@pipeworx/mcp-guam-code 0.0.0-stage → 0.1.0

package/src/index.ts

@@ -0,0 +1,1180 @@
+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,
+        ),
+      );
+    }
+    // Fleet #2382. Everything that isn't a timeout/abort here is a genuine
+    // NETWORK-LEVEL failure — DNS resolution, connection refused, TLS handshake,
+    // Cloudflare's own "Network connection lost." — meaning `fetch()` itself
+    // threw and no HTTP response of any kind was ever received. Until this fix
+    // that raw exception was rethrown VERBATIM: a bare `TypeError: fetch failed`
+    // (or the Workers-runtime equivalent) names no upstream, carries no class
+    // token, and reads exactly like a defect in OUR code — because it says
+    // nothing about the call at all. It landed in `error`, the tier that means
+    // "Pipeworx has a defect", for every one of the (at the time of writing)
+    // ~470 packs that call this helper directly with no wrapper of their own.
+    //
+    // `dexscreener` hit this independently (fleet #1579) and fixed it with a
+    // bespoke per-pack try/catch around `fetchWithTimeout`. That fix is correct
+    // but only covers one pack; every other caller of this shared helper still
+    // leaked the raw exception. Moving the same fix HERE — the one place that
+    // already carries the timeout case — covers every pack that uses
+    // `fetchWithTimeout` without a wrapper, for free, and without widening
+    // `classifyToolError`'s regex list: the fix is giving the message a proper
+    // `upstream_down:` token at the point the two facts (no response was ever
+    // received, and which host we were trying to reach) are actually in hand,
+    // not teaching the classifier to guess from prose after the fact.
+    //
+    // Safe on the same grounds as the timeout branch above: no argument a
+    // caller passes can make `fetch()` itself throw a connection-level error,
+    // so this is always an availability failure, never a caller mistake. Same
+    // `markInternalOrigin` treatment — an origin we run that never answered is
+    // still ours, not a third party's outage.
+    const raw = err instanceof Error ? err.message : String(err);
+    throw new Error(
+      markInternalOrigin(
+        `upstream_down: could not reach ${name} at all (${raw.slice(0, 160)}). ` +
+          `No request reached ${name}, so this says NOTHING about whether the arguments you passed ` +
+          'are valid — do not re-check them on the strength of this error. Retry shortly.',
+        url,
+      ),
+    );
+  }
+}
+
+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();
+}
+/**
+ * Guam Code Annotated (GCA) — section lookup and search. Fleet #2888.
+ *
+ * SOURCE AND SHAPE. The Compiler of Laws (Judiciary of Guam,
+ * col.guamcourts.gov) publishes the GCA as ~800 per-CHAPTER PDFs, one static
+ * HTML page (22 collapsible Title panels, each listing its Chapters with a
+ * link to that chapter's PDF) and NOTHING lighter — no per-title listing, no
+ * search API, no site-wide section index (confirmed live). That page is
+ * baked ONCE into `gu-index-data.ts` (title/chapter/catchline/PDF URL) by
+ * `scripts/bake-index.mjs`, the same "baked table" shape
+ * mcps/north-dakota-code uses for its own structurally-identical gap.
+ *
+ * UNLIKE North Dakota, the GCA index page does not also enumerate every
+ * SECTION. But every sampled chapter PDF opens with its own embedded table
+ * of contents — a run of "§ <chapter>.<nn>. Catchline." headings — so
+ * `scripts/bake-sections.mjs` fetches all ~800 chapter PDFs ONCE and bakes a
+ * section-level catchline index (`gu-section-data.ts`, 15,760 rows) from
+ * that TOC block. gu_search reads it; gu_statute fetches the live chapter
+ * PDF fresh on every call and re-derives the section boundaries itself
+ * (never trusts the baked catchline for the text it returns) — see
+ * `sliceGuamSection` below.
+ *
+ * TWO CITATION CONVENTIONS, BOTH REAL (not a bug in either): newer,
+ * recodified titles (9 GCA, the Criminal & Correctional Code, adapted from
+ * the 1977 Model Penal Code) cite "chapter.section" with a dot — "9 GCA §
+ * 16.30" is Aggravated Murder. Older titles (e.g. 21 GCA Ch. 33, Proof and
+ * Acknowledgments of Instruments) concatenate chapter and a zero-padded
+ * section number with NO separator — "21 GCA § 33101" is chapter 33,
+ * section 101. gu_statute accepts either form; see `parseGuCitation`.
+ *
+ * THE PDF EXTRACTOR NEEDED A PACK-LOCAL FORK, NOT A SHARED-FILE EDIT.
+ * `shared/src/pdf-text.ts` already reads Kentucky's and North Dakota's
+ * statute PDFs, but Guam's chapter PDFs defeated it in two new ways, and
+ * this pack was built the same week another pack (#2884, New Mexico) was
+ * independently editing that shared file — so the fixes live in
+ * `./gu-pdf-extract.ts`, a pack-local fork, with a loud note for whoever
+ * next touches the shared extractor to fold them in. Full detail in that
+ * file's header; short version: (1) Guam's producer writes some non-WinAnsi
+ * subset fonts' text as parenthesized LITERAL strings rather than hex, which
+ * the shared extractor's `decodeRun` passes through unchanged instead of
+ * resolving through that font's ToUnicode map; (2) several older chapter
+ * PDFs declare a `/Length` a few bytes longer than their actual deflate
+ * stream, which Node/Workers' strict `DecompressionStream` rejects as
+ * trailing junk (Python's zlib tolerates it silently) — retried here with
+ * the tail trimmed a few bytes at a time; (3) a crash-adjacent fix: the
+ * shared `inflate()`'s fire-and-forget `void writer.write(...)` turns any
+ * bad deflate stream into an UNHANDLED promise rejection that no caller's
+ * try/catch can see, which crashed this pack's own bake script outright —
+ * worth checking whether the shared file has the same exposure.
+ *
+ * FOUR CAPABILITIES — two available, two are not:
+ *   - citation lookup              -> gu_statute
+ *   - topic/keyword search         -> gu_search (over official section and
+ *     chapter catchlines, not full statutory body text — see its own
+ *     description)
+ *   - amendments/enactment history -> embedded in gu_statute's `text` field
+ *     as the statute's own "SOURCE:" line (e.g. "Enacted 1977 as 16.30;
+ *     amended by P.L. 14-125..."), the same shape Kentucky's "History:" line
+ *     takes. Not split out as a separate field — the GCA does not mark it
+ *     up distinctly enough to parse reliably across both PDF eras.
+ *   - historical version (a section's text as it read before an amendment)
+ *     -> NOT AVAILABLE. col.guamcourts.gov serves the CURRENT text only; no
+ *     per-year or per-session archive was found.
+ *
+ * Keyless, no signup.
+ */
+import { extractGuamPdfText } from './gu-pdf-extract.js';
+import type { GuChapterRow, GuTitleRow } from './gu-index-data.js';
+import type { GuSectionRow } from './gu-section-data.js';
+
+const UA = 'pipeworx-mcp-guam-code/1.0 (+https://pipeworx.io)';
+const UPSTREAM = 'Compiler of Laws, Judiciary of Guam (col.guamcourts.gov)';
+
+async function pwFetch(url: string | URL, init?: RequestInit): Promise<Response> {
+  const headers = { 'User-Agent': UA, ...(init?.headers ?? {}) };
+  return fetchWithTimeout(url, { ...init, headers }, UPSTREAM);
+}
+
+// fleet #2888, same shape as north-dakota-code (fleet #2754): the baked
+// index module is uploaded to KV at deploy and injected as args._bakedIndex
+// on the first call into this pack, rather than value-imported here — a
+// value import would put the whole ~900KB table into every gateway
+// isolate's retained heap whether or not it ever serves a guam-code call.
+// NEVER value-import gu-index-data.ts or gu-section-data.ts here — a
+// type-only import (above) is erased at build time and is fine.
+let GU_TITLES: GuTitleRow[] = [];
+let GU_CHAPTERS: GuChapterRow[] = [];
+let GU_SECTIONS: GuSectionRow[] = [];
+let GU_INDEX_CAPTURED_AT = '';
+let GU_SECTION_INDEX_CAPTURED_AT = '';
+
+function ensureIndex(args: Record<string, unknown>): void {
+  if (GU_CHAPTERS.length) return;
+  const idx = args._bakedIndex as Record<string, unknown> | null | undefined;
+  const titles = idx?.GU_TITLES;
+  const chapters = idx?.GU_CHAPTERS;
+  const sections = idx?.GU_SECTIONS;
+  if (
+    !Array.isArray(titles) ||
+    !Array.isArray(chapters) ||
+    chapters.length === 0 ||
+    !Array.isArray(sections) ||
+    typeof idx?.GU_INDEX_CAPTURED_AT !== 'string' ||
+    typeof idx?.GU_SECTION_INDEX_CAPTURED_AT !== 'string'
+  ) {
+    // Loud on purpose: answering from an empty table would read as "no such
+    // section" / "no match", a silent wrong answer.
+    throw new Error(
+      `guam-code: the baked GCA index is unavailable (${idx == null ? 'not injected' : 'malformed'}). ` +
+        'The gateway loads it from KV on the first call into this pack; this is a server-side fault, not a bad citation or query — retry shortly.',
+    );
+  }
+  GU_TITLES = titles as GuTitleRow[];
+  GU_CHAPTERS = chapters as GuChapterRow[];
+  GU_SECTIONS = sections as GuSectionRow[];
+  GU_INDEX_CAPTURED_AT = idx.GU_INDEX_CAPTURED_AT as string;
+  GU_SECTION_INDEX_CAPTURED_AT = idx.GU_SECTION_INDEX_CAPTURED_AT as string;
+}
+
+/** Strip the per-page running header ("9 GCA CRIMES AND CORRECTIONS CH. 16
+ *  CRIMINAL HOMICIDE <pagenum>") and the "COL<digits>" document-management
+ *  stamp that every chapter PDF repeats on every page — both are page
+ *  furniture, not statutory text, and would otherwise land mid-section
+ *  whenever a section spans a page break (confirmed live: 9gc016.pdf splits
+ *  "§ 16.30. Aggravated Murder Defined." from its own body text across
+ *  exactly this furniture). The header is self-derived from the chapter
+ *  text's own first occurrence rather than reconstructed from the title/
+ *  chapter catchline strings, which may not match the PDF's own casing or
+ *  punctuation exactly. */
+function stripChapterNoise(text: string): string {
+  let out = text.replace(/COL\d{6,10}/g, ' ');
+  const headerMatch = /^(\d+[A-Za-z]?\s+GCA\s+.+?\s+CH\.\s+\S+\s+.+?)\s+\d+\s+/.exec(out);
+  if (headerMatch) {
+    const prefix = headerMatch[1].replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+    out = out.replace(new RegExp(`${prefix}\\s+\\d+\\s*`, 'g'), ' ');
+  }
+  return out.replace(/\s+/g, ' ').trim();
+}
+
+interface ParsedCitation {
+  title: number | null;
+  chapter: string;
+  /** The citation's own numeric suffix, in whichever style the input used:
+   *  dotted ("16.30") keeps the dot; concatenated ("33101") does not. */
+  full: string;
+  dotted: boolean;
+}
+
+/** Parse free-form Guam citation input. Accepts "9 GCA § 16.30", "9 GCA
+ *  16.30", "16.30" (needs `titleArg`), "21 GCA § 33101", "33101" (needs
+ *  `titleArg`). Returns null if no digits were found at all. The
+ *  chapter/section SPLIT for the no-dot (concatenated) form is genuinely
+ *  ambiguous from the string alone (chapter numbers vary in length) — it is
+ *  resolved later, against the known chapter list for the given title, not
+ *  here. */
+function parseGuCitation(raw: string, titleArg: number | null): ParsedCitation | null {
+  let s = raw.trim();
+  let title = titleArg;
+  const titleMatch = /^(?:title\s+)?(\d+)\s*GCA\b\s*/i.exec(s);
+  if (titleMatch) {
+    title = parseInt(titleMatch[1], 10);
+    s = s.slice(titleMatch[0].length);
+  }
+  s = s.replace(/^§+\s*/, '').trim();
+  const dottedMatch = /^(\d+[A-Za-z]?)\.(\d+[A-Za-z]?)$/.exec(s);
+  if (dottedMatch) {
+    return { title, chapter: dottedMatch[1], full: `${dottedMatch[1]}.${dottedMatch[2]}`, dotted: true };
+  }
+  const concatMatch = /^(\d+[A-Za-z]?)$/.exec(s);
+  if (concatMatch) {
+    // Chapter unknown yet — resolved against the title's chapter list by
+    // the caller (longest-known-chapter-number-prefix match).
+    return { title, chapter: '', full: concatMatch[1], dotted: false };
+  }
+  return null;
+}
+
+/** For the concatenated (no-dot) citation style, the chapter/section
+ *  boundary is ambiguous from the digit string alone (e.g. is "33101"
+ *  chapter 3 + section 3101, or chapter 33 + section 101?). Resolve it by
+ *  picking the LONGEST known chapter number, for the given title, that is a
+ *  literal prefix of the citation with a non-empty remainder — the correct
+ *  answer in every case sampled, and the same disambiguation the baked
+ *  gu-section-data.ts table independently confirms for every chapter it
+ *  covers. */
+function resolveConcatenatedChapter(titleIdx: number, full: string): string | null {
+  let best: string | null = null;
+  for (const [tIdx, chapter] of GU_CHAPTERS) {
+    if (tIdx !== titleIdx) continue;
+    if (full.startsWith(chapter) && full.length > chapter.length && (best === null || chapter.length > best.length)) {
+      best = chapter;
+    }
+  }
+  return best;
+}
+
+function findChapter(titleIdx: number, chapter: string): { idx: number; row: GuChapterRow } | null {
+  for (let i = 0; i < GU_CHAPTERS.length; i++) {
+    const row = GU_CHAPTERS[i]!;
+    if (row[0] === titleIdx && row[1] === chapter) return { idx: i, row };
+  }
+  return null;
+}
+
+/** Cut one section's text out of a whole-chapter extraction. Heading
+ *  candidates are built from the TARGET citation itself (both the exact
+ *  string the caller is asking for AND a chapter-local "any section in this
+ *  chapter" pattern, used only to find the END boundary) rather than a
+ *  generic number pattern — the same defense mcps/north-dakota-code's
+ *  sliceNdSection uses against a cross-reference impostor, adapted for
+ *  Guam's two citation styles (see parseGuCitation). Among occurrences of
+ *  the TARGET citation, picks whichever has the most text before the next
+ *  heading — the same most-text-wins tie-break ND and Iowa use. */
+function sliceGuamSection(chapterText: string, chapterNum: string, targetFull: string, dotted: boolean): string | null {
+  const chapterEsc = chapterNum.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const anyHeadingRe = dotted
+    ? new RegExp(`§\\s*${chapterEsc}\\.\\d+[A-Za-z]?\\.`, 'g')
+    : new RegExp(`§\\s*${chapterEsc}\\d{2,4}[A-Za-z]?\\.`, 'g');
+  const targetEsc = targetFull.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const targetRe = new RegExp(`§\\s*${targetEsc}\\.`, 'g');
+
+  const allHeadings: { at: number; end: number }[] = [];
+  for (const m of chapterText.matchAll(anyHeadingRe)) allHeadings.push({ at: m.index, end: m.index + m[0].length });
+  allHeadings.sort((a, b) => a.at - b.at);
+
+  let best: { at: number; end: number } | null = null;
+  let bestLen = -1;
+  for (const m of chapterText.matchAll(targetRe)) {
+    const headingEnd = m.index + m[0].length;
+    const next = allHeadings.find((h) => h.at > m.index);
+    const bodyEnd = next ? next.at : chapterText.length;
+    const len = bodyEnd - headingEnd;
+    if (len > bestLen) {
+      bestLen = len;
+      best = { at: m.index, end: bodyEnd };
+    }
+  }
+  if (!best) return null;
+  return chapterText.slice(best.at, best.end).trim();
+}
+
+const tools: McpToolExport['tools'] = [
+  {
+    name: 'gu_statute',
+    description:
+      'Get the FULL TEXT of a section of the Guam Code Annotated (GCA) — Guam territorial law — by citation. "9 GCA 16.30" (or "9 GCA § 16.30") is Aggravated Murder Defined; "21 GCA 33101" is who may take an acknowledgment of a deed. Guam uses TWO citation styles depending on the title: newer titles write "chapter.section" with a dot (9 GCA, the Criminal Code); older titles concatenate chapter and a zero-padded section number with no separator (e.g. 21 GCA). Either form works. If the citation omits "N GCA", pass `title` separately (1-22). Returns the current statutory text, the official catchline, chapter/title context, source PDF URL, and data_as_of. Guam publishes the CURRENT text only — no per-year historical archive; this tool says so rather than guessing. Keyless. Use for "what does 9 GCA 16.30 say" or to check a Guam statute a case or contract relies on.',
+    inputSchema: {
+      type: 'object' as const,
+      properties: {
+        citation: {
+          type: 'string',
+          description: 'Section citation, e.g. "9 GCA 16.30", "9 GCA § 16.30", "21 GCA 33101", or a bare "16.30" / "33101" with `title` given separately.',
+        },
+        title: {
+          type: 'number',
+          description: 'GCA Title number (1-22), only needed if not already included in `citation` (e.g. citation="16.30", title=9).',
+        },
+      },
+      required: ['citation'],
+    },
+  },
+  {
+    name: 'gu_search',
+    description:
+      'Search the Guam Code Annotated BY TOPIC or keyword rather than citation — "murder", "security deposit", "concealed weapon permit", "homestead". Matches against official section and chapter CATCHLINES (not full statutory body text — Guam does not publish a full-text search, and this pack derives its section catchlines from each chapter PDF\'s own table of contents, not a separate official index, so coverage is best-effort). Optionally filter to one Title (1-22) with `title`. Returns matching citations, catchlines, and title/chapter context; call gu_statute with a citation for the full text. Keyless.',
+    inputSchema: {
+      type: 'object' as const,
+      properties: {
+        query: { type: 'string', description: 'Topic or keywords, e.g. "aggravated murder".' },
+        title: { type: 'number', description: 'Restrict to one GCA Title (1-22), e.g. 9 for Crimes and Corrections.' },
+        limit: { type: 'number', description: 'Max results to return, default 10, max 50.' },
+      },
+      required: ['query'],
+    },
+  },
+];
+
+async function guStatute(args: Record<string, unknown>) {
+  const raw = String(args.citation ?? '').trim();
+  if (!raw) {
+    return { found: false, reason: 'empty_citation', message: 'gu_statute needs a citation, e.g. "9 GCA 16.30" or "21 GCA 33101".' };
+  }
+  const titleArg = args.title !== undefined && args.title !== null ? Number(args.title) : null;
+  const parsed = parseGuCitation(raw, titleArg);
+  if (!parsed || !parsed.title) {
+    return {
+      found: false,
+      reason: 'unparseable_citation',
+      citation: raw,
+      message: `Could not parse "${raw}" as a GCA citation. Give it as "9 GCA 16.30" / "9 GCA § 16.30" / "21 GCA 33101", or pass a bare section with a separate title, e.g. citation="16.30", title=9.`,
+    };
+  }
+
+  const titleIdx = GU_TITLES.findIndex((t) => t.title === parsed.title);
+  if (titleIdx === -1) {
+    return {
+      found: false,
+      reason: 'title_not_found',
+      citation: raw,
+      message: `No GCA Title ${parsed.title} (valid range is 1-22).`,
+      source: `${UPSTREAM}, keyless`,
+    };
+  }
+  const titleRow = GU_TITLES[titleIdx]!;
+
+  let chapterNum = parsed.chapter;
+  let fullCitation = parsed.full;
+  if (!parsed.dotted && !chapterNum) {
+    const resolved = resolveConcatenatedChapter(titleIdx, parsed.full);
+    if (!resolved) {
+      return {
+        found: false,
+        reason: 'citation_not_found',
+        citation: raw,
+        title: parsed.title,
+        title_name: titleRow.name,
+        message: `No chapter of ${parsed.title} GCA (${titleRow.name}) has a number that is a prefix of "${parsed.full}".`,
+        hint: 'Call gu_search with a topic instead if you do not have an exact citation.',
+        source: `${UPSTREAM}, keyless`,
+      };
+    }
+    chapterNum = resolved;
+  }
+
+  const chapterHit = findChapter(titleIdx, chapterNum);
+  if (!chapterHit) {
+    return {
+      found: false,
+      reason: 'chapter_not_found',
+      citation: raw,
+      title: parsed.title,
+      title_name: titleRow.name,
+      message: `${parsed.title} GCA has no Chapter ${chapterNum}.`,
+      hint: 'Call gu_search with a topic instead if you do not have an exact citation.',
+      source: `${UPSTREAM}, keyless`,
+    };
+  }
+  const [, , chapterName, pdfUrl] = chapterHit.row;
+
+  const pdfRes = await pwFetch(pdfUrl);
+  if (!pdfRes.ok) throw await httpError(pdfRes, 'Guam Compiler of Laws chapter PDF');
+  const extracted = await extractGuamPdfText(await pdfRes.arrayBuffer());
+  if (!extracted.text.trim()) {
+    return {
+      found: true,
+      citation: `${parsed.title} GCA § ${fullCitation}`,
+      chapter: chapterNum,
+      chapter_name: chapterName,
+      title: parsed.title,
+      title_name: titleRow.name,
+      text: null,
+      reason: 'upstream_parse',
+      message: 'The chapter PDF was fetched but no text could be extracted from it — an extraction defect, not evidence the section does not exist.',
+      warnings: extracted.warnings,
+      pdf_url: pdfUrl,
+      source: `${UPSTREAM}, keyless`,
+    };
+  }
+
+  const cleaned = stripChapterNoise(extracted.text);
+  const sliced = sliceGuamSection(cleaned, chapterNum, fullCitation, parsed.dotted);
+  if (!sliced) {
+    return {
+      found: false,
+      reason: 'section_not_found_in_pdf',
+      citation: `${parsed.title} GCA § ${fullCitation}`,
+      chapter: chapterNum,
+      chapter_name: chapterName,
+      message: `"${fullCitation}" was not located as a heading in Chapter ${chapterNum}'s own PDF text — it may be vacant, repealed, or renumbered.`,
+      hint: 'Call gu_search with a topic, or gu_statute with a nearby citation in the same chapter, to check.',
+      pdf_url: pdfUrl,
+      extraction_warnings: extracted.warnings.length ? extracted.warnings : undefined,
+      source: `${UPSTREAM}, keyless`,
+    };
+  }
+
+  // The catchline is the heading text itself, up to its own trailing period
+  // — re-derived from THIS extraction rather than trusted from the baked
+  // gu-section-data.ts table, so a stale bake never produces a mismatched
+  // catchline next to fresh text.
+  const catchlineMatch = /^\s*([^.]+)\./.exec(sliced.replace(/^§\s*[\d.A-Za-z]+\.\s*/, ''));
+
+  return {
+    found: true,
+    citation: `${parsed.title} GCA § ${fullCitation}`,
+    catchline: catchlineMatch ? catchlineMatch[1].trim() : undefined,
+    chapter: chapterNum,
+    chapter_name: chapterName,
+    title: parsed.title,
+    title_name: titleRow.name,
+    text: sliced,
+    text_chars: sliced.length,
+    chapter_index_captured_at: GU_INDEX_CAPTURED_AT,
+    historical_versions: 'not available — col.guamcourts.gov publishes only the current online edition; no per-year or per-session archive was found.',
+    extraction_warnings: extracted.warnings.length ? extracted.warnings : undefined,
+    pdf_url: pdfUrl,
+    data_as_of: new Date().toISOString(),
+    source: `${UPSTREAM}, keyless`,
+    attribution: 'Guam statutes are public record.',
+  };
+}
+
+interface SearchHit { citation: string; catchline: string; chapterIdx: number; score: number }
+
+/** IDF-weighted word overlap over section catchlines (+ their chapter's own
+ *  caption, for recall when the word lives in the chapter name rather than
+ *  any one section's) — same scoring shape as mcps/north-dakota-code's
+ *  nd_search / mcps/iowa-code's ia_code_search. */
+function wordWeights(words: string[], haystacks: string[]): Map<string, number> {
+  const weight = new Map<string, number>();
+  for (const w of words) {
+    let df = 0;
+    for (const h of haystacks) if (h.includes(w)) df++;
+    weight.set(w, 1 / Math.max(1, df));
+  }
+  return weight;
+}
+
+async function guSearch(args: Record<string, unknown>) {
+  const q = String(args.query ?? '').trim();
+  if (!q) {
+    return {
+      found: false,
+      reason: 'empty_query',
+      message: 'gu_search needs a "query" — a topic or a few keywords, e.g. "aggravated murder" or "security deposit".',
+      hint: 'To look up a known citation directly, use gu_statute instead.',
+    };
+  }
+  const limit = Math.min(50, Math.max(1, Number(args.limit) || 10));
+  const titleFilter = args.title !== undefined && args.title !== null ? Number(args.title) : null;
+  const titleIdxFilter = titleFilter !== null ? GU_TITLES.findIndex((t) => t.title === titleFilter) : null;
+  if (titleFilter !== null && titleIdxFilter === -1) {
+    return { found: false, reason: 'title_not_found', title: titleFilter, message: `No GCA Title ${titleFilter} (valid range is 1-22).` };
+  }
+
+  const words = q.toLowerCase().split(/\s+/).filter((w) => w.length >= 3);
+  if (!words.length) {
+    return { found: false, reason: 'query_too_short', query: q, message: `"${q}" has no word of 3+ characters to match against section catchlines.` };
+  }
+
+  const candidateRows = titleIdxFilter !== null ? GU_SECTIONS.filter((r) => GU_CHAPTERS[r[2]]![0] === titleIdxFilter) : GU_SECTIONS;
+  const haystacks = candidateRows.map((r) => `${r[1]} ${GU_CHAPTERS[r[2]]![2]}`.toLowerCase());
+  const weight = wordWeights(words, haystacks);
+
+  const hits: SearchHit[] = [];
+  for (let i = 0; i < candidateRows.length; i++) {
+    const h = haystacks[i]!;
+    let score = 0;
+    for (const w of words) if (h.includes(w)) score += weight.get(w) ?? 1;
+    if (score > 0) hits.push({ citation: candidateRows[i]![0], catchline: candidateRows[i]![1], chapterIdx: candidateRows[i]![2], score });
+  }
+  hits.sort((a, b) => b.score - a.score);
+
+  if (!hits.length) {
+    return {
+      found: false,
+      query: q,
+      title: titleFilter ?? undefined,
+      message: `No section or chapter catchline in the Guam Code Annotated${titleFilter ? ` Title ${titleFilter}` : ''} matched "${q}".`,
+      hint: 'This matches official CATCHLINES (section and chapter names), not full statutory text — try shorter or more literal terms, or drop the `title` filter.',
+      sections_scanned: candidateRows.length,
+      data_as_of: GU_SECTION_INDEX_CAPTURED_AT,
+    };
+  }
+
+  const results = hits.slice(0, limit).map((h) => {
+    const chapterRow = GU_CHAPTERS[h.chapterIdx]!;
+    const titleRow = GU_TITLES[chapterRow[0]]!;
+    const dotted = h.citation.includes('.');
+    return {
+      citation: `${titleRow.title} GCA § ${h.citation}`,
+      section: h.citation,
+      catchline: h.catchline,
+      chapter: chapterRow[1],
+      chapter_name: chapterRow[2],
+      title: titleRow.title,
+      title_name: titleRow.name,
+      citation_style: dotted ? 'chapter.section' : 'concatenated',
+    };
+  });
+
+  return {
+    found: true,
+    query: q,
+    title: titleFilter ?? undefined,
+    count: results.length,
+    total_matched: hits.length,
+    results,
+    hint: 'Call gu_statute with a citation above for the full text.',
+    match_level: 'official_catchline',
+    source: `${UPSTREAM} GCA chapter table-of-contents index, keyless`,
+    data_as_of: GU_SECTION_INDEX_CAPTURED_AT,
+  };
+}
+
+async function callTool(name: string, args: Record<string, unknown>): Promise<unknown> {
+  ensureIndex(args);
+  switch (name) {
+    case 'gu_statute':
+      return guStatute(args);
+    case 'gu_search':
+      return guSearch(args);
+    default:
+      throw new Error(`Unknown tool: ${name}`);
+  }
+}
+
+export default { tools, callTool, meter: { credits: 1 } } satisfies McpToolExport;