@pipeworx/mcp-fred 0.1.0 → 0.1.2

package/src/index.ts

@@ -1,18 +1,675 @@
 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();
+}
 /**
  * FRED MCP — Federal Reserve Economic Data (St. Louis Fed)
  *
@@ -37,13 +694,52 @@ function extractKey(args: Record<string, unknown>): string {
   return key;
 }
 
-async function fredFetch(url: string): Promise<unknown> {
-  const res = await fetch(url);
+// FRED's upstream (api.stlouisfed.org) periodically slows/stalls. Without a
+// timeout the fetch hangs until Cloudflare kills the worker and returns a raw
+// 504 to the agent — the failure isn't catchable, so the gateway circuit
+// breaker can't trip. Abort at 8s instead: a clean error returns fast, and the
+// breaker can short-circuit the pack after repeated failures.
+//
+// In-pack retry on 5xx + timeout (revised 2026-06-08). Initial single-retry
+// landed 2026-06-07; production analytics next day showed FRED at 27 calls /
+// 9 errors (33% — only a modest improvement from the pre-retry 37.5%). Per-
+// attempt error rate is ~55%, suggesting FRED's CF origin has sustained
+// flakiness, not isolated flakes. Bumped to 3 attempts total with
+// exponential backoff (250ms, 1000ms). At p=0.55 per attempt, P(all 3 fail)
+// drops to ~17%, vs ~30% with 2 attempts and 55% with 1. Attempt count is
+// included in the final error message so we can see whether the cap should
+// be raised further.
+const FRED_RETRY_DELAYS_MS = [250, 1000]; // delay before attempts 2 and 3
+async function fredFetch(url: string, attempt = 1): Promise<unknown> {
+  const controller = new AbortController();
+  const t = setTimeout(() => controller.abort(), 8000);
+  let res: Response;
+  try {
+    res = await fetch(url, { signal: controller.signal });
+  } catch (e) {
+    if ((e as Error)?.name === 'AbortError') {
+      if (attempt <= FRED_RETRY_DELAYS_MS.length) {
+        await new Promise((r) => setTimeout(r, FRED_RETRY_DELAYS_MS[attempt - 1]));
+        return fredFetch(url, attempt + 1);
+      }
+      throw new Error(`upstream_down: FRED API timeout — api.stlouisfed.org did not respond within 8s on any of ${attempt} attempts.`);
+    }
+    throw e;
+  } finally {
+    clearTimeout(t);
+  }
+  if (res.status >= 500 && res.status < 600 && attempt <= FRED_RETRY_DELAYS_MS.length) {
+    await new Promise((r) => setTimeout(r, FRED_RETRY_DELAYS_MS[attempt - 1]));
+    return fredFetch(url, attempt + 1);
+  }
   if (!res.ok) {
-    const text = await res.text();
-    throw new Error(`FRED API error (${res.status}): ${text}`);
+    // FRED answers a throttle with an Akamai HTML page, so pasting the body
+    // handed the caller `<HTML><HEAD><TITLE>Access Denied</TITLE>…` verbatim.
+    // httpErrorMessage keeps the title and drops the markup (fleet #712).
+    const attemptNote = attempt > 1 ? ` (after ${attempt} attempts)` : '';
+    throw new Error(`${await httpErrorMessage(res, 'FRED API')}${attemptNote}`);
   }
-  return res.json();
+  return parseJson(res, 'FRED');
 }
 
 // ── Tool definitions ────────────────────────────────────────────────────
@@ -52,15 +748,23 @@ const tools: McpToolExport['tools'] = [
   {
     name: 'fred_get_series',
     description:
-      'Get observations (data points) for a FRED series. Key housing series: MORTGAGE30US (30-year mortgage rate), HOUST (housing starts), CSUSHPISA (Case-Shiller home price index), MSPUS (median home sale price), PERMIT (building permits), RRVRUSQ156N (rental vacancy rate).',
+      'AUTHORITATIVE historical time-series for US economic indicators from FRED (Federal Reserve Bank of St. Louis — the official US macroeconomic data repository, 800k+ series). Every series is an AGGREGATE for a whole country, state, or METROPOLITAN AREA (e.g. the Phoenix or Oklahoma City metro), reported monthly or quarterly. Pass a series ID like "MORTGAGE30US" (30y mortgage rate), "UNRATE" (unemployment), "CPIAUCSL" (CPI), "GDP", "FEDFUNDS" (Fed funds rate), "HOUST" (housing starts), "RHORUSQ156N" (US homeownership rate — the percent of households that OWN their home, distinct from the mortgage interest rate). Returns dates + values + the indicator\'s units. Use for macro, Fed, and metro-level US indicator questions. DO NOT GUESS a non-US series ID by pattern-matching a US one — "DEUUNRATE", "JPNUNRATE", "CANUNRATE" and similar <ISO3><METRIC> codes do NOT exist on FRED and 400 with "the series does not exist". For a NON-US country\'s unemployment/GDP/inflation/population, use worldbank_get_indicator (ISO country code + indicator, e.g. SL.UEM.TOTL.ZS for unemployment) instead — it is the reliable source for foreign macro data. If you don\'t know a FRED series ID, call fred_search first rather than inventing one.',
+    summary: 'A US economic indicator\'s full history, from the St. Louis Fed\'s FRED database.',
     inputSchema: {
       type: 'object' as const,
       properties: {
         series_id: { type: 'string', description: 'FRED series ID (e.g., "MORTGAGE30US", "HOUST", "CSUSHPISA")' },
         observation_start: { type: 'string', description: 'Start date in YYYY-MM-DD format (optional)' },
         observation_end: { type: 'string', description: 'End date in YYYY-MM-DD format (optional)' },
-        frequency: { type: 'string', description: 'Frequency aggregation: d, w, bw, m, q, sa, a (optional)' },
+        frequency: { type: 'string', description: 'Frequency aggregation (optional). Codes: d, w, bw, m, q, sa, a — English words also accepted (daily/weekly/biweekly/monthly/quarterly/semiannual/annual).' },
         units: { type: 'string', description: 'Data transformation: lin (levels), chg (change), ch1 (change from year ago), pch (% change), pc1 (% change from year ago), pca (compounded annual rate of change), cch (continuously compounded rate of change), cca (continuously compounded annual rate of change), log (natural log). Default: lin' },
+        // Read at runtime since the newest-first fix, but never declared — so it
+        // was invisible to callers and to every schema-based check.
+        limit: { type: 'number', description: 'Max observations, newest first (1-100000, default 20)' },
+        // Same gap `limit` had: read at runtime, never declared. Declared now so
+        // a caller who wants a series in chronological order can ask for it
+        // instead of reversing a truncated tail (fleet #718).
+        sort_order: { type: 'string', description: 'Observation order: desc = newest first (default), asc = oldest first.', enum: ['asc', 'desc'] },
         _apiKey: { type: 'string', description: 'FRED API key' },
       },
       required: ['series_id', '_apiKey'],
@@ -69,7 +773,8 @@ const tools: McpToolExport['tools'] = [
   {
     name: 'fred_search',
     description:
-      'Search for FRED series by keyword. Useful for discovering series IDs for housing, employment, inflation, and other economic data.',
+      'Search for economic data series by keyword. Returns series IDs, titles, and descriptions to identify the right indicator.',
+    summary: 'US economic data series matching a keyword search, from the St. Louis Fed\'s FRED database.',
     inputSchema: {
       type: 'object' as const,
       properties: {
@@ -85,7 +790,8 @@ const tools: McpToolExport['tools'] = [
   {
     name: 'fred_series_info',
     description:
-      'Get metadata about a FRED series: title, units, frequency, seasonal adjustment, notes, and date range.',
+      'Get metadata for a series: title, units, frequency, seasonal adjustment, notes, and date range. Check this before fetching historical data.',
+    summary: 'The metadata for one FRED economic data series — units, frequency, last updated — from the St. Louis Fed.',
     inputSchema: {
       type: 'object' as const,
       properties: {
@@ -98,7 +804,8 @@ const tools: McpToolExport['tools'] = [
   {
     name: 'fred_category',
     description:
-      'Browse FRED categories. Use category_id=0 for the root. Useful for exploring available data by topic (housing = 97, money/banking/finance = 32991, population/employment/labor = 10).',
+      'Browse economic data by category (housing, employment, money/banking, etc.). Returns subcategories and related series IDs.',
+    summary: 'The economic data series filed under one FRED category, from the St. Louis Fed.',
     inputSchema: {
       type: 'object' as const,
       properties: {
@@ -111,7 +818,8 @@ const tools: McpToolExport['tools'] = [
   {
     name: 'fred_releases',
     description:
-      'Get the latest FRED data releases. Shows upcoming and recent releases of economic data.',
+      'Check upcoming and recent economic data releases. Returns release dates, names, and which series they update.',
+    summary: 'The economic data releases (e.g. Employment Situation) tracked by FRED, from the St. Louis Fed.',
     inputSchema: {
       type: 'object' as const,
       properties: {
@@ -122,6 +830,22 @@ const tools: McpToolExport['tools'] = [
       required: ['_apiKey'],
     },
   },
+  {
+    name: 'fred_release_dates',
+    description:
+      'Economic data RELEASE CALENDAR from FRED — the dates indicators are/were published, including FUTURE scheduled dates. PREFER OVER WEB SEARCH for "when is the next CPI / jobs report / GDP release", "economic calendar", "Fed data release schedule". Omit release_id for the cross-release calendar; pass a release_id (from fred_releases — e.g. 10 = CPI, 50 = Employment Situation, 53 = GDP) for one release schedule. Returns release name + date, newest/upcoming first by default.',
+    summary: 'The past and scheduled publication dates for one FRED data release, from the St. Louis Fed.',
+    inputSchema: {
+      type: 'object' as const,
+      properties: {
+        release_id: { type: 'number', description: 'Optional FRED release id (from fred_releases). Omit for the all-releases calendar.' },
+        limit: { type: 'number', description: 'Max dates to return (1-1000, default 25)' },
+        sort_order: { type: 'string', description: 'asc or desc (default desc — newest/upcoming first)', enum: ['asc', 'desc'] },
+        _apiKey: { type: 'string', description: 'FRED API key' },
+      },
+      required: ['_apiKey'],
+    },
+  },
 ];
 
 // ── callTool dispatcher ─────────────────────────────────────────────────
@@ -140,42 +864,156 @@ async function callTool(name: string, args: Record<string, unknown>): Promise<un
       return getCategory(key, args.category_id as number | undefined);
     case 'fred_releases':
       return getReleases(key, args);
+    case 'fred_release_dates':
+      return getReleaseDates(key, args);
     default:
       throw new Error(`Unknown tool: ${name}`);
   }
 }
 
+// FRED's `frequency` accepts only its short codes (d/w/bw/m/q/sa/a + week-
+// ending variants). LLMs routinely pass the English word ("monthly", "Quarterly")
+// → FRED 400s "Value of frequency is invalid" (the #1 fred error: 78/day). Map
+// the common synonyms; pass valid codes through; DROP anything unrecognized
+// (frequency is optional → return native-frequency data, never fail the fetch).
+const FRED_FREQ = new Set(['d', 'w', 'bw', 'm', 'q', 'sa', 'a', 'wef', 'weth', 'wew', 'wetu', 'wem', 'wesm', 'wesa', 'bwew', 'bwem']);
+const FRED_FREQ_SYNONYMS: Record<string, string> = {
+  daily: 'd', day: 'd',
+  weekly: 'w', week: 'w', wk: 'w',
+  biweekly: 'bw', 'bi-weekly': 'bw', fortnightly: 'bw',
+  monthly: 'm', month: 'm', mo: 'm', mon: 'm',
+  quarterly: 'q', quarter: 'q', qtr: 'q', qtrly: 'q',
+  semiannual: 'sa', 'semi-annual': 'sa', semiannually: 'sa', biannual: 'sa', 'half-yearly': 'sa',
+  annual: 'a', annually: 'a', yearly: 'a', year: 'a', yr: 'a', y: 'a',
+};
+function normalizeFredFrequency(v: unknown): string | null {
+  const s = String(v).trim().toLowerCase();
+  if (FRED_FREQ.has(s)) return s;
+  return FRED_FREQ_SYNONYMS[s] ?? null;
+}
+
 // ── Tool implementations ────────────────────────────────────────────────
 
 async function getSeries(key: string, args: Record<string, unknown>) {
+  const seriesId = args.series_id as string;
   const params = new URLSearchParams({
-    series_id: args.series_id as string,
+    series_id: seriesId,
     api_key: key,
     file_type: 'json',
   });
   if (args.observation_start) params.set('observation_start', args.observation_start as string);
   if (args.observation_end) params.set('observation_end', args.observation_end as string);
-  if (args.frequency) params.set('frequency', args.frequency as string);
+  if (args.frequency) {
+    const f = normalizeFredFrequency(args.frequency);
+    if (f) params.set('frequency', f); // unknown → drop, return native frequency (don't 400)
+  }
   if (args.units) params.set('units', args.units as string);
+  // FRED's observations API defaults to sort_order=asc + ALL observations, so a
+  // bare fred_get_series(UNRATE) returned the full series from 1948 OLDEST-first
+  // and ignored the documented `limit: 20` — agents read a decades-old value,
+  // surfacing as "stale unemployment data" (user feedback 2026-06-24). Default to
+  // NEWEST-first + honor the limit so the latest observation leads.
+  params.set('sort_order', (args.sort_order as string) || 'desc');
+  params.set('limit', String(Math.min(100000, Math.max(1, (args.limit as number) ?? 20))));
 
-  const data = (await fredFetch(`${BASE}/series/observations?${params}`)) as {
-    realtime_start: string;
-    realtime_end: string;
-    observation_start: string;
-    observation_end: string;
-    count: number;
-    observations: { date: string; value: string }[];
-  };
+  const infoParams = new URLSearchParams({ series_id: seriesId, api_key: key, file_type: 'json' });
+  const [data, infoRaw] = await Promise.all([
+    fredFetch(`${BASE}/series/observations?${params}`) as Promise<{
+      realtime_start: string;
+      realtime_end: string;
+      observation_start: string;
+      observation_end: string;
+      count: number;
+      observations: { date: string; value: string }[];
+    }>,
+    fredFetch(`${BASE}/series?${infoParams}`)
+      .then((d) => ((d as { seriess?: { title: string; units: string; frequency: string; observation_start?: string; observation_end?: string }[] }).seriess?.[0] ?? null))
+      .catch(() => null),
+  ]);
+
+  // Empty-window rescue. A "what is the current rate" question routed here
+  // often arrives with a days-wide observation window, but weekly/monthly
+  // series have no observation inside it — FRED answers 200 with an empty
+  // array and the caller concludes "no data" (live case: MORTGAGE15US asked
+  // with a 4-day window that started after the last weekly print). Refetch
+  // without the window and lead with the latest observations, saying so.
+  let obs = data;
+  let windowNote: string | undefined;
+  let usedWindowRescue = false;
+  if (data.observations.length === 0 && (args.observation_start || args.observation_end)) {
+    const p2 = new URLSearchParams({
+      series_id: seriesId,
+      api_key: key,
+      file_type: 'json',
+      sort_order: 'desc',
+      limit: String(Math.min(20, Math.max(1, (args.limit as number) ?? 4))),
+    });
+    if (args.frequency) {
+      const f = normalizeFredFrequency(args.frequency);
+      if (f) p2.set('frequency', f);
+    }
+    if (args.units) p2.set('units', args.units as string);
+    const retry = (await fredFetch(`${BASE}/series/observations?${p2}`)) as typeof data;
+    if (retry.observations.length > 0) {
+      obs = retry;
+      usedWindowRescue = true;
+      windowNote = `No observations between ${args.observation_start ?? 'series start'} and ${args.observation_end ?? 'today'} — this series publishes ${infoRaw?.frequency ? `${infoRaw.frequency.toLowerCase()}, ` : ''}so narrow recent windows can be empty. Returning the latest observations instead.`;
+    }
+  }
+
+  // The observations endpoint ECHOES observation_start/observation_end back as
+  // it received them, and when the caller sets neither, FRED substitutes its own
+  // open bounds: 1600-01-01 and 9999-12-31. Emitting those under names that read
+  // as coverage told every caller that US GDP runs from 1600 through 9999
+  // (fleet #718). The real range is in the series metadata we already fetch
+  // alongside, so publish THAT under these names — matching what
+  // fred_series_info and fred_search already mean by them — and surface what the
+  // caller asked for separately, only when they actually asked. If the metadata
+  // call failed, omit the range rather than fall back to the sentinels.
+  const rows = obs.observations.map((o) => ({
+    date: o.date,
+    value: o.value === '.' ? null : o.value,
+  }));
+  const returned = rows.length;
+  const sortOrder = usedWindowRescue ? 'desc' : ((args.sort_order as string) || 'desc');
+  const order = sortOrder === 'asc' ? 'oldest_first' : 'newest_first';
+  // `count` is FRED's total for the requested window; `limit` (default 20) caps
+  // what comes back. Reporting the total beside a short list with nothing
+  // marking the cut is how a caller reads 20 rows as the whole series.
+  const truncated = typeof obs.count === 'number' && returned < obs.count;
+  const truncationNote = truncated
+    ? `Showing ${returned} of ${obs.count} observations (${order.replace('_', ' ')}); raise \`limit\` for more.`
+    : undefined;
+  const notes = [windowNote, truncationNote].filter(Boolean);
+
+  // Say what the series IS, in words (fleet #2238): a bare id plus numbers
+  // reads as nothing to the gateway's entity check, so 62/72 fred_get_series
+  // answers scored `unverifiable` even when the caller asked by exact id.
+  const newest = order === 'newest_first' ? rows[0] : rows[rows.length - 1];
+  const oldest = order === 'newest_first' ? rows[rows.length - 1] : rows[0];
+  const statement =
+    `FRED ${seriesId}${infoRaw?.title ? ` (${infoRaw.title})` : ''}: ${infoRaw?.units ?? 'units not reported'}, ${infoRaw?.frequency ?? 'native frequency'}, ` +
+    `from the Federal Reserve Bank of St. Louis (FRED). ${returned} of ${obs.count} observation(s) returned ${order.replace('_', ' ')}` +
+    (oldest && newest ? `, ${oldest.date} to ${newest.date}; latest ${newest.value ?? 'missing'} on ${newest.date}` : '') +
+    '.';
 
   return {
-    series_id: args.series_id,
-    count: data.count,
-    observation_start: data.observation_start,
-    observation_end: data.observation_end,
-    observations: data.observations.map((o) => ({
-      date: o.date,
-      value: o.value === '.' ? null : o.value,
-    })),
+    series_id: seriesId,
+    ...(infoRaw?.title ? { series_title: infoRaw.title, series_name: infoRaw.title } : {}),
+    source: 'Federal Reserve Bank of St. Louis (FRED)',
+    statement,
+    ...(infoRaw?.units ? { units: infoRaw.units } : {}),
+    ...(infoRaw?.frequency ? { frequency: infoRaw.frequency } : {}),
+    ...(notes.length ? { note: notes.join(' ') } : {}),
+    count: obs.count,
+    returned,
+    truncated,
+    observation_order: order,
+    ...(infoRaw?.observation_start ? { observation_start: infoRaw.observation_start } : {}),
+    ...(infoRaw?.observation_end ? { observation_end: infoRaw.observation_end } : {}),
+    ...(args.observation_start ? { requested_start: args.observation_start as string } : {}),
+    ...(args.observation_end ? { requested_end: args.observation_end as string } : {}),
+    observations: rows,
   };
 }
 
@@ -205,9 +1043,15 @@ async function searchSeries(key: string, args: Record<string, unknown>) {
     }[];
   };
 
+  const hits = data.seriess ?? [];
   return {
     total_matches: data.count,
-    series: data.seriess.map((s) => ({
+    returned: hits.length,
+    truncated: typeof data.count === 'number' && hits.length < data.count,
+    ...(typeof data.count === 'number' && hits.length < data.count
+      ? { note: `Showing ${hits.length} of ${data.count} matching series; raise \`limit\` (max 1000) or narrow search_text.` }
+      : {}),
+    series: hits.map((s) => ({
       series_id: s.id,
       title: s.title,
       units: s.units,
@@ -308,9 +1152,15 @@ async function getReleases(key: string, args: Record<string, unknown>) {
     }[];
   };
 
+  const rels = data.releases ?? [];
   return {
     total_releases: data.count,
-    releases: data.releases.map((r) => ({
+    returned: rels.length,
+    truncated: typeof data.count === 'number' && rels.length < data.count,
+    ...(typeof data.count === 'number' && rels.length < data.count
+      ? { note: `Showing ${rels.length} of ${data.count} releases; raise \`limit\` (max 1000) or page with \`offset\`.` }
+      : {}),
+    releases: rels.map((r) => ({
       id: r.id,
       name: r.name,
       press_release: r.press_release,
@@ -320,4 +1170,41 @@ async function getReleases(key: string, args: Record<string, unknown>) {
   };
 }
 
+async function getReleaseDates(key: string, args: Record<string, unknown>) {
+  const cap = Math.min(1000, Math.max(1, (args.limit as number) ?? 25));
+  // Fetch the full schedule (desc + future scheduled dates), then partition by
+  // today so the NEXT release is the headline — FRED has no date>=today filter,
+  // and a raw asc/desc sort buries the imminent release. 1000 rows comfortably
+  // spans the ~1y of future scheduled dates back through recent actuals.
+  const params = new URLSearchParams({
+    api_key: key,
+    file_type: 'json',
+    limit: '1000',
+    sort_order: 'desc',
+  });
+  const rid = args.release_id;
+  const single = rid != null && Number.isFinite(Number(rid));
+  // Future scheduled dates only for a SINGLE release — across ALL releases the
+  // far-future flood (300+ releases x ~12 months) overflows any fetch window
+  // and makes a cross-release "next" unreliable, so cross-release returns recent actuals.
+  if (single) { params.set('release_id', String(Math.floor(Number(rid)))); params.set('include_release_dates_with_no_data', 'true'); }
+  const path = single ? 'release/dates' : 'releases/dates';
+  const data = (await fredFetch(`${BASE}/${path}?${params}`)) as {
+    release_dates?: { release_id?: number; release_name?: string; date?: string }[];
+  };
+  const all = (data.release_dates ?? []).filter((d) => d.date);
+  const today = new Date().toISOString().slice(0, 10);
+  const shape = (d: { release_id?: number; release_name?: string; date?: string }) => ({ release_id: d.release_id ?? null, release_name: d.release_name ?? null, date: d.date ?? null });
+  const upcoming = all.filter((d) => (d.date as string) >= today).sort((a, b) => (a.date as string).localeCompare(b.date as string));
+  const recent = all.filter((d) => (d.date as string) < today).sort((a, b) => (b.date as string).localeCompare(a.date as string));
+  return {
+    release_id: single ? Math.floor(Number(rid)) : null,
+    as_of: today,
+    next_release: upcoming[0] ? shape(upcoming[0]) : null,
+    note: single ? undefined : 'Cross-release calendar returns RECENT releases. For an upcoming schedule / next release date, pass a release_id (from fred_releases).',
+    upcoming: upcoming.slice(0, cap).map(shape),
+    recent: recent.slice(0, cap).map(shape),
+  };
+}
+
 export default { tools, callTool, meter: { credits: 5 } } satisfies McpToolExport;