Imported from rianvdm/product-ai-public (music/finding-cd-bundles/SKILL.md). Install upstream with npx skills add rianvdm/product-ai-public --skill finding-cd-bundles. Copyright stays with the author.

Buying wantlist CDs on eBay

Overview

Price Rian's Discogs wantlist against live eBay listings — cheapest copy of each album, and whether any seller holds enough of them to be worth bundling.

Core insight: a CD remaster carries its own UPC — Thriller is 074643811224 original, 074646607329 for the 2001 Special Edition, 888750438621 for the 2014 reissue. So searching eBay by the barcode of the exact wantlist pressing returns that master and skips the remasters automatically. The barcode does the same job finding-original-cds does by hand. That's the whole trick; don't re-derive it.

Corollary that matters: this tool inherits its master judgement entirely from the wantlist. If a remaster got wantlisted by mistake, this faithfully finds remasters. The wantlist is the filter — curate it with finding-original-cds first.

And check the other marketplace. Every album is also priced against the Discogs marketplace floor, because eBay loses that comparison more often than not — see The Discogs cross-check below. Never report an eBay price without it.

Buy the subset, and message the seller

Two things about bundles that are easy to get wrong, both learned the hard way on 2026-08-05:

1. Never buy a whole bundle — buy the subset that wins. A seller's price is competitive on some of their stock and bad on the rest. resellingroly held 5 wantlist albums; four of them beat the market but their Nevermind was $21.99 against $13.20 elsewhere. Buy those four, get Nevermind somewhere else. Evaluating bundles all-or-nothing made every one of 51 look like a loss — the script now computes the optimal subset (include an item whenever the seller's price undercuts the cheapest landed price elsewhere; shipping is paid once, which is what creates the room).

2. The combined-shipping discount is negotiable, not a fixed property of the seller. eBay exposes no combined-shipping rule through the API, and plenty of sellers have none configured — resellingroly charged 4 × per-item shipping to the cent, $18.98 on $24.23 of CDs. Rian messaged them and they refunded the excess. So the "if they combine" figure isn't hypothetical; it's what you get by asking, and asking worked first try. That purchase saved ~$8.66 over buying the same four separately.

Scale expectations: US CD shipping is a near-flat ~$4.39 Media Mail (median of 152 listings; 17% ship free), so savings are real but modest — roughly 4 of 50 bundles clear $1, topping out around $6.50. --cheapest is still the right default for "what does my wantlist cost"; the bundle view is for "is there a message worth sending."

# what it costs to buy the wantlist, cheapest listing per album
node ~/git/product-ai/.opencode/skills/finding-cd-bundles/find-bundles.mjs --cheapest

# bundle view — every seller now carries a verdict vs buying separately
node ~/git/product-ai/.opencode/skills/finding-cd-bundles/find-bundles.mjs --min 3

The Discogs cross-check has no flag. It runs in every mode.

The Discogs cross-check

Added 2026-08-05 after Rian asked the obvious question nobody had asked: is eBay even the cheaper venue? Mostly it isn't. On the run that prompted it, 9 of 19 albums were cheaper on Discogs, $151.25 of eBay listings against $26.91 of Discogs floors.

Every album gets GET /marketplace/stats/<release_id>?curr_abbr=USD, printed under its eBay row as Discogs: N for sale from $X + shipping with a /sell/release/<id> link. Always on — there's no flag and no reason to want the worse picture.

The ⚠ marks rows where eBay's landed price exceeds the Discogs floor by more than $10 (DISCOGS_SHIP_GUESS in the script). That margin is a deliberate hedge, not a measurement — see the three blind spots below. Read the marker as "worth opening the tab", never as "definitely cheaper".

Three things the floor does not tell you

The Discogs figure is not comparable to an eBay landed price, in three specific ways. Say so whenever you quote it:

1. No shipping. eBay's numbers here are landed to the configured ZIP; the Discogs floor is item-only.
2. No seller country. The stats endpoint takes no location filter, so the global floor skews toward Europe and Japan — exactly the sellers whose shipping erases the gap. This is why the margin is $10 and not $5.
3. No condition. Every eBay row carries a grade; the floor carries none. A $0.50 Sonicpraise could be anything.

This is the only marketplace data the API exposes — Discogs closed per-release listing enumeration years ago, so there's no way to see the actual listings, their conditions, or their shipping. Don't go looking for a better endpoint; there isn't one.

It flags, it never filters

Nothing is dropped or hidden on the strength of a Discogs price. A floor that's really a Japanese seller quoting $22 shipping would silently delete a perfectly good US listing, and you'd never know. Show both numbers; let Rian judge.

Bundle math stays eBay-internal. The ✓/✗ subset logic still compares against the cheapest eBay alternative, which is what its verdict claims. Flagged items just carry a ⚠ so a "winning" bundle item that's still worse than Discogs is visible.

The coverage gaps are where it pays off most

Albums eBay can't see used to be a dead end. Now they carry a price and a link — and they're often the cheapest things on the list:

No US eBay listings found for:
  Massive Attack - Mezzanine  —  Discogs: 25 for sale from $2.30 + shipping
  The Postal Service - Give Up  —  Discogs: 25 for sale from $4.01 + shipping
No barcode on the Discogs release (invisible to eBay's gtin search):
  Delta Blue - Turn  —  Discogs: 1 for sale from $39.53 + shipping

Read this section out loud in any summary. It's the part that changes what Rian actually buys.

Caching

Barcodes cache forever (they never change); prices cache for 12 hours and live under a stats key in the same ~/.cache/cd-bundles/releases.json. First run of the day adds ~25 Discogs calls at the existing 1.1s throttle (~28s); re-runs add none. --refresh clears both. A Discogs failure is swallowed and deliberately not cached — the cross-check is an enrichment, so a hiccup can't sink an otherwise good eBay run, and the next run retries instead of going 12h blind.

When to use

• "Who has several of my wantlist CDs?" · "find me a bundle" · "what can I buy together"
• "Is eBay or Discogs cheaper for these?"
• Before placing an order, to check whether waiting to batch is worth it

Not for: choosing which pressing to want (→ finding-original-cds), vinyl, or actually placing the order — see Reviewing and buying below.

Tool reality — what returns what

Credentials live in ~/git/product-ai/.env (and mirrored in ~/.config/ebay/credentials.env): EBAY_APP_ID, EBAY_CERT_ID, DISCOGS_TOKEN. The script reads only those keys and never prints them.

If eBay auth returns invalid_client, the production keyset has been deactivated. Fix at Application Keys → Notifications → Marketplace Account Deletion → toggle "Not persisting eBay data" → Confirm. That gate — not a human approval queue — is what holds up new keysets, and it activates the key immediately.

Steps

1. Run --cheapest unless the question is specifically about one seller or one parcel. It handles wantlist → barcodes → searches → Discogs floors, and caches both so repeat runs make almost no Discogs calls.
2. Lead with the ⚠ summary line, not the eBay total. $244.33 for 19 albums on its own is misleading when 9 of them are cheaper on Discogs. Quote both numbers and the three caveats (no shipping, any country, no condition).
3. Read the coverage gaps it prints at the end. These are the albums eBay can't see — and they now carry Discogs prices, which is often where the best buys are. Never imply full eBay coverage.
4. Run the bundle view at the default --min 2 and quote the verdict line, not the album count. It names the subset to buy (✓/✗ per item, with where to get the ✗ ones instead) and what the message is worth. Expect ~4 in 50 to clear $1. Only raise --min if the output is genuinely too long to read — see Don't raise --min on a short wantlist.
5. If Rian is picking rather than browsing, re-run with --html and hand him the page — condition and cover art are what he'll actually decide on.
6. Hand off the listing links. Order placement is his; see Reviewing and buying.

Don't raise --min on a short wantlist

Learned the expensive way on 2026-08-05: the bundle view was run at --min 3, reported three sellers and no bundle worth messaging — and Rian then bought two CDs from boris32 in one order, a seller holding exactly 2 wantlist albums that the threshold had filtered out entirely.

The arithmetic is unforgiving. A 25-album wantlist simply doesn't produce many 3+ sellers; that run had three, all of them duds. The number of qualifying sellers falls off a cliff as the wantlist shrinks, so a threshold that was reasonable at 50 albums hides most of the real overlap at 25.

Use the default --min 2. Raise it only when the output is too long to read, and if you do, say which threshold you used so a missed 2-album pair is attributable rather than invisible. The earlier "3 is the useful floor" guidance was written when the wantlist was larger and is now wrong.

Filter the seller's location, always

The default --country US exists because without it the top result is misleading. On 2026-08-05 the biggest apparent bundle was dahub_australia holding 9 wantlist albums — shipping from AU erases the saving that makes a bundle worth assembling. koreadisco had the same problem. Only drop to --country ANY if Rian asks, and if you do, say where the sellers are.

Output

Per seller, best first:

An AUCTION tag does not mean "can't buy it now." buyingOptions is an array, and a listing carrying both FIXED_PRICE and AUCTION is a buy-it-now with bidding — price is the BIN price and currentBidPrice is the live bid. On 2026-08-05 an ad-hoc OK Computer search that filtered out everything tagged AUCTION reported the best US copy as $28.24 when a dual-format listing was sitting at $16.90 BIN — a 67% overstatement. Only exclude a listing when buyingOptions is AUCTION alone. The main script gets this right (it tags but never drops); ad-hoc searches are where the mistake happens. | Discogs: N for sale from $X + shipping | The other venue's floor, with a ⚠ when eBay loses by >$10. Advisory — see The Discogs cross-check | | Coverage gaps at the end | Which albums eBay returned nothing for, now with Discogs prices — always report these |

--html renders the same data as a contact sheet with cover thumbnails, with the Discogs floor as a clickable badge under each card. That's the one to use when the next step is choosing, not scanning.

Ordering: --cheapest lists albums alphabetically, so ⚠ rows fall wherever the album name lands. Only the bundle view ranks by economics (sellers by net saving, best first). Don't describe the cheapest list as prioritized.

Reviewing and buying — the actual workflow

There is no third-party cart API on eBay, and Order API guest/member checkout is Limited Release requiring Developer Technical Support (checked against live docs 2026-08-04). So the tool ends at links. The flow from there:

1. Open the ⚠ items on Discogs first. That's where the sell page shows what the API won't: the actual listings, their conditions, and each seller's country. A $2.00 floor from Portland is a buy; the same floor from Osaka usually isn't. Do this before building an eBay cart, not after.
2. Review the rest with --html. Photos, condition, and the seller's title in one page. Discard anything where the title contradicts the wantlist pressing, or the condition is worse than you'd accept.
3. Add each item to the eBay cart from its listing page. The cart holds items from multiple sellers, but combined shipping only ever applies within a single seller.
4. Check the cart's shipping line, and if it's charging per item, ask. eBay doesn't expose combined-shipping rules through the API, and many sellers have none configured — the cart will happily quote 4 × Media Mail. Message them; it works. On 2026-08-05 resellingroly quoted $18.98 shipping on $24.23 of CDs and refunded the excess when asked. That single message is worth more than any of the automated savings this tool finds.
5. Buy the ✓ items only. The ✗ ones are cheaper elsewhere and the script prints where. Adding them "since I'm already ordering" is how a winning bundle turns into a losing one.
6. Check buyingOptions before skipping an AUCTION item. Auction-only listings can't be bundled into a straightforward checkout, but a listing that is FIXED_PRICE and AUCTION has a Buy It Now and is bundleable like any other — and is often the cheapest copy on the page.

Calibration — what a healthy run looks like

Measured 2026-08-05 against a 32-item wantlist (US, used):

• 26 of 32 albums returned at least one UPC-matched listing; 1 had no barcode on Discogs; 4 had no US listings at all.
• --cheapest totalled $282.32 for 23 albums, ~$12.30 each landed.
• ~50 sellers with 2+, 13 with 3+ — of which 4 are worth messaging, saving $1.58–$6.46 on the optimal subset. Judged all-or-nothing instead, zero would have looked worth it.
• Precision is excellent — zero remasters/deluxe/anniversary editions across 50 hits for No Jacket Required; the barcode filter genuinely works.
• Title-mismatch drops run ~3%, and inspecting them showed only true junk (a Blind Melon CD mis-tagged with Phil Collins' UPC, two "you pick" lot listings).
• Shipping: median $4.39, 17% free, max $24.

Re-measured 2026-08-05 on the 28-entry wantlist (25 unique albums) after the resellingroly purchase:

• 19 of 25 priced on eBay, totalling $244.33; 4 had no US listings, 1 no barcode.
• Only 3 sellers with 3+ albums, and none worth messaging — every one came back "only 1 item here beats buying elsewhere". But a 2-album seller Rian did buy from was hidden by that threshold; see Don't raise --min.
• 9 of 19 flagged cheaper on Discogs — $151.25 of eBay listings against $26.91 of floors. If a run flags almost nothing, suspect the stats call is failing silently rather than eBay suddenly being competitive.
• All 5 coverage-gap albums had Discogs stock, 4 of them under $5.

If a run comes back dramatically worse than this, suspect the wantlist changed or the location filter, not the approach.

Common mistakes

• Reporting recall as a percentage of keyword-search hits. Tried on 2026-08-05 and it's the wrong metric — Nevermind's UPC matches "only 10.5% of keyword hits" but that's 75 real listings, which is abundant. What matters is absolute listings per album and whether sellers overlap. The percentage makes a working tool look broken.
• Passing a catalog number to gtin. It accepts UPC/EAN only. Catalog numbers are for the human eyeball step in finding-original-cds.
• Trusting the raw hit list without a title check. eBay's gtin index carries some mis-tagged listings; the script requires a shared distinctive token with artist or title. If a result looks wrong, it probably is.
• Implying full wantlist coverage. Roughly one album in six returns nothing. Name them.
• Recommending a bundle on album count. The largest bundle is routinely the worst buy — randrcollectables had 6 albums at a flat $21.95 each. Quote the verdict line, which is computed on the optimal subset.
• Raising --min to shorten the output. A 2-album seller is a real bundle, and on a short wantlist most of them are. --min 3 hid the order Rian actually placed on 2026-08-05.
• Judging a bundle all-or-nothing. One overpriced item shouldn't sink four good ones; it should just be bought elsewhere. This mistake made the tool report "no bundle is ever worth it," which was flatly wrong — Rian bought 4 of 5 from resellingroly the same day and saved ~$8.66.
• Treating combined shipping as a fixed property of the seller. It's negotiable. Ask before concluding a bundle doesn't pay.
• Letting a "you choose / combine" listing win a price comparison. These pick-from-a-list listings match a barcode and quote a headline price that isn't the price of that disc. They're filtered now; if a suspiciously cheap result appears, read the title.
• Quoting the Discogs floor as if it were a landed price. It has no shipping, no seller country, and no condition. "Cheaper on Discogs" is a lead to follow, not a conclusion — always attach the caveats.
• Reporting the eBay total without the ⚠ count. $244.33 for 19 albums reads as the answer when half the list is cheaper elsewhere. The two numbers belong in the same sentence.
• Forgetting the coverage gaps now have prices. They used to be a shrug; they're now often the best buys on the list (Mezzanine at $2.30 against no US eBay listing at all).

Scanning a named Discogs seller — seller-scan.mjs

Everything above prices the wantlist against eBay, with Discogs present only as an aggregate floor. This section is the Discogs-side counterpart: name a seller, get their actual listings — condition, shipping origin, seller comments and all.

node ~/git/product-ai/.opencode/skills/finding-cd-bundles/seller-scan.mjs philadelphiamusic [seller2 ...]

Keep the seller list and their shipping policies in [[discogs-seller-registry]]. Shipping policy is the whole strategy — see below.

Why this exists, and why it looks the way it does

Discogs closed release → sellers years ago and the sell pages sit behind a Cloudflare challenge, so finding-cd-bundles can only ever show num_for_sale + lowest_price. But seller → listings is open, so naming sellers inverts the dead end. Three facts, all found by probing on 2026-08-05, none documented:

1. GET /users/{seller}/inventory?q=<free text> filters that seller's stock. Undocumented, and the only reason this is feasible — q=annie lennox medusa cut philadelphiamusic's 99,068 listings to 7.
2. Pagination is capped at 100 pages for anyone else's inventory — 403 {"message":"Pagination above 100 disabled for inventories besides your own"}. A 10,000-item ceiling that sort/sort_order do not move, so you cannot partition past it. Enumeration is out; q= is the only route.
3. Inventory listings return release.master_id: null, and q= matches loosely (michael jackson thriller returns Aretha Franklin and David Sanborn). So every candidate is resolved against /releases/<id> — which also supplies the authoritative formats[] for the remaster check.

It matches the master, not the release

Unlike the eBay path — which searches the barcode of the exact wantlisted pressing — this matches on master_id, so any pressing of a wantlist album counts, then filters remasters out afterward. That's usually what Rian wants from a big seller ("do they have a non-remastered copy of this at all"), but it means the two halves of this skill answer subtly different questions. Say which one you ran.

When the wantlist entry was chosen for its packaging, master-level matching actively misleads. On 2026-08-16 the fixed filter surfaced Across A Wire at five sellers from $3.00 — every copy the US DGCD2-25222, which is a 4-panel digipak, when the wantlisted release is the EU GED 25226 jewel case picked deliberately for that reason. None carried [exact wantlist pressing], which is the only signal separating them.

So read that marker before quoting a price, and say plainly when a hit is a different pressing than the one wantlisted. A cheap hit on the wrong packaging is not a find.

Shipping policy is the strategy

Record it in the registry for every seller, and read the results through it:

• Flat rate + unlimited combining (philadelphiamusic: $5.00 flat, cart states "add up to 997 more discs/tapes at no additional shipping cost"): only the first item carries shipping. Judge each album on its item price against another venue's landed price, and batch aggressively — re-scan whenever the wantlist grows, because anything they stock is worth adding at item cost alone. This flipped the Cranberries from "skip, $10 landed vs $9.98 on eBay" to "buy, $5 marginal".
• Per-item shipping: treat albums independently, and ask about combining before checkout — it's negotiable and asking has worked.
• Unknown: assume per-item. The API never exposes shipping rules; only the cart does, so confirm there before concluding.

Calibration

philadelphiamusic, 2026-08-05, 32-item wantlist: 4 albums matched across 9 listings. Cart of 3 came to $16.00 + $5.00 shipping against $40.63 for the same three on eBay — $19.63 cheaper, with sealed NM where eBay had VG and Good. (Cart built, not placed — the saving is quoted, not realised.) The standout was The Police — Greatest Hits at $7.00 sealed against $22.00 on eBay, on an album with only 3 US eBay listings.

Known gap

The remaster filter reads the resolved formats[], but Special Edition is not tagged Remastered and will pass — a 2009 Nevermind Special Edition came through at $69.73. Obvious at that price; a cheap one could slip past. For anything expensive, confirm the master with the matrix runout per finding-original-cds.

raw > 0 -> 0 cand is the signature of a format-filter bug

Fixed 2026-08-15: DESC_CD was /\bCD\b/i, which does not match "HDCD" — there is no word boundary between the D and the C. HDCD is an ordinary CD that plays in any player, so every HDCD listing was fetched from the seller's inventory and thrown away one step before the master check. It cost an exact-pressing hit at $1.88 against $6.00 landed on eBay, and Rian found it by eye. Now /\b(?:HD)?CD\b/i; SACD, CDr, DVD and vinyl stay excluded via DESC_OTHER.

The scan line is the tell. 1 raw -> 0 cand -> 0 confirmed means the seller has something matching the query that the format filter rejected — which is indistinguishable from an honest loose-match rejection unless you look. A true miss reads 0 raw. When Rian can see stock the scan can't, check the raw > 0 -> 0 cand rows first, and print the offending release.description before assuming the search is fine.

It happened again on 2026-08-16, wider: \b(?:HD)?CD\b also fails on 2xCD. Same cause — the x is a word character. Every multi-disc release had been invisible to every seller scan ever run. The fix is /\b\d*x?(?:HD)?CD\b/i, and re-scanning eight sellers raised the hit count at six of them: Across A Wire appeared at five sellers and The Wall at three, those being the only two 2xCD albums on the wantlist. SACD, CDr and 2xLP stay correctly excluded.

The lesson is the regex family, not the two instances. \bCD\b fails against anything that prefixes CD without a separator — HDCD, 2xCD, 3xCD, SHM-CD is fine but SHMCD would not be. When adding a format token, test it against a list, and treat any raw > 0 -> 0 cand row as a bug report until proven otherwise.

A zero is trustworthy — verified 2026-08-06

q= does a loose full-text match, so a one-word query looks alarmingly broken: q=the cure returned 29 listings containing no Cure release, q=queen returned 249 with no Queen. Tempting to read as a failing search. It isn't.

The script queries artist + title together, and that form has good recall — tested against six albums confirmed present in a seller's stock (Jagged Little Pill Acoustic, Hours..., Boys For Pele, Tin Machine's Live: Oy Vey Baby, Echobelly's Lustra) it found 6 of 6, usually as the only hit. Trailing punctuation didn't matter.

So a reported 0 means the seller genuinely doesn't stock it. Sanity-check the inventory's format mix before writing off a whole scan, but don't assume the search is broken just because single-word queries look noisy — and check the raw > 0 -> 0 cand rows above before concluding anything.

Hit count is not buy count

The most forgettable rule here. A 6-album hit looks like a 6-album order and was a 3-album order.

Scanning answers "who has my wantlist albums". It does not answer "who should I buy them from" — for that, every hit still has to clear both alternatives:

1. The eBay landed price for that album, from --cheapest.
2. The Discogs floor for the exact release the seller is listing — not the wantlisted release ID, which is usually a different pressing with a different floor. GET /marketplace/stats/<release_id>?curr_abbr=USD.

Read the floor's depth, not just its price: 84 copies from $0.01 is real and reachable; 5 copies from $6.00 may be a foreign seller whose shipping erases the gap.

A seller can hold six of your albums and deserve three of them. Report the split, not the hit count.

Sealed listings need a catalog-number check

A big seller clearing retail stock (CD_WAREHOUSE_817) lists mostly "Brand New, Factory Sealed" — and one listing said outright "Cannot confirm matrix/pressing, as it is sealed." Two of its sealed Mint hits carried later catalog numbers than the wantlisted originals (88875043862 for Thriller instead of EK 38112; B0015887-02 for Nevermind instead of DGCD-24425) and passed the remaster filter only because neither was tagged Remastered.

Sealed + later catalog + unreadable matrix is the worst combination in this whole system — it looks like the best listing on the page and can't be verified. Compare the catalog number against the wantlisted release on every sealed hit.

Discovering new sellers — seller-discover.mjs

seller-scan.mjs needs a name. When Rian asks for new big sellers, there is no endpoint that lists them and every HTML route is Cloudflare-walled (marketplace sell pages, Discogs forums, Steve Hoffman — all 403 to a fetcher).

Don't guess usernames. Tried on 2026-08-06: 146 guessed names yielded 25 live sellers, almost all record stores already known, and nothing new above 30k listings. Guessing can only surface sellers you can already name.

node ~/git/product-ai/.opencode/skills/finding-cd-bundles/seller-discover.mjs [probes]   # default 1800

Marketplace listing IDs are sequential integers (currently to ~4.35e9), and /marketplace/listings/<id> returns seller.username, ships_from, condition and release.description. Sampling random IDs samples the marketplace weighted by listing count, so big sellers surface by construction. Hit rate is ~3%; 1,800 probes nets ~55 live listings and ~54 distinct sellers in ~35 min. Run it in the background and touch nothing else on the Discogs API meanwhile — the throttle already sits near the 60 req/min cap.

It's a candidate generator, not a ranking — the biggest sellers appear once or twice at that sample size, so the script resolves every name against /users/<name> for the true inventory size. This is how LIVINGISEASYNJ (146k listings, third-largest US seller) was found after guessing missed it.

Most of the marketplace is not US. Of 54 sellers sampled, only 8 shipped from the US — against UK 16, Germany 6, Spain 5, Japan 4, Netherlands 2, Sweden 2, France 2, Canada 2, and one each from Denmark, Ireland, Italy, Greece, Brazil, Portugal, Ukraine. A random Discogs listing is only ~15% likely to be US, which is why --country US matters on the eBay side and why the /marketplace/stats floor skews cheap-but-foreign.

The largest CD sellers are European and therefore out of scope for shipping: recordsale-de (Berlin, 1,338,208 listings) is 5.4× the biggest US seller. Also large: KUPIKU-COM (Japan, 474k), Lot.Of.Music (Netherlands, 321k), RecordCityDubStoreUK (204k), echogrove (Japan, 173k), www.hhv.de (154k). Worth knowing when something is unobtainable domestically.

Does size predict hits?

Loosely, and only within the right inventory shape: 250k listings → 6 albums, 146k → 2, 99k → 4, 46k → 6, 15k of the wrong kind → 0. But devman242 at 1% CD held the best copy of a gap album on two consecutive scans, and CD_WAREHOUSE_817 at ~19% CD tied the best hit rate in the registry.

So size and shape predict hit rate, never whether a specific album is present. A big wantlist gap justifies scanning a seller the screen would otherwise skip.

Ask Rian who he's already bought from — it beats both scripts

dailybookstore (32,281 listings, 100% across 11,959 ratings, 59% CD, US, and $5.00 for 1–5 CDs then +$0.25) was missed by 146 guessed names and by marketplace sampling. It surfaced on 2026-08-15 only because Rian mentioned he'd ordered there before.

Order history is a pre-filtered candidate list — a seller he's bought from has already cleared shipping, condition accuracy and communication, none of which any API field reports. Ask before probing, then confirm size and rating with one GET /users/{name} call.

Screening a candidate before a full scan

• Read ships_from off sampled listings, never the profile location field — it's free text and often blank or wrong. BagelRecords shows a blank location and ships from the Netherlands.
• CD share screens out vinyl shops. Under ~15% CD means a vinyl business listing a few discs: MUSICMKT 12%, academyrecords 13%, mightyvinyl 3%, audiophileusa 0%. All scored zero wantlist hits.
• Never reject on control-album queries. Querying a handful of specific mainstream albums and rejecting sellers that lack them would have rejected bordentownrecords, the best seller in the registry, because its 6 hits were six other albums. That test also scored LIVINGISEASYNJ as "skip" and devman242 as "marginal" — both then produced real hits on a full scan, including the best Nevermind found anywhere. Any control set cheap enough to run is too narrow to reject on.

Screen on format mix and origin, then run the full scan — it's ~35 calls and it's the only decisive test. And note that devman242, at 1% CD, held the best copy of a gap album: shape predicts hit rate, not whether a specific album is present.

New-retail as a third venue — narrow, but real

eBay and Discogs are both used markets. A new-CD retailer is a different shape, and it wins in exactly one case: a recent album whose current pressing is the only pressing.

Tested 2026-08-16 on deepdiscount.com (Alliance Entertainment): Slowdive Everything Is Alive at $13.65 new, in stock, against a $24.53 Discogs floor and no US eBay listing at all. Free shipping over $25.

A full 29-album sweep on 2026-08-16 found 6 worth buying, 5 of them on the exact wantlisted barcode. Full per-album results: [[deepdiscount]].

The UPC is the last path segment of the product URL — deepdiscount.com/<slug>/<upc> — so it is readable straight off the search results page with no product-page visit. Compare it against the wantlisted release's barcode, which is already cached in ~/.cache/cd-bundles/releases.json because the eBay half of this skill searches by barcode. A match means it is the wanted pressing, by the same logic --cheapest already relies on.

"Reissue" does not mean "remastered", and a different UPC is not a remaster. An earlier version of this section claimed new retail was structurally wrong for a used-market wantlist; Rian pushed back and the barcode data disproved it. The hard line is the Remastered tag (see finding-original-cds), and 3 of the differing-UPC hits carried no such tag on any release sharing the barcode. Check, don't assume.

When resolving a UPC on Discogs, read every release sharing that barcode, not results[0]. The top hit for Dummy's barcode is a Russian unofficial pressing and for Nevermind a Brazilian one — judging on the first result would have been badly wrong in both directions.

Still true: a catalogue title whose catalog number differs from the wantlisted original deserves scepticism even with no remaster tag, because Discogs tags are user-submitted and a new catalog number is the classic remaster tell. Absence of a tag is the absence of a claim.

It needs a browser, not a fetcher

deepdiscount.com is behind Cloudflare Bot Management — every path including the homepage returns 403 with a 12-byte body and sets __cf_bm. Both curl-with-a-browser-UA and WebFetch get the same 403. The claude-in-chrome skill works, because it drives a real Chrome session.

• Search URL is https://www.deepdiscount.com/search?q=<query>&mod=AP. Not /catalogsearch/result/?q= — that's the Magento default and it 404s here.
• Search results show title, format and price but no catalog number or UPC, so pressing identity needs the product page — which is the whole question for this wantlist.
• Filter to CD rows (/\bCD (List )?Price:/). Without it, t-shirts, jigsaws and socks flood the result and truncate the useful rows.
• Don't wire this into the scan pipeline. Driving a browser across 30+ albums is slow and breaks on markup changes; use it for one-off checks on recent albums.

Related

• Skill finding-original-cds — which pressing to want in the first place, and the matrix-runout check for confirming a master. This skill assumes that one has already run.

• [[discogs-seller-registry]] — who to scan and what shipping costs. Durable reference, rewritten in place. Read it before scanning; update it when a seller's policy or standing changes.

• [[discogs-scan-results]] — current hits, baskets and what to buy. Living document, overwritten each run, never appended. Write results here, not into the registry.

  Every seller name in that file must be a link, and every buy row must carry its listing link — the file exists to be acted on, and an unlinked username means opening Discogs and typing it in. seller-scan.mjs already prints https://www.discogs.com/sell/item/<id> per listing; the seller profile is https://www.discogs.com/seller/<name>/profile. Link once per seller (the heading), not three times on the same page.

• [[discogs-original-cd-finder]] — the human-readable version of the pressing playbook, for when Rian wants to eyeball the rules himself.

Keep those two files apart. They were one file until 2026-08-16 and it rotted: seller notes became append-only logs where durable facts (who a seller is, what shipping costs) were interleaved with dated state (this cart, that basket), so a single seller took five stacked paragraphs across three dates to read, and nothing stale ever got deleted — it just got another paragraph underneath. Durable facts go in the registry and get rewritten; dated state goes in scan-results and gets overwritten; mechanics go here.