@general-liquidity/sharpebench-mcp 0.1.0 → 0.2.0

package/dist/server.js

@@ -7,6 +7,7 @@
  * Every tool is read-only and deterministic (the kernel has no I/O), so the
  * server is safe to expose without sandboxing.
  */
+import { createRequire } from "node:module";
 import { fileURLToPath } from "node:url";
 import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
@@ -22,9 +23,21 @@ function run(fn) {
         return { content: [{ type: "text", text: `error: ${message}` }], isError: true };
     }
 }
+// The server's version is the package's version, read from package.json at
+// load time so a release bump (cargo-release rewrites npm/mcp/package.json from
+// the workspace version) is the single source of truth and nothing here drifts.
+const { version: PACKAGE_VERSION } = createRequire(import.meta.url)("../package.json");
+function kernelRegimeCompare() {
+    const fn = sb.regimeCompare;
+    if (typeof fn !== "function") {
+        throw new Error("regime_compare is not exported by this build of @general-liquidity/sharpebench; " +
+            "use `sharpebench regime <a.csv> <b.csv> <regimes.csv>` from the CLI");
+    }
+    return fn;
+}
 /** Build the SharpeBench MCP server with all kernel tools registered. */
 export function createServer() {
-    const server = new McpServer({ name: "sharpebench", version: "0.0.3" });
+    const server = new McpServer({ name: "sharpebench", version: PACKAGE_VERSION });
     server.tool("score", "Rank a field of agent submissions on the luck-robust composite (deflated Sharpe + pass^k + process discipline). Raw return is reported but is never the rank key. Returns ranked CompositeScore[].", { submissions: z.array(z.any()), config: z.any().optional() }, async ({ submissions, config }) => run(() => sb.score(submissions, config)));
     server.tool("score_agent", "Score a single submission → one CompositeScore (deflated Sharpe, pass^k verdict, process score, rolling worst-case Sharpe).", { submission: z.any(), config: z.any().optional() }, async ({ submission, config }) => run(() => sb.scoreAgent(submission, config)));
     server.tool("self_audit", "Fire the known gaming attacks at the scorer and report whether each is demoted (the benchmark's anti-gaming proof). No input.", async () => run(() => sb.selfAudit()));
@@ -53,6 +66,18 @@ export function createServer() {
         borderline,
         srBenchmark: sr_benchmark,
     })));
+    server.tool("regime_compare", "Compare two strategies' per-period returns WITHIN each market regime rather than pooled: per regime, the zero/no-trade mass vs the continuous part (ZAGA split), mean/sd/median, sign share, moment-matched gamma on magnitudes, a two-sample KS statistic, and whether the pooled mean gap hides a sign reversal. Regime labels are an input (one per period); nothing is inferred and no GAMLSS is fitted.", {
+        returns_a: z.array(z.number()),
+        returns_b: z.array(z.number()),
+        regimes: z.array(z.string()),
+        zero_tol: z.number().optional(),
+        min_periods: z.number().optional(),
+        tie_tol: z.number().optional(),
+    }, async ({ returns_a, returns_b, regimes, zero_tol, min_periods, tie_tol }) => run(() => kernelRegimeCompare()(returns_a, returns_b, regimes, {
+        zeroTol: zero_tol,
+        minPeriods: min_periods,
+        tieTol: tie_tol,
+    })));
     return server;
 }
 async function main() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@general-liquidity/sharpebench-mcp",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "MCP server exposing the SharpeBench luck-robust scoring kernel as agent-callable tools (deflated Sharpe, pass^k, process discipline, briefing audit, options Greeks).",
   "keywords": [
     "mcp",
@@ -30,7 +30,7 @@
     "prepublishOnly": "tsc"
   },
   "dependencies": {
-    "@general-liquidity/sharpebench": "^0.1.0",
+    "@general-liquidity/sharpebench": "^0.2.0",
     "@modelcontextprotocol/sdk": "^1.0.0",
     "zod": "^4.0.0"
   },