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