The badge counted lines nothing on the page identified. Each shaky line
(vision confidence below high, nothing verified yet — the badge's exact
predicate, served per-line so page and badge can't disagree) now shows
a dashed gold "shaky read" chip, a filterbar toggle shows just them,
and a "✓ looks right" button confirms a correct read in one click —
recorded as a confidence-high edit record, so verification is as
durable as any other curation and, unlike a real edit, re-queues
nothing. Saving an unchanged edit form on a shaky line counts as
confirming it. Help's chip legend explains the chip and its three ways
to clear (confirm, edit, or a BGG match).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g
Eric's observation: cleaning up raw reads happens BEFORE resolve, but
the app ordered Review ahead of Catalog and never said proofreading was
a step at all. The pipeline has two human checkpoints — proofread the
reads (after extract), decide the matches (after resolve) — and now the
app says so: the sidebar runs Pipeline, Photos, Titles, Review, Queue,
Library, Help in true workflow order; the Titles page (né Catalog — the
old name suggested a finished collection, which is the Library's job;
/catalog redirects) gets a badge counting unresolved shaky reads; an
edit marks its entry human-verified so the badge drains as you
proofread; the extract stage card nudges toward the proofread before
resolving; and the Help flow is rewritten as six stages + two
checkpoints with the loop called out.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g
README: the Catalog page description and a new "Fixing what the model
gets wrong" section cover the curation actions (edit/split/remove) and
their durability contract. Spec: amended-as-built notes — duplicate
same-edition copies are now supported via explicit splits, review is
both a TUI and the web app, and the data model lists the three curation
stores plus the dismissed-tickets file.
New Help page in the app (nav entry on every page): the six-stage flow,
what each page is for, the curation semantics with undo instructions, a
status-chip legend rendered with the real chips, keyboard shortcuts,
and where the data lives on disk.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g
Titles that aren't games (misread box art, out-of-scope items) can now
be removed outright: a danger button in the catalog's edit panel posts
/api/remove-title, which drops the line's matches rows (veto'd ones
too — removal is the human explicitly discarding the line), records the
decision photo-scoped in data/title_removals.json, and replays
titles.json. Every rebuild filters removed sightings after edits and
before dedupe, so re-extraction cannot resurrect them; undo by deleting
the record from the store. The three stores now share one scoped-record
parser and recorder.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g
The standing post-feature audit over a7f0cfe. Correctness (data): splits
become photo-scoped store records so splitting one edition no longer
force-splits same-named editions, and renaming a split copy migrates its
protection to the corrected title instead of silently re-merging copies.
Correctness (web): edit scoping now counts siblings by NORMALIZED title
(matching how stored edits apply), same-title-same-photos edits are
refused rather than corrupting the sibling entry, split copies serve
their real per-photo cues to the edit form instead of blanks, and a
split whose row vanished underneath returns 409 instead of a false 200.
Silent failures: replay_titles refuses to rebuild from a PARTIAL raw
cache (fresh clone + one --only extract would have truncated the
committed titles.json); the edit endpoint writes in crash-safe order
(cull, record, replay); corrupt curation stores fail loud naming the
file; retried edits don't double-record. Review-decision durability:
drop_rows never drops dedupe_veto rows — a rename retitles them in
place — and writes through a no-reload path so a concurrent rewrite
can't silently discard the cull. Style: catalog action cells get their
own class (.rowactions' flex display broke table alignment), editor
inputs match the design system and stop overriding the global
focus-visible outline, EditBody's clear-semantics docstring scoped to
cue fields, "nothing to change" derived from the record itself.
Tests: 8 new (photo-scoped splits, veto preservation, photo-narrowed
drops, 409s on both curation endpoints under a running job, partial-raw
replay guard, rename-keeps-protection lifecycle, corrupt-store error,
cue-field editing) and the dead edition_hint key in the edit test now
exercises real cue fields. 259 passing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g
Wiz-War had no split button: can_split required a matches row, but fresh
extractions leave multi-photo titles rowless until resolve runs. Splits
are now a title-level decision persisted in data/title_splits.json,
honored by extract's dedupe and resolve's dedupe on every rebuild, with
the button on any multi-photo line — resolved or not.
Same mechanism carries human corrections: data/title_edits.json stores
fixed misreads and known cues (publisher/edition/year/language), applied
before dedupe on every titles.json rebuild, editable from a new inline
form on every catalog line. An edit drops the title's stale matches rows
so resolve re-queries with the corrected data.
The catalog page now sorts alphabetically (case-insensitive; split
copies stay adjacent) instead of extraction order.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g
The re-run confirmed round 1 held and then caught second-order bugs in
its own fixes plus two long-standing ones everyone missed. TUI decisions
after a mid-session reload were counted but never written (rows are now
re-adopted into the fresh list on every save, preferring undecided slots
on duplicate keys); row_ix was computed by equality so duplicate rows
shared an ordinal (identity now, merges included, veto sends it); upload
job keys collided for two same-version copies (completions are counted
per key, so --limit or an interrupt can no longer strand the second
copy); diff consumes collids on exact-version matches (a vetoed
same-version second copy was silently swallowed) and splits mismatches:
report-only disagreement while an unclaimed copy exists, second-copy add
only when every copy is claimed.
Also: XML responses are validated and written atomically before caching
(a torn or truncated 200 body can never poison a re-run), JSON artifacts
write atomically, thing/search parsers refuse missing ids like the
collection parser, empty game names are refused by the upload queue, a
never-rendering version picker fails retryably instead of terminally,
the systemic-failure abort compares exception types, blocked same-title
entries defer as a group so positional pairing can't misalign,
truncation heads pick the earliest separator, diff messages tell the
truth when a token exists without a username, and the shared-constant
sweep now actually covers every module (statuses, search types, marker
names, client_for, ports). pydantic declared as a direct dependency.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Correctness: review vetoes persist via a dedupe_veto column (resolve
re-runs no longer overturn humans); diff emits second copies whose
confident version matches no owned copy (spec: pairs own only on both
ids) and fetches the live collection with refresh; resolve pairs
titles.json entries to rows by title so a reshoot photo updates
provenance instead of duplicating rows; version lookups survive empty
/thing results; publisher tie-break now honors the mixed
base/expansion veto and refuses multi-candidate picks; empty-normalized
(non-Latin) titles never count as exact.
Upload: LoginError aborts a run instead of logging N bogus failures
(and 3 identical consecutive failures abort as systemic); Cloudflare
interstitials are detected; added-without-version gets its own logged
status that verify understands; same-game updates run one per pass so
the name-targeted row edit can't overwrite a fresh version; absent
diff outputs fail loudly; pagination clicks are paced.
Web review: a lock serializes freshen/decide (threadpool race dropped
decisions); failed saves roll memory back and always alert the browser
(non-JSON 500s included); session warnings reach the page instead of a
StringIO; state-load failures and dead servers show banners instead of
a blank page; duplicate (title, photos) rows are addressable by
ordinal.
Consistency: shared CONFIDENT_VERSION_STATUSES, client_for(),
Config paths for every artifact, one review-port constant, named
matching thresholds, strict collection-id parsing, error-doc responses
never cached, unknown config keys warn, extract reports dropped vision
entries, fixture generators share escaping + marker text.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ReviewSession re-reads matches.csv/titles.json on mtime change so
external extract/resolve runs show up per-request (and stale in-memory
rows can no longer overwrite them); the page polls state every 3s,
re-rendering only on change and never mid-typing. --web --dev adds
uvicorn source-watch restarts, scoped to the package dir so decision
writes to data/ don't trigger them.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Original art (a bagpiper whose bag is a board game box) lives in
assets/; web-sized derivatives ship in the package: a face-crop avatar
in the header and favicon, and the framed full-length piper on the
review done screen. Served via an allowlisted /static route. README
leads with the full portrait, credited to Juniper.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rows resolving to the same (bgg_id, version_id — or both version-
unknown) are the same physical game read twice unless their extraction
cues conflict (two editions stay separate). The survivor is the read
whose transcription matches the BGG name; losers are marked
match_status=merged with a new merged_into column — no row is ever
deleted, and older matches.csv files without the column still read.
Downstream: diff skips merged rows but folds their photos into the
survivor's to_add provenance; enrich and the review passes ignore them.
The web UI gains a Merges section ("Jokin Ha... merged into Joking
Hazard") with a veto (v key) that restores the row as a distinct
approved match, plus a merged catalog chip and header tally.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The done screen claimed "diff-ready" while 104 extracted titles had
never been resolved (they're invisible to matches.csv until the BGG
token arrives). The state now counts titles.json entries with no
matches row: the header tally shows "awaiting resolve", the done screen
says "Resolved set fully reviewed" with the real extracted total and
what to run on token day, and a read-only Catalog section lists every
extracted title with its status chip (auto/approved/rejected/awaiting
BGG), matched game, version, and source photos.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
One self-contained page (inline CSS/JS, system fonts, works offline):
match cards show source photos, extracted cues, and candidates with
cached-XML thumbnails (placeholder tiles until real fixtures exist);
actions are pick / manual BGG id / reject, plus a skippable editions
pass (pick or unknown). Keyboard-first: j/k navigate, 1-9 pick, r
reject, m manual, u unknown, d dismiss. Every decision writes
matches.csv through the same ReviewSession methods the TUI now shares —
the TUI remains as the no-flag fallback. unidentified.json renders as
visually distinct reshoot work-orders with dismissals persisted in
data/unidentified_dismissed.json (survives extract rebuilds). Progress
tally and a diff-ready done screen; photo serving is allowlisted to
photos/ contents; server binds 127.0.0.1 only. Layout leaves room for
a later games.json browse view.
Provenance guard: fixture generators now write STUB_FIXTURES.marker
into their cache dirs, and CLAUDE.md gains the hard rule that stub-
resolved version_ids are placeholders — upload must refuse to run
while data/bgg_cache/STUB_FIXTURES.marker exists.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Eric Wagoner