@pipeworx/mcp-chrome-ux-report 0.1.0

package/src/index.ts

@@ -0,0 +1,1047 @@
+interface McpToolDefinition {
+  name: string;
+  description: string;
+  /** Human-facing one-liner (fleet #1967). Optional; consumers fall back to
+   *  description. Kept in step with shared/src/types.ts — scripts/lib/
+   *  check-inlined-types.mjs reports drift at publish time. */
+  summary?: string;
+  inputSchema: {
+    type: 'object';
+    properties: Record<string, unknown>;
+    required?: string[];
+    anyOf?: Array<{ required: string[] }>;
+    oneOf?: Array<{ required: string[] }>;
+    allOf?: Array<{ required: string[] }>;
+  };
+  outputSchema?: Record<string, unknown>;
+}
+
+interface McpToolExport {
+  tools: McpToolDefinition[];
+  callTool: (name: string, args: Record<string, unknown>) => Promise<unknown>;
+  meter?: { credits: number };
+  cost?: Record<string, unknown>;
+  provider?: string;
+}
+
+/**
+ * Was this failure OUR OWN web service? — the other half of `internal-db-class.ts`.
+ *
+ * fleet #1089 pulled failures from our own Postgres out of `upstream_down` by
+ * keying on the SQLSTATE inside PostgREST's four-key error envelope. That
+ * covered the majority and structurally could not cover the rest: the rest
+ * never reach Postgres, so they carry no SQLSTATE. What was left, measured over
+ * the 24h to 2026-09-02T15:00Z (fleet #1096):
+ *
+ *     5  pipeworx-catalog  get_pack_tools     Pipeworx catalog error: 522 — error code: 522
+ *     3  fleet             fleet_list_open …  upstream_down: Fleet task queue did not respond within 25s
+ *
+ * 521/522/523/526 are Cloudflare saying its edge could not reach an ORIGIN, and
+ * in both of those rows the origin is ours — `gateway.pipeworx.io` for the
+ * catalog pack (it self-fetches when the gateway hasn't injected a manifest),
+ * our own Supabase for fleet. There is no third party anywhere in either call.
+ * Same defect as #1089: our own outage filed under `upstream_down`, the one
+ * class that means "the source is unreachable and there is nothing for us to
+ * fix", which is why the problem-tools triage skips it.
+ *
+ * WHY NOT A WORDING RULE. The obvious fix is to match `fleet db error:` and
+ * `Pipeworx catalog error:` in classifyToolError. Each is emitted from exactly
+ * one site today, so it would work today. It would also rot the first time
+ * somebody rewords a label — silently, and in the direction of hiding our own
+ * outage, which is worse than the bug being fixed. Every prose rule in
+ * error-class.ts has needed widening as packs invented new wording (#409/#450/
+ * #584); that history is most of that file's comment budget.
+ *
+ * WHAT THIS KEYS ON INSTEAD: **the host the call actually reached.** A URL's
+ * hostname is a fact about the call, not a guess about its prose. Two
+ * consequences that a pack-level flag could not give us, and the reason the
+ * flag was rejected:
+ *
+ *   - It describes the CALL, not the pack. `govcon-intel` fans out to our own
+ *     Supabase AND to genuine third parties; `court-listener` holds our cache
+ *     in Supabase and fetches courtlistener.com. An `internallyHosted: true` on
+ *     either pack would relabel a real third-party outage as ours — inventing
+ *     work, which is the same class of error in the opposite direction.
+ *   - It covers every future internal pack for free, instead of one declared
+ *     slug at a time.
+ *
+ * WHY IT SURVIVES A REWORD. The marker below is not matched as a literal by two
+ * separate files. `markInternalOrigin()` writes it and `internalHostMetricsClass()`
+ * reads it, both from the single exported `INTERNAL_ORIGIN_MARKER` constant in
+ * this module — so changing the wording changes both sides in the same edit and
+ * cannot desynchronise them. The pack's own label (`fleet db error:`,
+ * `Pipeworx catalog error:`) is not read at all: reword it freely, the class is
+ * unaffected. That is the property `stripClassPrefix` lacked when it drifted
+ * from its own classifier three times and needed a CI gate to hold them
+ * together.
+ *
+ * WHERE THE 5xx TEST LIVES. `markInternalOrigin` is called from the places that
+ * hold the real `Response` — `httpError`/`httpErrorMessage` and the timeout
+ * branch of `fetchWithTimeout` in `shared/src/http.ts` — so "is this an
+ * availability failure" is decided from the actual status code, never re-derived
+ * by scraping a number out of a sentence. A 404 from our own registry for a slug
+ * that does not exist is a caller's bad argument and is deliberately NOT marked.
+ */
+
+/**
+ * OUR OWN web service was unreachable — not an upstream, and never `upstream_down`.
+ *
+ * ONE value, not three, unlike `internal_db_*`. That split existed because a
+ * slow query, an exhausted pool and an unknown SQLSTATE have different owners
+ * and different fixes. Here there is only one story to tell — an origin we run
+ * did not answer the edge — and one owner. A bucket with no distinct owner per
+ * value is decoration; #724 is what happens when a class holds several
+ * situations, and inventing sub-values ahead of a reason to act on them
+ * differently is the same mistake with the sign flipped.
+ *
+ * METRICS ONLY, exactly like PLATFORM_KEY_ERROR_CLASS and the internal_db
+ * values. `classifyToolError` still answers `upstream_down` for the retry and
+ * hint paths, which only care whether retrying or a sibling tool might work —
+ * and it might. Nothing a caller sees or is charged changes here.
+ *
+ * READ SIDE: this value is in BROKEN_TOOL_CLASSES, FAULT_CLASSES and
+ * ALL_ERROR_CLASSES in `workers/registry-api/src/index.ts`. All three, or it
+ * lands on no dashboard — fleet #721 is the warning, where the #719 split
+ * worked on the write side and was invisible for weeks.
+ */
+const INTERNAL_SERVICE_UNREACHABLE_CLASS = 'internal_service_unreachable';
+
+/**
+ * The token that carries "this origin is ours" from the call site to the
+ * classifier.
+ *
+ * Appended to the error message rather than attached to the Error object,
+ * because the object does not survive the trip: 275 packs return `{ error:
+ * string }` instead of throwing, the gateway reads `observedError` as a string,
+ * and the fleet pack rebuilds its error from a captured status + body across a
+ * retry loop. A property on an Error would be dropped by every one of those
+ * paths and the class would work in tests and vanish in production.
+ *
+ * WORDING IS LOAD-BEARING, same rule as labelAge's note in authority.ts. This
+ * string is appended to a pack's thrown Error message (shared/src/http.ts),
+ * and a thrown Error's message is exactly what the gateway hands back to the
+ * caller as `content[0].text` when nothing rewrites it (workers/gateway/src
+ * catches the throw and sets `rawResult.message = stripClassPrefix(error)`,
+ * which does not touch this suffix) — so the original wording,
+ * " [pipeworx-hosted origin — our own service, not a third party]", was not a
+ * theoretical leak: it shipped live on pipeworx-catalog's 522s, 7 times in 6
+ * hours on 2026-09-02 (see tests/golden-internal-service.test.ts), verbatim
+ * naming Pipeworx as the host. check:hosting-claims never caught it because it
+ * did not scan shared/ at all (task #2009). Reworded to describe the
+ * OBSERVATION (the origin did not answer) without a claim about who runs it —
+ * the identical fix labelAge got: drop the possessive, keep the fact.
+ */
+const INTERNAL_ORIGIN_MARKER = ' [origin did not respond — retry before concluding the named source is down]';
+
+/**
+ * Supabase's data plane for a project is `<ref>.supabase.co`, where the ref is
+ * exactly twenty lowercase letters (ours is `pqauisounztsgdgfkhke`).
+ *
+ * Matching the shape rather than listing the ref keeps this correct when we add
+ * a project — `supabaseEnv` on a pack entry already points some packs at a
+ * second one — while still excluding `status.supabase.co`, which is Supabase's
+ * own status page and emphatically not our database. Verified 2026-09-02 by
+ * `grep -rhoE '[a-z0-9-]+\.supabase\.(co|in)' mcps shared workers scripts`: the
+ * only real project ref anywhere in the tree is ours, the rest are doc
+ * placeholders (`abc`, `xyz`, `example`) which this pattern also excludes. Same
+ * finding internal-db-class.ts relies on for the PostgREST envelope being ours
+ * by construction.
+ */
+const SUPABASE_PROJECT_HOST = /^[a-z]{20}\.supabase\.(co|in)$/;
+
+/**
+ * Is this a host WE run?
+ *
+ * Deliberately NOT including `*.workers.dev`: plenty of third-party APIs are
+ * hosted on workers.dev, so the suffix says where something runs and not who
+ * owns it. Every internal call we actually make goes to a `pipeworx.io`
+ * hostname or to our Supabase project, both of which are ownership facts.
+ *
+ * `workers/gateway/src/provenance.ts`'s `OUR_HOSTS` answers the same
+ * question and DOES include `workers.dev` — a documented divergence
+ * (task #2051), not a bug to converge. That list decides what a response may
+ * cite as a data SOURCE, where a false negative (citing our own worker as an
+ * external source) is the hosting-disclosure leak this whole file exists to
+ * prevent, so it errs broad. This one decides who gets BLAMED for a 5xx in
+ * outage metrics read by on-call, where a false positive (crediting our own
+ * infra with a third party's outage) hides the real failure, so it errs
+ * narrow. Same suffix, opposite direction, because they are never called for
+ * the same reason.
+ *
+ * Returns false on anything unparseable rather than throwing — this runs inside
+ * an error path, and an error path that can itself throw turns a diagnosable
+ * failure into a mystery.
+ */
+function isPipeworxOrigin(url: string | URL | undefined | null): boolean {
+  if (!url) return false;
+  let host: string;
+  try {
+    host = new URL(url instanceof URL ? url.href : url).hostname.toLowerCase();
+  } catch {
+    return false;
+  }
+  if (host === 'pipeworx.io' || host.endsWith('.pipeworx.io')) return true;
+  return SUPABASE_PROJECT_HOST.test(host);
+}
+
+/**
+ * Append the marker when this failure was OUR origin failing to answer.
+ *
+ * `status` is the HTTP status when there is one, and omitted for a timeout —
+ * where there is no response at all, and "the origin did not answer" is the
+ * whole observation. Statuses below 500 are left alone: a 404 from our own
+ * registry for a slug that does not exist is the caller's argument, not our
+ * outage, and marking it would put ordinary 404s on the incident dashboard.
+ *
+ * Idempotent, so a message that is wrapped and re-marked on the way up (the
+ * fleet pack's retry loop re-throws through two layers) carries the marker once.
+ */
+function markInternalOrigin(
+  message: string,
+  url: string | URL | undefined | null,
+  status?: number,
+): string {
+  if (status !== undefined && status < 500) return message;
+  if (!isPipeworxOrigin(url)) return message;
+  if (message.includes(INTERNAL_ORIGIN_MARKER)) return message;
+  return message + INTERNAL_ORIGIN_MARKER;
+}
+
+/**
+ * Which blob4 value a failure from our own web services books as, or undefined
+ * if this is not one.
+ *
+ * Ordered AFTER `internalDbMetricsClass` at the call site: a PostgREST envelope
+ * from our own Supabase is a strictly more specific statement about the same
+ * row (which of our services, and why), and the two cannot disagree about
+ * whether the failure is ours.
+ */
+function internalHostMetricsClass(error: string): string | undefined {
+  return error.includes(INTERNAL_ORIGIN_MARKER) ? INTERNAL_SERVICE_UNREACHABLE_CLASS : undefined;
+}
+
+
+/**
+ * One place to turn a failed `fetch` into an error a caller can act on.
+ *
+ * Nearly every pack was written the same way:
+ *
+ *     if (!res.ok) throw new Error(`Unsplash: ${res.status}`);
+ *
+ * which discards the response body — and the body is usually where the upstream
+ * says what was actually wrong ("**symbol** not found: GBP", "parameter `year`
+ * out of range", "unknown taxonomy id"). The caller gets a number, cannot
+ * self-correct, and retries the same broken call. A 2026-07-31 sweep found this
+ * shape in 481 of 1,400 packs, 47 of them PLATFORM-keyed.
+ *
+ * It also hides bugs one level down. Two of the first three packs audited had a
+ * second defect that only existed because of this line: unsplash's rate-limit
+ * branch sat BELOW a catch-all and was unreachable, and bea-gov parsed
+ * `BEAAPI.Error.APIErrorDescription` below a `!res.ok` throw that made the
+ * parsing dead code for every non-200.
+ *
+ * DELIBERATELY NOT A CLASSIFIER. It does not add `user_error:` /
+ * `upstream_down:` prefixes. Those decide which tier a failure lands in, and the
+ * `error` tier is what the daily problem-tools list is built from — it means
+ * "Pipeworx has a defect". A 400 is genuinely ambiguous: often a caller's bad
+ * argument, but sometimes a query WE built wrong (ted-eu comma-joined its CPV
+ * values into something TED rejected, and that bug was found only because it sat
+ * in `error`). Blanket-classifying 400s as caller mistakes would have hidden it.
+ * A pack that KNOWS which it is should keep saying so explicitly; this helper is
+ * for the 481 that say nothing at all.
+ */
+
+/** Longest upstream explanation we'll pass through. Enough for a real message,
+ *  short enough that an HTML page or a stack trace can't swamp the error. */
+
+const MAX_DETAIL = 300;
+
+/**
+ * Default bound for `fetchWithTimeout` when a pack doesn't state its own.
+ *
+ * 25s mirrors the number `epo-ops` landed on after measuring the real failure:
+ * a degraded upstream that doesn't error, it just never answers, and a Worker
+ * sits in `await fetch()` until ITS OWN execution budget kills the request —
+ * which can take minutes, not seconds (epo_ops_search_patents measured 4-8
+ * MINUTE hangs before this existed). 25s is short enough that a caller gets a
+ * fast, actionable error instead of holding the connection, and long enough
+ * that it doesn't false-trip on a merely-slow-but-alive upstream.
+ */
+const DEFAULT_FETCH_TIMEOUT_MS = 25_000;
+
+/**
+ * Read the body of a failed response and fold it into a throwable Error.
+ *
+ * Usage — note the `await`, which is the one thing that makes this a mechanical
+ * change rather than a drop-in:
+ *
+ *     if (!res.ok) throw await httpError(res, 'Unsplash');
+ *
+ * Safe to call on any non-ok response: a body that is missing, empty, unreadable
+ * or HTML degrades to exactly the old `Name: 404` string rather than throwing
+ * something new from inside the error path.
+ */
+async function httpError(res: Response, name: string): Promise<Error> {
+  return new Error(await httpErrorMessage(res, name));
+}
+
+/** The message text without constructing an Error — for packs that need to wrap
+ *  it in their own envelope or add an explicit classification prefix. */
+async function httpErrorMessage(res: Response, name: string): Promise<string> {
+  // The one place a 5xx from a host WE run gets stamped as ours. `res.url` is
+  // the URL the fetch actually resolved to (after redirects), so this is a fact
+  // about the call rather than a guess from the `name` the pack passed in —
+  // reword that label freely, the class does not move. See
+  // internal-host-class.ts; no-op for every third-party upstream, which is why
+  // this touches 481 packs' error text and changes none of it.
+  return markInternalOrigin(
+    `${name}: ${res.status}${detailSuffix(await readDetail(res))}`,
+    res.url,
+    res.status,
+  );
+}
+
+/**
+ * Just the upstream's own explanation — no name, no status.
+ *
+ * For a pack that has already said both in its own sentence. epo-ops reads
+ * `EPO rejected this search as too large (HTTP 413) — ${httpErrorMessage(…)}`,
+ * which rendered as `… (HTTP 413) — EPO: 413.` once the XML detail was being
+ * dropped: the upstream named twice, the status twice, and the one thing EPO
+ * actually said ("Not enough characters before truncation character") nowhere
+ * (fleet #712). Returns '' when the body carries nothing readable, so a caller
+ * can fall back to its own wording.
+ */
+async function upstreamDetail(res: Response): Promise<string> {
+  return readDetail(res);
+}
+
+/**
+ * Read a SUCCESSFUL response as JSON, failing loudly when it isn't JSON.
+ *
+ * `httpError` above only ever runs on `!res.ok`, which leaves the nastier half
+ * of the problem unhandled: an upstream that answers **HTTP 200 with an HTML
+ * page**. A bot wall, a login redirect, a maintenance interstitial and a CDN
+ * error page are all 200s, so `res.ok` is true, and `res.json()` then throws
+ * `Unexpected token '<', "<!DOCTYPE "... is not valid JSON`.
+ *
+ * That string is the problem. It names no upstream, carries no status, and
+ * reads like a parser bug in Pipeworx — so it lands in the `error` tier, which
+ * means "we have a defect", and the caller is told nothing they can act on.
+ * data.govt.nz sat dead behind an Imperva challenge this way and every
+ * status-code health check we own reported it green (7889a845). A zero-length
+ * body has the same shape: `Unexpected end of JSON input`, seen this week on
+ * uk-gazette (83% of external calls) and census.
+ *
+ * UNLIKE `httpError`, this one DOES classify, and the asymmetry is deliberate.
+ * A 400 is genuinely ambiguous — often the caller's bad argument, sometimes a
+ * query we built wrong — so blanket-classifying it would hide our own bugs.
+ * There is no such ambiguity here: **no argument a caller can pass makes a JSON
+ * API return an HTML page.** It is always the upstream, so `upstream_down:` is
+ * a statement of fact rather than a guess, and it keeps these out of the
+ * problem-tools list where they crowd out real defects.
+ *
+ *     const data = await parseJson<Feed>(res, 'UK Gazette');
+ *
+ * Call it only after the `!res.ok` check — on a failed response you want
+ * `httpError`, which mines the body for the upstream's own explanation.
+ */
+async function parseJson<T>(res: Response, name: string): Promise<T> {
+  let raw: string;
+  try {
+    raw = await res.text();
+  } catch {
+    throw new Error(
+      `upstream_down: ${name} returned a body that could not be read (HTTP ${res.status}). ` +
+        'The connection most likely dropped mid-response; retrying is reasonable.',
+    );
+  }
+
+  const type = res.headers.get('content-type') ?? 'no content-type';
+
+  if (!raw.trim()) {
+    throw new Error(
+      `upstream_down: ${name} answered HTTP ${res.status} with an EMPTY body where JSON was expected (${type}). ` +
+        'Nothing about the request can cause this — it is an upstream fault, and the same call may well work on retry.',
+    );
+  }
+
+  // Checked before parsing rather than in the catch, because knowing it is
+  // markup is what turns "we failed to parse something" into "they served a
+  // web page" — the second is diagnosable, the first is not.
+  const head = raw.slice(0, 200).trimStart().toLowerCase();
+  if (head.startsWith('<!doctype') || head.startsWith('<html') || head.startsWith('<?xml')) {
+    const kind = head.startsWith('<?xml') ? 'an XML document' : 'an HTML page';
+    // The summary, not the source. Pasting the first 120 characters of a web
+    // page handed the agent `<!DOCTYPE html><html lang="en"…` — the same leak
+    // this branch exists to describe (fleet #712).
+    throw new Error(
+      `upstream_down: ${name} answered HTTP ${res.status} with ${kind} instead of JSON (${type}). ` +
+        'That is typically a bot wall, a login redirect or a maintenance page — it is returned as a SUCCESS, ' +
+        `so status-code health checks read it as fine. No argument change will get past it. ` +
+        `The page says: ${summarizeErrorBody(raw) || 'nothing readable'}`,
+    );
+  }
+
+  try {
+    return JSON.parse(raw) as T;
+  } catch {
+    throw new Error(
+      `upstream_down: ${name} answered HTTP ${res.status} with a body that is not valid JSON (${type}). ` +
+        `It begins: ${stripMarkup(raw).slice(0, 120) || '(unreadable)'}`,
+    );
+  }
+}
+
+/**
+ * `fetch`, but bounded — the fix for a systemic gap found 2026-08-30: a grep
+ * audit of every pack's `mcps/*\/src/index.ts` found 1,339 of ~1,500 call
+ * `fetch()` with NO timeout guard anywhere in the file. Two of those
+ * (epo-ops, statcan) were confirmed live-hanging for 4-8 minutes before this
+ * existed — every unguarded call carries the same risk, just unconfirmed.
+ *
+ * Mirrors the `epoFetch` wrapper `mcps/epo-ops/src/index.ts` shipped first:
+ * bound the request with `AbortSignal.timeout`, and on a timeout/abort throw
+ * an `upstream_down:` error that names the upstream and the bound rather than
+ * letting the raw `TimeoutError`/`AbortError` (which names neither) propagate.
+ * `upstream_down:` is deliberate, same reasoning as `parseJson` above — no
+ * argument a caller passes can make an upstream hang, so it is always the
+ * upstream's fault, and marking it that way keeps a slow API off the
+ * problem-tools list where it would crowd out our own defects.
+ *
+ * Usage — a mechanical swap for a bare `fetch(url, init)`:
+ *
+ *     const res = await fetchWithTimeout(url, init, 'Some API');
+ *
+ * Pass `timeoutMs` as a fourth argument to override the default for a pack
+ * with a known-slower upstream; the label should be the same short name you'd
+ * pass to `httpError`/`httpErrorMessage` for that call.
+ */
+async function fetchWithTimeout(
+  url: string | URL,
+  init: RequestInit = {},
+  name: string,
+  timeoutMs: number = DEFAULT_FETCH_TIMEOUT_MS,
+): Promise<Response> {
+  try {
+    return await fetch(url, { ...init, signal: AbortSignal.timeout(timeoutMs) });
+  } catch (err) {
+    if (err instanceof Error && (err.name === 'TimeoutError' || err.name === 'AbortError')) {
+      // States the OBSERVATION (no response in N seconds), not a diagnosis.
+      // "appears to be degraded" is an inference about the vendor that we have
+      // not checked, and it is wrong in a way that misdirects whoever reads it:
+      // a timeout from a Worker can equally mean OUR egress is blocked.
+      //
+      // Measured today (2026-09-01, fleet #1047): every call to
+      // mainnet.base.org failed from the x402 facilitator while the identical
+      // request from a laptop returned 200. Base was entirely healthy; the
+      // public RPC refuses Cloudflare Worker egress. Had this message fired
+      // there it would have blamed Base by name, and the next person would have
+      // waited for a vendor outage to clear that did not exist.
+      // A timeout has no status to test — there is no response at all — so
+      // `markInternalOrigin` is called without one: an origin we run that never
+      // answered is an availability failure by definition. This is the half of
+      // fleet #1096 with neither a SQLSTATE nor a status code to key on.
+      throw new Error(
+        markInternalOrigin(
+          `upstream_down: ${name} did not respond within ${timeoutMs / 1000}s. ` +
+            `That can be ${name} being slow or down, or this environment being unable to reach it ` +
+            `(some hosts refuse datacenter/Worker egress) — retry shortly, and check reachability ` +
+            `from elsewhere before concluding ${name} is down.`,
+          url,
+        ),
+      );
+    }
+    throw err;
+  }
+}
+
+function detailSuffix(detail: string): string {
+  return detail ? ` — ${detail}` : '';
+}
+
+async function readDetail(res: Response): Promise<string> {
+  let raw: string;
+  try {
+    raw = await res.text();
+  } catch {
+    // Body already consumed, or the connection died mid-read. The status alone
+    // is still worth throwing — never let the error path throw its own error.
+    return '';
+  }
+  return summarizeErrorBody(raw);
+}
+
+/**
+ * Turn ANY error body — JSON, HTML, XML or plain text — into one short phrase
+ * that never contains markup.
+ *
+ * This used to just drop an HTML or XML body on the floor, on the reasoning
+ * that markup crowds out the status. That was half right. Dropping it loses the
+ * one sentence a caller could have acted on: an `Access Denied` title, an SDMX
+ * `<message:Error>` text, an OPS fault string. A 2026-08-30 support sweep
+ * measured 13 of 291 caller-facing error rows carrying a raw page or document
+ * verbatim, across 11 packs, and in every one of them the useful content —
+ * "Access Denied", "Invalid country code", "SCRAPE_TIMEOUT" — was in there,
+ * buried in markup the agent had to parse out of a string (fleet #712).
+ *
+ * So: extract the meaning, discard the markup. The output is passed through
+ * `stripMarkup` unconditionally, which is what lets `check:error-body-leak`
+ * assert mechanically that no caller-facing message can contain `<?xml`,
+ * `<!DOCTYPE` or `<html`.
+ */
+function summarizeErrorBody(raw: string): string {
+  if (!raw || !raw.trim()) return '';
+
+  const head = raw.slice(0, 400).trimStart().toLowerCase();
+
+  // An HTML error page (Cloudflare interstitial, nginx default, a login
+  // redirect) says what it is in its <title>, and almost nowhere else.
+  if (head.startsWith('<!doctype') || head.startsWith('<html')) {
+    const title = htmlTitle(raw);
+    return title
+      ? `${title} (upstream returned an HTML error page, not an API response)`
+      : 'upstream returned an HTML error page, not an API response';
+  }
+
+  // XML fault documents — EPO OPS, SDMX (`<message:Error>`), SOAP faults. The
+  // human sentence sits in a child element whose tag name says what it is.
+  if (head.startsWith('<?xml') || head.startsWith('<')) {
+    const fault = xmlFaultText(raw);
+    return fault
+      ? `${stripMarkup(fault).slice(0, MAX_DETAIL)} (from the upstream's XML error document)`
+      : 'upstream returned an XML error document with no readable message';
+  }
+
+  // Most JSON error bodies bury one human sentence among ids and echoed request
+  // params. Prefer that sentence; fall back to the whole body when the shape is
+  // unfamiliar, since an unfamiliar shape is exactly when we can least afford to
+  // guess wrong and show nothing.
+  const fromJson = messageFromJson(raw);
+  return stripMarkup(fromJson ?? raw).slice(0, MAX_DETAIL);
+}
+
+/** The `<title>` of an HTML error page, or its first `<h1>` — the two places a
+ *  bot wall, a 502 and an "Access Denied" all state what happened. */
+function htmlTitle(raw: string): string | null {
+  const head = raw.slice(0, 4000);
+  for (const re of [/<title[^>]*>([\s\S]*?)<\/title>/i, /<h1[^>]*>([\s\S]*?)<\/h1>/i]) {
+    const m = re.exec(head);
+    const text = m ? stripMarkup(m[1]) : '';
+    if (text) return text.slice(0, 160);
+  }
+  return null;
+}
+
+/** Tag names that carry the explanation in an XML fault document, namespace
+ *  prefix optional (`<message:Error>`, `<com:Text>`, `<faultstring>`). */
+const XML_FAULT_TAG_RE =
+  /<(?:[A-Za-z0-9_.-]+:)?(?:text|message|description|faultstring|reason|detail|title|errormessage|error)\b[^>]*>([^<]{2,400})</i;
+
+function xmlFaultText(raw: string): string | null {
+  const head = raw.slice(0, 8000);
+  const tagged = XML_FAULT_TAG_RE.exec(head);
+  if (tagged && tagged[1].trim()) return tagged[1];
+
+  // Nothing conventionally named — take the longest text node instead. A fault
+  // document with one sentence in an oddly named element is still readable;
+  // returning nothing at all is not.
+  let best = '';
+  for (const m of head.matchAll(/>([^<>]{8,400})</g)) {
+    const text = m[1].trim();
+    if (text.length > best.length) best = text;
+  }
+  return best || null;
+}
+
+/**
+ * Remove every tag and stray angle bracket, then collapse whitespace.
+ *
+ * Applied to everything on the way out, including the JSON and plain-text
+ * paths, because an upstream is free to embed markup in a JSON string field —
+ * and a leak is a leak regardless of which branch produced it.
+ */
+function stripMarkup(s: string): string {
+  return collapse(decodeEntities(s.replace(/<[^>]*>/g, ' ')).replace(/[<>]/g, ' '));
+}
+
+/** The handful of entities that show up in error-page titles. Decoded AFTER
+ *  tags are stripped and BEFORE the angle-bracket sweep, so `&lt;script&gt;`
+ *  in a title cannot decode into markup that survives — EMBL-EBI's ChEMBL 500
+ *  page renders as `500 Internal Server Error &lt; EMBL-EBI` otherwise. */
+function decodeEntities(s: string): string {
+  return s
+    .replace(/&(?:amp|#0*38);/gi, '&')
+    .replace(/&(?:lt|#0*60);/gi, '<')
+    .replace(/&(?:gt|#0*62);/gi, '>')
+    .replace(/&(?:quot|#0*34);/gi, '"')
+    .replace(/&(?:#0*39|apos|#x0*27);/gi, "'")
+    .replace(/&nbsp;/gi, ' ');
+}
+
+/** The conventional "what went wrong" field, under any of the names upstreams
+ *  actually use. Checked in order; first non-empty string wins. */
+const MESSAGE_KEYS = [
+  'message', 'error_message', 'errorMessage', 'detail', 'details',
+  'description', 'error_description', 'reason', 'title', 'fault',
+];
+
+function messageFromJson(raw: string): string | null {
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(raw);
+  } catch {
+    return null;
+  }
+  return pickMessage(parsed, 0);
+}
+
+function pickMessage(node: unknown, depth: number): string | null {
+  // Two levels covers `{error: {message}}` and `{errors: [{detail}]}`, the two
+  // shapes that account for nearly all of them, without walking a large payload.
+  if (depth > 2 || node == null) return null;
+
+  if (typeof node === 'string') return node.trim() || null;
+
+  if (Array.isArray(node)) {
+    for (const item of node) {
+      const found = pickMessage(item, depth + 1);
+      if (found) return found;
+    }
+    return null;
+  }
+
+  if (typeof node !== 'object') return null;
+  const obj = node as Record<string, unknown>;
+
+  for (const key of MESSAGE_KEYS) {
+    const v = obj[key];
+    if (typeof v === 'string' && v.trim()) return v.trim();
+  }
+  // `{error: …}` where error is itself an object or a string — the single most
+  // common wrapper, so it is worth descending into by name rather than scanning
+  // every key and risking picking up an echoed request parameter.
+  for (const key of ['error', 'errors', 'fault', 'Error', 'data']) {
+    if (key in obj) {
+      const found = pickMessage(obj[key], depth + 1);
+      if (found) return found;
+    }
+  }
+  return null;
+}
+
+/** Errors are read in a single line of log output; newlines and runs of
+ *  whitespace make a multi-line body unreadable there. */
+function collapse(s: string): string {
+  return s.replace(/\s+/g, ' ').trim();
+}
+/**
+ * Real-user Core Web Vitals for any website with enough traffic, from Google's Chrome UX Report — the field measurements Google Search itself uses.
+ *
+ * CrUX aggregates performance data from opted-in Chrome users and serves it
+ * through the CrUX API at chromeuxreport.googleapis.com.
+ *
+ * This is FIELD data, not a lab test: the 75th-percentile Largest Contentful
+ * Paint, Interaction to Next Paint and Cumulative Layout Shift that actual
+ * visitors experienced on a site over the trailing 28 days, plus the
+ * good/needs-improvement/poor distribution behind each number. It is the
+ * dataset Google Search uses for its Core Web Vitals assessment, so it is the
+ * authoritative answer to "is this site passing Core Web Vitals", which no lab
+ * tool (Lighthouse included) can answer.
+ *
+ * CREDENTIALS
+ * -----------
+ * There is NO keyless path. The API answers an unkeyed request with HTTP 403
+ * and "Method doesn't allow unregistered callers" — verified live 2026-09-17 —
+ * so this pack requires an API key. A caller can bring their own as `_apiKey`;
+ * the gateway supplies `PLATFORM_GOOGLE_API_KEY` when that secret is set.
+ * A key is free and self-service: create one at
+ * https://console.cloud.google.com/apis/credentials and enable the Chrome UX
+ * Report API at
+ * https://console.cloud.google.com/apis/library/chromeuxreport.googleapis.com
+ * — no billing account, no approval step, 150 queries/minute.
+ *
+ * TRAPS, VERIFIED LIVE 2026-09-17
+ * -------------------------------
+ * 1. THE TWO KEY FAILURES LOOK NOTHING ALIKE. A missing key is HTTP 403
+ *    PERMISSION_DENIED "Method doesn't allow unregistered callers". A WRONG
+ *    key is HTTP 400 INVALID_ARGUMENT "API key not valid". A key whose project
+ *    has not enabled the CrUX API is a third shape again (SERVICE_DISABLED).
+ *    Collapsing all three into "auth failed" sends a caller looking for the
+ *    wrong problem, so each is reported separately below.
+ * 2. `origin` AND `url` ARE DIFFERENT QUESTIONS AND EXACTLY ONE MAY BE SENT.
+ *    `origin: "https://example.com"` is every page on the site aggregated;
+ *    `url: "https://example.com/pricing"` is that one page. Sending both is a
+ *    400. An origin must have no path — "https://example.com/" with the
+ *    trailing slash is accepted, "https://example.com/pricing" as an origin is
+ *    not.
+ * 3. NOT FOUND IS THE NORMAL ANSWER FOR MOST OF THE WEB. CrUX only publishes
+ *    an entry once a page or origin has enough opted-in Chrome traffic to be
+ *    both statistically sound and non-identifying. The API returns HTTP 404
+ *    for everything below that bar. That is not an outage and not a bad
+ *    request — it is the dataset saying "too little traffic", which is itself
+ *    the answer, and this pack reports it as such rather than as an error.
+ *    A URL can be absent while its origin is present.
+ * 4. THE HEADLINE NUMBER IS p75, NOT THE MEAN. Core Web Vitals is assessed at
+ *    the 75th percentile, and `percentiles.p75` is what Search uses. Averaging
+ *    the histogram bins gives a different, wrong number.
+ * 5. CLS IS UNITLESS; EVERY OTHER METRIC IS MILLISECONDS. CrUX returns CLS
+ *    p75 as a STRING like "0.05" while the timing metrics come back as
+ *    integers. Treating them uniformly silently mis-scales CLS by 1000x.
+ * 6. `crux_history` RETURNS ~40 WEEKLY COLLECTION PERIODS, and each is a
+ *    28-DAY TRAILING WINDOW, so consecutive points overlap by 27 days. A
+ *    week-on-week change in the series is a change in a rolling average, not
+ *    a change in that week's traffic.
+ */
+
+
+const API = 'https://chromeuxreport.googleapis.com/v1';
+const UA = 'pipeworx-mcp-chrome-ux-report/1.0 (+https://pipeworx.io)';
+const KEY_URL = 'https://console.cloud.google.com/apis/library/chromeuxreport.googleapis.com';
+
+/** Metrics the API serves, with what each one means and its unit. */
+const METRICS: Record<string, { label: string; unit: 'ms' | 'score'; good?: number; poor?: number }> = {
+  largest_contentful_paint: { label: 'Largest Contentful Paint', unit: 'ms', good: 2500, poor: 4000 },
+  interaction_to_next_paint: { label: 'Interaction to Next Paint', unit: 'ms', good: 200, poor: 500 },
+  cumulative_layout_shift: { label: 'Cumulative Layout Shift', unit: 'score', good: 0.1, poor: 0.25 },
+  first_contentful_paint: { label: 'First Contentful Paint', unit: 'ms', good: 1800, poor: 3000 },
+  experimental_time_to_first_byte: { label: 'Time to First Byte', unit: 'ms', good: 800, poor: 1800 },
+  round_trip_time: { label: 'Round-Trip Time', unit: 'ms' },
+};
+
+/** The three metrics Core Web Vitals is assessed on. Trap 4. */
+const CORE_WEB_VITALS = ['largest_contentful_paint', 'interaction_to_next_paint', 'cumulative_layout_shift'];
+
+const FORM_FACTORS = ['PHONE', 'DESKTOP', 'TABLET', 'ALL_FORM_FACTORS'];
+
+async function pwFetch(url: string, init: RequestInit): Promise<Response> {
+  return fetchWithTimeout(
+    url,
+    { ...init, headers: { 'User-Agent': UA, 'Content-Type': 'application/json', ...(init.headers ?? {}) } },
+    'Chrome UX Report API (Google)',
+  );
+}
+
+function requireKey(args: Record<string, unknown>): string {
+  const key = typeof args._apiKey === 'string' ? args._apiKey.trim() : '';
+  if (key) return key;
+  throw new Error(
+    `The Chrome UX Report requires an API key — Google refuses unregistered callers outright, so there is no keyless path. A key is free, instant and self-service: create one at https://console.cloud.google.com/apis/credentials, enable the Chrome UX Report API at ${KEY_URL}, and pass it as _apiKey. No billing account is needed.`,
+  );
+}
+
+interface CruxMetric {
+  histogram?: Array<{ start?: number | string; end?: number | string; density?: number }>;
+  percentiles?: { p75?: number | string };
+}
+interface CruxRecord {
+  key?: { origin?: string; url?: string; formFactor?: string };
+  metrics?: Record<string, CruxMetric>;
+  collectionPeriod?: { firstDate?: { year: number; month: number; day: number }; lastDate?: { year: number; month: number; day: number } };
+}
+
+function isoDate(d?: { year: number; month: number; day: number }): string | null {
+  if (!d || !d.year) return null;
+  return `${d.year}-${String(d.month).padStart(2, '0')}-${String(d.day).padStart(2, '0')}`;
+}
+
+/** Trap 5: CLS arrives as a decimal string, timings as integers. */
+function asNumber(v: unknown): number | null {
+  if (v === null || v === undefined) return null;
+  const n = Number(v);
+  return Number.isFinite(n) ? n : null;
+}
+
+function verdict(metric: string, p75: number | null): 'good' | 'needs-improvement' | 'poor' | null {
+  const def = METRICS[metric];
+  if (!def || p75 === null || def.good === undefined || def.poor === undefined) return null;
+  if (p75 <= def.good) return 'good';
+  if (p75 <= def.poor) return 'needs-improvement';
+  return 'poor';
+}
+
+function shapeMetrics(metrics: Record<string, CruxMetric> | undefined) {
+  const out: Record<string, unknown> = {};
+  for (const [name, m] of Object.entries(metrics ?? {})) {
+    const def = METRICS[name];
+    const p75 = asNumber(m.percentiles?.p75);
+    const bins = (m.histogram ?? []).map((b) => ({
+      start: asNumber(b.start),
+      end: b.end === undefined ? null : asNumber(b.end),
+      density: asNumber(b.density),
+    }));
+    out[name] = {
+      label: def?.label ?? name,
+      unit: def?.unit ?? null,
+      p75,
+      assessment: verdict(name, p75),
+      good_threshold: def?.good ?? null,
+      poor_threshold: def?.poor ?? null,
+      distribution: bins,
+    };
+  }
+  return out;
+}
+
+/**
+ * Trap 3: a 404 is the dataset's answer ("not enough traffic"), not a failure.
+ * Trap 1: the three credential failures are distinguished by status and body.
+ */
+async function queryCrux(path: string, body: Record<string, unknown>, key: string) {
+  const res = await pwFetch(`${API}/${path}?key=${encodeURIComponent(key)}`, {
+    method: 'POST',
+    body: JSON.stringify(body),
+  });
+  const text = await res.text();
+  let json: { error?: { code?: number; message?: string; status?: string }; record?: CruxRecord } = {};
+  try {
+    json = JSON.parse(text);
+  } catch {
+    throw new Error(`Chrome UX Report API returned HTTP ${res.status} with a non-JSON body: ${summarizeErrorBody(text)}`);
+  }
+
+  if (res.status === 404) return { notFound: true as const, record: null };
+  if (res.ok && json.record) return { notFound: false as const, record: json.record };
+
+  const status = json.error?.status ?? '';
+  const message = json.error?.message ?? summarizeErrorBody(text);
+  if (res.status === 403 && /unregistered callers/i.test(message)) {
+    throw new Error(
+      `The Chrome UX Report requires an API key and none reached Google. Upstream said: ${message}. Create a free key at https://console.cloud.google.com/apis/credentials and pass it as _apiKey.`,
+    );
+  }
+  if (res.status === 400 && /API key not valid/i.test(message)) {
+    throw new Error(
+      `The API key was rejected as invalid by Google (this is a WRONG key, not a missing one — a missing key returns 403 instead). Upstream said: ${message}`,
+    );
+  }
+  if (status === 'PERMISSION_DENIED' || /SERVICE_DISABLED|has not been used in project|is disabled/i.test(message)) {
+    throw new Error(
+      `The API key is valid but its Google Cloud project has not enabled the Chrome UX Report API. Enable it at ${KEY_URL}, then retry — Google caches the disabled state for a few minutes. Upstream said: ${message}`,
+    );
+  }
+  throw new Error(`Chrome UX Report API returned HTTP ${res.status} (${status || 'no status'}): ${message}`);
+}
+
+function buildTarget(args: Record<string, unknown>, mode: 'origin' | 'url'): Record<string, unknown> {
+  // Trap 2.
+  const raw = String(args[mode] ?? '').trim();
+  if (!raw) throw new Error(`\`${mode}\` is required.`);
+  if (!/^https?:\/\//i.test(raw)) {
+    throw new Error(`\`${mode}\` must be a full URL including the scheme, e.g. https://example.com — got "${raw}".`);
+  }
+  if (mode === 'origin') {
+    let parsed: URL;
+    try {
+      parsed = new URL(raw);
+    } catch {
+      throw new Error(`\`origin\` is not a valid URL: "${raw}".`);
+    }
+    if (parsed.pathname !== '/' && parsed.pathname !== '') {
+      throw new Error(
+        `An origin has no path. "${raw}" has one — use crux_query_url for a single page, or pass "${parsed.origin}" as the origin.`,
+      );
+    }
+  }
+
+  const body: Record<string, unknown> = { [mode]: raw };
+  const ff = args.form_factor ? String(args.form_factor).trim().toUpperCase() : null;
+  if (ff) {
+    if (!FORM_FACTORS.includes(ff)) {
+      throw new Error(`\`form_factor\` must be one of ${FORM_FACTORS.join(', ')} — got "${ff}".`);
+    }
+    if (ff !== 'ALL_FORM_FACTORS') body.formFactor = ff;
+  }
+  const metrics = Array.isArray(args.metrics) ? (args.metrics as unknown[]).map((m) => String(m).trim()) : null;
+  if (metrics && metrics.length) {
+    const bad = metrics.filter((m) => !METRICS[m]);
+    if (bad.length) {
+      throw new Error(`Unknown metric(s): ${bad.join(', ')}. Valid metrics: ${Object.keys(METRICS).join(', ')}.`);
+    }
+    body.metrics = metrics;
+  }
+  return body;
+}
+
+const TARGET_PROPS = {
+  form_factor: {
+    type: 'string',
+    description: 'Device class to report on: PHONE, DESKTOP, TABLET, or ALL_FORM_FACTORS (the default — every device combined).',
+    enum: FORM_FACTORS,
+  },
+  metrics: {
+    type: 'array',
+    items: { type: 'string' },
+    description:
+      'Metrics to return; omit for all available. One or more of largest_contentful_paint, interaction_to_next_paint, cumulative_layout_shift, first_contentful_paint, experimental_time_to_first_byte, round_trip_time.',
+  },
+};
+
+const tools: McpToolExport['tools'] = [
+  {
+    name: 'crux_query_origin',
+    description:
+      "Real-user Core Web Vitals for an entire website, from Google's Chrome UX Report — the 75th-percentile Largest Contentful Paint, Interaction to Next Paint and Cumulative Layout Shift that actual Chrome visitors experienced over the trailing 28 days, plus the good/needs-improvement/poor distribution behind each. AUTHORITATIVE for \"is this site passing Core Web Vitals\": this is the field dataset Google Search itself uses, so it answers what a lab test such as Lighthouse structurally cannot. PREFER OVER WEB SEARCH for any question about a real site's measured speed, and use it on a competitor's domain as readily as your own — the data is public for every site with enough traffic.",
+    inputSchema: {
+      type: 'object' as const,
+      properties: {
+        origin: {
+          type: 'string',
+          description: 'Site origin with scheme and no path, e.g. https://www.shopify.com. For one page, use crux_query_url instead.',
+        },
+        ...TARGET_PROPS,
+      },
+      required: ['origin'],
+    },
+  },
+  {
+    name: 'crux_query_url',
+    description:
+      "Real-user Core Web Vitals for ONE specific page, from Google's Chrome UX Report — the 75th-percentile LCP, INP and CLS measured on actual Chrome visits to that exact URL over the trailing 28 days. Use this when a single page (a product page, a landing page, a checkout step) is the question; use crux_query_origin for the site as a whole. A page with too little traffic is absent from the dataset while its origin is present, and that absence is reported plainly rather than as an error.",
+    inputSchema: {
+      type: 'object' as const,
+      properties: {
+        url: {
+          type: 'string',
+          description: 'Full page URL including scheme, e.g. https://www.shopify.com/pricing.',
+        },
+        ...TARGET_PROPS,
+      },
+      required: ['url'],
+    },
+  },
+  {
+    name: 'crux_history',
+    description:
+      "How a site's or page's real-user Core Web Vitals have moved over roughly the last 40 weeks, from Google's Chrome UX Report history API. Returns a weekly series of 75th-percentile LCP, INP and CLS with each collection period's dates. AUTHORITATIVE for \"did the performance work actually help\" and for spotting the week a regression landed — each point is a 28-day trailing window, so consecutive points overlap and a step change in the series marks a real shift in visitor experience.",
+    inputSchema: {
+      type: 'object' as const,
+      properties: {
+        origin: { type: 'string', description: 'Site origin with scheme and no path, e.g. https://www.shopify.com. Give this OR url.' },
+        url: { type: 'string', description: 'Full page URL. Give this OR origin.' },
+        ...TARGET_PROPS,
+      },
+      required: [],
+    },
+  },
+];
+
+function notFoundAnswer(kind: 'origin' | 'url', target: string, formFactor: string | null) {
+  return {
+    found: false,
+    [kind]: target,
+    form_factor: formFactor ?? 'ALL_FORM_FACTORS',
+    reason: 'insufficient_traffic',
+    note:
+      kind === 'url'
+        ? `The Chrome UX Report has no entry for this exact page. That is the dataset's normal answer for most URLs: a page is only published once it has enough opted-in Chrome traffic to be statistically sound and non-identifying. The page's ORIGIN may still be covered — try crux_query_origin on ${(() => { try { return new URL(target).origin; } catch { return 'the site origin'; } })()}.`
+        : `The Chrome UX Report has no entry for this origin, which means it is below the traffic threshold at which Google publishes data. This is not an outage; it is the answer. Narrower form factors (PHONE, DESKTOP) have separate, higher thresholds — omitting form_factor gives the best chance of a result.`,
+    source: 'Chrome UX Report (CrUX), Google',
+    source_url: 'https://developer.chrome.com/docs/crux',
+  };
+}
+
+async function queryRecord(args: Record<string, unknown>, mode: 'origin' | 'url') {
+  const key = requireKey(args);
+  const body = buildTarget(args, mode);
+  const { notFound, record } = await queryCrux('records:queryRecord', body, key);
+  const formFactor = (body.formFactor as string | undefined) ?? null;
+  if (notFound || !record) return notFoundAnswer(mode, String(args[mode]), formFactor);
+
+  const metrics = shapeMetrics(record.metrics);
+  const cwv = CORE_WEB_VITALS.filter((m) => metrics[m]);
+  const assessments = cwv.map((m) => (metrics[m] as { assessment: string | null }).assessment);
+  const allMeasured = cwv.length === CORE_WEB_VITALS.length && assessments.every((a) => a !== null);
+
+  return {
+    found: true,
+    [mode]: record.key?.[mode] ?? String(args[mode]),
+    form_factor: record.key?.formFactor ?? 'ALL_FORM_FACTORS',
+    collection_period: {
+      first_date: isoDate(record.collectionPeriod?.firstDate),
+      last_date: isoDate(record.collectionPeriod?.lastDate),
+    },
+    // Trap 4: the assessment is on p75, and Search's pass/fail needs all three.
+    core_web_vitals_assessment: allMeasured ? (assessments.every((a) => a === 'good') ? 'pass' : 'fail') : 'incomplete',
+    core_web_vitals_note:
+      'Google Search assesses Core Web Vitals at the 75th percentile of real visits, and requires LCP, INP and CLS all "good" to pass. CLS is a unitless score; every other metric here is milliseconds.',
+    metrics,
+    source: 'Chrome UX Report (CrUX), Google',
+    source_url: 'https://developer.chrome.com/docs/crux',
+  };
+}
+
+async function history(args: Record<string, unknown>) {
+  const key = requireKey(args);
+  const hasOrigin = !!String(args.origin ?? '').trim();
+  const hasUrl = !!String(args.url ?? '').trim();
+  if (hasOrigin === hasUrl) {
+    throw new Error('Give exactly one of `origin` (the whole site) or `url` (one page) — the CrUX API rejects both together and needs one.');
+  }
+  const mode: 'origin' | 'url' = hasOrigin ? 'origin' : 'url';
+  const body = buildTarget(args, mode);
+  const { notFound, record } = await queryCrux('records:queryHistoryRecord', body, key);
+  const formFactor = (body.formFactor as string | undefined) ?? null;
+  if (notFound || !record) return notFoundAnswer(mode, String(args[mode]), formFactor);
+
+  // The history shape differs from the point shape: collectionPeriods is an
+  // array, and every metric carries parallel arrays of the same length.
+  const raw = record as CruxRecord & {
+    collectionPeriods?: Array<{ firstDate?: { year: number; month: number; day: number }; lastDate?: { year: number; month: number; day: number } }>;
+    metrics?: Record<string, { percentilesTimeseries?: { p75s?: Array<number | string | null> } }>;
+  };
+  const periods = raw.collectionPeriods ?? [];
+  const series: Record<string, unknown> = {};
+  for (const [name, m] of Object.entries(raw.metrics ?? {})) {
+    const def = METRICS[name];
+    const p75s = m.percentilesTimeseries?.p75s ?? [];
+    const points = periods.map((p, i) => {
+      const v = asNumber(p75s[i]);
+      return {
+        period_start: isoDate(p.firstDate),
+        period_end: isoDate(p.lastDate),
+        p75: v,
+        assessment: verdict(name, v),
+      };
+    });
+    const measured = points.filter((p) => p.p75 !== null);
+    series[name] = {
+      label: def?.label ?? name,
+      unit: def?.unit ?? null,
+      points,
+      first_measured: measured.length ? measured[0].p75 : null,
+      latest_measured: measured.length ? measured[measured.length - 1].p75 : null,
+      change: measured.length > 1 ? (measured[measured.length - 1].p75 as number) - (measured[0].p75 as number) : null,
+    };
+  }
+
+  return {
+    found: true,
+    [mode]: record.key?.[mode] ?? String(args[mode]),
+    form_factor: record.key?.formFactor ?? 'ALL_FORM_FACTORS',
+    period_count: periods.length,
+    first_period_start: isoDate(periods[0]?.firstDate),
+    last_period_end: isoDate(periods[periods.length - 1]?.lastDate),
+    series,
+    // Trap 6.
+    window_note:
+      'Each point is a 28-day TRAILING window ending on period_end, so consecutive weekly points overlap by 27 days. A move in this series is a move in a rolling average — a one-week regression shows up damped and spread across four points.',
+    source: 'Chrome UX Report (CrUX) history API, Google',
+    source_url: 'https://developer.chrome.com/docs/crux/history-api',
+  };
+}
+
+async function callTool(name: string, args: Record<string, unknown>): Promise<unknown> {
+  switch (name) {
+    case 'crux_query_origin':
+      return queryRecord(args, 'origin');
+    case 'crux_query_url':
+      return queryRecord(args, 'url');
+    case 'crux_history':
+      return history(args);
+    default:
+      throw new Error(`Unknown tool: ${name}`);
+  }
+}
+
+export default { tools, callTool, meter: { credits: 1 } } satisfies McpToolExport;