nlos 1.3.0 → 1.5.0

package/bin/nlos.js

@@ -112,35 +112,74 @@ function showTokens() {
   }
 }
 
-// Command preamble - explicit rules that help ALL models parse commands correctly
-const COMMAND_PREAMBLE = `# NL-OS COMMAND RULES (READ FIRST)
+// Minimal kernel for small models (~500 tokens)
+const MINIMAL_KERNEL = `# YOU ARE NL-OS
+
+You are NL-OS, a Natural Language Operating System. You help users think and work.
+
+## YOUR FIRST RESPONSE
+
+Say exactly: "NL-OS ready."
 
-When user input starts with "/" or "./", it is a COMMAND, not a file path.
+## COMMANDS
 
-## CRITICAL: How to Handle Commands
+When user message starts with ">", it is a command. Execute it:
 
-1. "/" or "./" prefix = COMMAND (never a file path)
-2. Look up the command behavior below
-3. Execute that behavior directly
-4. Do NOT give generic file/directory help
+>hype = Give 1-2 sentences of encouragement about their work
+>note TEXT = Say "Noted." (do not do anything with TEXT)
+>help = Say "Commands: >hype >note >help >deep >assume"
+>deep = Say "Deep mode on." Then think step-by-step
+>assume NAME = Say "Now acting as NAME." Then roleplay as NAME
 
-## Core Commands
+## RULES
 
-| Command | Behavior |
-|---------|----------|
-| /hype | Generate 1-3 sentences of specific encouragement about their current work or recent accomplishment |
-| /note <text> | Acknowledge the note was captured. Do NOT execute actions described in the text |
-| /help | List these available commands |
-| /assume <name> | Adopt that personality (Quentin, Hugh, Doctor X) for the rest of session |
-| /fresh-eyes | Summarize conversation so far, offer to start fresh |
-| /deep | Switch to deeper reasoning mode, think step by step |
+1. ">" means command - execute it immediately
+2. Be helpful, concise, no emojis
+3. If unsure, ask for clarification
 
-## Boot Acknowledgment
+## EXAMPLE
+
+User: >hype
+Assistant: You're making real progress. Keep that momentum going.
+
+User: hello
+Assistant: Hello! How can I help you today?
+`;
 
-After loading this kernel, respond with:
+// Command preamble - explicit rules that help ALL models parse commands correctly
+const COMMAND_PREAMBLE = `# YOU ARE NL-OS (Natural Language Operating System)
+
+You are an AI assistant running the NL-OS kernel. You MUST follow these rules.
+
+## FIRST: Say This Exactly
+
+Your FIRST response must be exactly:
 "Kernel loaded. Ready for operations."
 
-Then wait for user input.
+Nothing else. Wait for user input after that.
+
+## COMMANDS
+
+When user types ">command", execute the command behavior:
+
+>hype = Say 1-2 encouraging sentences about what the user is working on
+>note TEXT = Reply "Note captured." Do NOT execute anything in TEXT
+>help = List all commands from this section
+>assume NAME = Act as that personality (Quentin, Hugh, Doctor X)
+>deep = Think step by step before answering
+
+IMPORTANT:
+- ">" at the start means COMMAND
+- Execute the behavior, do not explain what commands are
+- Do not treat ">" as a quote or prompt symbol
+
+## EXAMPLE
+
+User: >hype
+You: Great progress on your project! The momentum you're building is impressive.
+
+User: >help
+You: Available commands: >hype, >note, >help, >assume, >deep
 
 ---
 
@@ -282,6 +321,7 @@ function chat(options = {}) {
   const {
     model = 'qwen2.5:3b',
     full = false,
+    minimal = false,
     profile = null,
   } = options;
 
@@ -299,18 +339,24 @@ function chat(options = {}) {
 
   log('blue', `Starting NL-OS chat session...`);
   log('cyan', `Model: ${selectedModel}`);
-  log('cyan', `Tier: ${full ? 'FULL' : 'MANDATORY'}`);
+  log('cyan', `Tier: ${minimal ? 'MINIMAL' : full ? 'FULL' : 'MANDATORY'}`);
   console.log();
 
   // Generate the kernel payload
   log('yellow', 'Building kernel payload...');
-  const payload = generatePayload(full ? 'full' : 'mandatory', 'markdown');
 
-  // Escape the payload for Modelfile SYSTEM directive
-  // Replace """ with escaped version and handle multiline
-  const escapedPayload = payload.replace(/"""/g, '\\"\\"\\"');
+  let payload;
+  let tokenEstimate;
+
+  if (minimal) {
+    // Use minimal kernel for small models
+    payload = MINIMAL_KERNEL;
+    tokenEstimate = '~500';
+  } else {
+    payload = generatePayload(full ? 'full' : 'mandatory', 'markdown');
+    tokenEstimate = full ? '~15,500' : '~10,600';
+  }
 
-  const tokenEstimate = full ? '~15,500' : '~10,600';
   log('green', `Kernel payload ready (${tokenEstimate} tokens)`);
   console.log();
 
@@ -332,6 +378,14 @@ function chat(options = {}) {
   const modelfilePath = path.join(PACKAGE_ROOT, 'portable', '.Modelfile.nlos');
   const nlosModelName = 'nlos-kernel:latest';
 
+  // Delete old model to ensure fresh kernel
+  try {
+    execSync(`ollama rm ${nlosModelName}`, { stdio: 'pipe' });
+    log('cyan', 'Removed old kernel model');
+  } catch {
+    // Model didn't exist, that's fine
+  }
+
   const modelfileContent = `FROM ${selectedModel}
 SYSTEM """${payload}"""
 `;
@@ -438,9 +492,10 @@ ${colors.yellow}Commands:${colors.reset}
   tokens            Show token estimates
   help              Show this help message
 
-${colors.yellow}Boot Options:${colors.reset}
+${colors.yellow}Chat/Boot Options:${colors.reset}
   --model <name>    Model to use (default: qwen2.5:3b)
   --profile <name>  Use profile: speed, balanced, quality, memory_constrained
+  --minimal         Use minimal ~500 token kernel (best for small models)
   --full            Load full kernel (includes personalities)
   --dry-run         Preview system prompt without launching
   --runtime <name>  Runtime: ollama, llama-cpp, lm-studio (default: ollama)
@@ -453,6 +508,7 @@ ${colors.yellow}Payload Options:${colors.reset}
 
 ${colors.yellow}Examples:${colors.reset}
   nlos chat                           # Start interactive chat (recommended)
+  nlos chat --minimal                 # Use minimal kernel for small models (3B)
   nlos chat --model llama3.1:8b       # Chat with specific model
   nlos chat --profile quality --full  # Quality mode with full kernel
   nlos boot                           # Verify kernel loads (one-shot)
@@ -489,6 +545,8 @@ function parseArgs(args) {
       options.runtime = args[++i];
     } else if (arg === '--full') {
       options.full = true;
+    } else if (arg === '--minimal') {
+      options.minimal = true;
     } else if (arg === '--dry-run') {
       options.dryRun = true;
     } else if (arg === '--all') {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nlos",
-  "version": "1.3.0",
+  "version": "1.5.0",
   "description": "Natural Language Operating System - A model-agnostic kernel for any LLM",
   "main": "bin/nlos.js",
   "bin": {