@browserstack/mcp-server 1.5.1 → 2.0.0-beta.1

package/dist/tools/capability-registry/search.js

@@ -0,0 +1,907 @@
+/**
+ * Ranking capabilities against a plain-language query.
+ *
+ * Ported from the Python `discover._score`, including the two properties that were each
+ * fixed after a live mis-ranking:
+ *
+ *  * PENALTIES REORDER, THEY DO NOT EXCLUDE. `matched` is the pre-penalty term score and is
+ *    what decides inclusion; `ranked` carries the preferences. Conflating them dropped 40
+ *    legitimate matches outright, because a cardinality penalty took an otherwise-valid
+ *    score to zero and the caller saw "no such capability".
+ *  * CARDINALITY. A "list" query answered by a single-record getter sends the caller to a
+ *    capability needing an id it cannot possibly have yet.
+ */
+/**
+ * Every non-alphanumeric character separates, `_` included.
+ *
+ * `_` used to be a word character, which made `test_case` a single token while every
+ * haystack rendered it as "test case" — so the two could never match. That is the exact
+ * string `listProducts` hands back, so a caller following the documented flow searched with
+ * a term guaranteed to score zero: "list test_runs" matched 19 capabilities and put an
+ * admin settings endpoint first, where "list test runs" matched 103 and put the test-runs
+ * listing first.
+ */
+const WORD = /[a-z0-9]+/g;
+const STOPWORDS = new Set([
+    "a",
+    "an",
+    "and",
+    "are",
+    "as",
+    "at",
+    "be",
+    "by",
+    "can",
+    "do",
+    "for",
+    "from",
+    "how",
+    "i",
+    "in",
+    "is",
+    "it",
+    "me",
+    "my",
+    "of",
+    "on",
+    "or",
+    "has",
+    "have",
+    "that",
+    "the",
+    "these",
+    "this",
+    "those",
+    "to",
+    "want",
+    "what",
+    "which",
+    "with",
+    "you",
+]);
+// Verbs that reveal what the caller means to DO. A preference, not a filter — an explicit
+// `mode` argument is the filter.
+const READ_VERBS = new Set([
+    "list",
+    "get",
+    "show",
+    "find",
+    "fetch",
+    "read",
+    "count",
+    "search",
+    "view",
+    "which",
+    "how",
+]);
+const WRITE_VERBS = new Set([
+    "create",
+    "add",
+    "update",
+    "edit",
+    "delete",
+    "remove",
+    "move",
+    "copy",
+    "archive",
+    "assign",
+    "restore",
+    "reorder",
+    "bulk",
+    "set",
+    "upload",
+    "import",
+    "clone",
+]);
+// Words that mean "give me many", which is what makes a single-record getter the wrong answer.
+const PLURAL_INTENT = new Set([
+    "list",
+    "all",
+    "every",
+    "many",
+    "count",
+    "search",
+    "find",
+    "which",
+    "each",
+]);
+/**
+ * Query/haystack terms. Verbs are deliberately NOT stopwords — they carry the intent.
+ *
+ * camelCase is split before lowercasing, so `testRunId`, `test_run_id` and `test run id`
+ * all tokenize alike.
+ */
+/**
+ * Is this the browse wildcard rather than a search?
+ *
+ * Accepts `*` and `all` — both are what a caller reaches for when the intent is "show me
+ * everything", and treating only one of them as the wildcard means the other silently
+ * becomes a bad search. Whitespace is tolerated so a pasted value still works.
+ */
+export function isBrowseQuery(query) {
+    const q = (query || "").trim().toLowerCase();
+    return q === "*" || q === "all";
+}
+export function terms(text) {
+    return [
+        ...(text || "")
+            .replace(/([a-z0-9])([A-Z])/g, "$1 $2")
+            .toLowerCase()
+            .matchAll(WORD),
+    ]
+        .map((match) => match[0])
+        .filter((word) => !STOPWORDS.has(word));
+}
+/**
+ * A term plus its naive singular variants.
+ *
+ * QUERY SIDE ONLY, which is what makes this cheap and safe. Matching is one-directional
+ * substring containment, so indexed `attachments` already contains a query of `attachment`;
+ * only the reverse — a plural query against singular text — needs help. Stemming the
+ * indexed side too would mean rewriting the product's own vocabulary to guess at English,
+ * for no additional match.
+ */
+export function termForms(term) {
+    const forms = [term];
+    // A stripped form must still be three characters. `has` -> `ha` matched more than half
+    // the surface as a substring and pushed a correct answer out of the top 8 entirely;
+    // short fragments are noise, not variants.
+    const add = (form) => {
+        if (form.length >= 3)
+            forms.push(form);
+    };
+    if (term.endsWith("es"))
+        add(term.slice(0, -2));
+    if (term.endsWith("s") && !term.endsWith("ss"))
+        add(term.slice(0, -1));
+    return forms;
+}
+/**
+ * ONE canonical spelling of a word, so two spellings of a noun cannot look like two terms.
+ *
+ * `termForms` is deliberately generous — it offers every candidate and lets substring
+ * containment sort them out. Canonicalising needs the opposite: exactly one answer, and
+ * the right one. Taking the SHORTEST candidate is what a first cut did, and it folded
+ * `cases` to `cas`, so the vocabulary entry `test case` matched no query containing "test
+ * cases" — the phrase it exists for. The ambiguity check then fell through to `test`
+ * alone, reaching the right verdict by the wrong route.
+ *
+ * Strip `es` only after a sibilant, where English actually inserts it (`boxes`,
+ * `batches`, `statuses`). Otherwise strip the single `s`, which is right for the `-e`
+ * plurals this vocabulary is full of: case, suite, phase, template.
+ */
+export function singular(word) {
+    if (word.endsWith("ss"))
+        return word;
+    // histories -> history. Without this the alias `histories` and the entity word
+    // `history` are two unrelated terms, which is the same defect in another spelling.
+    if (word.endsWith("ies") && word.length > 4)
+        return `${word.slice(0, -3)}y`;
+    // Only where the `e` is genuinely inserted. `s` is NOT in this set: `cases` ends in
+    // `ses` and is `case` + `s`, not `cas` + `es`, and no rule can tell it from `statuses`
+    // by suffix alone — so the commoner reading wins and `-s` is stripped below.
+    if (/(?:ch|sh|x|z)es$/.test(word))
+        return word.slice(0, -2);
+    if (word.endsWith("s") && word.length > 3)
+        return word.slice(0, -1);
+    return word;
+}
+/** A haystack as one lowercased, space-separated string, ready for containment tests. */
+function haystack(text) {
+    return terms(text).join(" ");
+}
+export function modeHint(query) {
+    const words = new Set(terms(query));
+    const wantsWrite = [...words].some((word) => WRITE_VERBS.has(word));
+    if (wantsWrite)
+        return "write";
+    const wantsRead = [...words].some((word) => READ_VERBS.has(word));
+    return wantsRead ? "read" : "";
+}
+export function wantsCollection(query) {
+    return [...(query || "").toLowerCase().matchAll(WORD)].some((match) => PLURAL_INTENT.has(match[0]));
+}
+/**
+ * True when a capability answers with many records rather than one.
+ *
+ * Pagination is the reliable signal — a paged operation is a listing by construction. The
+ * plural terminal path segment is a weaker fallback for unpaged collections. (The Python
+ * side used the capability NAME here. tm now publishes one, but its verbs are not
+ * consistent — the endpoint that lists folders is `list_root_folders` — so the path's
+ * terminal noun remains the better signal.)
+ */
+export function isCollection(capability) {
+    if (capability.paginated)
+        return true;
+    const segments = capability.path
+        .split("/")
+        .filter((s) => s && !s.startsWith("{"));
+    const tail = segments[segments.length - 1] || "";
+    return tail.endsWith("s") && !tail.endsWith("ss");
+}
+/**
+ * Does this capability's mode answer what the caller asked to DO?
+ *
+ * `destructive` counts as a write. It has to: `modeHint` reads "delete" and "remove" as
+ * WRITE verbs, but a delete endpoint's mode is `destructive`, so a strict equality check
+ * penalised every destructive capability by 20 FOR BEING A DELETE — on exactly the queries
+ * that wanted one. "bulk delete test cases" scored 15.8 on terms, the highest of any missed
+ * query in the eval, and still fell out of the top 8.
+ *
+ * That accounted for 15 of the eval's 36 misses. The mode hint exists to separate reading
+ * from changing; destructive is changing.
+ *
+ * A read hint still refuses destructive, which is the useful half: "show me the test plans"
+ * must not surface a delete.
+ */
+function satisfiesHint(mode, hint) {
+    if (hint === "write")
+        return mode === "write" || mode === "destructive";
+    return mode === hint;
+}
+/** Everything a caller might say that lives on a parameter rather than in the prose. */
+function parameterText(capability) {
+    const parts = [];
+    for (const group of [
+        capability.path_params,
+        capability.query,
+        capability.body,
+    ]) {
+        for (const param of group || []) {
+            parts.push(param.name);
+            if (param.description)
+                parts.push(param.description);
+            if (param.values)
+                parts.push(param.values.map(String).join(" "));
+        }
+    }
+    return parts.join(" ");
+}
+/**
+ * Path words are the identity haystack. The published `name` is deliberately NOT scored.
+ *
+ * That is a measured result, not an oversight. tm now names every capability, and adding
+ * the name to the ranking was tried three ways against tests/fixtures/search-eval.json,
+ * which the pre-names index serves 18/18:
+ *
+ *   folded into this field, weight 6          17/18
+ *   its own field, weights 1 / 2 / 3 / 4 / 6  17 / 17 / 17 / 17 / 16
+ *   only the words the path lacks, 1..6       17 / 16 / 15 / 15 / 14
+ *
+ * Every variant loses, for two reasons. The nouns in a name are the route restated — the
+ * name is snake-cased from the operationId, itself derived from the path — so scoring them
+ * again rewards verbose names for repeating themselves: `list_test_run_test_cases_by_integer_id`
+ * displaced `create_test_result_for_test_case` on "record a pass or fail for a test case in
+ * a run", putting a read above the write that answers it. And what a name adds beyond the
+ * route is mostly its verb, which tm applies inconsistently (`list_root_folders` lists,
+ * `list_folder_test_cases` also lists), so the verb is noise as often as signal.
+ *
+ * The query this was meant to fix, "list all projects", only went 7 -> 4 even where it
+ * helped: `projects` is in 156 of 173 paths as a scope prefix, so rarity correctly values it
+ * near zero and no amount of name weighting recovers it. That one is fixed instead by the
+ * terminal-segment bonus in `score`, which tells "is that thing" from "is scoped by it".
+ *
+ * Revisit when a product ships a consistent verb convention — then the verb becomes signal.
+ */
+/**
+ * The terminal path segment — the thing this endpoint is actually ABOUT.
+ *
+ * A REST path mixes two different things: the resources it is SCOPED BY, and the resource it
+ * ADDRESSES. Only the last segment is the latter. See the bonus in `score` for why that
+ * distinction matters and why it is applied flat rather than weighted.
+ *
+ * Trailing placeholders are skipped, so `/test-cases/{id}` is still about test cases.
+ * Action tails (`close`, `edit`, `delete`) are kept rather than skipped: for "close a test
+ * run" the tail IS the most specific thing the caller said.
+ */
+function resourceText(capability) {
+    const segments = capability.path
+        .split("/")
+        .filter((s) => s && !s.startsWith("{") && s !== "api");
+    return (segments[segments.length - 1] || "").replace(/[-_]/g, " ");
+}
+function identityText(capability) {
+    return capability.path
+        .split("/")
+        .filter((segment) => segment && !segment.startsWith("{") && segment !== "api")
+        .join(" ")
+        .replace(/[-_]/g, " ");
+}
+/**
+ * How much one term is worth, by how rare it is.
+ *
+ * Containment made every project-scoped endpoint match the term `project` — ~150 of tm's
+ * 173 capabilities — so that word carried as much weight as `access`, which appears in
+ * exactly one. Rarity is what separates them: a term matching everything scores near zero,
+ * a term matching one capability scores near one.
+ *
+ * This is the IDF idea alone, not BM25. The term-frequency saturation and length
+ * normalisation BM25 adds would rescale every score, and the mode and cardinality
+ * adjustments below are absolute constants fitted against live mis-rankings. Bounding the
+ * factor to 0..1 keeps those constants meaningful.
+ */
+function rarity(documents, forms) {
+    let df = 0;
+    for (const text of documents) {
+        if (forms.some((form) => text.includes(form)))
+            df += 1;
+    }
+    const total = documents.length || 1;
+    return Math.log((total + 1) / (df + 1)) / Math.log(total + 1);
+}
+/**
+ * Sized to sit alongside the mode (+6) and cardinality (+8) constants, not to dwarf them.
+ * The pinned eval is unchanged at every value from 2 to 14 — the bonus only ever fires on
+ * queries it does not cover — so this was chosen on the wider sweep: at 10, "close a test
+ * run" starts pulling `close_exploratory_session` into second place on the tail match alone.
+ */
+const RESOURCE_BONUS = 6;
+/**
+ * How much a cardinality guess is worth. It used to be 8, and it should not have been.
+ *
+ * `isCollection` is a GUESS read off the URL, and the index carries nothing better: of 88
+ * read capabilities, zero declare an array in their response schema, and `paginated` — the
+ * one signal that looks trustworthy — is true for `get_report_detail`, a single-record
+ * endpoint. There is no derived cardinality in this artifact to appeal to.
+ *
+ * At ±8 that guess swung 16 points and buried 11 correct answers: `/test-runs/closed` and
+ * `/{entity}/search` are listings whose last path segment is not a plural noun, `users-v2`
+ * is plural with a version suffix in the way, and the count endpoints answer "how many"
+ * with a scalar. All were penalised for what they are called.
+ *
+ * The fix is not a better classifier — a qualifier word-list would encode tm's route
+ * conventions into generic code and rot on the next product. It is to stop betting so much
+ * on an unreliable signal. Swept against the eval, holding top1 at 138 with no ceiling
+ * violations: ±7 retires 1, ±6 retires 2, ±5 retires 5, ±4 retires 6. Below that it starts
+ * costing more than it returns — ±3.5 retires 8 but breaks a case, ±3 breaks two.
+ */
+const CARDINALITY = 4;
+/**
+ * The mode penalty stays at 20, unlike the cardinality one — it is earned.
+ *
+ * Where `isCollection` is a guess off the URL, `modeHint` is measured right: across the
+ * eval it agrees with the correct answer's mode 125 times and disagrees 3. A signal that
+ * accurate deserves to be decisive.
+ *
+ * Sweeping it 20 -> 6 retires none of the three misses filed against it and regresses
+ * nothing, which says the penalty is not what holds them back. It isn't: they sit at ranks
+ * 41, 62 and 78 of the matched set, far below anything a constant could lift. All three
+ * are vocabulary gaps wearing a mode-penalty label — "what gets removed" wants a path
+ * spelled `rm-summary`, "get rid of" wants `delete`, and "the option set" wants what the
+ * product calls a `dataset`. The target barely matches on TERMS; the hint is incidental.
+ *
+ * Left at 20 deliberately. There was no evidence for moving it, and an unjustified constant
+ * is how this scorer got into trouble in the first place.
+ */
+const MODE_PENALTY = 20;
+function score(capability, wanted, weights, aliases, hint, plural) {
+    if (wanted.length === 0)
+        return { matched: 1, ranked: 1 };
+    const haystacks = [
+        [identityText(capability), IDENTITY_WEIGHT],
+        [capability.entity, 4],
+        [(aliases[capability.entity] || []).join(" "), 4],
+        [capability.intent || "", 2],
+        // `returns` is scored BELOW identity, not gated on it. At parity with intent it put a
+        // projects listing at #2 for "list test cases in a project" (its returns carries
+        // `test_cases_count`); gating it on an identity match instead made a field reachable
+        // only through returns unreachable, which is worse.
+        [(capability.returns || []).join(" "), 1],
+        [(capability.guidance || []).join(" "), 1],
+        // Parameter names, their descriptions, and their enum values — 330 descriptions and 34
+        // value lists that the artifact already carries and nothing was reading. The vocabulary
+        // a caller uses is often the value they mean to send: `pass` and `fail` appear nowhere
+        // else in the index, only as the `status` enum on the test-result writes.
+        [parameterText(capability), 1],
+    ];
+    // CONTAINMENT, not set membership. A query of `attachment` has to reach an endpoint whose
+    // path says `attachments`; under exact token equality it did not, and that endpoint fell
+    // out of the results entirely. A term scores its field once however many forms match.
+    let ranked = 0;
+    for (const [text, weight] of haystacks) {
+        const blob = haystack(text);
+        if (!blob)
+            continue;
+        for (let i = 0; i < wanted.length; i += 1) {
+            if (wanted[i].some((form) => blob.includes(form)))
+                ranked += weight * weights[i];
+        }
+    }
+    // PHRASE. Adjacent query terms occurring together say more than the same two words
+    // scattered: "test case" is one noun in this vocabulary, "test" and "case" separately
+    // are two of the commonest words in the index. Scored at half the field's weight and
+    // still scaled by rarity, so it sharpens an existing match rather than creating one.
+    for (const [text, weight] of haystacks) {
+        const blob = haystack(text);
+        if (!blob)
+            continue;
+        for (let i = 0; i + 1 < wanted.length; i += 1) {
+            if (blob.includes(`${wanted[i][0]} ${wanted[i + 1][0]}`)) {
+                ranked += weight * 0.5 * (weights[i] + weights[i + 1]);
+            }
+        }
+    }
+    const matched = ranked;
+    // THE ENDPOINT IS THAT THING, not merely scoped by it.
+    //
+    // A REST path mixes the resources it is SCOPED BY with the one it ADDRESSES. `projects` is
+    // in 156 of tm's 173 paths but is the terminal segment in 3, so whole-corpus rarity —
+    // correctly — values it near nothing, and every project-scoped listing scored the same as
+    // the projects listing itself. The top 8 for "list all projects" spanned 15.9 to 14.8,
+    // where +8 collection and +6 mode already account for 14: the term signal was ~1 point of
+    // noise and the right answer sat 7th.
+    //
+    // Flat, and deliberately NOT rarity-scaled. Rarity would reintroduce the same problem in
+    // reverse — a rare scope noun outranking the real target, which is exactly how a weighted
+    // version of this put `/projects/{id}/folders` above `/folder/{id}/test-cases` for "tc list
+    // for a folder". This asks one yes/no question instead: is the caller's own word the last
+    // thing in the path?
+    //
+    // Equality is against the QUERY's forms, never the haystack's — the same one-directional
+    // rule as containment. `projects` is in forms("projects"), so the projects listing hits;
+    // `folders` is not in forms("folder"), so a folder-scoped query does not drag in the
+    // folders listing.
+    const tail = terms(resourceText(capability));
+    if (tail.length &&
+        wanted.some((forms) => tail.every((word) => forms.includes(word)))) {
+        ranked += RESOURCE_BONUS;
+    }
+    if (hint)
+        ranked += satisfiesHint(capability.mode, hint) ? 6 : -MODE_PENALTY;
+    if (plural)
+        ranked += isCollection(capability) ? CARDINALITY : -CARDINALITY;
+    return { matched, ranked };
+}
+/**
+ * How much of a perfect match the best hit actually achieved, below which we doubt it.
+ *
+ * A RATIO, because the raw score is corpus-relative and an absolute threshold therefore
+ * cannot travel. `rarity` is measured over whatever is being searched, so the same query
+ * scores differently depending on scope: "list my load tests" tops out at 1.9 against Load
+ * Testing alone and 10.1 against both products. The old absolute cut of 3 called the first
+ * of those weak and the second strong — the same query, the same right answer, opposite
+ * verdicts, with the smaller corpus penalised precisely because it is smaller.
+ *
+ * Both halves of this ratio move with the corpus, so it does not:
+ *
+ *     coverage = top_matched / (identity weight x Σ rarity of the query's terms)
+ *
+ * — "what fraction of a perfect identity match did the best hit manage". Above 1 is
+ * ordinary, since a good hit matches several fields, not just the route.
+ *
+ * Measured across tm-only, LT-only and both: genuine misses land at 0.00–0.18, queries
+ * that work at 0.44–1.94. The absolute score cannot separate those (misses reach 2.9,
+ * good queries drop to 1.9); this does, with the gap in the same place in all three
+ * corpora, which is the property that was missing.
+ *
+ * RECALIBRATED AT v1.14, and the reason matters more than the number. The ratio is
+ * corpus-relative and still is — "make a new bucket for my tests" scores 0.259 against tm
+ * alone and 0.258 against both products, which is the property this was built for. What
+ * moved is neither corpus size nor query: it is how much TEXT each capability carries.
+ * v1.13 gave all 242 entries guidance and tripled the length of `intent`, so the numerator
+ * — the best hit's term score — roughly doubled, while the denominator depends only on the
+ * rarity of the query's own terms and did not. Both bands shifted up together: genuine
+ * misses now land at 0.00–0.46 and working queries at 0.24–4.16.
+ *
+ * At the old 0.25 the hand-off had silently stopped firing — "bucket", which is nobody's
+ * word for a folder, came in at 0.259 and read as a confident answer. Re-measured over 173
+ * served eval cases against five queries the product has no words for, 0.50 catches 5 of 5
+ * while flagging 3 of 173 served queries (1.7%), which is the same trade the original cut
+ * made. Expect to move it again the next time the artifact's text volume changes; the
+ * threshold tracks prose per capability, not the number of capabilities.
+ */
+const WEAK_COVERAGE = 0.5;
+/** The heaviest field, and so the yardstick a perfect match is measured against. */
+const IDENTITY_WEIGHT = 6;
+// THE CROSS-PRODUCT FLOOR WAS HERE, and is gone with the thing that needed it.
+//
+// It reserved two slots per matching product so a small product could not be shut out by
+// a large one — tm has 173 capabilities to Load Testing's 20, and on a word both share
+// ("test", "config", "tag") tm took the whole page. "add a tag to xyz test" returned
+// tm tm LT tm tm, and an agent reading the first row went to the wrong product.
+//
+// `product` is now REQUIRED on searchCapability, so a search never spans products and the
+// floor could never fire. The bug it fixed is fixed harder: the agent chooses the product
+// deliberately against listProducts, or asks the user, instead of inferring it from which
+// row happened to rank first. Keeping the code would mean carrying a branch nothing
+// reaches and tests asserting behaviour the surface cannot produce.
+//
+// If searching across products ever returns — a "search everywhere" mode, or a caller
+// that legitimately cannot know the product — this comes back with it. It is in the
+// history at 617b4b9, and the reasoning above is why it was right.
+/** Which products a query could plausibly be about, when more than one could. */
+/** Words too common in this domain to settle a product on their own. */
+const GENERIC_ANCHORS = new Set(["test", "tests", "testing"]);
+/**
+ * Products whose OWN vocabulary the query hits, when no word in the query settles it.
+ *
+ * Requiring `product` made every CALL unambiguous and did nothing about the agent making
+ * two of them: "list all projects" was answered by searching tm, then Load Testing, then
+ * merging — a question about which product the user meant, answered by guessing both.
+ * Tool-description prose asking the agent to check with the user is a suggestion it is
+ * free to decline, and declining is cheaper than interrupting someone.
+ *
+ * WHOLE VOCABULARY ENTRIES, never their constituent words. Splitting them makes almost
+ * everything look shared: Load Testing's "load test" and tm's "test case" both yield
+ * `test`, which flagged 135 of the 198 eval queries. Matching entries whole flags 27.
+ *
+ * A UNIQUE TERM SETTLES IT. "create a test case in a folder" contains `folder`, which
+ * both products claim, and `test case`, which only tm does — so the query has already
+ * answered the question and there is nothing to ask. Only queries whose every recognised
+ * word is shared are genuinely undecided.
+ */
+export function ambiguousProducts(products, query) {
+    const none = { products: [], terms: [] };
+    if (terms(query).length === 0)
+        return none;
+    /**
+     * One spelling per word, so a plural cannot masquerade as a different term.
+     *
+     * Both halves need it. On the QUERY side, the entry is `project` and the user types
+     * "list all projects" — the reported case, which exact containment missed. On the
+     * VOCABULARY side the same fold is what keeps the answer honest: tm lists both `report`
+     * and `reports` as aliases, and treating them as two entries made `reports` look
+     * tm-exclusive, so "show me the report" read as settled when both products claim it.
+     */
+    const fold = (text) => terms(text).map(singular).join(" ");
+    const asked_ = ` ${fold(query || "")} `;
+    const owners = new Map();
+    for (const [product, entries] of Object.entries(vocabularyOf(products))) {
+        for (const entry of entries) {
+            for (const word of [entry.entity, ...(entry.aliases ?? [])]) {
+                const key = fold(String(word));
+                if (!key)
+                    continue;
+                if (!owners.has(key))
+                    owners.set(key, new Set());
+                owners.get(key).add(product);
+            }
+        }
+    }
+    // WORDS A PRODUCT OWNS WITHOUT MODELLING THEM. `routing_terms` is how a product claims
+    // what its users call it, as opposed to what it contains, and unlike a summary it MAY
+    // introduce a term: that is the whole point of it, since the words most in need of
+    // routing — `flaky` — are the ones no index names an entity after.
+    for (const [product, index] of Object.entries(products)) {
+        for (const word of index.routing_terms ?? []) {
+            const key = fold(String(word));
+            if (!key)
+                continue;
+            if (!owners.has(key))
+                owners.set(key, new Set());
+            owners.get(key).add(product);
+        }
+    }
+    // THE SUMMARY CAN ADD A CLAIMANT, NEVER A TERM.
+    //
+    // A product that models `view` as an entity and one that only describes views in its
+    // summary both answer questions about views, but only the first was visible here, so
+    // "delete the saved view" read as settled. Letting the summary claim a word the other
+    // product already models catches that, and refusing to mine NEW terms out of it is what
+    // keeps the prose from taking over: a summary is a paragraph of verbs and nouns written
+    // to sell the product, and every word in it would otherwise become routing vocabulary.
+    const summaries = Object.fromEntries(Object.entries(products).map(([product, index]) => [
+        product,
+        ` ${fold(index.summary ?? "")} `,
+    ]));
+    for (const [key, claimants] of owners) {
+        for (const [product, text] of Object.entries(summaries)) {
+            if (text.includes(` ${key} `))
+                claimants.add(product);
+        }
+    }
+    const shared = new Map();
+    const decided = new Set();
+    const decidedBy = new Map();
+    for (const [key, claimants] of owners) {
+        if (!asked_.includes(` ${key} `))
+            continue;
+        // A word this domain uses in every other sentence settles nothing and clashes over
+        // nothing: `test` is an alias on one side and appears throughout both summaries, so
+        // counting it would make almost any request look like a clash.
+        if (GENERIC_ANCHORS.has(key))
+            continue;
+        if (claimants.size < 2) {
+            // A term only one product claims POINTS somewhere; whether it decides the question
+            // depends on what else is in the sentence. It used to end the check outright, and
+            // that is what made this gate almost silent: an agent's paraphrase nearly always
+            // carries some incidental one-product word — `upload`, `saved report` — so the
+            // shared noun the user actually said stopped being examined. Worse, a sentence
+            // carrying one-product words for BOTH products read as settled, which is the case
+            // most in need of asking.
+            if (!GENERIC_ANCHORS.has(key)) {
+                const owner = [...claimants][0];
+                decided.add(owner);
+                decidedBy.set(owner, [...(decidedBy.get(owner) ?? []), key]);
+            }
+            continue;
+        }
+        shared.set(key, claimants);
+    }
+    // NOTHING RECOGNISED IS THE MOST AMBIGUOUS CASE, NOT THE LEAST.
+    //
+    // "Can we ship this?", "What should we fix first?", "Show me the failed tests." — no
+    // word in them is any product's, so the shared-term check finds nothing and used to
+    // report the same calm verdict it gives a request that genuinely settles itself. These
+    // are the requests with the least to go on, and answering them for a product nobody
+    // chose is exactly the guess this whole check exists to stop.
+    if (shared.size === 0 && decided.size === 0)
+        return { ...none, unknown: true };
+    // Decided, with no shared word anywhere in it — "bulk delete test cases". There is
+    // nothing to ask about, and saying WHICH product it landed on matters as much here as
+    // it does below: a caller that only hears "no clash" can still search the other one.
+    if (shared.size === 0)
+        return decided.size === 1
+            ? {
+                ...none,
+                settled: [...decided][0],
+                because: (decidedBy.get([...decided][0]) ?? []).sort(),
+            }
+            : none;
+    // One product, and only one, is pointed at by the unshared words: the sentence answers
+    // itself and there is nothing to ask. It is still worth SAYING which one, because the
+    // caller sees only that nothing needs asking — "tell me about this project, including
+    // its settings" settles on the product that models settings, and an answer of silence
+    // let the other one be searched instead, which is the same wrong turn by a quieter route.
+    if (decided.size === 1)
+        return {
+            ...none,
+            settled: [...decided][0],
+            because: (decidedBy.get([...decided][0]) ?? []).sort(),
+        };
+    const claimed = new Set();
+    for (const claimants of shared.values())
+        for (const product of claimants)
+            claimed.add(product);
+    return { products: [...claimed].sort(), terms: [...shared.keys()].sort() };
+}
+/**
+ * The product's vocabulary, for a caller whose words are not the product's words.
+ *
+ * THE ONE FAILURE LEXICAL SEARCH CANNOT FIX. "make a new bucket for my tests" wants the
+ * folder-create capability, and `bucket` appears nowhere in the index — no scoring change
+ * reaches it, because the word is simply absent. What CAN reach it is the caller: it is a
+ * language model, and given tm's entity list it maps bucket -> folder without effort. It
+ * just cannot guess the list unprompted.
+ *
+ * Aliases and a one-line `description`; `title` stays dropped as a near-duplicate of
+ * `entity` ("Test run" next to `test_run` buys nothing). Aliases alone route but do not
+ * DEFINE — `result -> outcome, execution-result, test-result` never says whether that is
+ * the per-case verdict inside a run or a run-level rollup — so an agent holding names and
+ * aliases has exactly one way to find out, which is describeEntity once per entity. For
+ * tm that is 19 calls at ~1.4KB each, ~27KB, to answer what ~1.6KB of description answers
+ * here for every entity at once. The `key_facts` remain out: ten times the size, and the
+ * caller can ask describeEntity for the one entity it settles on.
+ *
+ * Its share of the response grew when search became a shortlist: 3.2KB against a 38KB full
+ * search was 8%, against a 6.8KB shortlist it is nearly half. The absolute cost did not
+ * change and is still under a thousand tokens — far less than one wrong invoke — so this
+ * is budgeted in bytes rather than as a fraction of a baseline that now moves.
+ */
+export function vocabularyOf(products, only) {
+    const out = {};
+    for (const [name, bundle] of Object.entries(products)) {
+        if (only && name !== only)
+            continue;
+        const entries = [];
+        for (const [entity, doc] of Object.entries(bundle.entities || {})) {
+            const aliases = (doc.aliases || []);
+            const description = doc.description;
+            entries.push({
+                entity,
+                // Before the aliases, because it is what the reader needs first: what the thing
+                // is, then what else it is called.
+                ...(description ? { description } : {}),
+                ...(aliases.length ? { aliases } : {}),
+            });
+        }
+        if (entries.length)
+            out[name] = entries;
+    }
+    return out;
+}
+export function searchCapabilities(products, query, options = {}) {
+    const limit = options.limit && options.limit > 0 ? options.limit : 8;
+    const offset = options.offset && options.offset > 0 ? Math.floor(options.offset) : 0;
+    // BROWSE MODE: `*` lists the product instead of searching it.
+    //
+    // Scoring keeps only capabilities with matched > 0, so there was no query that returned
+    // everything — an agent that wanted to see what exists had to guess words until the
+    // shortlist looked complete, and could never tell when it was. `*` skips scoring
+    // entirely and pages the filtered set in a STABLE order (name, then path), because a
+    // ranked order is not a safe thing to paginate: ranking is query-dependent and ties
+    // break arbitrarily, so page 2 of a ranked list can repeat or skip rows. Browse is
+    // ordered by identity, so offset means the same thing on every call.
+    if (isBrowseQuery(query)) {
+        const all = Object.entries(products)
+            .filter(([name]) => !options.product || name === options.product)
+            .flatMap(([name, bundle]) => bundle.capabilities
+            .filter((capability) => (!options.entity || capability.entity === options.entity) &&
+            (!options.mode || capability.mode === options.mode))
+            .map((capability) => ({ product: name, capability })))
+            .sort((a, b) => (a.capability.name || "").localeCompare(b.capability.name || "") ||
+            a.capability.path.localeCompare(b.capability.path));
+        const page = all.slice(offset, offset + limit);
+        const end = offset + page.length;
+        return {
+            hits: page,
+            truncated: end < all.length,
+            total_matched: all.length,
+            top_matched: 0,
+            coverage: 1,
+            weak: false,
+            offset,
+            ...(end < all.length ? { next_offset: end } : {}),
+            browse: true,
+        };
+    }
+    // Forms are computed once per query, not per capability: 173 capabilities x 6 haystacks
+    // would otherwise rebuild the same handful of strings a thousand times.
+    const wanted = terms(query).map(termForms);
+    const hint = options.mode ? "" : modeHint(query);
+    const plural = wantsCollection(query);
+    const scored = [];
+    const searched = Object.entries(products).filter(([name]) => !options.product || name === options.product);
+    const aliasesByProduct = {};
+    for (const [name, bundle] of searched) {
+        const aliases = {};
+        for (const [entity, doc] of Object.entries(bundle.entities)) {
+            aliases[entity] = (doc.aliases || []);
+        }
+        aliasesByProduct[name] = aliases;
+    }
+    // ONE CORPUS ACROSS EVERY SEARCHED PRODUCT, not one per product.
+    //
+    // Measured per product, a term's rarity inverts across them: `load` appears in nearly
+    // every Load Testing capability, so it scored as noise there, while it appears in 16 of
+    // tm's 173 (`upload`, `download`, reached by containment), so it scored as gold there.
+    // The word that identifies a product was worth least inside it, and "list load tests"
+    // returned five tm results and no Load Testing ones at all.
+    //
+    // It is also measured against EVERY capability, not the entity- or mode-filtered subset:
+    // narrowing a search must not make a common word look rare.
+    const corpus = searched.flatMap(([name, bundle]) => bundle.capabilities.map((capability) => [
+        identityText(capability),
+        capability.entity,
+        (aliasesByProduct[name][capability.entity] || []).join(" "),
+        capability.intent || "",
+        (capability.returns || []).join(" "),
+        parameterText(capability),
+    ]
+        .map(haystack)
+        .join(" ")));
+    const weights = wanted.map((forms) => rarity(corpus, forms));
+    for (const [name, bundle] of searched) {
+        const aliases = aliasesByProduct[name];
+        for (const capability of bundle.capabilities) {
+            if (options.entity && capability.entity !== options.entity)
+                continue;
+            if (options.mode && capability.mode !== options.mode)
+                continue;
+            const { matched, ranked } = score(capability, wanted, weights, aliases, hint, plural);
+            if (matched > 0)
+                scored.push({ matched, ranked, product: name, capability });
+        }
+    }
+    scored.sort((a, b) => b.ranked - a.ranked || a.capability.path.localeCompare(b.capability.path));
+    // Ranked order decides the page; the best TERM score decides confidence. They are
+    // different questions: the mode and cardinality constants can lift a weakly-matched
+    // capability to the top of a page that is entirely wrong.
+    const topMatched = scored.reduce((best, s) => Math.max(best, s.matched), 0);
+    // What a perfect identity match would have scored for THIS query in THIS corpus. Both
+    // this and `topMatched` scale with the corpus, so their ratio is comparable across
+    // products and across `product`/`entity` narrowing — which the raw score is not.
+    const perfect = weights.reduce((sum, w) => sum + IDENTITY_WEIGHT * w, 0);
+    const coverage = perfect > 0 ? topMatched / perfect : 1;
+    const page = scored.slice(offset, offset + limit);
+    const end = offset + page.length;
+    return {
+        hits: page.map(({ product, capability }) => ({ product, capability })),
+        truncated: end < scored.length,
+        total_matched: scored.length,
+        offset,
+        ...(end < scored.length ? { next_offset: end } : {}),
+        top_matched: topMatched,
+        coverage,
+        weak: wanted.length > 0 && perfect > 0 && coverage < WEAK_COVERAGE,
+    };
+}
+/**
+ * Which products the USER'S OWN WORDS name, directly or through vocabulary only one
+ * product claims.
+ *
+ * This is the other half of `ambiguousProducts`. That one asks "is the query a word both
+ * products use?"; this asks "did the user say anything that settles it?" — and it reads
+ * the user's text, not a flag the caller sets about itself. A boolean saying the user
+ * confirmed is answered by whoever wants to proceed; a quote has to come from the
+ * conversation, and the words in it are checked against the index here.
+ *
+ * Anchors are of two kinds and both are needed. A product NAME ("test management",
+ * "observability") is the explicit case. A word only one product's vocabulary claims
+ * ("quality gate", "test plan") is the implicit one, and it is the common one: users say
+ * what they want, not which product owns it.
+ */
+export function productAnchors(products, text) {
+    const none = { products: [], terms: [] };
+    if (terms(text).length === 0)
+        return none;
+    const fold = (s) => terms(s).map(singular).join(" ");
+    const said = ` ${fold(text || "")} `;
+    const found = new Map();
+    const claim = (product, term) => {
+        if (!found.has(product))
+            found.set(product, new Set());
+        found.get(product).add(term);
+    };
+    // The product's own name and key, as a user would type them. `summary` is prose, not a
+    // name, so only the leading clause before ':' is used — "BrowserStack Test Management",
+    // "BrowserStack Test Reporting & Analytics (TRA, aka Test Observability)".
+    for (const [product, index] of Object.entries(products)) {
+        const names = [
+            product,
+            ...(index.summary ?? "").split(":")[0].split(/[(),]/),
+        ];
+        for (const name of names) {
+            // "aka Test Observability" is the product's other name with a connective in front
+            // of it; the user types the name, not the connective.
+            const key = fold(name.replace(/^\s*(browserstack|aka|formerly)\s+/i, ""));
+            if (key.length < 2)
+                continue;
+            if (said.includes(` ${key} `))
+                claim(product, key);
+        }
+    }
+    // Vocabulary only one product claims. Shared words settle nothing and are skipped —
+    // that is exactly what `ambiguousProducts` flagged in the first place.
+    const owners = new Map();
+    for (const [product, entries] of Object.entries(vocabularyOf(products))) {
+        for (const entry of entries) {
+            for (const word of [entry.entity, ...(entry.aliases ?? [])]) {
+                const key = fold(String(word));
+                if (!key)
+                    continue;
+                if (!owners.has(key))
+                    owners.set(key, new Set());
+                owners.get(key).add(product);
+            }
+        }
+    }
+    // The same words `ambiguousProducts` routes on, so a term that settles the question
+    // there also anchors the product here — otherwise `flaky` names a product in one half
+    // of the check and nothing at all in the other.
+    for (const [product, index] of Object.entries(products)) {
+        for (const word of index.routing_terms ?? []) {
+            const key = fold(String(word));
+            if (!key)
+                continue;
+            if (!owners.has(key))
+                owners.set(key, new Set());
+            owners.get(key).add(product);
+        }
+    }
+    const matched = [];
+    for (const [key, claimants] of owners) {
+        if (claimants.size !== 1)
+            continue;
+        if (!said.includes(` ${key} `))
+            continue;
+        matched.push({ key, product: [...claimants][0] });
+    }
+    for (const { key, product } of matched) {
+        // A word carried by nearly every request in this domain anchors nothing: tra's
+        // sp_test lists the bare word `test` as an alias, which would make "what test plans
+        // are active" read as tra on the strength of one noun that means nothing on its own.
+        if (GENERIC_ANCHORS.has(key))
+            continue;
+        // Longest match wins. "test management" and "test" both match the same sentence, and
+        // only the longer one carries the user's meaning — counting both makes a settled
+        // question look ambiguous again.
+        if (matched.some((other) => other.key !== key && ` ${other.key} `.includes(` ${key} `)))
+            continue;
+        claim(product, key);
+    }
+    return {
+        products: [...found.keys()].sort(),
+        terms: [...new Set([...found.values()].flatMap((s) => [...s]))].sort(),
+    };
+}