Source: task/chunk-large-thread-candidates (plan/steward-linker-quality) — content.js sent all candidates in one call, the server capped at 100, and names past the cap were permanently null-cached and never retried — silently breaking large threads.
- content.js: resolve the per-scan ask set in batches of BATCH (100, <= server MAX_CANDIDATES) via a new pure Extract.chunk helper, so every candidate is actually answered and none is dropped by the resolver's slice.
- content.js: cache only keys actually present in a response (hasOwnProperty); cap-overflow / partial-error / resolver-down names are left out of `resolved` and retried next scan instead of stuck at null.
- extract.js: add pure chunk(arr, size) on the RSLExtract global; extract.test.cjs covers exact-multiple, remainder, under-size, empty, and lossless in-order concatenation.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Source: task/album-cue-lower-run-converge (plan/steward-linker-design) — the 16d8f74 LOWER_RUN both leaked lowercase prose and truncated digit-led titles; this converges the precision/recall balance.
- Weak cue (album|record|ep|lp): admit a capital/digit-led run OR a SINGLE lowercase token only (LOWER_ONE, guarded by a negative lookahead so a second lowercase word reads as prose and is rejected) — closes the 'makes me cry' / 'changed everything' prose leak.
- Strong cue (called|titled|named): the ONLY place a multi-word lowercase run (LOWER_RUN) is admitted — 'an album called good kid maad city' → 'good kid maad city'.
- New DIGIT_RUN (digit lead absorbs trailing lowercase tokens, bounded 1-4 words, tried before TITLE_RUN) fixes the digit truncation — '36 chambers'/'1989 deluxe'/'808s and heartbreak' capture in full.
- Position-mapping invariant enforced in the album push helper; bounded regexes (ReDoS test on a multi-KB pathological input); comprehensive BOTH-direction tests (precision: non-stop-set lowercase leads emit nothing; recall: digit continuation, single token, strong cue).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Source: task/album-cue-recall-rebalance (plan/steward-linker-design) — the a35dd95 precision tightening over-corrected, dropping real number/lowercase-led and abbreviation-period album titles.
- Loosen the album-cue lead token to admit a DIGIT (and, after an explicit
"the album X" cue, a lowercase-styled run), so number-led ("the album 1989",
"1989 LP", "36 Chambers") and lowercase-styled ("good kid maad city") titles
survive; ALBUM_LEAD_STOP still rejects an is/was/by/the/a/an lead.
- Narrow QUOTED_RE's mid-span reject to SENTENCE-ENDING punctuation only (a .!?
before whitespace/EOL, abbreviation periods exempted), so "M.A.A.D City",
"Mr. Bungle", "Sgt. Pepper's" survive while prose quotes still reject; raise
the word cap 6 → 8.
- Add the curly quotes ”“ (U+201D/U+201C) to TRAILING_PUNCT so an LP/EP run that
grabbed a trailing curly quote ("Reek of Putrefaction”") is trimmed, with the
span index/length kept so text.substr(index,length)===name still holds.
- Bound the "the album X" branch to the same 1-4-token run as the LP/EP branch,
so "the album Money Store dropped" yields "Money Store", not the sentence tail.
- Tests assert BOTH directions: recovered recall AND retained precision, plus the
curly-quote-trim position-mapping invariant.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Source: task/album-cue-precision (plan/steward-linker-design) — album cue regexes over-generated junk candidates (no left anchors), feeding noise to the resolver/LLM and risking fallback mis-links.
- X LP/EP branch: anchor the capture to a 1-4 capitalized/quoted title run (mirrors CAND_RE) so 'I think their new LP' no longer yields the sentence prefix.
- 'the album X' branch: require the name to start with a capital/quote and reject a leading verb/article (is|was|by|the|a|an) so 'the album by Carcass' / 'the album is great' yield no candidate.
- QUOTED_RE: reject prose quotes — a span with mid-string sentence punctuation (. ! ?), over ~6 words, or a lone stopword interjection ('lol') — so prose quotes never become album candidates.
- Preserve position mapping (trim a captured leading quote and shift the index) and dedupe identical spans across the two signals; recall of real titles ('Reign in Blood', 'Human Jerky') unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Source: task/resolver-security-hardening (plan/steward-linker-security) — defense-in-depth: client trusted resolver href, outbound fetches had no timeout, CORS was blanket-*.
- F1: extension/content.js makeLink re-validates primary.url scheme via a new
pure Extract.isSafeHttpUrl helper (in extract.js, unit-tested); a non-http(s)
url (javascript:/data:/garbage) is dropped and the plain text is kept.
- F2: each of the three resolver outbound fetches (Spotify token, Spotify
search, ollama /api/chat) now carries signal: AbortSignal.timeout(5000) so a
hung upstream fails fast (caught per-candidate -> null), zero-dep.
- F3: resolver CORS reflects the request Origin only for moz-extension://* or
null (the extension's background-script origin), with Vary: Origin, instead
of blanket access-control-allow-origin: *.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Source: task/album-detection (objective/linker-improvements) — project brief asks for artist OR album linking; MVP linked artists only.
- service/lib.js: pickAlbum (the exact-name gate, named for the album path) + linkResult now carries kind:"artist"|"album" (additive; defaults to artist so older clients keep working).
- service/resolver.js: one Spotify type=artist,album search per candidate, exact-match against either; returns {url,kind}; cache key includes the kind hint. resolveOne/resolveAll thread the LLM kind hint through.
- service/llm.js: parseClassification now returns Map(name -> "artist"|"album") so album-kind candidates route to the album path (.has() still works like the old Set).
- extension/extract.js: albumMatchesIn over two precise signals (quoted strings + album cue phrases "the album X", "X LP/EP"), folded into allMatchesIn.
- extension/content.js + styles.css: album links render italicised (.rsl-album) with an album tooltip; artist behavior unchanged.
- Tests: new pure-logic coverage in lib.test.js / llm.test.js / extract.test.cjs (album shaping, Map-kind routing, album candidate extraction). scripts/check.sh green.
- README.md + AGENTS.md: document album linking.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Catch artist names written in lowercase right after a cue phrase ("check out devourment",
"by cattle decapitation", "fan of ...") — the Title-Case/ALL-CAPS pass missed those.
- extension/extract.js (new): pure extraction (matchesIn + new cueMatchesIn + allMatchesIn),
UMD so it loads as a content script AND require()s in node for tests. The cue pass uses the
regex `d` flag to map matches back to positions for in-place wrapping; a small cue-stoplist
drops "the/this/new/..." right after the cue.
- extension/content.js: extraction now comes from globalThis.RSLExtract.allMatchesIn
(Title-Case + cue union), keeping the existing position-based wrapping.
- extension/manifest.json: load extract.js before content.js.
- extension/extract.test.cjs (new) + scripts/check.sh: 6 node:test cases for the extraction
logic; check.sh now runs the extension tests too (16 total, green).
- README: updated the detection limitation note.
Deferred (design item 2): merging names split across inline formatting (cross-node) — more
involved; left as a follow-up.
Verified: scripts/check.sh green (16 tests). In-browser behavior is logic-verified only (no
headless Firefox); additive to the pending background-fix re-test — the [rsl] logs distinguish
"candidates found" from "resolve failed".
Source: task/smarter-candidate-extraction (objective/linker-improvements).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
