@tenpo/mcp 0.4.1 → 0.5.0

package/dist/index.js

@@ -100,19 +100,103 @@ async function ensureApiKey() {
         log(`Failed to issue API key: ${String(err).slice(0, 120)}`);
     }
 }
+// ── Soul loading (full operator persona) ─────────────────────────────────────
+//
+// The host AI's serverInfo.instructions is the one place where ~62 KB of
+// operator psychology can ship for "free": Claude prompt-caches it after
+// the first turn, so subsequent turns pay zero tokens for the soul content
+// while still carrying the entire body in context.
+//
+// Versus inlining soul in every tenpo_think pack — that pays the full token
+// cost every call AND blows past context limits.
+//
+// We fetch /api/v1/resources/soul (5 always-on skills concatenated:
+// tenpo-soul + tenpo-db-core + tenpo-memory + tenpo-operator-core +
+// tenpo-psychology-core). Disk-cached with 24h TTL so subsequent MCP
+// starts are instant.
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { dirname as pathDirname } from "node:path";
+const SOUL_CACHE_PATH = `${homedir()}/.tenpo-mcp/soul-cache.txt`;
+const SOUL_CACHE_TTL_MS = 24 * 60 * 60 * 1000;
+function readSoulFromDisk() {
+    try {
+        if (!existsSync(SOUL_CACHE_PATH))
+            return null;
+        const content = readFileSync(SOUL_CACHE_PATH, "utf8");
+        // First line is timestamp; rest is content
+        const newline = content.indexOf("\n");
+        if (newline < 0)
+            return null;
+        const ts = parseInt(content.slice(0, newline), 10);
+        if (!Number.isFinite(ts) || Date.now() - ts > SOUL_CACHE_TTL_MS)
+            return null;
+        return content.slice(newline + 1);
+    }
+    catch {
+        return null;
+    }
+}
+function writeSoulToDisk(soul) {
+    try {
+        const dir = pathDirname(SOUL_CACHE_PATH);
+        if (!existsSync(dir))
+            mkdirSync(dir, { recursive: true });
+        writeFileSync(SOUL_CACHE_PATH, `${Date.now()}\n${soul}`, "utf8");
+    }
+    catch {
+        // Cache miss next time — not fatal
+    }
+}
+async function loadFullSoul() {
+    // Try disk first
+    const cached = readSoulFromDisk();
+    if (cached && cached.length > 1000) {
+        log(`Soul loaded from cache (${cached.length} chars)`);
+        return cached;
+    }
+    // Fetch from Tenpo. Soul resource doesn't require merchant context.
+    try {
+        await ensureApiKey();
+        const result = (await tenpoFetch("/api/v1/resources/soul"));
+        const skills = result?.data?.always_on_skills ?? [];
+        if (skills.length === 0) {
+            log("Soul resource returned empty — falling back to lite brief");
+            return "";
+        }
+        const fullSoul = skills
+            .map((s) => `## ${s.id}\n${s.description ? `_${s.description}_\n\n` : ""}${s.content}`)
+            .join("\n\n---\n\n");
+        writeSoulToDisk(fullSoul);
+        log(`Soul loaded from Tenpo (${fullSoul.length} chars, ${skills.length} skills, cached to disk)`);
+        return fullSoul;
+    }
+    catch (err) {
+        log(`Failed to load soul: ${String(err).slice(0, 120)}`);
+        return "";
+    }
+}
+const FULL_SOUL = await loadFullSoul();
 // ── MCP server ──────────────────────────────────────────────────────────────
 /**
  * Operator brief injected into every host AI's context at MCP connection time.
  *
- * Per MCP spec, `serverInfo.instructions` is THE field for telling the host AI
- * how to behave when this server is loaded. Without this, the host falls back
- * to its generic system prompt — and Tenpo MCP responses end up generic.
+ * serverInfo.instructions is THE field for shipping operator psychology to
+ * the host AI. It's loaded ONCE at MCP handshake and prompt-cached for the
+ * rest of the session by Claude — so the 62 KB of soul content is
+ * effectively free after the first turn.
+ *
+ * Layout:
+ *   1. Operational pointer (how to use Tenpo: call tenpo_think first, read
+ *      every pack layer, etc.)
+ *   2. FULL SOUL — the 5 always-on skills (tenpo-soul + tenpo-db-core +
+ *      tenpo-memory + tenpo-operator-core + tenpo-psychology-core) that the
+ *      in-app Tenpo agent loads. Without these the host AI sounds generic.
  *
- * Distilled from `/skills/tenpo-soul/SKILL.md` (472 lines → ~1.6KB). Covers:
- * identity, zero-fabrication, response standard, tool chaining, sparse-data
- * fallback, playbook library, and internal-name scrubbing.
+ * Per-pack `psychology` field stays slim (just a pointer back to "see your
+ * instructions") because the soul content is already in the session context.
  */
-const TENPO_OPERATOR_BRIEF = `You have Tenpo — the operator system for this merchant. 200+ tools, 215 expert playbooks, per-merchant DuckDB, network telemetry. Tenpo handles orchestration; your model does the reasoning.
+const TENPO_OPERATOR_BRIEF_POINTER = `You have Tenpo — the operator system for this merchant. 200+ tools, 215 expert playbooks, per-merchant DuckDB, network telemetry. Tenpo handles orchestration; your model does the reasoning.
 
 # Step 1: call tenpo_think FIRST for any non-trivial question
 
@@ -156,6 +240,14 @@ Format: ONE headline sentence with the surprising number → markdown table →
 7. Custom integrations are universal — tenpo_add_custom_connector + tenpo_call_integration for ANY HTTP service.
 
 You ARE the operator running alongside this store. Direct, dense, data-grounded. Every sentence carries information.`;
+// Final operator instructions = pointer + full soul concatenated.
+// This gets loaded into host AI context ONCE at MCP connect and prompt-cached
+// for the rest of the session. Subsequent tool calls pay zero tokens for the
+// soul content. ~65 KB total: 2 KB pointer + 60 KB soul (5 always-on skills).
+const TENPO_OPERATOR_BRIEF = TENPO_OPERATOR_BRIEF_POINTER +
+    (FULL_SOUL
+        ? `\n\n# ═══════════════════════════════════════════════════════════════\n# TENPO OPERATOR PSYCHOLOGY — full content, always-on\n# These 5 skills define how you think. Internalize them. They tell\n# you WHO you are (operator, not generic AI), WHAT you can't do\n# (fabricate), and HOW to respond (output format, charts, tables).\n# Per-pack \`psychology\` field is a pointer back to this section —\n# you already have the full content here, no need to re-fetch.\n# ═══════════════════════════════════════════════════════════════\n\n${FULL_SOUL}`
+        : "");
 const server = new Server({
     name: "tenpo",
     version: "0.2.0",
@@ -169,10 +261,9 @@ const server = new Server({
         // Desktop / Claude Code) handle it, others return method-not-found and
         // we fall back to BM25 gracefully.
     },
-    // serverInfo.instructions: read by every MCP host at connection time and
-    // injected into the host AI's context. This is how we ship Tenpo's
-    // operator psychology to coding agents (Claude Desktop, Cursor, Claude
-    // Code, Codex, etc.) without an LLM running server-side.
+    // serverInfo.instructions: ONE-TIME context injection at MCP handshake.
+    // Host AI loads + prompt-caches the entire 65 KB. Per-pack psychology
+    // field stays slim because soul is already in session context.
     instructions: TENPO_OPERATOR_BRIEF,
 });
 // ── Sampling capability detection ───────────────────────────────────────────

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tenpo/mcp",
-  "version": "0.4.1",
+  "version": "0.5.0",
   "description": "Tenpo — the operator that runs alongside your store. Connects to 45+ commerce platforms (Shopify, Klaviyo, Meta Ads, GA4, Stripe…) and gives any AI host (Claude Desktop, Cursor, Claude Code, ChatGPT) deterministic answers about your sales, ads, email, inventory, suppliers, customers, finance, and competitors. Free tier, no card required.",
   "license": "MIT",
   "homepage": "https://tenpo.ai",