Product Requirements Documents

Eight comprehensive technical PRDs covering the complete Claude Code system architecture. Each document provides requirements, design patterns, and implementation guidance for building production-grade AI agent infrastructure.

Before You Read

This is an integrated development platform built for a single operator—not a distributable tool, not a SaaS product, and not a framework you install. The PRDs describe a working system where every component exists because it solved a real problem. The scale is deliberate: small enough to run on a few VPS instances, complex enough to enforce real governance.

Component Count Reality Check
Active plugins 4 Each is ~5 files
Conductor agents 28 Markdown persona files, not running processes
Docker containers 5 core Rest are separate projects
Hook executions per tool call 2–3 Most <50ms
Human approval prompts per session 0 typical Only at MAJOR tier + elevated tools

Where to Start

You don't need all 8. Start with what solves your problem:

  • Just want memory? The Vector Memory System PRD alone
  • Want governance? Governance Framework + Plugin Ecosystem PRDs
  • Want the full pipeline? All 8, in order listed below

Why This Complexity Exists

Component What It Does Why Not Simpler
Vector Memory
Qdrant + Ollama
Cross-session knowledge retention Flat files can't do semantic search; cloud embeddings violate local-first privacy
Governance Hooks Prevent agents from exfiltrating data Trust-but-verify doesn't work for autonomous agents
Conductor Orchestration Complex builds need sequenced quality gates Ad-hoc prompting produces incomplete code
Context Guard 200K context window fills during long sessions Without monitoring, compaction destroys state
Code Hardener AI-generated code has 62% dead code / 40% vulns Running 67 tools manually is impractical

Plugin Ecosystem

35+ requirements

Build extensible plugin architectures with lifecycle hooks, slash commands, and MCP integration. Define 8 hook events (SessionStart, PreToolUse, PostToolUse, etc.), implement plugin discovery and registration, and create a platform for extending Claude Code's capabilities.

Read PRD →

Multi-Agent Orchestration

27 agent personas

Design tiered agent workflows with 5-signal classification (complexity, uncertainty, domain depth, tool chains, intent sensitivity), quality gates, and intelligent routing. Manage 27 specialized agents across security, development, architecture, and compliance domains.

Read PRD →

Context Management

Multi-layer context

Master CLAUDE.md hierarchies (global, project, memory-injected), auto-memory integration, and session persistence. Implement context guards, directory awareness, and intelligent fallback strategies for maintaining coherent agent behavior across sessions.

Read PRD →

Vector Memory System

6 MCP tools

Implement cross-session persistent memory with Qdrant vector storage, semantic recall, and automated knowledge management. Deploy 6 MCP tools (memory_store, memory_recall, procedure, learning, trajectory, benchmark) with auto-link detection and governance integration.

Read PRD →

Agent Governance Framework

TRiSM-compliant

Deploy enterprise trust, risk, and security management frameworks. Implement 15 audit event types, LLM threat detection (jailbreak, hallucination, PII exposure), approval gates for external communications, and comprehensive compliance tracking for production AI systems.

Read PRD →

Memory Dashboard

Visual analytics

Build a visual web UI for your memory system. FastAPI + HTMX + D3.js knowledge graph + Chart.js analytics. Semantic search, paginated browsing, collection stats, relationship visualization, and drift monitoring across all memory collections.

Read PRD →

Markdown-for-Agents

Content proxy

Deploy an HTML-to-markdown proxy that gives agents clean, token-counted content from any URL. Three conversion modes (URL-path, nginx, sidecar), Playwright fallback for JS-heavy sites, and graceful degradation pipeline.

Read PRD →

Code Assurance Platform

67 integrated tools

Build a unified code assurance platform with 67 tools across 6 pillars. 12 scan profiles, 6-stage finding enrichment pipeline, 1000-point quality scoring with sqrt penalty, cryptographic attestation via Ed25519, and SLSA Level 3 provenance.

Read PRD →

How to Use These PRDs

Each PRD is designed to be fed to Claude Code as a blueprint. Don't copy-paste implementations from others—every environment has unique requirements. Instead:

  1. Read the PRD to understand the architectural domain
  2. Identify which requirements apply to your environment
  3. Use Claude Code to generate tailored implementation code
  4. Test, iterate, and adapt to your specific needs

These PRDs represent hundreds of hours of architectural research and production testing. They're not theoretical—they're battle-tested patterns from real-world deployment.