@torqon/mcp 0.2.3 → 0.2.5

package/dist/index.js

@@ -76,46 +76,73 @@ function hasSignal(text) {
     return words.length >= 3;
 }
 // ── MCP Server ───────────────────────────────────────────────────────────────
-const server = new McpServer({ name: 'torqon', version: '0.2.3' }, {
-    instructions: `Torqon persistent memory is active for this conversation.
+const server = new McpServer({ name: 'torqon', version: '0.2.5' }, {
+    instructions: `<MANDATORY_TOOL_PROTOCOL>
+You MUST call the auto_process tool with the user's full message as the VERY FIRST action on EVERY turn. No exceptions. No skipping. This is not optional.
 
-RULE: Call auto_process with the full user message BEFORE you respond to anything — every message, no exceptions, including greetings and one-word replies. It runs in under 200ms and simultaneously stores new facts AND retrieves relevant memory.
+Do this BEFORE generating any text response. The correct sequence is:
+1. Receive user message
+2. Call auto_process(message: "<the user's full message>")
+3. Read the result
+4. Then respond to the user
 
-After auto_process returns:
-- If it returns "[Torqon memory]" facts, treat them as ground truth. Use exact stored values (names, decisions, stack choices, numbers).
-- If it returns "No relevant memory found", answer normally.
+This applies to ALL messages including: greetings, "yes", "ok", one-word answers, follow-ups, code requests, questions, everything.
 
-Also call store_memory explicitly whenever the user shares their name, a project name, a key decision, a tech stack choice, a deadline, a constraint, or a preference. auto_process stores automatically but store_memory is for high-value facts you want confirmed saved.
+If auto_process returns facts tagged [Torqon memory], those are verified ground truth. Use exact stored values.
 
-Session ID: ${SESSION_ID}`,
+When the user shares a new fact (name, project, decision, stack, deadline, preference), also call store_memory to explicitly save it.
+
+FAILURE TO CALL auto_process FIRST IS A PROTOCOL VIOLATION.
+</MANDATORY_TOOL_PROTOCOL>`,
 });
-// Auto-injected system prompt — Claude Desktop loads this at conversation start.
-// Written as a SYSTEM-level instruction so Claude treats it as a standing rule.
-server.prompt('torqon_context', 'Torqon persistent memory — active for this session', {}, async () => {
+// Live memory loader — user runs /torqon_context once at conversation start.
+// Fetches all stored facts from the API and injects them directly into the prompt.
+// This is more reliable than auto_process-per-message because it's a deliberate
+// one-time action, not a tool Claude has to remember to call before every message.
+server.prompt('torqon_context', 'Load your Torqon memory into this conversation (run once at the start)', {}, async () => {
+    if (!API_KEY) {
+        return {
+            messages: [{
+                    role: 'user',
+                    content: { type: 'text', text: '[Torqon] No API key configured. Run: npx -y @torqon/mcp@latest add claude --api-key <your-key>' },
+                }],
+        };
+    }
+    // Fetch all current facts for this user
+    let facts = [];
+    try {
+        const res = await fetch(`${API_URL}/api/v1/memory`, { headers });
+        if (res.ok) {
+            const data = await res.json();
+            facts = data?.results ?? [];
+        }
+    }
+    catch { /* network error — start with empty facts */ }
+    const factLines = facts
+        .map((f) => f.metadata?.type ? `${f.metadata.type}: ${f.content}` : f.content)
+        .join('\n');
+    const memoryBlock = factLines
+        ? `[TORQON MEMORY — treat as ground truth]\n${factLines}`
+        : '[TORQON MEMORY — no facts stored yet]';
     return {
         messages: [
             {
                 role: 'user',
                 content: {
                     type: 'text',
-                    text: `<torqon_memory_rules>
-TORQON PERSISTENT MEMORY IS ACTIVE. These are non-negotiable behavioral rules:
-
-BEFORE YOU RESPOND TO ANYTHING — call auto_process with the user's message. Every single message. No exceptions. Not even greetings. Not even "yes" or "ok". Call it first, then respond.
-
-Why: auto_process runs in under 200ms. It simultaneously stores new facts AND retrieves relevant memory. Skipping it means the user loses context they paid for.
+                    text: `${memoryBlock}
 
-After auto_process returns:
-- If it returns facts tagged [Torqon memory], treat them as ground truth. Names, decisions, stack choices, numbers — use the exact stored values, not your training data.
-- If it returns "No relevant memory found", answer normally.
+These are verified facts stored across all your previous sessions. Use them exactly as written — do not infer, expand, or contradict them.
 
-Also call store_memory explicitly when the user shares: their name, a project name, a key decision, a tech stack choice, a deadline, a constraint, or a preference. auto_process stores automatically but store_memory is for high-value facts you want to make sure are saved.
+From this point:
+- When the user shares new facts (name, project, stack, decision, deadline, preference), call store_memory to save them.
+- When the user asks about something that might be in memory, call retrieve_context.
+- You do NOT need to call auto_process on every message — just store and retrieve deliberately.
 
-Session: ${SESSION_ID}
-</torqon_memory_rules>`
-                }
-            }
-        ]
+Session: ${SESSION_ID}`,
+                },
+            },
+        ],
     };
 });
 // ── auto_process ─────────────────────────────────────────────────────────────

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@torqon/mcp",
-    "version": "0.2.3",
+    "version": "0.2.5",
     "type": "module",
     "main": "dist/index.js",
     "bin": {