LLM Intelligence

A systematic multi-fire cognitive response system that separates AI thinking from communication through response files and organized conversation tracking.

Overview

Based on Google's research showing embedding geometry doesn't work at scale, this system enables LLMs to use multiple "fires" for full intelligence expression through cognitive separation:

• Conversation Channel: AI's thinking space (tool calls, analysis, exploration)
• Response Channel: AI's deliberate communication (curated response files)

Key Features

• Arbitrary Response Files: LLM writes response files anywhere, system organizes automatically
• Emergent Tracking: Free-form project hierarchy (project → feature → component → deliverable → subtask → task → workflow)
• STARLOG Integration: Logs to debug diary with structured format
• Cognitive Separation: Clean separation between thinking and communication
• JSON Safety: All content properly escaped through json module

Installation

pip install llm-intelligence

Usage

As MCP Server

llm-intelligence-server

Direct API Usage

from llm_intelligence import respond

# LLM writes response file anywhere
with open("/tmp/my_response.md", "w") as f:
    f.write("I implemented OAuth authentication...")

# System organizes everything automatically  
result = respond(
    qa_id="abc123",
    response_file_path="/tmp/my_response.md",  # Any path
    one_liner="OAuth implementation complete",
    key_tags=["oauth", "auth"],
    involved_files=["auth.py", "oauth.py"],
    project_id="auth_system",
    feature="oauth",
    component="middleware", 
    deliverable="auth_flow",
    subtask="jwt_validation",
    task="implement_verify",
    workflow_id="sprint_1"
)

Architecture

• Core Module: llm_intelligence.core - All business logic
• MCP Server: llm_intelligence.mcp_server - Thin wrapper for MCP integration
• Organized Storage: qa_sets/{qa_id}/responses/response_XXX/response.md
• JSON Tracking: Full conversation history with emergent metadata

Cognitive Flow

1. Fire 1: LLM writes curated response file
2. Fire 2: LLM does actual work (Read, Edit, Bash)
3. Fire 3: LLM reports tool usage (optional)
4. Fire 4: LLM harvests everything with respond()

System handles all organization, cleanup, and tracking automatically.