable published_at forced the feed keyset into a non-sargable
`... OR published_at IS NULL` disjunction: with the 0067 index the planner still
applied it as a FILTER and scanned from the top of the index, so a deep `newest`
page discarded every earlier row (O(depth)).
- Migration 0068: backfill any NULLs (0 in prod) -> SET NOT NULL + DEFAULT now().
- Model: Video.published_at is now non-Optional, nullable=False, server_default now().
- feedquery: the `pub` key is nullable=False, and keyset_predicate is rebuilt
recursively so a non-nullable leading key gets a non-strict range bound
(`published_at <= v`) as an AND conjunct — an index range Cond instead of a
filter. Nullable keys keep the NULL-aware OR form (a `<= v` bound would drop
their NULL group). Equivalence proven by test_feed_keyset (all 13 sorts,
cursor-walk == unpaged).
- Ingest paths (stub_from_playlist_item, apply_video_details, playlists sync)
fall back to now() so they never write an explicit NULL.
Measured on the prod-sized localdev reseed at a ~50k-deep cursor (0067 index +
prod planner): OLD `Rows Removed by Filter: 50001`, 25,687 buffers / 9.16 ms ->
NEW `Index Cond: published_at <= v`, 1 row filtered, 47 buffers / 0.037 ms. SET
NOT NULL on 257k rows: 0.32s. Gate green (334 pytest).
The feed's default `newest` sort is `published_at DESC NULLS LAST, id ASC`, but
the only published_at index was a plain ASC btree — read backwards it satisfies
DESC NULLS FIRST, not NULLS LAST — so every feed page fell back to a full
`videos` seq scan + top-N sort. Migration 0067 adds the two matching indexes
(global + per-channel) CONCURRENTLY and drops the now-redundant single-column
ix_videos_channel_id (the channel composite leads with channel_id, a strict
superset).
Measured on the prod-sized localdev reseed (257,312 videos, prod planner):
- global newest: 44,747 buffers / 154 ms -> 60 buffers / 0.32 ms
- per-channel(12.6k): 7,604 buffers / 96 ms -> 54 buffers / 0.14 ms
CONCURRENTLY build of both on 257k rows: 1.7s total (~0 downtime). This is the
likely root cause of the "prod feed is slow" symptom.
Model: drop the inline index=True on Video.channel_id (the composite supersedes
it); the composites are migration-defined like the search_vector GIN. Gate green
(334 pytest incl. the migration drift + downgrade/re-upgrade round-trip).
Bump VERSION 0.66.0 -> 0.67.0 and add the user-facing release-notes entry for
the E1 Foundation epic (data-safety fixes, account-switch + search-error fixes,
graceful key-mismatch degrade; plus the security/test-net/CI/repro chores).
Review round 2: the unconditional `command.upgrade(cfg, "head")` in finally ran
a redundant no-op on the success path and, on a genuinely broken re-upgrade,
replaced the try's original exception with its own. Gate the restore on a
`reupgraded` flag so it fires only when the upgrade didn't complete — the
cascade-prevention intent is kept without the double-run or error-masking.
Review round on the whole E1 stack (main..HEAD). 6 findings, all low/medium,
none blocking; fixed for a clean gate-green handoff.
- plex/sync (correctness): the un-enriched cast guard compared the stored cast
against THIS listing's length, so a trimmed Plex cast on a not-yet-enriched
row was never applied. Compare against the listing CAP (_LISTING_CAST_CAP=20)
instead: >cap = enriched (keep, survives a watermark reset), <=cap =
listing-sourced (track edits in both directions). +2 regression tests.
- main (security): add X-Frame-Options: DENY as the legacy-browser fallback for
the CSP frame-ancestors clickjacking guard.
- downloads/service (altitude): finalize_if_running now uses a SQLAlchemy 2.0
update() construct instead of a hand-built text() UPDATE string — same atomic
WHERE status='running', idiomatic, no identifier interpolation, no manual
StrEnum coercion. Drops the now-unused text import.
- auth/admin (reuse): extract the SEC-10 "drop untrusted preset password + bump
epoch" invariant into one helper (revoke_preset_password_if_unverified) so the
Google-callback and admin-activation paths can't drift.
- tests/migrations (robustness): guarantee the shared schema is restored to head
in a finally, so a downgrade/re-upgrade failure can't cascade into later tests.
- worker (simplification): align _finish_from_cache's display_name fill with the
_download path's guard (no behavior change).
Gate green: ruff clean, 334 pytest (0 skip), frontend unchanged.
SEC-10 (account pre-emption / "pre-hijack"): a password set on an account whose
email was never verified is untrusted — whoever registered the address (to
pre-empt the real owner) could have chosen it. Both the Google-adoption callback
and admin activation (_activate_user_by_email) now drop that pre-set password and
bump the session epoch when the account isn't email_verified, so a later Google
login / admin approval can't hand the pre-set password a working login. No-op for
a verified account (the legit no-SMTP flow). +2 tests.
SEC-11 (link/upgrade bound to the session default): the OAuth link/upgrade GETs
resolved to the session DEFAULT account, so a tab acting as a secondary account
would link Google / the YouTube grant to the wrong account. The frontend now
threads the active account through the navigation URL (withActiveAccount ->
?account=<active>), and _complete_link accepts any wallet account, not only the
default, so the identity lands on the account that started the flow.
The pip-audit step added in E1 S4 flagged cryptography 46.0.7 for PYSEC-2026-3552
/3553/3554 and GHSA-537c-gmf6-5ccf; all are fixed by 50.0.0. The requirements.txt
ceiling `<47` was blocking the upgrade — relaxed to `>=50.0.0,<51`. The lock is a
surgical bump (only the cryptography line changed; cffi 2.1.1 already satisfies
it, no other package moved). pip-audit -r requirements.lock is now clean.
Add tests/test_security.py (there was none for the crypto primitives): the
Fernet encrypt/decrypt round-trip, decrypt_optional degrading on garbage and on a
key mismatch (the B-core-2 path), and the argon2 password hash/verify round-trip
— so a future crypto-lib bump can't silently break the token-at-rest path. Gate
green: 330 pytest.
A vitest that fails the gate when the newest release-notes entry's version
diverges from the repo-root VERSION file (the two hand-edited version sources),
plus a no-duplicate-versions check. Catches the "bumped one, forgot the other"
release mistake.
requirements.txt stays the source (floor ranges); requirements.lock pins the
exact resolved versions (generated in the python:3.12-slim prod base, extras like
psycopg-binary/uvloop included). The prod Dockerfile, Dockerfile.test and CI now
install the lock, so the image is reproducible and the versions are the ones
pip-audit scans. CI gains a non-blocking pip-audit step over the lock. The
frontend build + CI move from the EOL node 20 to node 22.
A small, safe set applied to every response via middleware: X-Content-Type-Options
nosniff, Referrer-Policy strict-origin-when-cross-origin, CSP frame-ancestors
'none' (anti-clickjacking; unrelated to the YouTube/Plex embeds we host), and
Permissions-Policy disabling camera/mic/geolocation. A full CSP and HSTS are
deferred (HSTS belongs on the Caddy edge).
Add .gitea/workflows/check.yml so the quality gate runs on the lab act_runner,
machine-independent, on every push/PR — the review's "no server-side CI" gap.
Two parallel jobs mirror the two local lanes:
- frontend (node:20-bookworm, matching the prod build image): tsc + eslint +
prettier --check + knip + vitest.
- backend (nikolaik/python-nodejs python3.12+node): ruff + the full pytest suite
against a real postgres:16 service, with ffmpeg installed and
SIFTLODE_TEST_REQUIRE_DB=1 so a DB/ffmpeg that failed to come up is a hard
error, never a silent skip.
Notes for this runner: pytest runs natively against the service (the
docker-compose bind-mount is unreliable under a mounted-socket runner), and the
backend uses a baked python+node image rather than actions/setup-python, which
can't reach GitHub's python-versions download from the lab.
The drift check dropped every diff except table/column presence, so a column
type or nullability change would slip through. Add modify_type/modify_nullable
to the ops it fails on. That immediately surfaced a real pre-existing drift:
notifications.created_at was created NULLABLE by migration 0017 but the model
has always declared it NOT NULL (Mapped[datetime] + server_default=func.now()).
Migration 0066 enforces NOT NULL to match the model — safe, since the
server_default has filled every row since day one.
Address the /code-review findings on the S2 test-net:
- as_user no longer sets a global `lambda: user` override that two users in one
test would collide on; the shared current_user override now resolves the caller
from an x-test-user header, and each returned client tags its own requests, so
as_user(a)/as_user(b) are independent (#1).
- override_dep records the prior value per dependency key and restores it on
teardown instead of a blanket dependency_overrides.clear(), so fixtures compose
without clobbering each other (#5).
- a _configured fixture bakes the post-setup state into client/as_user, and the
per-test inline monkeypatch / autouse duplicates are removed (#4).
- the feed keyset walk asserts all seeded rows surfaced before comparing, so a
seeding regression can't make the sorts pass vacuously (#3).
The auth security mag had zero tests. Add a DB-lane suite over the real
auth/session stack (TestClient carries the signed cookie): password login ->
/api/me authenticated; a wrong password and an unknown email return a
byte-identical 401 (no enumeration oracle); an X-Siftlode-Account header naming
a non-wallet account falls back to the default (impersonation gate); logout-others
bumps the epoch so a second session is rejected (401 on a protected endpoint)
while the caller stays in; and password-reset/confirm is single-use (replay ->
400) and bumps the epoch so pre-reset sessions die.
The feed's cursor paging (sort_keys/order_by/keyset_predicate/encode/decode) is
the NULLS-LAST + tiebreaker algebra where a page-boundary bug hides as a
duplicated or skipped row on the busiest endpoint, and had zero tests. Seed 7
videos on 2 channels with duplicate sort values and NULLs across
published_at/view_count/duration/title/subscriber_count, then assert — for every
sort — that walking the feed in pages of 2 via the cursor reproduces exactly the
order a single unpaged query returns, with no duplicates.
Prod only ever runs `alembic upgrade head`; nothing checks that the migrated
schema still matches Base.metadata, and the downgrade path never runs. Add a
DB-lane test that (a) asserts no structural (table/column) drift between the ORM
models and the head schema — filtering the index/constraint/sequence noise that
compare_metadata reports for migration-only objects — and (b) exercises the head
migration's downgrade + re-upgrade so a broken downgrade is caught the sprint it
lands.
The DB-backed lane and the two ffmpeg tests skipped SILENTLY when the DB was
unreachable or ffmpeg was absent, so `siftlode check` could print green while
~60 tests never ran. Make the compose lane authoritative:
- SIFTLODE_TEST_REQUIRE_DB=1 in docker-compose.test.yml; conftest.pytest_configure
turns a would-be skip (no Postgres, or no ffmpeg) into a hard UsageError.
- install ffmpeg in Dockerfile.test so the thumbnail/poster tests actually run.
- `-rs` in pytest.ini so skip reasons are always visible.
Add the shared TestClient fixtures the risk-core suites build on: `make_user`
(create + commit a login-ready User), `client` (real auth/session stack, no
current_user override — for login/wallet/epoch), and `as_user` (a client
pre-authenticated as a given user, for tests that aren't exercising login).
The union form fixed the downgrade but removed the per-sync self-healing: a name
Plex removed from an unenriched title lingered in cast_names/people_text until
enrichment. Replace the union with a length guard — take the fresh listing
unless it is shorter than what is stored. A stored cast longer than the ~20-cap
listing can only be an enriched fuller list (kept, so a watermark reset never
downgrades it); at or below the cap the stored value is itself listing-sourced,
so replacing it with the current listing stays self-correcting.
Follow-up to the /code-review high on the S1 branch. The per-site
`if job.status == 'running'` guards were a non-atomic read-then-write (TOCTOU):
a cancel committing between the child's status read and its commit could still
be overwritten with done/error, re-opening the double-release window.
Introduce service.finalize_if_running(db, job_id, **fields): one atomic
`UPDATE download_jobs SET ... WHERE status='running'` that returns whether it
fired. It flushes pending asset writes first so the asset-then-job row-lock
order matches the cancel route (no deadlock), and runs in the caller's session
so it commits with the asset writes in one transaction. All five worker sites
(download/edit completion, both error paths, cache finish) now go through it,
so the invariant lives in one place. The 'job done' logs move inside the
finalized branch so a job stopped mid-completion no longer logs a false 'done'.
Add tests/test_downloads_finalize.py: running -> finalizes; canceled/paused ->
no-op.
POST /api/me/switch validated wallet membership, access and suspension but not
the target account's session epoch. Switching to an account revoked elsewhere (a
password reset/change or "log out everywhere" bumped its epoch) made it the
session default, so the next request's current_user hit the epoch mismatch on
the default identity and cleared the WHOLE wallet session — logging out every
account in the browser. Mirror the suspend guard: drop the stale account from
the switch list and return 403 instead of switching.
A rotated or missing TOKEN_ENCRYPTION_KEY makes every stored secret undecryptable.
sysconfig._raw_override called security.decrypt() with no guard, and it sits on
hot paths (/api/me via google_enabled, login, every YouTube scheduler job,
account purge) — so a key mismatch raised InvalidToken and 500'd the whole app,
the SPA wouldn't even boot, with no documented recovery.
Catch the decrypt failure in _raw_override -> warn once per key and fall back to
the env/config default. The YouTube client's refresh-token decrypt now raises a
clear YouTubeError ("reconnect") that every caller already handles, and
purge_user uses decrypt_optional so a best-effort revoke can't 500 the delete.
Document that the key must never change once secrets are stored (recovery: clear
the affected system_config rows and re-enter).
/api/search/youtube raised 502 on an upstream error, but the frontend treats
502/503/504 as a gateway dropout: it auto-retries the GET (a second quota-costed
search) and shows a "connection lost" banner instead of the real message. Every
sibling upstream-error site already returns 422 for exactly this reason. Reserve
5xx for the raw byte-proxies.
_apply_facet_fields seeded cast_names from the cheap top-~20 listing only while
people_enriched_at was NULL, relying on "a sync never clears the watermark" to
protect an enriched fuller cast. Migration 0065 broke that assumption by nulling
people_enriched_at on every movie/show row, so the first post-deploy sync
downgraded enriched casts back to the listing top-20 — silently dropping
person-page cast-filter names for weeks until the backfill caught up (the S5b
regression, re-opened by the migration).
Guard structurally instead of trusting the watermark: union the listing cast
onto whatever is stored, never dropping a name already present. A fresh row
still seeds from the listing; a row whose watermark was reset but whose fuller
cast survives keeps it. Robust to any future watermark reset, not just 0065.
The cancel route writes status 'canceled' but the supervisor compared against
'cancelled', so its "a route stopped it -> kill the child, free the asset"
branch was dead code. A canceled job's child ran to completion and wrote 'done'
unconditionally over the route's 'canceled'; since 'done' holds the asset but
'canceled' does not, a later delete released the asset a second time, dropping
ref_count below the live holder count and deleting a file still in use.
Add app/downloads/status.py as the single source of truth for the lifecycle
literals (JobStatus StrEnum + the HOLDING / STOPPED_BY_ROUTE sets) and use it at
the safety-critical sites: the supervisor stop check, the route cancel, and
release_asset's holding set. Make the worker's terminal done/error writes
conditional on the row still being 'running' so a child finishing just after a
route cancel/pause can no longer resurrect it. Add the missing 'canceled'
supervisor regression test.
(The full literal sweep across the downloads surface is left for the E3
downloads owner-unification; this is the data-safety subset.)
The codebase is fully AI-authored and developed across many agent sessions; the
canonical conventions have lived only in a private memory store, which the
2026-08-17 review flagged as a real gap (unwritten rules → asymmetric fixes and
comment-vs-code drift, e.g. the cancelled/canceled literal bug).
Add a short, auto-loaded rules-of-the-road: the local gate (never merge/ship
red, no --skip), the trunk workflow and no-AI-co-author commit rule, the
one-way architecture shape, a canonical-pattern table (the review's §6
consistency choices), and the house rules learned from real bugs. Kept
deliberately thin; slims to a pointer once docs/development.md lands.
peter