feat: spotify-only matching + optional local-LLM gate + google→bandcamp fallback (closes task/drop-musicbrainz-google-fallback, closes task/ollama-classifier)

Source: human-feedback note 01KVDYV3A35B57N15CRBQDXPCH (project/reddit-spotify-linker,
2026-06-18) — "remove the musicbrainz integration and just use spotify"; a google search
for "<name> bandcamp" fallback; and "would a local llm be better able to categorise these
... include instructions for running something suitable via ollama (16GB M2)".

- Remove MusicBrainz entirely (viaMusicBrainz + url-rels + rate-gate; lib relUrls/mbResult).
  Spotify is now the only music API; it both gates by exact-name match and supplies the
  direct artist link.
- Optional ollama LLM gate (OLLAMA_URL enables it, OLLAMA_MODEL default qwen2.5:3b) called
  over plain HTTP -> stays zero npm-dep. It classifies which candidates are real
  artists/albums (the matching-quality lever). Pure prompt/parse logic in new llm.js
  (unit-tested in llm.test.js); the fetch + resolveAll gate wiring live in resolver.js.
- A confirmed name links to its direct Spotify page when available, else a Google search for
  "<name> bandcamp". New "google" primary platform rendered by the extension (styles.css
  blue accent + platform-aware link title in content.js).
- README: drop MusicBrainz; document the two gates + a runnable local-LLM setup for a 16GB
  MacBook M2. AGENTS.md: flip the old "no ML/NER" standing decision per the human override.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

AGENTS.md

@@ -8,7 +8,10 @@ is tracked in the **towl** work-item store under project **`reddit-spotify-linke
 
 - `service/` — zero-dependency Node resolver (Node 18+, built-in `fetch`/`http`).
   `resolver.js` = the HTTP server; `lib.js` = pure, side-effect-free logic (the
-  artist-match precision lever) so it's unit-testable; `lib.test.js` = `node:test`.
+  artist-match precision lever) so it's unit-testable; `llm.js` = pure ollama
+  prompt/parse helpers (the optional local-LLM gate, HTTP-only — no in-process
+  model runtime); `lib.test.js` + `llm.test.js` = `node:test`. The only music API
+  is Spotify (MusicBrainz removed).
 - `extension/` — Firefox **MV2** content script (`content.js` + `manifest.json` +
   `styles.css`). No build step.
 - `scripts/check.sh` — the local check gate.
@@ -56,7 +59,13 @@ fixes small things directly.
 
 - Firefox **MV2** stays — Mozilla has no deprecation timeline; do not migrate to MV3
   speculatively.
-- **No ML/NER** in the resolver — keep it zero-dep; improve detection with cheap
-  heuristics + the API exact-match filter.
-- Detection precision lever: a candidate links **only** when a music API returns an
-  artist whose name matches it case-insensitively.
+- **Optional local LLM is the recommended better-matching path** (human override,
+  feedback 2026-06-18 — supersedes the old "no ML/NER" rule). The resolver may call
+  ollama over HTTP (`OLLAMA_URL` enables it, `OLLAMA_MODEL` defaults `qwen2.5:3b`)
+  to classify which candidates are real artists/albums. It stays **zero-NPM-dep**:
+  HTTP only, no in-process model runtime. The heuristic + Spotify exact-match is the
+  zero-dep default when no LLM is configured.
+- Detection precision lever (zero-dep default): a candidate links **only** when
+  Spotify returns an artist whose name matches it case-insensitively. A confirmed
+  name links to its Spotify artist page if available, else a "<name> bandcamp"
+  Google fallback.

README.md

@@ -8,15 +8,19 @@ list of clickable artists instead of names you have to copy-paste into search.
 
 The hard part is *"which words are artists and which aren't?"*. Rather than guess
 with client-side rules, the extension hands candidate names to a tiny local
-**resolver service** that checks each one against a real music database and only
-links the names that actually match an artist.
+**resolver service** that confirms each one and only links the names that are
+really artists. The resolver confirms names two ways (you enable either or both):
+**Spotify** (exact-name match → direct artist pages) and an optional **local LLM**
+via ollama (classifies which candidates are artists/albums). A confirmed name
+links to its Spotify artist page when available, otherwise to a Google search for
+"<name> bandcamp".
 
 ```
  reddit page                 content script                 resolver service
  ───────────                 ──────────────                 ────────────────
- comment text  ──extract──>  candidate names  ──POST /resolve──>  Spotify API  (direct links)
-                                                                     or
- <a> links     <──rewrite──  confirmed links  <─────────────────  MusicBrainz  (zero-setup)
+ comment text  ──extract──>  candidate names  ──POST /resolve──>  Spotify API  (exact-match + direct links)
+                                                                     and/or
+ <a> links     <──rewrite──  confirmed links  <─────────────────  local LLM    (ollama, optional)
 ```
 
 ## Layout
@@ -42,18 +46,18 @@ there are no dependencies.
 ```sh
 cd service
 node resolver.js
-# resolver listening on http://localhost:8787  (provider: musicbrainz)
+# resolver listening on http://localhost:8787  (spotify: off, llm: off)
 ```
 
-That's it. With no configuration the resolver validates names against
-**MusicBrainz** (no account required). When MusicBrainz knows an artist's official
-links (its `url-relations`), you get **direct** Spotify and Bandcamp artist pages;
-otherwise it falls back to Spotify/Bandcamp **search** URLs.
+For anything to link you need **at least one** of two gates configured —
+**Spotify credentials** or the **local LLM** (or both). With neither, the resolver
+warns at startup and nothing links.
 
-### Optional: direct Spotify links
+### Spotify credentials (one of the two gates)
 
-For **direct artist links** (instead of Spotify search), give the service Spotify
-API credentials:
+Spotify does two jobs: it **gates** a name (a candidate links only when Spotify has
+that artist by exact name match) and supplies the **direct** artist-page link. To
+enable it:
 
 1. Create a free app at <https://developer.spotify.com/dashboard> and copy its
    **Client ID** and **Client Secret**.
@@ -63,10 +67,11 @@ API credentials:
    SPOTIFY_CLIENT_ID=your_id
    SPOTIFY_CLIENT_SECRET=your_secret
    ```
-3. Restart the service. It will report `provider: spotify` on startup.
+3. Restart the service. Startup reports `(spotify: on, llm: off)`.
 
-(Bandcamp has no public API; direct Bandcamp links come from MusicBrainz
-`url-relations` when present, else a Bandcamp search URL.)
+With Spotify alone, names that aren't on Spotify don't link — precise, but it
+misses real artists Spotify doesn't carry. The local LLM gate (below) catches
+those.
 
 Quick check that it works:
 
@@ -77,6 +82,37 @@ curl -s localhost:8787/resolve \
 # {"Cerebral Bore":{...},"Devourment":{...},"Notarealband Xyzzy":null}
 ```
 
+A confirmed name links to its Spotify artist page when available, otherwise to a
+Google search for "<name> bandcamp" (which surfaces the artist's Bandcamp page).
+
+## Better matching with a local LLM (optional)
+
+The optional LLM gate is the **matching-quality lever**: it classifies which
+candidate strings are real artists/albums, so confirmed names that aren't on
+Spotify still link (via the "<name> bandcamp" Google fallback). The resolver calls
+ollama over **plain HTTP**, so it stays **zero npm dependencies**.
+
+On a 16GB MacBook (M2):
+
+1. Install ollama — <https://ollama.com/download> or `brew install ollama`.
+2. Pull a small model: `ollama pull qwen2.5:3b` (~2GB; runs comfortably on a 16GB
+   M2; the first resolve call warms the model so it's a little slow once, then
+   fast).
+3. In `service/.env`, point the resolver at ollama (and optionally override the
+   model):
+
+   ```
+   OLLAMA_URL=http://localhost:11434
+   OLLAMA_MODEL=qwen2.5:3b
+   ```
+4. Restart the resolver. Startup reports `(spotify: off, llm: qwen2.5:3b)` (or
+   `(spotify: on, llm: qwen2.5:3b)` if you also set Spotify creds).
+5. Confirm: `curl localhost:8787/health` should show `"llm":"qwen2.5:3b"`.
+
+With both gates, the LLM decides which names are artists and Spotify supplies the
+direct links; with the LLM alone, confirmed names get a Spotify direct link if one
+exists, else the Google fallback.
+
 ## 2. Install the extension in Firefox
 
 The extension is unpacked / unsigned, so load it as a **temporary add-on**:
@@ -104,7 +140,9 @@ with `xpinstall.signatures.required = false`.)
 If nothing gets linked:
 
 1. Confirm the resolver is running — `curl localhost:8787/health` should return
-   `{"ok":true,...}`.
+   `{"ok":true,"spotify":<bool>,"llm":<model|false>,"cached":<n>}`. If both
+   `spotify` and `llm` are off, nothing can link — configure at least one gate
+   above.
 2. On the reddit thread, open the browser console (F12) and look for `[rsl]` logs: they
    report the root element, candidate count, and resolver responses on each scan. A
    `resolve failed` line means the background script can't reach the resolver (is it on
@@ -117,7 +155,7 @@ If nothing gets linked:
 ## Configuration
 
 The resolver URL is the constant `RESOLVER_URL` at the top of
-`extension/content.js` (default `http://localhost:8787`). If you change the host
+`extension/background.js` (default `http://localhost:8787`). If you change the host
 or port, update **both** that constant **and** the matching entry in
 `extension/manifest.json` → `permissions` (Firefox needs the host listed there to
 let the content script reach it).
@@ -128,12 +166,14 @@ let the content script reach it).
   lowercase names right after "check out / by / fan of / recommend …". Names split across
   inline formatting (e.g. **bold** mid-name) can still be missed. The resolver's
   exact-name match keeps false *positives* low; the trade-off is some false *negatives*.
-- **Links are direct when MusicBrainz has the artist's `url-relations`** (official
-  Bandcamp / Spotify pages); otherwise they fall back to a search URL. (Each
-  confirmed artist costs one extra throttled MusicBrainz lookup, cached after.)
+- **Links are direct to Spotify when Spotify has the artist**; otherwise a confirmed
+  name links to a Google search for "<name> bandcamp" (which surfaces the artist's
+  Bandcamp page).
+- **Without Spotify creds or the LLM, nothing links** — at least one gate must be
+  configured.
+- The optional LLM adds a little latency on the **first** call (ollama warms the
+  model), then is fast.
 - The resolver caches results in memory; it resets when you restart it.
-- Without Spotify credentials the MusicBrainz path is rate-limited to ~1 lookup/
-  second, so the first scan of a large thread fills in gradually (cached after).
 - Built and tested as a Firefox MV2 extension.
 
 ## License

extension/content.js

@@ -46,17 +46,21 @@
   }
 
   // ---- link building / rewriting --------------------------------------------
-  // Per the user's "if on spotify, link that, else bandcamp" rule: one link to the
-  // primary platform, colour-coded so you can tell which at a glance.
+  // Per the user's "if on spotify, link that, else a Bandcamp Google-search fallback"
+  // rule: one link to the primary platform, colour-coded so you can tell which at a glance.
   function makeLink(label, links) {
     const primary = links.primary || { platform: "spotify", url: links.spotify };
+    const name = links.name || label;
+    const platform = primary.platform;
     const a = document.createElement("a");
-    a.className = `rsl-link rsl-${primary.platform}`;
+    a.className = `rsl-link rsl-${platform}`;
     a.textContent = label;
     a.href = primary.url;
     a.target = "_blank";
     a.rel = "noopener noreferrer";
-    a.title = `Open on ${primary.platform === "spotify" ? "Spotify" : "Bandcamp"}`;
+    a.title = platform === "spotify"
+      ? `Open ${name} on Spotify`
+      : `Find ${name} on Bandcamp (Google search)`;
     return a;
   }
 
@@ -80,7 +84,7 @@
   }
 
   // ---- resolver client ------------------------------------------------------
-  const resolved = new Map(); // candidate string -> { spotify, bandcamp } | null
+  const resolved = new Map(); // candidate string -> { name, spotify?, google?, primary } | null
 
   // The network call goes through the background script — a page-context fetch to
   // http://localhost is blocked by reddit's CSP / mixed-content rules.

extension/styles.css

@@ -1,5 +1,5 @@
 /* Injected links — kept subtle so reddit comments stay readable. One link per artist,
-   coloured by platform: green = Spotify, teal = Bandcamp. */
+   coloured by platform: green = Spotify, blue = Google (bandcamp-search fallback). */
 .rsl-link {
   text-decoration: underline dotted;
   text-underline-offset: 2px;
@@ -8,8 +8,8 @@
 .rsl-spotify {
   text-decoration-color: #1db954 !important; /* spotify green */
 }
-.rsl-bandcamp {
-  text-decoration-color: #629aa9 !important; /* bandcamp teal */
+.rsl-google {
+  text-decoration-color: #4a72d0 !important; /* google-ish blue */
 }
 .rsl-link:hover {
   text-decoration-style: solid !important;

service/.env.example

@@ -1,13 +1,18 @@
-# Copy to `.env` (it is gitignored). All values are OPTIONAL.
+# Copy to `.env` (it is gitignored).
 #
-# Without Spotify credentials the resolver validates names against MusicBrainz and links
-# to Spotify/Bandcamp SEARCH pages — works with zero setup.
-#
-# With credentials it uses the Spotify Web API for DIRECT artist links.
-# Create a (free) app at https://developer.spotify.com/dashboard to get these.
+# Spotify Web API (client-credentials) — gives DIRECT artist links and, on its own, gates
+# candidates by exact-name match (a name links only when Spotify has that artist). Create a
+# free app at https://developer.spotify.com/dashboard to get these.
 SPOTIFY_CLIENT_ID=
 SPOTIFY_CLIENT_SECRET=
 
-# Port the resolver listens on (default 8787). If you change it, also update
-# RESOLVER_URL in extension/content.js and the host in extension/manifest.json.
+# Optional local LLM (via ollama) for BETTER matching — it decides which candidate substrings
+# are real artists/albums. Set OLLAMA_URL to enable it; still zero npm deps (plain HTTP).
+# Confirmed names get a direct Spotify link when available, else a Google->Bandcamp fallback.
+#   ollama pull qwen2.5:3b   # ~2GB, runs comfortably on a 16GB MacBook M2
+# OLLAMA_URL=http://localhost:11434
+# OLLAMA_MODEL=qwen2.5:3b
+
+# Port the resolver listens on (default 8787). If you change it, also update RESOLVER_URL in
+# extension/background.js and the host in extension/manifest.json permissions.
 PORT=8787

service/lib.js

@@ -4,44 +4,22 @@
 
 export const norm = (s) => String(s).normalize("NFKC").trim().toLowerCase();
 
-export const bandcampSearch = (n) => `https://bandcamp.com/search?q=${encodeURIComponent(n)}&item_type=b`;
-export const spotifySearch = (n) => `https://open.spotify.com/search/${encodeURIComponent(n)}`;
+// Fallback link when we have no direct Spotify page: a Google search biased at Bandcamp,
+// which surfaces the artist's Bandcamp page (when it exists) plus general results.
+export const googleSearch = (n) => `https://www.google.com/search?q=${encodeURIComponent(`${n} bandcamp`)}`;
 
-// The precision lever: return the result whose name matches the candidate
-// case-insensitively, else null. `minScore` (MusicBrainz) additionally requires a
-// confidence score. Tolerates null/garbage items.
-export function pickArtist(name, items, { minScore = null } = {}) {
+// The precision lever: return the item whose name matches the candidate
+// case-insensitively, else null. Tolerates null/garbage items.
+export function pickArtist(name, items) {
   const want = norm(name);
-  return (
-    (items ?? []).find(
-      (a) => a && norm(a.name) === want && (minScore == null || (a.score ?? 0) >= minScore),
-    ) ?? null
-  );
+  return (items ?? []).find((a) => a && norm(a.name) === want) ?? null;
 }
 
 // A "/search" url means we don't have a direct artist page (just a search fallback).
 export const isSearchUrl = (url) => typeof url === "string" && url.includes("/search");
 
-// Choose the single best link to show: prefer Spotify when the artist is actually ON
-// Spotify (a direct, non-search url); else a direct Bandcamp; else the Spotify search.
-export function primaryOf({ spotify, bandcamp }) {
-  if (spotify && !isSearchUrl(spotify)) return { platform: "spotify", url: spotify };
-  if (bandcamp && !isSearchUrl(bandcamp)) return { platform: "bandcamp", url: bandcamp };
-  return { platform: "spotify", url: spotify };
-}
-
-// Link object for a confirmed Spotify hit (direct artist url when present, else search).
-export function spotifyResult(hit) {
-  const r = {
-    name: hit.name,
-    spotify: hit.external_urls?.spotify ?? spotifySearch(hit.name),
-    bandcamp: bandcampSearch(hit.name),
-  };
-  return { ...r, primary: primaryOf(r) };
-}
-
-// True for http/https URLs only — reject javascript:, ftp:, garbage. These values flow
-// into an injected link's href, so validate the third-party MB data defensively.
+// True for http/https URLs only — reject javascript:, ftp:, garbage. The Spotify
+// external_url flows into an injected link's href, so validate it defensively.
 export function isHttpUrl(u) {
   try {
     const p = new URL(u).protocol;
@@ -51,31 +29,14 @@ export function isHttpUrl(u) {
   }
 }
 
-// Extract direct links from a MusicBrainz artist's url-relations (zero-cred: MB stores
-// official Bandcamp / streaming URLs). Accepts only http(s) and matches by parsed hostname
-// (not substring, so "evilbandcamp.com" / "javascript:…bandcamp.com…" don't slip through).
-// Either field may be null.
-export function relUrls(relations) {
-  const urls = (relations ?? []).map((r) => r && r.url && r.url.resource).filter(isHttpUrl);
-  const byHost = (host) =>
-    urls.find((u) => {
-      try {
-        const h = new URL(u).hostname;
-        return h === host || h.endsWith("." + host);
-      } catch {
-        return false;
+// Build the single link to show for a confirmed name. Prefer the direct Spotify artist
+// url (validated http(s), not a /search url); otherwise fall back to a Google→Bandcamp
+// search. `primary` is what the extension renders ({ platform, url }).
+export function linkResult(name, spotifyUrl) {
+  const direct = spotifyUrl && isHttpUrl(spotifyUrl) && !isSearchUrl(spotifyUrl) ? spotifyUrl : null;
+  if (direct) {
+    return { name, spotify: direct, primary: { platform: "spotify", url: direct } };
   }
-    }) ?? null;
-  return { bandcamp: byHost("bandcamp.com"), spotify: byHost("open.spotify.com") };
-}
-
-// Link object for a confirmed MusicBrainz hit: prefer the direct url-rel links, fall back
-// to search urls when MB has no relation for that platform.
-export function mbResult(hit, links) {
-  const r = {
-    name: hit.name,
-    spotify: links?.spotify ?? spotifySearch(hit.name),
-    bandcamp: links?.bandcamp ?? bandcampSearch(hit.name),
-  };
-  return { ...r, primary: primaryOf(r) };
+  const url = googleSearch(name);
+  return { name, google: url, primary: { platform: "google", url } };
 }

service/lib.test.js

@@ -1,7 +1,7 @@
 // Unit tests for the resolver's pure logic (run with `node --test`). Zero-dep.
 import { test } from "node:test";
 import assert from "node:assert/strict";
-import { norm, bandcampSearch, spotifySearch, pickArtist, spotifyResult, relUrls, mbResult, primaryOf, isSearchUrl, isHttpUrl } from "./lib.js";
+import { norm, googleSearch, pickArtist, linkResult, isSearchUrl, isHttpUrl } from "./lib.js";
 
 test("norm lowercases, trims, NFKC-normalizes", () => {
   assert.equal(norm("  Cerebral Bore "), "cerebral bore");
@@ -15,78 +15,44 @@ test("pickArtist matches case-insensitively, rejects non-matches", () => {
   assert.equal(pickArtist("Notarealband", items), null);
 });
 
-test("pickArtist enforces minScore when given (MusicBrainz path)", () => {
-  const items = [{ name: "Devourment", score: 100 }, { name: "Devourment", score: 50 }];
-  assert.equal(pickArtist("Devourment", items, { minScore: 90 })?.score, 100);
-  assert.equal(pickArtist("Devourment", [{ name: "Devourment", score: 50 }], { minScore: 90 }), null);
-});
-
 test("pickArtist tolerates empty / garbage input", () => {
   assert.equal(pickArtist("x", null), null);
   assert.equal(pickArtist("x", []), null);
   assert.equal(pickArtist("x", [null, { name: "y" }]), null);
 });
 
-test("spotifyResult prefers the direct url, falls back to search", () => {
-  const direct = spotifyResult({
-    name: "Devourment",
-    external_urls: { spotify: "https://open.spotify.com/artist/abc" },
-  });
-  assert.equal(direct.spotify, "https://open.spotify.com/artist/abc");
-  assert.match(direct.bandcamp, /^https:\/\/bandcamp\.com\/search\?q=Devourment/);
-
-  assert.deepEqual(direct.primary, { platform: "spotify", url: "https://open.spotify.com/artist/abc" });
-
-  const noUrl = spotifyResult({ name: "Devourment" });
-  assert.equal(noUrl.spotify, "https://open.spotify.com/search/Devourment");
-  assert.equal(noUrl.primary.platform, "spotify"); // search fallback still points at spotify
+test("googleSearch points at a Google query for '<name> bandcamp', url-encoded", () => {
+  const u = googleSearch("Cerebral Bore");
+  assert.match(u, /^https:\/\/www\.google\.com\/search\?q=/);
+  assert.ok(u.includes("Cerebral%20Bore%20bandcamp"));
+  assert.ok(googleSearch("A&B").includes("A%26B"));
 });
 
-test("primaryOf / isSearchUrl: prefer direct spotify, else direct bandcamp, else spotify search", () => {
+test("linkResult prefers a direct Spotify url, primary = spotify", () => {
+  const r = linkResult("Devourment", "https://open.spotify.com/artist/abc");
+  assert.equal(r.spotify, "https://open.spotify.com/artist/abc");
+  assert.deepEqual(r.primary, { platform: "spotify", url: "https://open.spotify.com/artist/abc" });
+  assert.equal(r.google, undefined);
+});
+
+test("linkResult falls back to Google->Bandcamp when there is no direct Spotify url", () => {
+  const r = linkResult("Some Local Band", null);
+  assert.equal(r.spotify, undefined);
+  assert.equal(r.primary.platform, "google");
+  assert.match(r.primary.url, /^https:\/\/www\.google\.com\/search\?q=/);
+  assert.equal(r.google, r.primary.url);
+});
+
+test("linkResult rejects non-http(s) / search Spotify urls and falls back to Google", () => {
+  for (const bad of ["javascript:alert(1)", "ftp://x/y", "https://open.spotify.com/search/X", "not a url"]) {
+    const r = linkResult("X", bad);
+    assert.equal(r.primary.platform, "google", `expected fallback for ${bad}`);
+  }
+});
+
+test("isSearchUrl detects /search urls", () => {
   assert.equal(isSearchUrl("https://open.spotify.com/search/X"), true);
   assert.equal(isSearchUrl("https://open.spotify.com/artist/abc"), false);
-  assert.deepEqual(
-    primaryOf({ spotify: "https://open.spotify.com/artist/abc", bandcamp: "https://bandcamp.com/search?q=X&item_type=b" }),
-    { platform: "spotify", url: "https://open.spotify.com/artist/abc" },
-  );
-  assert.deepEqual(
-    primaryOf({ spotify: "https://open.spotify.com/search/X", bandcamp: "https://x.bandcamp.com/" }),
-    { platform: "bandcamp", url: "https://x.bandcamp.com/" },
-  );
-  assert.deepEqual(
-    primaryOf({ spotify: "https://open.spotify.com/search/X", bandcamp: "https://bandcamp.com/search?q=X&item_type=b" }),
-    { platform: "spotify", url: "https://open.spotify.com/search/X" },
-  );
-});
-
-test("relUrls picks direct bandcamp + spotify from MB relations, tolerates junk", () => {
-  const relations = [
-    { url: { resource: "https://cattledecapitation.bandcamp.com/" } },
-    { url: { resource: "https://open.spotify.com/artist/67ZMMtA88DDO0gTuRrzGjn" } },
-    { url: { resource: "https://en.wikipedia.org/wiki/Cattle_Decapitation" } },
-    null,
-    { url: null },
-  ];
-  const links = relUrls(relations);
-  assert.equal(links.bandcamp, "https://cattledecapitation.bandcamp.com/");
-  assert.equal(links.spotify, "https://open.spotify.com/artist/67ZMMtA88DDO0gTuRrzGjn");
-});
-
-test("relUrls returns nulls when relations absent/empty", () => {
-  assert.deepEqual(relUrls(null), { bandcamp: null, spotify: null });
-  assert.deepEqual(relUrls([]), { bandcamp: null, spotify: null });
-});
-
-test("relUrls rejects non-http(s) and look-alike hosts (hardening)", () => {
-  const relations = [
-    { url: { resource: "javascript:alert('bandcamp.com')" } }, // not http(s)
-    { url: { resource: "ftp://files.bandcamp.com/x" } }, // not http(s)
-    { url: { resource: "https://evilbandcamp.com/" } }, // host doesn't end in .bandcamp.com
-    { url: { resource: "https://x.bandcamp.com/" } }, // valid subdomain
-  ];
-  const links = relUrls(relations);
-  assert.equal(links.bandcamp, "https://x.bandcamp.com/");
-  assert.equal(links.spotify, null);
 });
 
 test("isHttpUrl accepts http/https only", () => {
@@ -97,26 +63,3 @@ test("isHttpUrl accepts http/https only", () => {
   assert.equal(isHttpUrl("not a url"), false);
   assert.equal(isHttpUrl(null), false);
 });
-
-test("mbResult prefers direct links, falls back to search per-platform", () => {
-  const direct = mbResult(
-    { name: "Cattle Decapitation" },
-    { bandcamp: "https://cattledecapitation.bandcamp.com/", spotify: "https://open.spotify.com/artist/x" },
-  );
-  assert.equal(direct.bandcamp, "https://cattledecapitation.bandcamp.com/");
-  assert.equal(direct.spotify, "https://open.spotify.com/artist/x");
-  assert.deepEqual(direct.primary, { platform: "spotify", url: "https://open.spotify.com/artist/x" });
-
-  const fallback = mbResult({ name: "Cerebral Bore" }, null);
-  assert.equal(fallback.spotify, "https://open.spotify.com/search/Cerebral%20Bore");
-  assert.equal(fallback.bandcamp, "https://bandcamp.com/search?q=Cerebral%20Bore&item_type=b");
-
-  const partial = mbResult({ name: "X" }, { bandcamp: null, spotify: "https://open.spotify.com/artist/y" });
-  assert.equal(partial.spotify, "https://open.spotify.com/artist/y");
-  assert.match(partial.bandcamp, /bandcamp\.com\/search/);
-});
-
-test("link builders url-encode", () => {
-  assert.equal(spotifySearch("A B"), "https://open.spotify.com/search/A%20B");
-  assert.ok(bandcampSearch("A&B").includes("A%26B"));
-});

service/llm.js

@@ -0,0 +1,51 @@
+// Pure helpers for the OPTIONAL local-LLM (ollama) classifier. The HTTP call lives in
+// resolver.js; the prompt builder + reply parser live here so they unit-test with
+// `node --test` (no network). The LLM answers one question: "which of these candidate
+// substrings are music artists/albums?" — the precision gate when it is configured.
+
+import { norm } from "./lib.js";
+
+// Small, fast, runs comfortably on a 16GB M2 (3.1B params, Q4 ~2GB). Override with OLLAMA_MODEL.
+export const DEFAULT_MODEL = "qwen2.5:3b";
+
+// Chat messages for an ollama /api/chat call (used with stream:false, format:"json").
+export function buildClassifyMessages(candidates) {
+  const system =
+    "You are a music-name classifier for a tool that links artist/album mentions in Reddit " +
+    "music discussions. You are given a JSON array of candidate text snippets pulled from reddit " +
+    "comments. For EACH candidate decide whether it names a music ARTIST/BAND, a music " +
+    "ALBUM/release, or NEITHER. The snippets come from music threads, so prefer the music reading " +
+    "when a candidate is a plausible band/album even if it is also an ordinary word. Mark ordinary " +
+    "words, sentence fragments, generic phrases, place names, and non-music proper nouns as \"none\". " +
+    'Reply with ONLY a JSON object of the form {"results":[{"name":<candidate verbatim>,' +
+    '"kind":"artist"|"album"|"none"}]}. Use each candidate string VERBATIM as its name.';
+  const user = JSON.stringify({ candidates });
+  return [
+    { role: "system", content: system },
+    { role: "user", content: user },
+  ];
+}
+
+// Parse the model's JSON reply into a Set of norm(name) the model marked artist/album.
+// Defensive: only names that were actually in the candidate list count (guards against the
+// model inventing names); an unparseable reply yields an empty Set, and the caller then
+// falls back to the Spotify-exact-match gate.
+export function parseClassification(text, candidates) {
+  const allowed = new Set(candidates.map(norm));
+  const confirmed = new Set();
+  let obj;
+  try {
+    obj = JSON.parse(text);
+  } catch {
+    return confirmed;
+  }
+  const rows = Array.isArray(obj?.results) ? obj.results : Array.isArray(obj) ? obj : [];
+  for (const row of rows) {
+    if (!row || typeof row.name !== "string") continue;
+    const k = norm(row.name);
+    if (!allowed.has(k)) continue;
+    const kind = String(row.kind ?? "").toLowerCase();
+    if (kind === "artist" || kind === "album") confirmed.add(k);
+  }
+  return confirmed;
+}

service/llm.test.js

@@ -0,0 +1,55 @@
+// Unit tests for the optional LLM classifier's pure logic (run with `node --test`). Zero-dep.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { buildClassifyMessages, parseClassification, DEFAULT_MODEL } from "./llm.js";
+
+test("DEFAULT_MODEL is a small ollama model", () => {
+  assert.equal(DEFAULT_MODEL, "qwen2.5:3b");
+});
+
+test("buildClassifyMessages embeds the candidate list as JSON for the user turn", () => {
+  const msgs = buildClassifyMessages(["Tool", "the best"]);
+  assert.equal(msgs.length, 2);
+  assert.equal(msgs[0].role, "system");
+  assert.match(msgs[0].content, /artist/i);
+  assert.equal(msgs[1].role, "user");
+  assert.deepEqual(JSON.parse(msgs[1].content), { candidates: ["Tool", "the best"] });
+});
+
+test("parseClassification keeps only artist/album candidates that were in the list", () => {
+  const candidates = ["Cattle Decapitation", "Devourment", "the best part"];
+  const reply = JSON.stringify({
+    results: [
+      { name: "Cattle Decapitation", kind: "artist" },
+      { name: "Devourment", kind: "ARTIST" }, // case-insensitive kind
+      { name: "the best part", kind: "none" },
+    ],
+  });
+  const got = parseClassification(reply, candidates);
+  assert.ok(got.has("cattle decapitation"));
+  assert.ok(got.has("devourment"));
+  assert.ok(!got.has("the best part"));
+});
+
+test("parseClassification accepts album kind and a bare array reply", () => {
+  const got = parseClassification(
+    JSON.stringify([{ name: "Reek of Putrefaction", kind: "album" }]),
+    ["Reek of Putrefaction"],
+  );
+  assert.ok(got.has("reek of putrefaction"));
+});
+
+test("parseClassification ignores hallucinated names not in the candidate list", () => {
+  const got = parseClassification(
+    JSON.stringify({ results: [{ name: "Some Band I Invented", kind: "artist" }] }),
+    ["Real Candidate"],
+  );
+  assert.equal(got.size, 0);
+});
+
+test("parseClassification returns an empty set on unparseable / junk replies", () => {
+  assert.equal(parseClassification("not json at all", ["X"]).size, 0);
+  assert.equal(parseClassification("", ["X"]).size, 0);
+  assert.equal(parseClassification(JSON.stringify({ nope: true }), ["X"]).size, 0);
+  assert.equal(parseClassification(JSON.stringify({ results: [null, { kind: "artist" }] }), ["X"]).size, 0);
+});

service/resolver.js

@@ -1,16 +1,20 @@
 // Tiny resolver service: candidate names -> artist links, for CONFIRMED artists only.
 //
-// Zero npm deps; Node 18+ (built-in fetch). The "which words are real artists" problem
-// is delegated to a music API rather than guessed client-side:
-//   * with SPOTIFY_CLIENT_ID/SECRET set -> Spotify Web API (client-credentials) for
-//     DIRECT artist links;
-//   * without creds -> MusicBrainz validation + Spotify/Bandcamp SEARCH links (zero setup).
-// Either way a candidate only resolves when an artist's name matches it case-insensitively
-// (the precision lever). The pure matching/link logic lives in lib.js (unit-tested).
+// Zero npm deps; Node 18+ (built-in fetch). The "which words are real artists" problem is
+// not guessed client-side — it is gated one of two ways:
+//   * SPOTIFY_CLIENT_ID/SECRET set -> a candidate resolves only when Spotify has an artist
+//     whose name matches it case-insensitively (the exact-match precision lever), linking to
+//     the DIRECT Spotify artist page.
+//   * OLLAMA_URL set -> a local LLM (via ollama HTTP, still zero npm deps) classifies which
+//     candidates are artists/albums; confirmed names get a Spotify direct link when available,
+//     else a Google->Bandcamp search fallback. This is the better-matching path.
+// The two combine: with both set, the LLM gates and Spotify supplies direct links.
+// The pure matching/link/prompt logic lives in lib.js + llm.js (unit-tested).
 
 import { createServer } from "node:http";
 import { readFileSync } from "node:fs";
-import { norm, pickArtist, spotifyResult, mbResult, relUrls } from "./lib.js";
+import { norm, pickArtist, linkResult } from "./lib.js";
+import { buildClassifyMessages, parseClassification, DEFAULT_MODEL } from "./llm.js";
 
 // --- minimal .env loader (no dependency) ------------------------------------
 try {
@@ -25,12 +29,16 @@ const PORT = Number(process.env.PORT) || 8787;
 const SPOTIFY_ID = process.env.SPOTIFY_CLIENT_ID;
 const SPOTIFY_SECRET = process.env.SPOTIFY_CLIENT_SECRET;
 const USE_SPOTIFY = Boolean(SPOTIFY_ID && SPOTIFY_SECRET);
-const UA = "reddit-spotify-linker/0.1 (https://git.yapplesauce.com/paul/linker)";
+const OLLAMA_URL = (process.env.OLLAMA_URL || "").replace(/\/$/, ""); // presence enables the LLM gate
+const OLLAMA_MODEL = process.env.OLLAMA_MODEL || DEFAULT_MODEL;
+const USE_LLM = Boolean(OLLAMA_URL);
 const MAX_CANDIDATES = 100;
 
-const cache = new Map(); // norm(name) -> { spotify, bandcamp, name } | null  (negatives cached too)
+// Cache only the network half (Spotify direct-url lookup); the final link object is rebuilt
+// per request so a name's result never depends on a stale gate decision.  norm(name) -> url|null
+const spCache = new Map();
 
-// --- Spotify (optional) -----------------------------------------------------
+// --- Spotify (optional): direct-artist-link lookup --------------------------
 let spToken = null;
 let spTokenExp = 0;
 async function spotifyToken() {
@@ -48,72 +56,85 @@ async function spotifyToken() {
   return spToken;
 }
 
-async function viaSpotify(name) {
+// The direct Spotify artist url for an exact name match, or null (no creds / no exact hit).
+async function spotifyDirect(name) {
+  if (!USE_SPOTIFY) return null;
   const token = await spotifyToken();
   const url = `https://api.spotify.com/v1/search?type=artist&limit=5&q=${encodeURIComponent(name)}`;
   const r = await fetch(url, { headers: { authorization: `Bearer ${token}` } });
   if (!r.ok) throw new Error(`spotify search http ${r.status}`);
   const j = await r.json();
   const hit = pickArtist(name, j.artists?.items);
-  return hit ? spotifyResult(hit) : null;
+  return hit?.external_urls?.spotify ?? null;
 }
 
-// --- MusicBrainz (zero-cred path) -------------------------------------------
-// MusicBrainz asks for <=1 request/second, so every request is serialized + spaced.
-const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
-let mbGate = Promise.resolve();
-let mbLast = 0;
-function mbSchedule() {
-  const gate = mbGate.then(async () => {
-    const wait = 1100 - (Date.now() - mbLast);
-    if (wait > 0) await sleep(wait);
-    mbLast = Date.now();
-  });
-  mbGate = gate.catch(() => {}); // a failure never wedges the queue
-  return gate;
-}
-async function mbFetchJson(url) {
-  await mbSchedule();
-  const r = await fetch(url, { headers: { "user-agent": UA, accept: "application/json" } });
-  if (!r.ok) throw new Error(`musicbrainz http ${r.status}`);
-  return r.json();
-}
-
-async function viaMusicBrainz(name) {
-  const q = encodeURIComponent(`artist:"${name}"`);
-  const search = await mbFetchJson(`https://musicbrainz.org/ws/2/artist?fmt=json&limit=5&query=${q}`);
-  const hit = pickArtist(name, search.artists, { minScore: 90 });
-  if (!hit) return null;
-  // Enrich with MB url-relations -> DIRECT bandcamp/spotify links (best-effort).
-  let links = null;
-  try {
-    const detail = await mbFetchJson(`https://musicbrainz.org/ws/2/artist/${hit.id}?inc=url-rels&fmt=json`);
-    links = relUrls(detail.relations);
-  } catch (e) {
-    console.error(`mb url-rels "${name}": ${e.message}`); // enrichment is best-effort
-  }
-  return mbResult(hit, links);
-}
-
-// --- resolve one candidate (cached) -----------------------------------------
-async function resolveOne(name) {
+async function spotifyDirectCached(name) {
   const key = norm(name);
-  if (!key) return null;
-  if (cache.has(key)) return cache.get(key);
+  if (spCache.has(key)) return spCache.get(key);
+  const url = await spotifyDirect(name);
+  spCache.set(key, url);
+  return url;
+}
+
+// --- LLM gate (optional): which candidates are artists/albums? --------------
+async function ollamaClassify(candidates) {
+  const r = await fetch(`${OLLAMA_URL}/api/chat`, {
+    method: "POST",
+    headers: { "content-type": "application/json" },
+    body: JSON.stringify({
+      model: OLLAMA_MODEL,
+      messages: buildClassifyMessages(candidates),
+      stream: false,
+      format: "json",
+      options: { temperature: 0 },
+    }),
+  });
+  if (!r.ok) throw new Error(`ollama http ${r.status}`);
+  const j = await r.json();
+  return parseClassification(j.message?.content ?? "", candidates);
+}
+
+// --- resolve --------------------------------------------------------------
+// allowFallback: when the gate has already confirmed this name is an artist/album, a missing
+// Spotify direct link still yields a Google->Bandcamp fallback link (instead of null).
+async function resolveOne(name, allowFallback) {
+  if (!norm(name)) return null;
   try {
-    const result = USE_SPOTIFY ? await viaSpotify(name) : await viaMusicBrainz(name);
-    cache.set(key, result); // cache hits AND confirmed misses
-    return result;
+    const sp = await spotifyDirectCached(name);
+    if (sp) return linkResult(name, sp); // direct Spotify page
+    return allowFallback ? linkResult(name, null) : null; // google fallback, or unconfirmed
   } catch (e) {
-    console.error(`resolve "${name}": ${e.message}`); // transient -> don't cache
+    console.error(`resolve "${name}": ${e.message}`); // transient -> caller sees null this time
     return null;
   }
 }
 
 async function resolveAll(candidates) {
   const uniq = [...new Set(candidates.map((c) => String(c).trim()).filter(Boolean))].slice(0, MAX_CANDIDATES);
+
+  // Gate: which names may link? null => no LLM gate, the Spotify exact-match is the gate
+  // (only names actually on Spotify link). A Set => the LLM's verdict; names not in it are
+  // dropped, names in it may use the Google fallback when not on Spotify.
+  let confirmed = null;
+  if (USE_LLM && uniq.length) {
+    try {
+      confirmed = await ollamaClassify(uniq);
+    } catch (e) {
+      console.error(`llm classify: ${e.message}`); // LLM down -> fall back to the Spotify gate
+      confirmed = null;
+    }
+  }
+
   const out = {};
-  await Promise.all(uniq.map(async (name) => { out[name] = await resolveOne(name); }));
+  await Promise.all(
+    uniq.map(async (name) => {
+      if (confirmed && !confirmed.has(norm(name))) {
+        out[name] = null; // LLM said this is not an artist/album
+        return;
+      }
+      out[name] = await resolveOne(name, confirmed != null);
+    }),
+  );
   return out;
 }
 
@@ -142,7 +163,7 @@ const server = createServer(async (req, res) => {
   if (req.method === "OPTIONS") { res.writeHead(204); return res.end(); }
 
   if (req.method === "GET" && req.url === "/health") {
-    return json(res, 200, { ok: true, provider: USE_SPOTIFY ? "spotify" : "musicbrainz", cached: cache.size });
+    return json(res, 200, { ok: true, spotify: USE_SPOTIFY, llm: USE_LLM ? OLLAMA_MODEL : false, cached: spCache.size });
   }
 
   if (req.method === "POST" && req.url === "/resolve") {
@@ -159,5 +180,10 @@ const server = createServer(async (req, res) => {
 });
 
 server.listen(PORT, () => {
-  console.log(`resolver listening on http://localhost:${PORT}  (provider: ${USE_SPOTIFY ? "spotify" : "musicbrainz"})`);
+  const sp = USE_SPOTIFY ? "on" : "off";
+  const llm = USE_LLM ? OLLAMA_MODEL : "off";
+  console.log(`resolver listening on http://localhost:${PORT}  (spotify: ${sp}, llm: ${llm})`);
+  if (!USE_SPOTIFY && !USE_LLM) {
+    console.warn("  ! no gate configured: set SPOTIFY_CLIENT_ID/SECRET and/or OLLAMA_URL, or nothing will link.");
+  }
 });