The 'copy this wholesale' docker-compose.yml block in the README had
drifted from the real make-of-record file: it hardcoded 5432/8765
(missing the DB_PORT/SERVER_PORT configurability from de8f71b2), lost
the loopback-only db publishing (b03ed9a2), carried its own hand-tailored
db-init service with inline SQL instead of the z_rls_role.sh script
version shipped in 313268a7, and still contained the vestigial
./uploads mount on the db service (dropped in the previous commit).
Replace the whole yaml block with a byte-for-byte copy of
docker-compose.yml so the quickstart can no longer drift — the
documented optional envs, loopback bind, db-init script path, and
env_file wiring all now match what actually ships. Verified by diffing
the fenced block against the file.
The db service mounted ./uploads:/app/uploads, which Postgres has never
used and nothing in the repo relies on: no SQL COPY TO/pg_dump path, no
doc/test reference to an uploads path inside the db container (the only
/app/uploads mentions in TEST_DATA.md are the app/test containers), and
the backup scan is app-side.
Forensics: the mount was introduced by accident in 6ebe9749 ("Simplify
docker-compose healthcheck configuration"), which incidentally REPLACED
the db service's volumes with the app service's uploads mount. The
follow-up 91456d11 ("restore essential database configuration")
restored postgres_data and the schema-init mount but nobody noticed the
stray uploads mount riding along, and it has survived every compose
refactor since.
The tests service in docker-compose.dev.yml keeps its own ./uploads
mount — that is the app-like test runner (TEST_UPLOAD_PATH=/app/uploads)
and is legitimate. The dev db inherits its volumes from the base file,
so it is fixed automatically.
The Installation section assumed 'git clone the repo'. Self-hosted
operators only need three files. The install is now: mkdir + copy this
compose block + .env (three openssl secrets) + uploads dir + pull/up.
The compose is fully self-contained: the runtime-role bootstrap is
inlined into db-init's command (same SQL as
database/schema/z_rls_role.sh — keep in sync) — no ./database/schema
bind mount, because the app itself executes the schema as the runtime
role (owner) on first boot; with the role as DATABASE owner the trusted
pg_trgm extension needs no superuser.
Verified end to end in a throwaway project: fresh volume, pasted block,
.env with generated secrets, db-init converged role+ownership, app ran
the full schema and came up healthy — no repo checkout anywhere.
Compose-quoting lessons encoded (the reason for odd dollars): compose
interpolates '532016' to '$' before the container sees it, so the plpgsql
532016 tags in the inline SQL are written 532016532016, and PGUSER must be set
explicitly (POSTGRES_USER is not a libpq variable).
Approves and lands the doctrine rewrite from the verified rollout:
RLS on media_items IS the enforcement (policy joins library_visibility
on app.user_id, declared per transaction by ScopedDB); new
media-listing queries and endpoints need no visibility plumbing.
Documents the three identity bridge sites, background-job identities,
the FORCE/owner role and db-init provisioning, per-boot schema
idempotency, and the two auditors (TestRlsInvariant + tests/ui).
Item-scoped gates may answer 404 as well as 403 (contract_pins
2026-10-29). History records the convention-era leak the dashboard
section queries carried.
Deleting any device whose stats rows still referenced it failed with
500: reading_history.device_id is a NO-ACTION FK by design (the rows
are real captured reading stats — Users reads them), so a blind DELETE
violated the constraint. Surfaced by the Android app's new logout
deregister call, but the admin Devices page Deletes hit the same stone.
Fix: DetachReadingHistoryFromDevice — NULL the device association on
the user's history rows before the device DELETE (single handler
pass). device_id is nullable; the stats survive, the device frees its
identifier. sqlc regenerated.
The role split previously required a one-time manual script run on
existing volumes (initdb.d only fires on fresh ones). This folds the
bootstrap into the compose file the way split-account deployments
standardly do it: a one-shot db-init service runs z_rls_role.sh
(idempotent, network mode via PGHOST=db + PGPASSWORD from DBPASS) on
every 'compose up', and the app gates on
db-init: condition: service_completed_successfully.
- Zero manual steps for fresh AND existing deployments: update compose,
add DBSRVPASS to .env, 'compose pull && up -d'.
- Security posture improves: the superuser password is now held ONLY by
the init service — the app container connects exclusively as the
non-superuser runtime role.
- z_rls_role.sh gains dual connection modes (local socket for docker
exec, network for the init service) with the DBPASS requirement
enforced only in network mode.
- Verified on the dev stack: init runs and exits clean, app gates on it
and comes up healthy, second 'up' re-converges idempotently, invariant
test passes through the new dependency chain.
The initdb.d mount stays as-is (harmless double coverage on fresh
volumes; both paths are idempotent).
.env.example is the template prod/new deployments copy; it was missing
the runtime role password the app now requires (DATABASE_USER
bookhoard_srv). Notes distinguish it from the bootstrap DBPASS and
point existing-volume operators at the one-time z_rls_role.sh run.
Phase 5.5 (plan §7.5) first increment: the SSR/HTMX frontend is swept
in a real browser (official Playwright image, no host browsers) for
BOTH personas — testuser (admin) and rls-phone (restricted; Audiobooks
hidden from them via the visibility API).
Green checks: auth rejection + session establishment; admin dashboard
renders sections incl. the hidden-from-others library; restricted
dashboard renders real sections (Recently Added present — not a
vacuous landing-page pass) with ZERO hidden-library traces; search and
media payloads leak nothing for the restricted persona.
Harness lessons encoded: / is a landing page (dashboard lives at
/dashboard); the login endpoint rate-limits rapid attempts so sessions
are established once per persona and reused via storageState — a sweep
must not look like a brute-force run; getByText visibility assertions
are wrong for <option> text (library switcher) — body-textContent
assertions instead.
COVERAGE.md tracks the full 14-area sweep per the plan; the remaining
areas (reader, collections CRUD, admin matrix, upload roundtrip) are
enumerated and pending their sweep runs.
Phase 5 (plan §7): the entire integration suite passes with the RLS
policies live. The suite was the triage enumerator the plan predicted:
- device_auth middleware: the THIRD identity bridge — KOReader/Kobo
device-authenticated requests read media as the DEVICE OWNER
(plan A9 called this path out; the flip was 404s on sync metadata).
- conflicts/analytics/progress handlers: detached
context.Background() reads threaded to the request context (the
requesting user's identity) — 12+ sites; resolve helpers take ctx.
- scanner: ScanFolders self-declares its configured admin identity
(server maintenance, product-correct for any caller); generic job
creation stamps the requesting admin.
- harness: setup.DBPool and the sync-test pool declare
rls.SeedIdentity at SESSION scope via BeforeAcquire — the pool never
serves a second identity, so no A3 bleed; setup.DB falls back to the
seed identity (NewSeedDB) for ungoverned-context calls; filter tests
use seedExec for raw governed-table statements.
- contract pin WIDENED (docs/contract_pins.md 2026-10-29): under RLS,
item-scoped gates may answer 404 (row invisible — no existence leak)
as well as 403 (Go-level checks on listings keep strict 403).
- TestWorker_JobStatusTracking — the pre-existing red from the Phase 0
baseline — passes with the harness seed work; no product change.
The full suite enumerated the remaining triage class: tests drive the
scanner directly (ScanFolders(context.Background())) and read seeded
media via Queries with bare contexts — every governed read raised.
- MediaScanner.ScanFolders self-declares its configured admin
(SetAdminID — set by main AND the test helpers): a scan is server
maintenance, product-correct regardless of caller.
- ScopedDB gains a Fallback identity, exposed as NewSeedDB for the
TEST HARNESS ONLY: setup.DB falls back to rls.SeedIdentity when the
context carries none, so direct seeding/verification helpers work
uniformly instead of per-site context threading. rls.SeedIdentity is
the shared constant (no library_visibility rows => full visibility).
- generic job creation (jobs.go) stamps the requesting admin into the
job Context, same as the scan endpoint.
The sync queue processor stays identity-free: sync_queue is not an
RLS-governed table.
Phase 4 of the RLS plan: detached contexts (context.Background()) that
reach media_items now carry identities — with the policies live, a bare
context made governed reads raise (fresh connections: 42704 unrecognized
parameter; reused pool connections: 22P02 empty-string cast after the
LOCAL GUC reverted — the plan's A2 signature).
- scanner jobs (API-triggered, mtime-poll, initial scan): the job's
Context now carries the acting admin's identity — the library owner,
first-admin fallback (the job's UserID already encoded this trust
decision; the Context just needed to carry it). performInitialScan
uses the scanner's configured admin.
- websocket initial-sync read (GetUserMediaItemsForSync): runs as the
connection's own user.
- hash backfill: runs as first admin (server self-heal over all media).
Verified on the dev stack: forced rescan of the 1199-file Audiobooks
library completed with 0 errors under RLS (previously every SHA lookup
and create raised); hash-backfill refusals gone. The sync queue
processor needs no identity — sync_queue is not an RLS-governed table.
Known remaining detached reads (conflicts/analytics handlers using
context.Background()) are HTTP-path triage, enumerated by the full
suite next.
The dashboard leak is closed at the database engine. The black-box
invariant (restricted user + hidden library => absent from every
media-returning surface) went RED->GREEN with this change set; clause-
protected surfaces never regressed, proving enforcement came from the
policies, not the test.
Phase 3 of plan ~/.opencode/plan/server-rls-library-visibility.md:
- schema.sql: idempotent RLS block — ENABLE + FORCE on media_items,
SELECT policy mirroring COALESCE(is_visible,true) exactly, permissive
per-command write stand-ins (FOR ALL would OR into SELECT).
- internal/rls + internal/database/scoped.go: ScopedDB declares the
acting user per statement (LOCAL GUC). DESIGN CORRECTION during
execution: the decorator never refuses — the ENGINE is the fail-loud
layer (policies raise on governed tables; auth lookups and settings,
ungoverned, simply work). A decorator refusal would have been a new
convention gate — plan B thinking — and broke login (chicken-and-egg:
credentials are checked before any identity exists).
- database/schema.go: the migration tx declares a synthetic identity —
schema.sql's dedup DO block SELECTs media_items, which raises once
policies are live (plan A11). Caught by the two-boot gate: boot 1
created the policies after the block ran; boot 2 raised.
- router.go: RLS bridge at BOTH c.Set("user") sites (JWT + device
auth) — echo-local c.Set never reaches the stdlib ctx the DB layer
receives.
- search: the search service's raw DBTX path receives the ScopedDB
itself, never the bare pool.
- harness: the server under test is wired exactly like main.go; the
seed handle is also ScopedDB-wired but carries seedCtx() — a
synthetic identity with no visibility rows — because sqlc :one
inserts use RETURNING, which evaluates the SELECT policy even on
INSERT. Regex UUID-scraping in the invariant replaced with a
structured JSON decode.
- invariant test: RED -> GREEN across every surface incl. View All;
403 pins and admin-scoping pins hold.
Design correction discovered by the auth chicken-and-egg: the
decorator-level refusal was re-created a convention gate (plan B
thinking). ScopedDB now NEVER refuses — it declares the GUC when the
context carries an identity and declares nothing otherwise; the ENGINE
is the fail-loud layer (policies raise on governed tables, ungoverned
tables like auth lookups simply work). rls.WithInternal removed as a
dead concept.
- search service's raw DBTX path now receives the ScopedDB itself
(main.go + harness) — a bare pool bypassed the GUC.
- database/schema.go: migration tx declares a synthetic identity
(schema.sql's dedup DO block SELECTs media_items; with policies
live, boot 2 raised 'unrecognized configuration parameter' — A11).
- harness: TestServerSetup.DB is a raw-pool SEED handle (write
policies permissive, other seeded tables ungoverned); the server
under test keeps the scoped wiring. 33 CreateMediaItem sites use
seedCtx() — a seed identity with no visibility rows — because
RETURNING requires SELECT visibility of the new row.
- invariant test: regex UUID-scraping replaced with a structured JSON
decode of the sections payload.
INTERRUPTION NOTE: an edit to this same effect (seedCtx) was applied,
confirmed by the tool, and subsequently vanished from disk while an
'opencode serve' process (PID 84551) is alive on the machine. This
commit pins the tree against further loss; implementation is PAUSED
pending the user killing that process.
Phase 3 of the RLS plan, pinned mid-flight because a parallel writer
was actively editing cmd/server/tests/rls_visibility_test.go (A14):
the file flip-flopped between two implementations and gained a compile
error after the previous commit. This snapshot makes the whole tree
compile; functional verification (two-boot gate, invariant green,
full suite) lands in the next commit(s).
- internal/rls: typed ctx keys — user identity (bridged from echo by
the router) and an INTERNAL marker whose semantics grant nothing:
no GUC is declared, so policy'd tables still raise at the engine;
only RLS-ungoverned tables (system_settings, users counts) work for
user-less startup paths (settings registry Load, CountAdmins).
- internal/database/scoped.go: DBTX decorator — per-statement tx,
commit-on-Close (:many) / commit-on-Scan (:one), error paths always
terminate the tx, refusal before SQL when no identity and not
internal.
- database/schema/schema.sql: idempotent policy block (DROP IF EXISTS
+ CREATE — schema.sql re-runs every boot): ENABLE + FORCE, read
policy exactly mirroring COALESCE(is_visible,true), permissive
per-command write stand-ins (FOR ALL would OR into SELECT).
- router.go: RLS bridge at BOTH c.Set("user") sites (JWT + device
auth) — echo-local storage never reaches the stdlib ctx.
- main.go + test harness: database.New over ScopedDB at both wiring
sites; startup reads use rls.WithInternal.
Three harness corrections found while executing the invariant test:
- library type is 'ebooks' (plural) in library_types, not 'ebook'
- cleanup ordering: t.Cleanup is LIFO — register the pool Close BEFORE
the fixture's library deletes, or the deletes run against a closed
pool and fail silently into ignored errors (leaving RLSMARKER rows
behind; found because a second run saw the first run's 'deleted'
books still leaking through Recently Added)
- collection detail is addressed by the section's collection_id (the
UUID in view_all_url), not the section name
The app connected as 'postgres' — the bootstrap SUPERUSER — which
bypasses row-level security unconditionally (FORCE RLS does not apply
to superusers), making any future policy a silent no-op. This is the
prerequisite commit for the RLS visibility work (plan
~/.opencode/plan/server-rls-library-visibility.md REV 2, §4.0).
What changed:
- database/schema/z_rls_role.sh: creates the bookhoard_srv role
(LOGIN, non-superuser, non-bypassrls) and transfers ownership of
every public table/sequence/function + schema + database to it.
Ownership (not mere grants) because executeSchema re-runs the whole
schema.sql on EVERY boot and needs owner DDL. Idempotent, catalog-
only — no user data is read or written.
- docker-compose.yml app service + docker-compose.dev.yml tests
service: DATABASE_USER postgres -> bookhoard_srv, DATABASE_PASSWORD
DBPASS -> DBSRVPASS (new .env key, generated locally).
- POSTGRES_USER: postgres stays as the container bootstrap account;
on FRESH volumes z_rls_role.sh runs automatically via
docker-entrypoint-initdb.d AFTER schema.sql (alphabetical: z > s).
Verified on the dev stack: role has rolsuper=f, rolbypassrls=f, owns
all 40 tables and every public function; single boot as bookhoard_srv
executed schema.sql cleanly (CREATE EXTENSION IF NOT EXISTS pg_trgm
skips fine as non-superuser when the extension exists); health +
authenticated smoke pass. Behavior otherwise unchanged — no policies
yet.
Adds the permanent auditor for per-user library visibility: a restricted
user with a hidden library sees none of its items on any media-returning
surface (media list, search, dashboard sections, series, facets, OPDS
device catalog, collection detail, scoped/detail access).
RED now, on exactly the surfaces the audit predicted: the dashboard
section queries (GetNotStartedItems / GetRecentlyAddedItems /
GetCollectionItemsForDashboard et al.) never carried the
visible_library_ids clause, and the collection-detail endpoint (mobile
View All) exposes the full section query results. Clause-protected
surfaces (media list, search, series) pass, proving the test itself
isn't a false positive.
Implementation-agnostic on purpose: it asserts the user-level invariant
through the public API, so it stays valid when RLS replaces the
per-query clauses and catches any future regression regardless of how
the code is written. Also pins: hidden library_id=403 (never silently
passed), item detail denial, and that the hide is USER-scoped (admin
still sees the hidden library).
Test-lesson encoded: cleanup ordering is LIFO — the pool Close must be
registered BEFORE fixture cleanups, or library deletes run against a
closed pool and fail silently into ignored errors.
Book cards on the library (all-books) page were oversized at full
window width, and both the all-books and series grids showed only
2 columns on phones, making covers unnecessarily large on mobile.
Shift the all-books column ladder up one at every breakpoint
(2/4/5/6 -> 3/5/6/7) so the responsive reflow the grid is liked for
is preserved while every window size gets smaller cards; the series
grid gains a column at mobile only (2 -> 3), leaving md/lg untouched
per request. Dashboard carousel sizing (w-36/w-40) and the series
stacked-covers component are deliberately unchanged.
templ codegen and the tailwind build output commit with the source
change that produced them (repo convention).
Two independent dead-ends, confirmed on prod's /admin/archived
(2026-10): Purge All rendered but clicking it did nothing at all, and
per-row Delete never even showed its confirm dialog.
Purge All is Alpine-bound (@click="purgeArchivedItems()"), but this
page's <body> lacked x-data="admin" — the only admin page without it.
Alpine evaluates directives only inside an x-data tree, so the
attribute was inert: the admin component (and its method, registered
in admin.ts) never instantiated here. The body now carries
x-data="admin", matching admin.templ and admin_library.templ.
Per-item Delete was a second, subtler dead-end: admin.ts's API helpers
guard with `if (!token) return;` BEFORE the confirm dialog, and the
reporting session had no localStorage token — SSR pages run on the
cookie session, so the web UI works while every Bearer-
authenticated admin action (purge, delete, restore, merge, scan-all,
cancel-scan) silently no-ops with zero feedback. Those guards now go
through requireToken, which toasts "Session expired — sign out and
back in, then retry" when the token is absent. The two passive
page-load probes (loadWatchStatus, checkActiveScans) deliberately keep
the silent return: a toast there would fire on every visit.
Pinned by TestAdminArchivedPageHasAlpineScope: renders the page and
asserts the body scope and the action buttons coexist in one render.
After deploy: hard-refresh prod, click Purge All (confirm appears);
if the token toast fires on any admin action, log out and back in once
to stash a fresh API token.
The archive pass looped over the scan's folders and, inside each
iteration, compared the ENTIRE library's item list against that one
folder's walk. With two registered folders every item living under the
other folder was flagged "missing (1/2)" on every reconciling scan —
prod's audiobook library (the only multi-folder one) showed 743/743
flagged — and nothing reached the two-consecutive-miss threshold only
because each scan's walk clears missing_scan_count before the pass
re-flags it. With three or more folders the defect stops bouncing: a
later iteration sees an earlier iteration's count-1 marks on items its
own walk lacks and archives them mid-pass — one registered folder away
from archiving a third of a library per scan. Single-folder libraries
were unaffected, which is why comics/ebooks/manga showed zero flags.
Folder-book presence was suspected alongside it (audioBookRoots read
during the comparison), but the map is written for every >=2-track
folder-book root on every walk (processAudioBookFile, since 6216bd9b);
the flagged folder-books were victims of the same cross-folder
comparison. The roots attribution did carry the same unchecked
strings.HasPrefix boundary bug as the rest of the scanner (8th site,
variable named `root`, missed by the earlier sweep); it now uses
folderContains.
The pass now groups folders by owning library and, per library,
compares ListMediaItemsByLibraryIncludingArchived against the UNION of
all that library's folders' walks. The coverage (partial-scope) guard,
the walk-failure skip, and the mark/archive/purge item loop are
unchanged.
Regression test (archive_multifolder_test.go) builds a two-folder
audiobooks library — single-file book in one folder, folder-of-tracks
book in the other — and scans twice: pre-fix both items are flagged
missing on both scans (verified by running the test against the old
code), post-fix neither is ever flagged or archived.
None of the three save surfaces on the admin settings page gave usable
feedback: the Base URL form re-rendered into an identical-looking form
(its only indicator was the Go-source twin's banner — markup that
diverged from the templ original), the scan form was WORSE — its handler
responds JSON, so htmx's outerHTML swap replaced the whole scan card
with literal JSON text — and the tunable rows showed only a tiny
per-row label.
Client side (toasts are the sanctioned client-local island; no new
inline scripts, CSP unchanged):
- web/src/settings.ts: one delegated htmx:afterRequest/sendError
listener keyed on a data-settings-save attribute stamped on each
form. Success → green toast per surface ("Settings saved" / "Scan
settings saved" / "Setting saved"); failure → error toast carrying
the server's JSON error field, degrading to a status line for the
tunable endpoint's HTML error body (markup must never reach a toast).
Network-level failures (htmx:sendError) toast instead of silently
re-enabling the button.
- htmx 2.x quirk handled: for FORM submissions, afterRequest's
detail.elt is the event-dispatch ancestor, not the form — the
requesting element lives on detail.requestConfig.elt (verified
empirically; saveLabelFor reads the right one).
- hx-disabled-elt="this" on all save buttons: the click is
acknowledged and double-clicks cannot double-save.
Server side:
- Scan form: hx-swap="none" (the inputs already hold the saved values;
nothing needs re-rendering) — kills the JSON-over-the-card swap.
- The Base URL form's save response is a Go-source copy of the templ
form (handlers cannot import templates: templates/utils.go imports
handlers types). The twin now carries the functional attributes
(data-settings-save, hx-disabled-elt) so feedback survives the swap,
with a drift-warning comment; the double-save e2e pins this.
Dev-DB safety: every success save in the spec re-submits the value
already on the page (zero drift), and the spec restores
system_config.updated_by after itself — the save handler stamps the
requesting admin's id there, and that FK (no cascade) otherwise blocks
the integration harness's transient-user sweep (it broke
TestUpdateProfileAdminMode/TestDeleteUserConsolidated on this run's
first pass; pollution cleaned up by hand and now self-restoring).
Tests: web/src/settings.test.ts pins the form→label map and the
error-extraction contract (including the no-markup-in-toast rule);
e2e/settings-feedback.spec.ts drives all three surfaces through the
real admin UI including the error path and the twin-drift double-save.
A bare string-prefix test (strings.CutPrefix / strings.HasPrefix) let a
registered folder whose path was a string prefix of a sibling claim the
sibling's files: "/app/audiobooks" matched "/app/audiobookssd/Brian
Jacques/...", so getRelativePath sliced the remnant "sd/" into every
stored relative path for that library (prod scan log: "[AUDIO] Created
folder book 'Taggerung' (41 tracks) at sd/Brian Jacques/Redwall/Book 14").
Scanner lookups built their keys with the same function
(getMediaItemByFilePath, scanStateLookup), so writes and lookups were
consistently poisoned and no scan ever corrected them; everything
outside the scanner resolves paths by joining back to the roots
(ResolveMediaPath), so cover/track URLs 404ed for files sitting one
directory away.
The same unchecked HasPrefix pattern also drove five folder-membership
sites, so a sibling-prefix pair misattributed more than paths:
owning-library resolution (processMediaFile, maybeProcessAudioBook),
per-library-type extension allowlisting (isScannableFile), and
folder-structure metadata extraction on both create and update paths —
a file could be scanned under the wrong library's rules entirely.
folderContains(folder, path) is now the single boundary-aware
primitive: path must equal the folder or continue with a
filepath.Separator (after Clean). All seven sites go through it, so no
folder name — however similar to another — can interfere with any
other's files again, in any order of registration; trailing slashes on
registered folders are normalized. Registration order still decides for
genuinely overlapping (nested) registered roots: pre-existing behavior,
deliberately unchanged.
Regression tests (media_scanner_path_test.go, pure unit — no DB) pin the
prod trap (audiobooks/audiobookssd), order-independence, trailing
slashes, mid-word siblings, and the helper itself.
Repair of already-poisoned rows is deliberately separate and safe only
under this fix: a rescan repoints them via the SHA-256 move path
(MoveMediaItemFilePath) once lookups stop producing poisoned keys.
The "top search bar doesn't work" report had TWO stacked causes, found
by reproducing with real keypress-by-keypress typing instead of
Playwright's fill() (which fires a single input event and masked both):
1. htmx.min.js was included PER PAGE. The header's live search is
htmx-driven on every page, but five pages rendering @Header —
series, admin, admin hash-conflicts, admin processing-issues,
browse-detail — never loaded it: the input's hx-* attributes were
inert (htmx-internal-data absent), so typing produced no request at
all. The include now lives in the header partial itself (classic
script, before the module bundle); the sixteen per-page copies on
@Header pages are removed (double-processing guard). Non-header
pages (login/register/setup/reader/index) keep their own.
2. htmx's delay: modifier on a PROGRAMMATICALLY dispatched custom event
(htmx.trigger) loses the request entirely under rapid successive
triggers — every re-fire canceled the pending timer and the last
never issued, so fast typing (the normal kind) never opened the
dropdown while one-shot fill() worked. The debounce moved into the
Alpine component (client-local state, where it belongs); htmx now
executes the bare `search` event immediately.
Also made bundle staleness structurally impossible: each bundle's
script stamp is a content hash computed at startup (templates/
bundles.go), so a rebuild that changes bundle bytes changes its cache
key with no manual ?v=N bump — the manual stamp was how the first
dropdown fix stayed invisible for a deploy. The reader bundle (never
stamped at all) is covered too. HTML pages additionally send
Cache-Control: no-cache (they are dynamic), so a stale page can no
longer reference a stale stamp; /api, /static and /uploads keep their
native caching, pinned by a middleware test.
Two follow-ups from live testing:
- "recent" is BOTH a base toolbar key and a full registry key (where it
means desc). ResolveSeriesSort's full-key passthrough ran first, so
sort=recent&dir=asc silently resolved back to desc — the direction
chevron flipped but Date Added could never sort ascending. "recent"
now composes with the direction before the passthrough, unknown bases
fall to the default entirely (a hostile dir can't ride an unknown
sort), and the composition table is unit-pinned plus an HTTP-level
row in the series filter test.
- The header search "still broken" report was a stale-cache artifact:
the fixed code shipped but main.js kept its ?v=7 stamp, so browsers
with a cached copy never refetched (fresh-profile e2e passed while
the user's browser didn't — heuristically cached module). Stamp bumped
to ?v=8 in header + docs; the e2e suite re-verified against the
rebuilt server.
Per the owner's testing bar: a Playwright spec registers an admin,
seeds one fixture per write-capable type (EPUB, PDF, CBZ, MP3, FLAC,
M4A, M4B — generated at runtime, no binary media in git), drives the
REAL metadata-editor flow (Edit → change title → in-app dialog →
Save & write) for each, and then reads every artifact back
INDEPENDENTLY of the app — EPUB/CBZ/sidecar via python zip+XML parse,
audio via ffprobe — asserting the written title landed in the FILE and
the old one is gone.
Writing the e2e exposed a real bug: atomicWrite left server-written
sidecars at 0600 root (os.CreateTemp's default), so host-side tools
sharing the volume — Calibre reading metadata.opf — were locked out.
The temp file is now chmod'd 0644 before the rename, matching the
scanner's cover-file convention; the e2e's host-side reads prove it.
Three defects from real-world use:
- The m4b rewrite assumed every stream survives -map 0 -c copy; files
carrying text-ish tracks (the movenc 'text' fourcc) fail the ipod
muxer with "Tag text incompatible with output codec id" — exactly the
error a write attempt hit in the field. The rewrite now attempts full
fidelity first, then retries mapping audio + attached pictures only
(-dn -sn): chapters live in the container header (-map_chapters 0)
so they survive; every attempt still ffprobe-validates before the
atomic rename and the original is never touched on total failure.
- The browser confirm() is replaced by an in-app dialog
(WriteToFileConfirmDialog, Alpine + templ — no native dialogs).
- "Write to file" now SAVES the pending editor edits first (persisted
through the same PUT the Save button uses), then writes the file
from those DB values, and closes the editor + refreshes the page
exactly like Save — previously the endpoint wrote the LAST-SAVED DB
values, so edits sitting in the modal never reached the file, and
the page didn't visibly update.
Test coverage per the owner's bar: EVERY single-file container (MP3,
FLAC, M4A, M4B) gets its own runtime-generated fixture, gets written,
and the FILE's metadata is then verified independently — ffprobe reads
the rewritten container and every mapped key is asserted per format,
including the subtitle round-trip split (MP3/FLAC yes, M4A/B no) — and
the EPUB/CBZ/sidecar assertions now PARSE the written XML (dc titles/
creators, ComicInfo fields, calibre meta) instead of substring
matching. A smart-genre unit test also pins the mixed-blob contract: a
recognized shelf wins the genre column even when demoted format words
ride along in the same blob.
Toolbar feedback round: the series search box now matches whatever the
main search matches for a group — series NAME, any book's AUTHOR, and
any book's TITLE inside the series (WHERE EXISTS per group-row, so
counts stay true); the separate author input is gone from the toolbar
(the author= param stays live for the JSON API and deep links). The
sort control becomes a base-key select (Date added/Name/Author/Books/
Progress/Last read) plus a chevron direction toggle that flips and
resubmits; sort/dir compose through ResolveSeriesSort (full registry
keys keep working for back-compat, unknown base keys fall to recent).
The search/apply button moved next to the search box as requested, and
default URLs stay clean (sort/dir only emitted once the user chooses).
The dropdown never opened on typing: the search-dropdown migration left
./search un-imported after dropping the initializeSearch call from
header.ts, so the bundler tree-shook the module away and the
x-data="headerSearch" scope never registered — every handler binding
(input/focus/keys) silently died. This is a client-wiring defect that
HTTP-level integration tests structurally cannot see; a Playwright e2e
spec (e2e/search-scrubber.spec.ts) now drives the real bundle: sub-2-char
input stays closed, 2+ chars opens the swapped templ dropdown.
The first grab implementation degraded Shift+grab to a quarter-speed
relative crawl — losing the ±5-minute rail the user expects when they
grab the handle with Shift held. The rail is back: Shift+grab anchors
railWindow at the CURRENT position (not the nearest chapter line, and
never teleports the playhead on pointerdown); plain grab stays relative
1:1. GRAB_FINE_SCALE is deleted as dead code, and the dialog guide
documents the restored gesture.
TestAudiobookTagFallbacks still asserted the verbatim "Audiobook" genre
column, but 885d6140 demoted format words to tags-only — and that
commit's message explicitly says the pinned expectation was updated to
the demoted contract. The code change landed; the test edit it
describes never did, so the full suite has been red since 13:21 today
on a test my changes never touched.
Pin the new contract: the genre column stays empty for demoted format
words and the raw candidate ("Audiobook") lands in the tags array —
exactly the tags-only behavior the fix's message documents. Full
integration suite green after this.
Phases 5b and 5c of the metadata writer.
5b — single-file audio (MP3/FLAC/M4A/M4B): the container's tags are
rewritten with ffmpeg -c copy (no re-encode; -map 0 keeps cover art,
-map_chapters 0 keeps chapters). The key set is exactly what the
scanner's applyAudioTagFallbacks reads back — title+album, subtitle,
artist/album_artist, composer (narrators, Audible's convention),
grouping + iTunes series-part, date, genre — so a write + rescan
round-trips. Output is ffprobe-VALIDATED (duration within 1s/0.5% of
the source) before the atomic rename; a lossy container rewrite never
replaces a good book. Fields the format can't carry per our own read
chain are deliberately not written, and fields_written reports only
what really round-trips (subtitle rides MP3/FLAC — this image's mp4
muxer drops subtitle keys, discovered empirically and pinned by test).
5c — CBZ: ComicInfo.xml injected (or replaced) with the same atomic zip
rewrite as the EPUB path, extended with an append-if-missing mode
since a comic may not carry one yet. Field map mirrors the scanner's
comic block: Title/Series/Number/Count/Volume, Writer (authors),
Summary, Publisher, Year/Month/Day, LanguageISO, Web, Genre+Tags;
fractional series numbers are omitted rather than truncated to a lie.
CBR/CB7 stay on the sidecar path (RAR/7z can't be rewritten in place).
Integration tests generate media at test runtime with ffmpeg (m4b) and
in-memory zips (cbz) — no binary fixtures — pinning the tag round-trip
through ffprobe, post-write rescan titles, the ComicInfo injection and
the validation-gate behavior.
Phase 5a of the metadata-writer feature: the book editor gains a
"Write to file" action (admin) that persists the item's CURRENT
database metadata into the artifact on disk, so the library stays
portable to Calibre and other apps — the thing Save alone can't do.
Dispatch (services/metadata_writer.go): EPUB/KEPUB gets its embedded
OPF spliced in place (fresh dc:* block generated inside the ORIGINAL
metadata open tag; manifest/spine/guide untouched; every other zip
entry copied byte-identical in original order so mimetype stays first
and STORED; original dc:identifiers preserved — the package's
unique-identifier must keep resolving). Everything else takes a
Calibre-style sidecar: metadata.opf when the book is ALONE in its
directory (Calibre's own convention — round-trips through Calibre
too), <basename>.opf when siblings share the folder, so one book's
sidecar can never poison dir-mates' scans (the scanner now reads both,
basename first).
Safety contract: temp file + atomic rename — the original is never
truncated on failure; the generated OPF is unit-tested to ROUND-TRIP
through the scanner's own parser (write side and read side can never
disagree); metadata_overrides are deliberately kept (file and DB agree
after the write; overrides still guard against stale sidecars). The
endpoint re-extracts inline (RescanMediaItem) and reports rescan
failure separately — a failed rescan doesn't undo the write. SHA-256
changes are inherent and documented for users (Kobo/KOReader re-key).
UI: the metadata editor modal carries the file path as a data
attribute; the button confirms before mutating, labels the target
(embedded vs sidecar) and toasts the outcome. Integration tests pin
the 403, missing-file error, EPUB embed + post-write rescan title, the
lone-book metadata.opf case, and the shared-dir <basename>.opf case
with no dir-scoped litter.
Search results now show where you are in a book: a completed check at
the 0.999 threshold (same cut as the filter status arms), a partial
progress bar along the cover while mid-read, nothing when you never
opened it. Rendered in BookCard (any grid whose adapter loads viewer
progress) and in the header-search dropdown rows.
The progress rides a LEFT JOIN reading_progress on the SAME user_id
param SearchMediaItemsUnified already binds for library visibility —
one requester param, no new wire inputs, row-neutral join (the table
is UNIQUE per media_item+user). SearchMediaItemsUnifiedRow and BookInfo
gain the keys additively (viewer_percentage null when unknown,
library_name empty on feeds that don't load it); contract_pins.md
carries the changelog entry and the pin tests widen in lockstep.
The header dropdown itself migrates off the innerHTML string renderer
in search.ts (a rendering-ownership violation: a templ twin existed via
the HX-Request branch). The input now hx-gets layout=dropdown and swaps
the new templ SearchResults partial — titles/authors highlighted
server-side (escape-first, safe by construction), rows with the same
progress overlays — while a slim Alpine component owns only
client-local state (min-length gating, open/close, keyboard selection,
htmx lifecycle). web/src/rendering-allowlist.txt drops the search.ts
baseline to zero accordingly.
Integration test pins the per-user join (fresh user sees 1.0/0.42/null
on the fixture; the shared admin sees null everywhere — read-only) and
the dropdown partial's overlays/count/row count.
The /series page rendered its grid in one fixed order (newest first)
with no way to find a series beyond eyeballing cards. The page now has
a toolbar — series-name search, author filter, viewer-progress filter
and a 12-key sort select — wired as a plain GET form (the page's
existing model: the URL is the single source of truth, and the library
switcher now carries q/author/status/sort along on switch).
Extensibility is deliberate (user request): sorts live in a registry
(services.ResolveSeriesSort) mirrored by a CASE cascade in
GetDistinctSeries and a <select> in series.templ — a new sort is one
entry in each. The sort value is a BOUND PARAM compared inside the
cascade, never interpolated; unknown keys resolve to the default.
Filters: series-name matches in WHERE (the group key, so counts stay
true); author (BOOL_OR any-book match) and status (finished/reading/
unstarted vs the viewer's progress, same 0.999 threshold as the
filter_query status arms) in HAVING so COUNT(*) keeps reporting the
whole series. GetDistinctSeriesCount shares the JOIN + HAVING so
pagination totals match. The reading_progress join can't multiply rows
(the table is UNIQUE per media_item+user).
Cards gain the viewer's finished-count ("X of N read" + progress bar),
and the empty state distinguishes "no series at all" from "no matches —
clear filters". The JSON /api/series endpoint takes the same params and
gains read_count/last_read_at additively.
Integration test pins every sort key against a three-series fixture
with distinct created/author/count/progress answers, the filter matrix,
composition, unknown-sort fallback, card read-counts and the JSON
contract — all as the fresh regular user so the live reader's data is
untouched.
A pointerdown on the scrub bar unconditionally mapped the click X to a
position, so grabbing the handle itself made the playhead jump — worst
with Shift, whose branch force-snapped to the nearest chapter line
before the drag even started. The user's complaint: the handle click
flowed through to the track-click behavior behind it.
The bar now hit-tests the thumb first (GRAB_SLOP_PX, a pixel test like
the chapter magnetism — a seconds threshold goes sub-pixel on long
books). A grab enters a new "grab" drag mode: the position stays where
it is until the pointer moves, then advances RELATIVELY via the new
applyGrab helper (incremental like the touch cursor, fixed scale — the
mouse is always on the bar, so Shift = quarter-speed fine crawl instead
of a pull-away zoom).
Track clicks keep today's behavior exactly: click = jump with chapter
magnetism, Shift+click = nearest chapter line, Shift+drag = precision
rail, touch = incremental pull-away. MOUSE_GUIDE gains the two grab
rows so the dialog documents the gesture; new pure helpers are unit-
tested alongside the existing scrub math.
Found while double-checking /login after the redirect fix: a wrong
password left the form completely silent. Two stacked causes, both
pre-dating the CSP work:
1. The two HX-Request branches for invalid credentials returned JSON,
unlike every sibling error branch — there was no HTML partial to
show at all. They now return a shared invalidCredentialsHTML div
(the JSON shape is kept for non-htmx API callers).
2. htmx does not swap 4xx/5xx responses by default, so even the HTML
error partials that DID exist (429 lockout, 500s, registration
failures like "Email already exists") never landed in the target.
auth_forms.ts now swaps them itself on htmx:beforeSwap when the
error response is text/html — JSON errors stay ignored.
Verified live: /login success → /dashboard; wrong password shows
"Invalid email/username or password." and stays on /login; duplicate
email on /register shows "Email already exists". Gates: make test 9
pkgs, vitest 155, Playwright 20.
The htmx login/registration partials answered success with an inline
<script> that stashed the auth token and set window.location — refused
by hash-mode script-src (no 'unsafe-inline'), so the user saw "Login
successful! Redirecting..." and nothing else (reported from the home-
page login form; /login was equally broken). The same pattern lived in
collections.go (delete → redirect) and devices.go (token regen →
reload, whose script tag was also unclosed).
Replacement, per the AGENTS.md CSP rules:
- auth.go login/registration: the partial now carries data-auth-token +
data-redirect (redirect is path-only — open-redirect hygiene — and
escaped); web/src/auth_forms.ts (new, armed from main.ts on every
page) listens for htmx:afterSwap, stashes the token, and navigates.
- collections.go delete: data-redirect, same listener.
- devices.go token regen: htmx's native HX-Refresh response header.
- The dead localStorage 'user'/'refreshToken' writes are dropped — no
reader anywhere in web/src; only 'token' is load-bearing (19 modules).
Tripwire extension: internal/middleware/csp_go_sources_test.go scans
non-test Go sources for literal <script> tags and native on*= attributes
(hand-written files only — templ codegen is owned by the .templ scan),
verified to fire on a reintroduced probe. AGENTS.md CSP section notes
the Go-emitted-partial vector.
Verified live: home-page login form → /dashboard with token stashed and
dashboard loading clean; /register form → /dashboard; collections and
devices share the same mechanisms (data-redirect / HX-Refresh). Gates:
make test (9 pkgs incl. both tripwires), vitest 155, Playwright 20,
tsc/vp clean.
Standing rule: commit per related edit with detailed conventional-commit
messages as each project section finishes, instead of batching a whole
session into one giant commit. Committing is local bookkeeping; pushing
stays a separate, explicit decision. Ordering guidance added so stacked
trees compile (service before handler, guard before dependent feature).
TestDocsMatchRoutes flagged the example response as missing a key the
handler emits — drift introduced by c645b8e2 (the wake-lock preference),
surfaced while regenerating the API mirrors for the retry endpoint.
- TestRetryProcessingIssueRowless: a row-less import_failed entry retries
to a real import and resolves; a missing file retries to 409 with the
real reason and stays open; an item-backed issue retries through the
rescan mechanism and resolves.
- TestScanLibraryFolderPathValidation: an unregistered folder_path is a
400; a registered one scopes and enqueues (202).
- TestScanSingleFolderSkipsArchiveReconciliation: scanning one folder of
a two-folder library leaves the other folder's healthy item at
missing_scan_count 0 / unarchived — pinning the partial-scope guard
against the false-archive regression it exists to prevent.
Per issue card: a Retry button — error-severity issues previously had no
action at all (only warning/info got Dismiss). Right of the Errors stat
card: Retry all, which builds a queue of every issue currently on the
page and runs them strictly one at a time through the per-issue retry
endpoint. The browser is the queue — resource use is bounded to a single
book in flight, and no batch endpoint exists to maintain.
While the queue runs, the button shows an n/m counter and cards tint via
client-local state classes (the sanctioned island shape; the cards
themselves stay templ-rendered). After the queue drains, the page swaps
in the new /admin/libraries/:id/issues/content templ partial — the
unlinked-books htmx.ajax pattern with a reload fallback — so resolved
issues drop off server-side while failures remain listed, with a
summary toast and per-item error toasts carrying the real failure
reasons.
CSP discipline holds throughout: Alpine @click component methods, no
inline handlers, no js:-prefixed hx-vals.
POST /api/libraries/:id/scan accepts an optional {"folder_path": "..."}:
the scan then covers exactly that one registered folder's subtree instead
of the whole library. The path is validated against the library's
registered folders — anything else is a 400, never trusted. An omitted
body keeps today's full-library behavior, and coalescing (409 with the
active job) is unchanged.
The LibraryPanel folder rows get a per-folder Scan button: pure htmx
attributes posting to the same scan endpoint, with hxValsFolderScan
emitting plain eval-free JSON for hx-vals alongside the existing
hxValsFolderDelete precedent. Progress surfaces through the existing
active-scan polling; no new client code.
Safety: the scanner's archive-reconciliation guard (previous commit)
skips partially-covered libraries, so a one-folder scan cannot mark the
library's other folders' items missing.
POST /api/libraries/:id/issues/:issueId/retry (admin, family-lens gated
via the existing gateIssue) re-runs the work that produced an open issue
— deliberately NOT a library-wide rescan. Two shapes mirror the two
ledger entry kinds:
- item-backed issues rescan their media item through the book-detail
RescanMediaItem mechanism (self-resolving folders, multi-file
audiobook routing, write-after-success guards) and then close the
issue (RetryMediaItemIssue);
- row-less import_failed issues re-import the single file through the
watcher's standalone processMediaFile path (RetryFailedImport: folders
resolved RescanMediaItem-style so paths stay library-root-relative),
whose create-path self-heal plus an explicit path resolve closes the
issue even when the content was already imported by a prior attempt.
Failures return the REAL underlying error so the UI can show why the
issue still exists; a vanished file is a 409 (ErrRetryFileMissing);
already-resolved issues reject re-retry.
Docs + bruno mirrors regenerated for the new route (the generator's seq
renumbering touches three unrelated bruno files as fallout; openapi
gains the path and the date bump), and the docs schema registry gains
the retry entry so TestDocsMatchRoutes passes.
Two scanner-integrity changes supporting the Processing Issues retry work:
- The size-change and sidecar-change update paths only printf'd their
failures, so a book whose metadata update kept failing (the Crusader
genre-22001 loop) degraded silently on every scan with nothing on the
Processing Issues page. Both paths now record an item-backed
metadata_extraction_failed issue — the same type and shape the
sidecar-extraction guard already used — and resolveIssuesForItem
clears it when an update finally succeeds.
- The archive reconciliation pass compares every library item against
what one scan saw on disk. The new per-folder scan scope runs a scan
covering a single folder, which would mark every item under the
library's OTHER folders missing and archive them on a later pass.
The pass now skips a library whose registered folders the scan only
partially covered (GetLibraryFolders vs the scan's folder set),
logging the skip.
applyAudioTagFallbacks captured the ffprobe genre atom verbatim into
media_items.genre — no splitting, no shelf resolution, no cap. That raw
capture is what fed 100+-char Audible category blobs into the column
(the 2026-10 Crusader import/update failures).
The ABS sidecar path already has the correct doctrine (ab2a5568): a known
lexicon shelf wins the genre column, everything else is tags-only. Give
the tag fallback the same treatment — splitGenreCandidates +
pickSmartGenre for the column, split candidates appended to tags
(NormalizeTags dedups). Demoted genres like "Audiobook" stop masquerading
as a shelf, and an unknown blob can no longer manufacture an overflowing
value at all.
Fixtures pin the new behavior: the real 111-char Crusader blob stays out
of the column entirely and lands as six separate tags; a shelf-carrying
blob still wins the column; the previously-pinned verbatim "Audiobook"
expectation is updated to the demoted (tags-only) contract.
The genre column was VARCHAR(100), and the audio tag fallback stores the
file's raw genre atom verbatim. Libation writes Audible's category ladder
as one comma-joined string ("Politicians, Presidents & Heads of State,
United States, Russia, Communism & Socialism, International Relations" —
111 chars), so any such book failed its import or metadata update with
SQLSTATE 22001 (value too long), taking the whole row down with it.
VARCHAR -> TEXT is a metadata-only change in Postgres and safe to re-run;
the in-place ALTER follows the media_highlights position-widening
precedent so existing databases converge on restart (schema.sql runs on
every boot).
The "This Week" reading-stats strip leaves its full-width row below the
subbar and now lives inside the library bar itself, between the switcher
and the customize/refresh actions. Expanding it opens an absolutely
positioned popover that FLOATS over the collections — opening the stats
no longer shifts the page down.
- LibrarySwitcher gains an optional `center` component slot rendered
between the left group and the actions group (nil = unused). Series
and collections call sites pass nil and render byte-identical output.
- DashboardWeekWidget redesigned for the slot: compact inline summary
(icon, "This Week", minutes; pages/sessions/books/streak reveal at
md/lg/xl) and a card popover panel with the labeled grid, sync note,
and the Full stats link moved in from the old summary row.
- Panel anchoring is responsive: below `sm` it anchors to the sticky
bar itself (details goes static; bar-anchored left-4/right-4, i.e.
screen-centered); at `sm`+ the details is relative again and the
panel centers on the widget (w-[30rem], translate-x centering).
- On phones the summary collapses to the chart icon + chevron (bar row
width is scarce); "This Week"/minutes reveal at `sm`, and the
"Your Library" title hides below `sm` to keep the row from
overlapping. Icon-only toggle carries an aria-label.
- web/src/dashboard.ts: click-outside and Escape close the popover,
re-querying #week-widget per event because htmx replaces it on
library switches. Rendering untouched (behavior-only island).
- id="week-widget" and all htmx wiring unchanged: week-refresh refetch
+ the /dashboard/week-widget fragment still drive library-scoped
refreshes behind the switch fade; stats-off suppression now yields an
empty slot via weekWidgetComponent instead of a hidden strip.
- web/static/style.css rebuilt for the new utilities.
Known behavior (unchanged or deliberate): the bar is sticky, so stats
stay visible while scrolling; a library switch re-renders the details
closed; scopes with no current-week reading show zeros (e.g. Manga).
Unidentified sync files used to be stuck: every KOReader re-sync
re-queues them, and there was no way to say "this isn't a book" or "I
don't want this one on the server".
Ignore is the durable dismiss: a new ignored_at column hides the row
from the page list, dashboard badge, and per-device API while the sync
keeps refreshing its metadata (the upsert never clears the column).
A collapsible "Ignored" section lists dismissed rows with Restore and
Delete so a misclick is recoverable. Delete is the hard exit — the row
is gone, and honestly re-queued if the device syncs that file again
(the confirm dialog says so).
Actions are page-scoped SSR routes: htmx swaps in the re-rendered
section so counts and buckets can never drift from the DB; ownership
is row → device → user (admins may act on any row). Integration tests
pin the lifecycle (hide everywhere, survive the sync upsert, restore,
delete), foreign-user 403s, and the render surface.
New wake-lock.ts engine (sanctioned island, dependency-injected for
tests): feature-detected navigator.wakeLock, silent no-op when
unsupported or denied, and the spec-mandated force-release-on-hide →
re-acquire-on-visible loop handled via the sentinel release event.
The player acquires on play and releases on pause (sleep-timer fades
ride the same path); the reader holds the lock for the session. Both
gate on the keep_screen_awake preference — data-keep-awake on the
player root, keepScreenAwake in readerInitExpr.
Hidden-tab playback intentionally does not keep the screen awake:
the spec forbids screen wake locks for hidden documents.
John O'Keefe