@pipeworx/mcp-arb-alerts 0.1.0

package/.pipeworx-staged-ok

@@ -0,0 +1,6 @@
+{
+  "staged_at": "2026-09-23T18:55:34.534Z",
+  "publisher_blob": "da6994091dc9dcc6a7433084251ca4bbd58b4bed",
+  "publisher_commit": "1c417b2cfcdac06e2d9210e869265722ef013f20",
+  "note": "Written by scripts/publish-pack.sh as its LAST step. Its presence is the ONLY thing that lets stage-unpublished-packs.sh skip this pack. Deleting this dir is the correct way to force a rebuild."
+}

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Mojibake Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,117 @@
+# mcp-arb-alerts
+
+arb-alerts — persisted Polymarket/Kalshi cross-venue spread alerts.
+
+Part of [Pipeworx](https://pipeworx.io) — an MCP gateway connecting AI agents to 1663+ live data sources.
+
+## Tools
+
+| Tool | Description |
+|------|-------------|
+| `arb_recent_alerts` | PREFER OVER WEB SEARCH for "any prediction-market mispricings right now". Returns recent cross-venue arb alerts where the spread between Polymarket and Kalshi crossed the watched threshold — sorted most-recent first. Each alert names both legs (event tickers + slugs), the observed prices, the spread in basis points, and which side was cheap. Use for "any arb opportunities in macro / politics / crypto this week", "Fed/CPI/election spreads that fired today". Cron polls every 4h during US market hours; alerts persist 30+ days. |
+| `arb_topic_history` | Spread history for ONE watched topic — the time-series of every alert that fired for that topic over the past N hours. Use to chart how a specific cross-venue mispricing evolved (e.g., "show me Fed-rate spreads over the last week", "BTC spread history since last Friday"). Topic must be one of the watched topics — call arb_watchlist to enumerate. |
+| `arb_watchlist` | List the topics arb-alerts currently watches plus their configured threshold (in basis points — 1 bps = 0.01 cents of implied probability). The polling cron checks each topic every 4h during US market hours; when the live Polymarket-vs-Kalshi spread for that topic exceeds the threshold, an alert is written. Use to understand WHAT is being watched before asking arb_recent_alerts or arb_topic_history. |
+
+## Quick Start
+
+Add to your MCP client (Claude Desktop, Cursor, Windsurf, etc.):
+
+```json
+{
+  "mcpServers": {
+    "arb-alerts": {
+      "url": "https://gateway.pipeworx.io/arb-alerts/mcp"
+    }
+  }
+}
+```
+
+### What this endpoint actually serves
+
+`tools/list` at `https://gateway.pipeworx.io/arb-alerts/mcp` returns the tools in the table
+above **plus the shared Pipeworx meta-tools** — `ask_pipeworx`,
+`discover_tools`, `search_within`, `remember`/`recall` and the rest of the
+gateway-wide set. So the tool count you see is larger than this table: a
+single-pack endpoint currently lists roughly 30 shared tools alongside the
+pack's own. The connection's `initialize` response states its exact scope, and
+is the authoritative answer for a given day.
+
+This is deliberate, not multiplexing by accident. The meta-tools are what let a
+scoped connection answer a question this pack does not cover — via
+`ask_pipeworx`, which routes across the whole catalog — without you adding a
+second MCP server. There is currently no way to mount a pack endpoint without
+them; if the extra schemas cost you more context than the routing is worth,
+connect to the full gateway once rather than to several pack endpoints.
+
+Or connect to the full Pipeworx gateway to get every pack's tools listed
+directly, instead of just this one's:
+
+```json
+{
+  "mcpServers": {
+    "pipeworx": {
+      "url": "https://gateway.pipeworx.io/mcp"
+    }
+  }
+}
+```
+
+Both URLs reach the same gateway and the same 1663+ data sources. The
+only difference is which pack's tools are listed **directly**; `ask_pipeworx`
+reaches all of them from either one.
+
+## No MCP client? Call it over HTTP
+
+```bash
+curl -X POST https://gateway.pipeworx.io/v1/tools/arb_recent_alerts \
+  -H 'Content-Type: application/json' \
+  -d '{}'
+```
+
+No account needed for the first calls. Inspect any tool: `GET https://gateway.pipeworx.io/v1/tools/arb_recent_alerts`. Find one: `POST https://gateway.pipeworx.io/v1/tools/search_packs` with `{"query":"..."}`.
+
+## Standalone (no gateway account)
+
+This package also runs as a local stdio MCP server — no Pipeworx account, no
+gateway round-trip:
+
+```json
+{
+  "mcpServers": {
+    "arb-alerts": {
+      "command": "npx",
+      "args": ["-y", "@pipeworx/mcp-arb-alerts"]
+    }
+  }
+}
+```
+
+Or run it directly to confirm it starts:
+
+```bash
+npx -y @pipeworx/mcp-arb-alerts
+```
+
+It speaks MCP over stdin/stdout and answers `initialize`/`tools/list`/`tools/call`
+for **only** this pack's tools — none of the shared meta-tools the gateway
+connection above adds. Same source, same tools, no ask_pipeworx routing.
+
+## Using with ask_pipeworx
+
+Instead of calling tools directly, you can ask questions in plain English —
+this works on the pack endpoint above as well as on the full gateway:
+
+```
+ask_pipeworx({ question: "your question about Arb Alerts data" })
+```
+
+The gateway picks the right tool and fills the arguments automatically.
+
+## More
+
+- [Docs and guides](https://pipeworx.io/docs)
+- [pipeworx.io](https://pipeworx.io)
+
+## License
+
+MIT

package/bin/cli.js

@@ -0,0 +1,17 @@
+#!/usr/bin/env node
+//
+// Entry point for `npx @pipeworx/mcp-<slug>`.
+//
+// Packs ship as raw TypeScript (no build step — see publish-pack.sh for why:
+// tsx sidesteps every extensionless-import / bare-JSON-import edge case a
+// per-pack tsc build would have to solve one pack at a time). This file
+// registers tsx's ESM loader programmatically, then hands off to src/server.ts,
+// which wraps the pack's {tools, callTool} export in a stdio MCP server.
+//
+// Copied verbatim into every published pack repo by scripts/publish-pack.sh —
+// edit this file, not a per-pack copy.
+import { register } from 'tsx/esm/api';
+
+register();
+
+await import('../src/server.ts');

package/package.json

@@ -0,0 +1,28 @@
+{
+  "name": "@pipeworx/mcp-arb-alerts",
+  "version": "0.1.0",
+  "description": "arb-alerts — persisted Polymarket/Kalshi cross-venue spread alerts.",
+  "type": "module",
+  "main": "src/index.ts",
+  "types": "src/index.ts",
+  "bin": {
+    "mcp-arb-alerts": "bin/cli.js"
+  },
+  "keywords": ["mcp", "mcp-server", "model-context-protocol", "pipeworx", "arb-alerts"],
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/pipeworx-io/mcp-arb-alerts.git"
+  },
+  "scripts": {
+    "typecheck": "tsc --noEmit"
+  },
+  "dependencies": {
+    "@modelcontextprotocol/sdk": "^1.30.0",
+    "tsx": "^4.19.0"
+  },
+  "devDependencies": {
+    "typescript": "^5.9.3",
+    "@cloudflare/workers-types": "^4.20260405.1"
+  }
+}

package/server.json

@@ -0,0 +1,18 @@
+{
+  "$schema": "https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json",
+  "name": "io.github.pipeworx-io/arb-alerts",
+  "title": "Arb Alerts",
+  "description": "arb-alerts — persisted Polymarket/Kalshi cross-venue spread alerts.",
+  "version": "0.1.0",
+  "websiteUrl": "https://pipeworx.io/packs/arb-alerts",
+  "repository": {
+    "url": "https://github.com/pipeworx-io/mcp-arb-alerts",
+    "source": "github"
+  },
+  "remotes": [
+    {
+      "type": "streamable-http",
+      "url": "https://gateway.pipeworx.io/arb-alerts/mcp"
+    }
+  ]
+}

package/src/index.ts

@@ -0,0 +1,874 @@
+interface McpToolDefinition {
+  name: string;
+  description: string;
+  /** Human-facing one-liner (fleet #1967). Optional; consumers fall back to
+   *  description. Kept in step with shared/src/types.ts — scripts/lib/
+   *  check-inlined-types.mjs reports drift at publish time. */
+  summary?: string;
+  inputSchema: {
+    type: 'object';
+    properties: Record<string, unknown>;
+    required?: string[];
+    anyOf?: Array<{ required: string[] }>;
+    oneOf?: Array<{ required: string[] }>;
+    allOf?: Array<{ required: string[] }>;
+  };
+  outputSchema?: Record<string, unknown>;
+}
+
+interface McpToolExport {
+  tools: McpToolDefinition[];
+  callTool: (name: string, args: Record<string, unknown>) => Promise<unknown>;
+  meter?: { credits: number };
+  cost?: Record<string, unknown>;
+  provider?: string;
+}
+
+/**
+ * Was this failure OUR OWN web service? — the other half of `internal-db-class.ts`.
+ *
+ * fleet #1089 pulled failures from our own Postgres out of `upstream_down` by
+ * keying on the SQLSTATE inside PostgREST's four-key error envelope. That
+ * covered the majority and structurally could not cover the rest: the rest
+ * never reach Postgres, so they carry no SQLSTATE. What was left, measured over
+ * the 24h to 2026-09-02T15:00Z (fleet #1096):
+ *
+ *     5  pipeworx-catalog  get_pack_tools     Pipeworx catalog error: 522 — error code: 522
+ *     3  fleet             fleet_list_open …  upstream_down: Fleet task queue did not respond within 25s
+ *
+ * 521/522/523/526 are Cloudflare saying its edge could not reach an ORIGIN, and
+ * in both of those rows the origin is ours — `gateway.pipeworx.io` for the
+ * catalog pack (it self-fetches when the gateway hasn't injected a manifest),
+ * our own Supabase for fleet. There is no third party anywhere in either call.
+ * Same defect as #1089: our own outage filed under `upstream_down`, the one
+ * class that means "the source is unreachable and there is nothing for us to
+ * fix", which is why the problem-tools triage skips it.
+ *
+ * WHY NOT A WORDING RULE. The obvious fix is to match `fleet db error:` and
+ * `Pipeworx catalog error:` in classifyToolError. Each is emitted from exactly
+ * one site today, so it would work today. It would also rot the first time
+ * somebody rewords a label — silently, and in the direction of hiding our own
+ * outage, which is worse than the bug being fixed. Every prose rule in
+ * error-class.ts has needed widening as packs invented new wording (#409/#450/
+ * #584); that history is most of that file's comment budget.
+ *
+ * WHAT THIS KEYS ON INSTEAD: **the host the call actually reached.** A URL's
+ * hostname is a fact about the call, not a guess about its prose. Two
+ * consequences that a pack-level flag could not give us, and the reason the
+ * flag was rejected:
+ *
+ *   - It describes the CALL, not the pack. `govcon-intel` fans out to our own
+ *     Supabase AND to genuine third parties; `court-listener` holds our cache
+ *     in Supabase and fetches courtlistener.com. An `internallyHosted: true` on
+ *     either pack would relabel a real third-party outage as ours — inventing
+ *     work, which is the same class of error in the opposite direction.
+ *   - It covers every future internal pack for free, instead of one declared
+ *     slug at a time.
+ *
+ * WHY IT SURVIVES A REWORD. The marker below is not matched as a literal by two
+ * separate files. `markInternalOrigin()` writes it and `internalHostMetricsClass()`
+ * reads it, both from the single exported `INTERNAL_ORIGIN_MARKER` constant in
+ * this module — so changing the wording changes both sides in the same edit and
+ * cannot desynchronise them. The pack's own label (`fleet db error:`,
+ * `Pipeworx catalog error:`) is not read at all: reword it freely, the class is
+ * unaffected. That is the property `stripClassPrefix` lacked when it drifted
+ * from its own classifier three times and needed a CI gate to hold them
+ * together.
+ *
+ * WHERE THE 5xx TEST LIVES. `markInternalOrigin` is called from the places that
+ * hold the real `Response` — `httpError`/`httpErrorMessage` and the timeout
+ * branch of `fetchWithTimeout` in `shared/src/http.ts` — so "is this an
+ * availability failure" is decided from the actual status code, never re-derived
+ * by scraping a number out of a sentence. A 404 from our own registry for a slug
+ * that does not exist is a caller's bad argument and is deliberately NOT marked.
+ */
+
+/**
+ * OUR OWN web service was unreachable — not an upstream, and never `upstream_down`.
+ *
+ * ONE value, not three, unlike `internal_db_*`. That split existed because a
+ * slow query, an exhausted pool and an unknown SQLSTATE have different owners
+ * and different fixes. Here there is only one story to tell — an origin we run
+ * did not answer the edge — and one owner. A bucket with no distinct owner per
+ * value is decoration; #724 is what happens when a class holds several
+ * situations, and inventing sub-values ahead of a reason to act on them
+ * differently is the same mistake with the sign flipped.
+ *
+ * METRICS ONLY, exactly like PLATFORM_KEY_ERROR_CLASS and the internal_db
+ * values. `classifyToolError` still answers `upstream_down` for the retry and
+ * hint paths, which only care whether retrying or a sibling tool might work —
+ * and it might. Nothing a caller sees or is charged changes here.
+ *
+ * READ SIDE: this value is in BROKEN_TOOL_CLASSES, FAULT_CLASSES and
+ * ALL_ERROR_CLASSES in `workers/registry-api/src/index.ts`. All three, or it
+ * lands on no dashboard — fleet #721 is the warning, where the #719 split
+ * worked on the write side and was invisible for weeks.
+ */
+const INTERNAL_SERVICE_UNREACHABLE_CLASS = 'internal_service_unreachable';
+
+/**
+ * The token that carries "this origin is ours" from the call site to the
+ * classifier.
+ *
+ * Appended to the error message rather than attached to the Error object,
+ * because the object does not survive the trip: 275 packs return `{ error:
+ * string }` instead of throwing, the gateway reads `observedError` as a string,
+ * and the fleet pack rebuilds its error from a captured status + body across a
+ * retry loop. A property on an Error would be dropped by every one of those
+ * paths and the class would work in tests and vanish in production.
+ *
+ * WORDING IS LOAD-BEARING, same rule as labelAge's note in authority.ts. This
+ * string is appended to a pack's thrown Error message (shared/src/http.ts),
+ * and a thrown Error's message is exactly what the gateway hands back to the
+ * caller as `content[0].text` when nothing rewrites it (workers/gateway/src
+ * catches the throw and sets `rawResult.message = stripClassPrefix(error)`,
+ * which does not touch this suffix) — so the original wording,
+ * " [pipeworx-hosted origin — our own service, not a third party]", was not a
+ * theoretical leak: it shipped live on pipeworx-catalog's 522s, 7 times in 6
+ * hours on 2026-09-02 (see tests/golden-internal-service.test.ts), verbatim
+ * naming Pipeworx as the host. check:hosting-claims never caught it because it
+ * did not scan shared/ at all (task #2009). Reworded to describe the
+ * OBSERVATION (the origin did not answer) without a claim about who runs it —
+ * the identical fix labelAge got: drop the possessive, keep the fact.
+ */
+const INTERNAL_ORIGIN_MARKER = ' [origin did not respond — retry before concluding the named source is down]';
+
+/**
+ * Supabase's data plane for a project is `<ref>.supabase.co`, where the ref is
+ * exactly twenty lowercase letters (ours is `pqauisounztsgdgfkhke`).
+ *
+ * Matching the shape rather than listing the ref keeps this correct when we add
+ * a project — `supabaseEnv` on a pack entry already points some packs at a
+ * second one — while still excluding `status.supabase.co`, which is Supabase's
+ * own status page and emphatically not our database. Verified 2026-09-02 by
+ * `grep -rhoE '[a-z0-9-]+\.supabase\.(co|in)' mcps shared workers scripts`: the
+ * only real project ref anywhere in the tree is ours, the rest are doc
+ * placeholders (`abc`, `xyz`, `example`) which this pattern also excludes. Same
+ * finding internal-db-class.ts relies on for the PostgREST envelope being ours
+ * by construction.
+ */
+const SUPABASE_PROJECT_HOST = /^[a-z]{20}\.supabase\.(co|in)$/;
+
+/**
+ * Is this a host WE run?
+ *
+ * Deliberately NOT including `*.workers.dev`: plenty of third-party APIs are
+ * hosted on workers.dev, so the suffix says where something runs and not who
+ * owns it. Every internal call we actually make goes to a `pipeworx.io`
+ * hostname or to our Supabase project, both of which are ownership facts.
+ *
+ * `workers/gateway/src/provenance.ts`'s `OUR_HOSTS` answers the same
+ * question and DOES include `workers.dev` — a documented divergence
+ * (task #2051), not a bug to converge. That list decides what a response may
+ * cite as a data SOURCE, where a false negative (citing our own worker as an
+ * external source) is the hosting-disclosure leak this whole file exists to
+ * prevent, so it errs broad. This one decides who gets BLAMED for a 5xx in
+ * outage metrics read by on-call, where a false positive (crediting our own
+ * infra with a third party's outage) hides the real failure, so it errs
+ * narrow. Same suffix, opposite direction, because they are never called for
+ * the same reason.
+ *
+ * Returns false on anything unparseable rather than throwing — this runs inside
+ * an error path, and an error path that can itself throw turns a diagnosable
+ * failure into a mystery.
+ */
+function isPipeworxOrigin(url: string | URL | undefined | null): boolean {
+  if (!url) return false;
+  let host: string;
+  try {
+    host = new URL(url instanceof URL ? url.href : url).hostname.toLowerCase();
+  } catch {
+    return false;
+  }
+  if (host === 'pipeworx.io' || host.endsWith('.pipeworx.io')) return true;
+  return SUPABASE_PROJECT_HOST.test(host);
+}
+
+/**
+ * Append the marker when this failure was OUR origin failing to answer.
+ *
+ * `status` is the HTTP status when there is one, and omitted for a timeout —
+ * where there is no response at all, and "the origin did not answer" is the
+ * whole observation. Statuses below 500 are left alone: a 404 from our own
+ * registry for a slug that does not exist is the caller's argument, not our
+ * outage, and marking it would put ordinary 404s on the incident dashboard.
+ *
+ * Idempotent, so a message that is wrapped and re-marked on the way up (the
+ * fleet pack's retry loop re-throws through two layers) carries the marker once.
+ */
+function markInternalOrigin(
+  message: string,
+  url: string | URL | undefined | null,
+  status?: number,
+): string {
+  if (status !== undefined && status < 500) return message;
+  if (!isPipeworxOrigin(url)) return message;
+  if (message.includes(INTERNAL_ORIGIN_MARKER)) return message;
+  return message + INTERNAL_ORIGIN_MARKER;
+}
+
+/**
+ * Which blob4 value a failure from our own web services books as, or undefined
+ * if this is not one.
+ *
+ * Ordered AFTER `internalDbMetricsClass` at the call site: a PostgREST envelope
+ * from our own Supabase is a strictly more specific statement about the same
+ * row (which of our services, and why), and the two cannot disagree about
+ * whether the failure is ours.
+ */
+function internalHostMetricsClass(error: string): string | undefined {
+  return error.includes(INTERNAL_ORIGIN_MARKER) ? INTERNAL_SERVICE_UNREACHABLE_CLASS : undefined;
+}
+
+
+/**
+ * One place to turn a failed `fetch` into an error a caller can act on.
+ *
+ * Nearly every pack was written the same way:
+ *
+ *     if (!res.ok) throw new Error(`Unsplash: ${res.status}`);
+ *
+ * which discards the response body — and the body is usually where the upstream
+ * says what was actually wrong ("**symbol** not found: GBP", "parameter `year`
+ * out of range", "unknown taxonomy id"). The caller gets a number, cannot
+ * self-correct, and retries the same broken call. A 2026-07-31 sweep found this
+ * shape in 481 of 1,400 packs, 47 of them PLATFORM-keyed.
+ *
+ * It also hides bugs one level down. Two of the first three packs audited had a
+ * second defect that only existed because of this line: unsplash's rate-limit
+ * branch sat BELOW a catch-all and was unreachable, and bea-gov parsed
+ * `BEAAPI.Error.APIErrorDescription` below a `!res.ok` throw that made the
+ * parsing dead code for every non-200.
+ *
+ * DELIBERATELY NOT A CLASSIFIER. It does not add `user_error:` /
+ * `upstream_down:` prefixes. Those decide which tier a failure lands in, and the
+ * `error` tier is what the daily problem-tools list is built from — it means
+ * "Pipeworx has a defect". A 400 is genuinely ambiguous: often a caller's bad
+ * argument, but sometimes a query WE built wrong (ted-eu comma-joined its CPV
+ * values into something TED rejected, and that bug was found only because it sat
+ * in `error`). Blanket-classifying 400s as caller mistakes would have hidden it.
+ * A pack that KNOWS which it is should keep saying so explicitly; this helper is
+ * for the 481 that say nothing at all.
+ */
+
+/** Longest upstream explanation we'll pass through. Enough for a real message,
+ *  short enough that an HTML page or a stack trace can't swamp the error. */
+
+const MAX_DETAIL = 300;
+
+/**
+ * Default bound for `fetchWithTimeout` when a pack doesn't state its own.
+ *
+ * 25s mirrors the number `epo-ops` landed on after measuring the real failure:
+ * a degraded upstream that doesn't error, it just never answers, and a Worker
+ * sits in `await fetch()` until ITS OWN execution budget kills the request —
+ * which can take minutes, not seconds (epo_ops_search_patents measured 4-8
+ * MINUTE hangs before this existed). 25s is short enough that a caller gets a
+ * fast, actionable error instead of holding the connection, and long enough
+ * that it doesn't false-trip on a merely-slow-but-alive upstream.
+ */
+const DEFAULT_FETCH_TIMEOUT_MS = 25_000;
+
+/**
+ * Read the body of a failed response and fold it into a throwable Error.
+ *
+ * Usage — note the `await`, which is the one thing that makes this a mechanical
+ * change rather than a drop-in:
+ *
+ *     if (!res.ok) throw await httpError(res, 'Unsplash');
+ *
+ * Safe to call on any non-ok response: a body that is missing, empty, unreadable
+ * or HTML degrades to exactly the old `Name: 404` string rather than throwing
+ * something new from inside the error path.
+ */
+async function httpError(res: Response, name: string): Promise<Error> {
+  return new Error(await httpErrorMessage(res, name));
+}
+
+/** The message text without constructing an Error — for packs that need to wrap
+ *  it in their own envelope or add an explicit classification prefix. */
+async function httpErrorMessage(res: Response, name: string): Promise<string> {
+  // The one place a 5xx from a host WE run gets stamped as ours. `res.url` is
+  // the URL the fetch actually resolved to (after redirects), so this is a fact
+  // about the call rather than a guess from the `name` the pack passed in —
+  // reword that label freely, the class does not move. See
+  // internal-host-class.ts; no-op for every third-party upstream, which is why
+  // this touches 481 packs' error text and changes none of it.
+  return markInternalOrigin(
+    `${name}: ${res.status}${detailSuffix(await readDetail(res))}`,
+    res.url,
+    res.status,
+  );
+}
+
+/**
+ * Just the upstream's own explanation — no name, no status.
+ *
+ * For a pack that has already said both in its own sentence. epo-ops reads
+ * `EPO rejected this search as too large (HTTP 413) — ${httpErrorMessage(…)}`,
+ * which rendered as `… (HTTP 413) — EPO: 413.` once the XML detail was being
+ * dropped: the upstream named twice, the status twice, and the one thing EPO
+ * actually said ("Not enough characters before truncation character") nowhere
+ * (fleet #712). Returns '' when the body carries nothing readable, so a caller
+ * can fall back to its own wording.
+ */
+async function upstreamDetail(res: Response): Promise<string> {
+  return readDetail(res);
+}
+
+/**
+ * Read a SUCCESSFUL response as JSON, failing loudly when it isn't JSON.
+ *
+ * `httpError` above only ever runs on `!res.ok`, which leaves the nastier half
+ * of the problem unhandled: an upstream that answers **HTTP 200 with an HTML
+ * page**. A bot wall, a login redirect, a maintenance interstitial and a CDN
+ * error page are all 200s, so `res.ok` is true, and `res.json()` then throws
+ * `Unexpected token '<', "<!DOCTYPE "... is not valid JSON`.
+ *
+ * That string is the problem. It names no upstream, carries no status, and
+ * reads like a parser bug in Pipeworx — so it lands in the `error` tier, which
+ * means "we have a defect", and the caller is told nothing they can act on.
+ * data.govt.nz sat dead behind an Imperva challenge this way and every
+ * status-code health check we own reported it green (7889a845). A zero-length
+ * body has the same shape: `Unexpected end of JSON input`, seen this week on
+ * uk-gazette (83% of external calls) and census.
+ *
+ * UNLIKE `httpError`, this one DOES classify, and the asymmetry is deliberate.
+ * A 400 is genuinely ambiguous — often the caller's bad argument, sometimes a
+ * query we built wrong — so blanket-classifying it would hide our own bugs.
+ * There is no such ambiguity here: **no argument a caller can pass makes a JSON
+ * API return an HTML page.** It is always the upstream, so `upstream_down:` is
+ * a statement of fact rather than a guess, and it keeps these out of the
+ * problem-tools list where they crowd out real defects.
+ *
+ *     const data = await parseJson<Feed>(res, 'UK Gazette');
+ *
+ * Call it only after the `!res.ok` check — on a failed response you want
+ * `httpError`, which mines the body for the upstream's own explanation.
+ */
+async function parseJson<T>(res: Response, name: string): Promise<T> {
+  let raw: string;
+  try {
+    raw = await res.text();
+  } catch {
+    throw new Error(
+      `upstream_down: ${name} returned a body that could not be read (HTTP ${res.status}). ` +
+        'The connection most likely dropped mid-response; retrying is reasonable.',
+    );
+  }
+
+  const type = res.headers.get('content-type') ?? 'no content-type';
+
+  if (!raw.trim()) {
+    throw new Error(
+      `upstream_down: ${name} answered HTTP ${res.status} with an EMPTY body where JSON was expected (${type}). ` +
+        'Nothing about the request can cause this — it is an upstream fault, and the same call may well work on retry.',
+    );
+  }
+
+  // Checked before parsing rather than in the catch, because knowing it is
+  // markup is what turns "we failed to parse something" into "they served a
+  // web page" — the second is diagnosable, the first is not.
+  const head = raw.slice(0, 200).trimStart().toLowerCase();
+  if (head.startsWith('<!doctype') || head.startsWith('<html') || head.startsWith('<?xml')) {
+    const kind = head.startsWith('<?xml') ? 'an XML document' : 'an HTML page';
+    // The summary, not the source. Pasting the first 120 characters of a web
+    // page handed the agent `<!DOCTYPE html><html lang="en"…` — the same leak
+    // this branch exists to describe (fleet #712).
+    throw new Error(
+      `upstream_down: ${name} answered HTTP ${res.status} with ${kind} instead of JSON (${type}). ` +
+        'That is typically a bot wall, a login redirect or a maintenance page — it is returned as a SUCCESS, ' +
+        `so status-code health checks read it as fine. No argument change will get past it. ` +
+        `The page says: ${summarizeErrorBody(raw) || 'nothing readable'}`,
+    );
+  }
+
+  try {
+    return JSON.parse(raw) as T;
+  } catch {
+    throw new Error(
+      `upstream_down: ${name} answered HTTP ${res.status} with a body that is not valid JSON (${type}). ` +
+        `It begins: ${stripMarkup(raw).slice(0, 120) || '(unreadable)'}`,
+    );
+  }
+}
+
+/**
+ * `fetch`, but bounded — the fix for a systemic gap found 2026-08-30: a grep
+ * audit of every pack's `mcps/*\/src/index.ts` found 1,339 of ~1,500 call
+ * `fetch()` with NO timeout guard anywhere in the file. Two of those
+ * (epo-ops, statcan) were confirmed live-hanging for 4-8 minutes before this
+ * existed — every unguarded call carries the same risk, just unconfirmed.
+ *
+ * Mirrors the `epoFetch` wrapper `mcps/epo-ops/src/index.ts` shipped first:
+ * bound the request with `AbortSignal.timeout`, and on a timeout/abort throw
+ * an `upstream_down:` error that names the upstream and the bound rather than
+ * letting the raw `TimeoutError`/`AbortError` (which names neither) propagate.
+ * `upstream_down:` is deliberate, same reasoning as `parseJson` above — no
+ * argument a caller passes can make an upstream hang, so it is always the
+ * upstream's fault, and marking it that way keeps a slow API off the
+ * problem-tools list where it would crowd out our own defects.
+ *
+ * Usage — a mechanical swap for a bare `fetch(url, init)`:
+ *
+ *     const res = await fetchWithTimeout(url, init, 'Some API');
+ *
+ * Pass `timeoutMs` as a fourth argument to override the default for a pack
+ * with a known-slower upstream; the label should be the same short name you'd
+ * pass to `httpError`/`httpErrorMessage` for that call.
+ */
+async function fetchWithTimeout(
+  url: string | URL,
+  init: RequestInit = {},
+  name: string,
+  timeoutMs: number = DEFAULT_FETCH_TIMEOUT_MS,
+): Promise<Response> {
+  try {
+    return await fetch(url, { ...init, signal: AbortSignal.timeout(timeoutMs) });
+  } catch (err) {
+    if (err instanceof Error && (err.name === 'TimeoutError' || err.name === 'AbortError')) {
+      // States the OBSERVATION (no response in N seconds), not a diagnosis.
+      // "appears to be degraded" is an inference about the vendor that we have
+      // not checked, and it is wrong in a way that misdirects whoever reads it:
+      // a timeout from a Worker can equally mean OUR egress is blocked.
+      //
+      // Measured today (2026-09-01, fleet #1047): every call to
+      // mainnet.base.org failed from the x402 facilitator while the identical
+      // request from a laptop returned 200. Base was entirely healthy; the
+      // public RPC refuses Cloudflare Worker egress. Had this message fired
+      // there it would have blamed Base by name, and the next person would have
+      // waited for a vendor outage to clear that did not exist.
+      // A timeout has no status to test — there is no response at all — so
+      // `markInternalOrigin` is called without one: an origin we run that never
+      // answered is an availability failure by definition. This is the half of
+      // fleet #1096 with neither a SQLSTATE nor a status code to key on.
+      throw new Error(
+        markInternalOrigin(
+          `upstream_down: ${name} did not respond within ${timeoutMs / 1000}s. ` +
+            `That can be ${name} being slow or down, or this environment being unable to reach it ` +
+            `(some hosts refuse datacenter/Worker egress) — retry shortly, and check reachability ` +
+            `from elsewhere before concluding ${name} is down.`,
+          url,
+        ),
+      );
+    }
+    throw err;
+  }
+}
+
+function detailSuffix(detail: string): string {
+  return detail ? ` — ${detail}` : '';
+}
+
+async function readDetail(res: Response): Promise<string> {
+  let raw: string;
+  try {
+    raw = await res.text();
+  } catch {
+    // Body already consumed, or the connection died mid-read. The status alone
+    // is still worth throwing — never let the error path throw its own error.
+    return '';
+  }
+  return summarizeErrorBody(raw);
+}
+
+/**
+ * Turn ANY error body — JSON, HTML, XML or plain text — into one short phrase
+ * that never contains markup.
+ *
+ * This used to just drop an HTML or XML body on the floor, on the reasoning
+ * that markup crowds out the status. That was half right. Dropping it loses the
+ * one sentence a caller could have acted on: an `Access Denied` title, an SDMX
+ * `<message:Error>` text, an OPS fault string. A 2026-08-30 support sweep
+ * measured 13 of 291 caller-facing error rows carrying a raw page or document
+ * verbatim, across 11 packs, and in every one of them the useful content —
+ * "Access Denied", "Invalid country code", "SCRAPE_TIMEOUT" — was in there,
+ * buried in markup the agent had to parse out of a string (fleet #712).
+ *
+ * So: extract the meaning, discard the markup. The output is passed through
+ * `stripMarkup` unconditionally, which is what lets `check:error-body-leak`
+ * assert mechanically that no caller-facing message can contain `<?xml`,
+ * `<!DOCTYPE` or `<html`.
+ */
+function summarizeErrorBody(raw: string): string {
+  if (!raw || !raw.trim()) return '';
+
+  const head = raw.slice(0, 400).trimStart().toLowerCase();
+
+  // An HTML error page (Cloudflare interstitial, nginx default, a login
+  // redirect) says what it is in its <title>, and almost nowhere else.
+  if (head.startsWith('<!doctype') || head.startsWith('<html')) {
+    const title = htmlTitle(raw);
+    return title
+      ? `${title} (upstream returned an HTML error page, not an API response)`
+      : 'upstream returned an HTML error page, not an API response';
+  }
+
+  // XML fault documents — EPO OPS, SDMX (`<message:Error>`), SOAP faults. The
+  // human sentence sits in a child element whose tag name says what it is.
+  if (head.startsWith('<?xml') || head.startsWith('<')) {
+    const fault = xmlFaultText(raw);
+    return fault
+      ? `${stripMarkup(fault).slice(0, MAX_DETAIL)} (from the upstream's XML error document)`
+      : 'upstream returned an XML error document with no readable message';
+  }
+
+  // Most JSON error bodies bury one human sentence among ids and echoed request
+  // params. Prefer that sentence; fall back to the whole body when the shape is
+  // unfamiliar, since an unfamiliar shape is exactly when we can least afford to
+  // guess wrong and show nothing.
+  const fromJson = messageFromJson(raw);
+  return stripMarkup(fromJson ?? raw).slice(0, MAX_DETAIL);
+}
+
+/** The `<title>` of an HTML error page, or its first `<h1>` — the two places a
+ *  bot wall, a 502 and an "Access Denied" all state what happened. */
+function htmlTitle(raw: string): string | null {
+  const head = raw.slice(0, 4000);
+  for (const re of [/<title[^>]*>([\s\S]*?)<\/title>/i, /<h1[^>]*>([\s\S]*?)<\/h1>/i]) {
+    const m = re.exec(head);
+    const text = m ? stripMarkup(m[1]) : '';
+    if (text) return text.slice(0, 160);
+  }
+  return null;
+}
+
+/** Tag names that carry the explanation in an XML fault document, namespace
+ *  prefix optional (`<message:Error>`, `<com:Text>`, `<faultstring>`). */
+const XML_FAULT_TAG_RE =
+  /<(?:[A-Za-z0-9_.-]+:)?(?:text|message|description|faultstring|reason|detail|title|errormessage|error)\b[^>]*>([^<]{2,400})</i;
+
+function xmlFaultText(raw: string): string | null {
+  const head = raw.slice(0, 8000);
+  const tagged = XML_FAULT_TAG_RE.exec(head);
+  if (tagged && tagged[1].trim()) return tagged[1];
+
+  // Nothing conventionally named — take the longest text node instead. A fault
+  // document with one sentence in an oddly named element is still readable;
+  // returning nothing at all is not.
+  let best = '';
+  for (const m of head.matchAll(/>([^<>]{8,400})</g)) {
+    const text = m[1].trim();
+    if (text.length > best.length) best = text;
+  }
+  return best || null;
+}
+
+/**
+ * Remove every tag and stray angle bracket, then collapse whitespace.
+ *
+ * Applied to everything on the way out, including the JSON and plain-text
+ * paths, because an upstream is free to embed markup in a JSON string field —
+ * and a leak is a leak regardless of which branch produced it.
+ */
+function stripMarkup(s: string): string {
+  return collapse(decodeEntities(s.replace(/<[^>]*>/g, ' ')).replace(/[<>]/g, ' '));
+}
+
+/** The handful of entities that show up in error-page titles. Decoded AFTER
+ *  tags are stripped and BEFORE the angle-bracket sweep, so `&lt;script&gt;`
+ *  in a title cannot decode into markup that survives — EMBL-EBI's ChEMBL 500
+ *  page renders as `500 Internal Server Error &lt; EMBL-EBI` otherwise. */
+function decodeEntities(s: string): string {
+  return s
+    .replace(/&(?:amp|#0*38);/gi, '&')
+    .replace(/&(?:lt|#0*60);/gi, '<')
+    .replace(/&(?:gt|#0*62);/gi, '>')
+    .replace(/&(?:quot|#0*34);/gi, '"')
+    .replace(/&(?:#0*39|apos|#x0*27);/gi, "'")
+    .replace(/&nbsp;/gi, ' ');
+}
+
+/** The conventional "what went wrong" field, under any of the names upstreams
+ *  actually use. Checked in order; first non-empty string wins. */
+const MESSAGE_KEYS = [
+  'message', 'error_message', 'errorMessage', 'detail', 'details',
+  'description', 'error_description', 'reason', 'title', 'fault',
+];
+
+function messageFromJson(raw: string): string | null {
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(raw);
+  } catch {
+    return null;
+  }
+  return pickMessage(parsed, 0);
+}
+
+function pickMessage(node: unknown, depth: number): string | null {
+  // Two levels covers `{error: {message}}` and `{errors: [{detail}]}`, the two
+  // shapes that account for nearly all of them, without walking a large payload.
+  if (depth > 2 || node == null) return null;
+
+  if (typeof node === 'string') return node.trim() || null;
+
+  if (Array.isArray(node)) {
+    for (const item of node) {
+      const found = pickMessage(item, depth + 1);
+      if (found) return found;
+    }
+    return null;
+  }
+
+  if (typeof node !== 'object') return null;
+  const obj = node as Record<string, unknown>;
+
+  for (const key of MESSAGE_KEYS) {
+    const v = obj[key];
+    if (typeof v === 'string' && v.trim()) return v.trim();
+  }
+  // `{error: …}` where error is itself an object or a string — the single most
+  // common wrapper, so it is worth descending into by name rather than scanning
+  // every key and risking picking up an echoed request parameter.
+  for (const key of ['error', 'errors', 'fault', 'Error', 'data']) {
+    if (key in obj) {
+      const found = pickMessage(obj[key], depth + 1);
+      if (found) return found;
+    }
+  }
+  return null;
+}
+
+/** Errors are read in a single line of log output; newlines and runs of
+ *  whitespace make a multi-line body unreadable there. */
+function collapse(s: string): string {
+  return s.replace(/\s+/g, ' ').trim();
+}
+/**
+ * arb-alerts — persisted Polymarket/Kalshi cross-venue spread alerts.
+ *
+ * Companion to the live polymarket_kalshi_spread tool. Where that
+ * tool reports a snapshot of the current spread for ONE topic, this
+ * pack exposes the TIME DIMENSION: every alert that has fired (spread
+ * crossed threshold) across all watched topics, and per-topic history
+ * to see when spreads widen and converge.
+ *
+ * Data lives in arb_alerts (supabase/migrations/032). A cron in the
+ * scraper worker polls every 4h during US market hours, writes a row
+ * when a topic's spread crosses its configured threshold. These MCP
+ * tools query the table.
+ *
+ * arb_alerts only gets a row when a spread CROSSES threshold, so a quiet
+ * market and a dead collector both render as a clean empty response — the
+ * defect fleet #2041 exists for (docs/silent-zero-policy.md). Every response
+ * from arb_recent_alerts / arb_topic_history now carries a `collector`
+ * block (from migration 204's arb_scan_runs log, which gets one row per
+ * cron firing regardless of whether anything fired) so an empty result can
+ * say WHY it is empty instead of just being empty.
+ */
+
+
+// Bound every fetch() in this pack to a fixed timeout — an upstream that
+// degrades without erroring would otherwise hold the Worker in `await fetch()`
+// until its own execution budget kills the request (minutes, not seconds).
+// Mirrors the epoFetch / usaspending retryFetch pattern (fleet #685).
+async function pwFetch(url: string | URL, init?: RequestInit): Promise<Response> {
+  return fetchWithTimeout(url, init ?? {}, 'arb-alerts');
+}
+
+interface AlertRow {
+  topic: string;
+  fired_at: string;
+  spread_bps: number;
+  kalshi_yes_price: number | null;
+  polymarket_yes_price: number | null;
+  cheap_venue: string;
+  kalshi_market_ticker?: string | null;
+  polymarket_market_slug?: string | null;
+  description?: string;
+  threshold_at_fire_bps?: number;
+}
+
+interface WatchlistRow {
+  topic: string;
+  description: string;
+  threshold_bps: number;
+  enabled: boolean;
+}
+
+interface CollectorStatusRow {
+  last_collected_at: string | null;
+  runs_in_window: number;
+  legs_evaluated_in_window: number;
+  alerts_fired_in_window: number;
+}
+
+interface TopicCollectorStatusRow {
+  last_collected_at: string | null;
+  topic_last_scanned_at: string | null;
+  runs_in_window: number;
+  legs_evaluated_in_window: number;
+  max_spread_bps_in_window: number;
+}
+
+// Human-readable collector health, computed the same way for both tools so
+// "quiet market" and "dead collector" read the same regardless of which one
+// you called. The scan cron fires 3x/day on market days (13/17/21 UTC,
+// Mon-Fri) — 26h is a little over one weekday's worth of slack for a single
+// missed firing before calling it stale, and generous over a weekend gap.
+function collectorHealth(lastCollectedAt: string | null): { status: 'healthy' | 'stale' | 'never_run'; note: string } {
+  if (!lastCollectedAt) {
+    return { status: 'never_run', note: 'No arb_scan_runs row exists yet — the collector has never logged a run since migration 204 shipped, or the table is empty for another reason. Not the same as "dead": check deploy history before assuming outage.' };
+  }
+  const ageHours = (Date.now() - new Date(lastCollectedAt).getTime()) / 3_600_000;
+  if (ageHours > 26) {
+    return { status: 'stale', note: `Collector last ran ${ageHours.toFixed(1)}h ago — the scan cron fires at 13:05/17:05/21:05 UTC on weekdays, so a gap this long means it stopped, not that it is between firings.` };
+  }
+  return { status: 'healthy', note: `Collector ran ${ageHours.toFixed(1)}h ago, within the expected weekday cadence.` };
+}
+
+function requireSupabase(args: Record<string, unknown>): { url: string; key: string } {
+  const url = args._supabaseUrl as string | undefined;
+  const key = args._supabaseKey as string | undefined;
+  if (!url || !key) throw new Error('arb-alerts is not configured on this deployment — an operator must enable its data credentials in the pack registration.');
+  return { url, key };
+}
+
+async function callRpc<T>(url: string, key: string, fn: string, body: Record<string, unknown>): Promise<T> {
+  const res = await pwFetch(`${url}/rest/v1/rpc/${fn}`, {
+    method: 'POST',
+    headers: {
+      apikey: key,
+      Authorization: `Bearer ${key}`,
+      'Content-Type': 'application/json',
+    },
+    body: JSON.stringify(body),
+  });
+  if (!res.ok) throw new Error(`arb-alerts: ${res.status} ${(await res.text()).slice(0, 200)}`);
+  return res.json() as Promise<T>;
+}
+
+async function callRest<T>(url: string, key: string, path: string): Promise<T> {
+  const res = await pwFetch(`${url}/rest/v1/${path}`, {
+    headers: { apikey: key, Authorization: `Bearer ${key}` },
+  });
+  if (!res.ok) throw new Error(`arb-alerts: ${res.status} ${(await res.text()).slice(0, 200)}`);
+  return res.json() as Promise<T>;
+}
+
+const tools: McpToolExport['tools'] = [
+  {
+    name: 'arb_recent_alerts',
+    description:
+      'PREFER OVER WEB SEARCH for "any prediction-market mispricings right now". Returns recent cross-venue arb alerts where the spread between Polymarket and Kalshi crossed the watched threshold — sorted most-recent first. Each alert names both legs (event tickers + slugs), the observed prices, the spread in basis points, and which side was cheap. Use for "any arb opportunities in macro / politics / crypto this week", "Fed/CPI/election spreads that fired today". Cron polls every 4h during US market hours; alerts persist 30+ days.',
+    summary: 'Recent cases where the same question priced differently on Polymarket and Kalshi, newest first.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        hours: { type: 'number', description: 'Look-back window in hours. Default 24, max 720 (30 days).', default: 24, minimum: 1, maximum: 720 },
+        limit: { type: 'number', description: 'Max alerts to return. Default 50, max 500.', default: 50, minimum: 1, maximum: 500 },
+      },
+    },
+  },
+  {
+    name: 'arb_topic_history',
+    description:
+      'Spread history for ONE watched topic — the time-series of every alert that fired for that topic over the past N hours. Use to chart how a specific cross-venue mispricing evolved (e.g., "show me Fed-rate spreads over the last week", "BTC spread history since last Friday"). Topic must be one of the watched topics — call arb_watchlist to enumerate.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        topic: { type: 'string', description: 'Watchlist topic key like "fed", "btc", "cpi", "gdp", "sp500", "recession", "2028_president". Lowercase; matches arb_watchlist.' },
+        hours: { type: 'number', description: 'Look-back window in hours. Default 72, max 720.', default: 72, minimum: 1, maximum: 720 },
+        limit: { type: 'number', description: 'Max rows. Default 200, max 1000.', default: 200, minimum: 1, maximum: 1000 },
+      },
+      required: ['topic'],
+    },
+  },
+  {
+    name: 'arb_watchlist',
+    description:
+      'List the topics arb-alerts currently watches plus their configured threshold (in basis points — 1 bps = 0.01 cents of implied probability). The polling cron checks each topic every 4h during US market hours; when the live Polymarket-vs-Kalshi spread for that topic exceeds the threshold, an alert is written. Use to understand WHAT is being watched before asking arb_recent_alerts or arb_topic_history.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        enabled_only: { type: 'boolean', description: 'If true (default), only show currently-enabled topics.', default: true },
+      },
+    },
+  },
+];
+
+async function callTool(name: string, args: Record<string, unknown>): Promise<unknown> {
+  const { url, key } = requireSupabase(args);
+
+  switch (name) {
+    case 'arb_recent_alerts': {
+      const hours = Math.min(720, Math.max(1, (args.hours as number) ?? 24));
+      const limit = Math.min(500, Math.max(1, (args.limit as number) ?? 50));
+      const [rows, statusRows] = await Promise.all([
+        callRpc<AlertRow[]>(url, key, 'arb_recent_alerts', { p_hours: hours, p_limit: limit }),
+        callRpc<CollectorStatusRow[]>(url, key, 'arb_collector_status', { p_hours: hours }).catch(() => []),
+      ]);
+      const status = statusRows[0] ?? null;
+      const health = collectorHealth(status?.last_collected_at ?? null);
+      return {
+        hours,
+        count: rows.length,
+        alerts: rows,
+        // Present on EVERY response, including count:0 — this is what tells
+        // "the venues agreed, nothing to alert" apart from "the pipeline has
+        // been dead since May" (fleet #2041). observations_in_window is the
+        // number of matched Kalshi/Polymarket legs the collector actually
+        // evaluated across ALL watched topics in this window, whether or not
+        // any of them crossed threshold.
+        collector: {
+          last_collected_at: status?.last_collected_at ?? null,
+          runs_in_window: status?.runs_in_window ?? 0,
+          observations_in_window: status?.legs_evaluated_in_window ?? 0,
+          alerts_fired_in_window: status?.alerts_fired_in_window ?? 0,
+          status: health.status,
+          note: health.note,
+        },
+      };
+    }
+
+    case 'arb_topic_history': {
+      const topic = String(args.topic ?? '').toLowerCase().trim();
+      if (!topic) throw new Error('topic is required (e.g., "fed", "btc")');
+      const hours = Math.min(720, Math.max(1, (args.hours as number) ?? 72));
+      const limit = Math.min(1000, Math.max(1, (args.limit as number) ?? 200));
+      const [rows, statusRows] = await Promise.all([
+        callRpc<AlertRow[]>(url, key, 'arb_topic_history', { p_topic: topic, p_hours: hours, p_limit: limit }),
+        callRpc<TopicCollectorStatusRow[]>(url, key, 'arb_topic_collector_status', { p_topic: topic, p_hours: hours }).catch(() => []),
+      ]);
+      const status = statusRows[0] ?? null;
+      // Collector-wide health (is the whole pipeline alive) is a different
+      // question from topic-specific health (was THIS topic ever matched to
+      // a market on both venues) — surface both rather than collapsing them,
+      // since "collector healthy but this topic has 0 legs every run" points
+      // at a resolver/matching bug in polymarket_kalshi_spread, not an outage.
+      const health = collectorHealth(status?.last_collected_at ?? null);
+      return {
+        topic,
+        hours,
+        count: rows.length,
+        history: rows,
+        collector: {
+          last_collected_at: status?.last_collected_at ?? null,
+          topic_last_scanned_at: status?.topic_last_scanned_at ?? null,
+          runs_in_window: status?.runs_in_window ?? 0,
+          observations_in_window: status?.legs_evaluated_in_window ?? 0,
+          max_spread_bps_in_window: status?.max_spread_bps_in_window ?? 0,
+          status: health.status,
+          note: status && status.topic_last_scanned_at === null
+            ? `${health.note} This topic specifically has never been matched to legs by the collector — check it exists and is enabled in arb_watchlist, and that polymarket_kalshi_spread resolves it to markets on both venues.`
+            : health.note,
+        },
+      };
+    }
+
+    case 'arb_watchlist': {
+      const enabledOnly = args.enabled_only !== false;
+      const path = enabledOnly
+        ? 'arb_topic_watchlist?enabled=eq.true&order=topic.asc'
+        : 'arb_topic_watchlist?order=topic.asc';
+      const rows = await callRest<WatchlistRow[]>(url, key, path);
+      return { count: rows.length, watchlist: rows };
+    }
+
+    default:
+      throw new Error(`Unknown tool: ${name}`);
+  }
+}
+
+export default { tools, callTool, meter: { credits: 2 } } satisfies McpToolExport;

package/src/server.ts

@@ -0,0 +1,45 @@
+/**
+ * Stdio MCP server entry point for @pipeworx/mcp-arb-alerts.
+ * Generated by scripts/publish-pack.sh — do not hand-edit in the pack repo;
+ * edit scripts/publish-pack.sh (the server.ts heredoc) and republish instead.
+ */
+import { Server } from '@modelcontextprotocol/sdk/server/index.js';
+import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
+import { CallToolRequestSchema, ListToolsRequestSchema } from '@modelcontextprotocol/sdk/types.js';
+import pack from './index.js';
+
+const server = new Server(
+  { name: '@pipeworx/mcp-arb-alerts', version: '0.1.0' },
+  { capabilities: { tools: {} } },
+);
+
+server.setRequestHandler(ListToolsRequestSchema, async () => ({
+  tools: pack.tools.map((t) => ({
+    name: t.name,
+    description: t.description,
+    inputSchema: t.inputSchema,
+  })),
+}));
+
+server.setRequestHandler(CallToolRequestSchema, async (request) => {
+  const { name, arguments: args } = request.params;
+  try {
+    const result = await pack.callTool(name, (args ?? {}) as Record<string, unknown>);
+    return { content: [{ type: 'text', text: JSON.stringify(result, null, 2) }] };
+  } catch (err) {
+    return {
+      content: [{ type: 'text', text: err instanceof Error ? err.message : String(err) }],
+      isError: true,
+    };
+  }
+});
+
+async function main() {
+  const transport = new StdioServerTransport();
+  await server.connect(transport);
+}
+
+main().catch((err) => {
+  console.error('Fatal error running server:', err);
+  process.exit(1);
+});

package/tsconfig.json

@@ -0,0 +1,18 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "ESNext",
+    "moduleResolution": "bundler",
+    "lib": ["ES2022"],
+    "types": ["@cloudflare/workers-types"],
+    "strict": true,
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "resolveJsonModule": true,
+    "outDir": "dist",
+    "rootDir": "src",
+    "declaration": true
+  },
+  "include": ["src"],
+  "exclude": ["src/server.ts"]
+}