@opennous/mcp 0.60.0 → 0.61.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@opennous/mcp",
-  "version": "0.60.0",
+  "version": "0.61.0",
   "description": "Nous — the Context Graph for AI Agents.",
   "license": "Apache-2.0",
   "repository": {

package/src/server.js

@@ -105,15 +105,18 @@ export function createServer() {
     ],
   });
 
-  // Nous is fully a revenue plugin: this server exposes EXACTLY the 7 primitives and
+  // Nous is fully a revenue plugin: this server exposes EXACTLY these primitives and
   // nothing else (the ~30 legacy tools were pruned). Writes: record · record_insight.
   // Reads: get_context · get_account · query (search folds in via facts:true) · score.
-  // Identity: whoami (who the key acts as, scope, GTM role[s]) — the governance
-  // substrate. record_signal folds into record; save_note is out. The `tool()` guard
-  // keeps the surface locked to these eight — a stray non-primitive registration is
-  // dropped, not exposed. See docs/revenue-plugin/README.md §2–3.
+  // ICP: set_icp writes the hypothesis; record_closed_deals trains it on real
+  // won/lost outcomes (contrastive lift). Identity: whoami (who the key acts as,
+  // scope, GTM role[s]) — the governance substrate. record_signal folds into record;
+  // save_note is out. The `tool()` guard keeps the surface locked to these nine — a
+  // stray non-primitive registration is dropped, not exposed. See
+  // docs/revenue-plugin/README.md §2–3.
   const PLUGIN_TOOLS = new Set([
     "record", "record_insight", "get_context", "get_account", "query", "score", "whoami", "set_icp",
+    "record_closed_deals",
   ]);
   const _tool = server.tool.bind(server);
   const tool = (name, ...rest) => {
@@ -484,6 +487,57 @@ export function createServer() {
     }
   );
 
+  // ===========================================================================
+  // TOOL: record_closed_deals  —  POST /v2/workspace/closed-deals
+  // Train the ICP on REAL outcomes. set_icp writes the hypothesis; this upgrades
+  // it to an outcome-graded model: feed the won and lost domains (the CRM/Stripe
+  // closed cohorts), the engine runs contrastive lift — what separates wins from
+  // losses — links known contacts, resolves their open predictions with the real
+  // outcome, and re-scores every open account. The onboarding + win-loss payoff.
+  // ===========================================================================
+  tool(
+    "record_closed_deals",
+    "Train this workspace's ICP on REAL closed deals. Pass the won and lost deals (each a bare " +
+    "domain, or an object {domain, amount, currency, closed_at}). The engine enriches each domain, " +
+    "links the contacts you already have there, resolves their open ICP predictions with the real " +
+    "outcome, runs contrastive lift (the signals that separate wins from losses), and re-scores every " +
+    "open account against the upgraded model. Use it in onboarding once the backfill has landed the " +
+    "closed cohorts (pull them with query scope.property:'stage' — accounts at 'closed_won' and " +
+    "'closed_lost'), and whenever new deals close. Needs at least one deal; both sides (some won AND " +
+    "some lost) make the lift trustworthy — one side alone is directional only. The ICP is the one " +
+    "company model, so this is an admin/founder action.",
+    {
+      won:  z.array(CLOSED_DEAL).optional().describe("Closed-WON deals — bare domains or {domain, amount, currency, closed_at}. Pull from accounts at stage 'closed_won'."),
+      lost: z.array(CLOSED_DEAL).optional().describe("Closed-LOST deals — same shape. Pull from accounts at stage 'closed_lost'."),
+    },
+    async ({ won = [], lost = [] }) => {
+      const r = await post("/v2/workspace/closed-deals", { won, lost });
+      if (r.need_more_deals || r.error === "need_more_deals") {
+        return { content: [{ type: "text", text:
+          "Need at least one closed deal (a won or lost domain) to train the ICP. Pull the closed cohorts with query scope.property:'stage', then pass their domains." }] };
+      }
+      const lines = [
+        `ICP trained on real outcomes — ${r.won || 0} won · ${r.lost || 0} lost` +
+        (r.enriched ? ` · ${r.enriched} companies enriched` : "") +
+        (r.surfaced ? ` · ${r.surfaced} predictions resolved` : "") + ".",
+      ];
+      if ((r.discovered || []).length) {
+        lines.push("", "Signals the model learned (contrastive lift, won vs lost):");
+        for (const d of r.discovered) {
+          const dir = (d.weight ?? 0) > 0 ? "+" : "";
+          lines.push(`  • ${d.label}  [${dir}${d.weight}]${d.note ? ` — ${d.note}` : ""}`);
+        }
+      }
+      if ((r.linked || []).length) {
+        const names = r.linked.map(l => l.name).filter(Boolean);
+        lines.push("", `Linked ${r.linked.length} known contact${r.linked.length !== 1 ? "s" : ""} to these deals` +
+          (names.length ? `: ${names.slice(0, 8).join(", ")}${names.length > 8 ? "…" : ""}` : "") + ".");
+      }
+      lines.push("", "Open accounts have been re-scored. score / get_context now reflect the outcome-graded ICP.");
+      return { content: [{ type: "text", text: lines.join("\n") }] };
+    }
+  );
+
   // ===========================================================================
   // TOOL: query  —  POST /v2/query
   // Retrieve a corpus of activity across many people. You do the analysis.