System Architecture

OSTwin is built from four independent subsystems that communicate through the filesystem. There is no central process, no message broker, and no shared database. This section maps the complete system architecture.

Architecture Overview

┌──────────────────────────────────────────────────────────┐
│                      User / CLI                          │
└──────────────┬───────────────────────────┬───────────────┘
               │                           │
    ┌──────────▼──────────┐     ┌──────────▼──────────┐
    │   Engine (.agents/)       │     │  Dashboard (Py+JS)  │
    │                     │     │                     │
    │  Invoke-Agent.ps1   │     │  FastAPI backend    │
    │  Build-PlanningDAG.ps1      │◄───►│  Next.js frontend   │
    │  New-WarRoom.ps1    │     │  SSE streaming      │
    │  Resolve-*.ps1      │     │                     │
    └──────────┬──────────┘     └──────────┬──────────┘
               │                           │
    ┌──────────▼───────────────────────────▼──────────┐
    │              Filesystem (.agents/)               │
    │  plans/ war-rooms/ roles/ skills/ ledger.jsonl   │
    └──────────┬───────────────────────────┬──────────┘
               │                           │
    ┌──────────▼──────────┐     ┌──────────▼──────────┐
    │  MCP Servers (Py)   │     │     Bot (TS)        │
    │                     │     │                     │
    │  memory server      │     │  Discord adapter    │
    │  warroom server     │     │  Telegram adapter   │
    │  dashboard server   │     │  Slack adapter      │
    │  skills server      │     │                     │
    └─────────────────────┘     └─────────────────────┘

Subsystem 1: Engine (PowerShell)

The engine is the orchestration core. Written in PowerShell, it manages plan execution, agent invocation, and lifecycle enforcement.

Key Scripts

Script	Purpose
roles/_base/Invoke-Agent.ps1	Universal agent runner – assembles prompt, resolves role/skills/MCP, launches LLM session
plan/Build-PlanningDAG.ps1	Kahn’s algorithm – converts epic dependencies into executable waves
war-rooms/New-WarRoom.ps1	War-room scaffolding – creates directory structure, config, lifecycle
plan/Start-Plan.ps1	Wave resolution – determines which epics can execute next
roles/_base/Resolve-Role.ps1	5-tier role discovery
roles/_base/Resolve-RoleSkills.ps1	3-tier skill resolution with union merge
mcp/config_resolver.py	4-tier MCP configuration merge
war-rooms/Get-WarRoomStatus.ps1	Lifecycle transition with validation
plan/Start-Plan.ps1	Timeout enforcement loop
roles/_base/New-DynamicRole.ps1	Dynamic role creation
plan/Start-Plan.ps1	Plan review orchestration and DAG generation

Why PowerShell?

PowerShell provides native JSON handling, cross-platform support (runs on macOS, Linux, Windows via PowerShell Core), pipeline composition, and robust error handling. It also integrates cleanly with both Python (MCP servers) and Node.js (dashboard) processes.

Engine Responsibilities

• Parse PLAN.md and build the DAG
• Create war-rooms for each wave of epics
• Invoke agents in the correct sequence within each room
• Monitor timeouts and enforce lifecycle transitions
• Handle retries and escalations
• Coordinate cross-room memory injection

Subsystem 2: Dashboard (FastAPI + Next.js)

The dashboard provides real-time visibility into plan execution.

Backend (FastAPI)

The Python backend exposes REST and SSE endpoints:

Endpoint	Method	Purpose
/api/plans	GET	List all plans with status
/api/plans/{id}/dag	GET	DAG visualization data
/api/rooms	GET	List all war-rooms with current state
/api/rooms/{id}/channel	GET	Read room’s message channel
/api/rooms/{id}/progress	GET	Current progress snapshot
/api/stream	SSE	Real-time updates for all rooms
/api/search	POST	Semantic search across memories
/api/stats	GET	Aggregate statistics

Frontend (Next.js)

The Next.js frontend renders:

• Plan overview – DAG visualization with wave grouping
• War-room dashboard – grid of active rooms with status badges
• Room detail – channel messages, progress bars, artifact list
• Memory explorer – search and browse the shared ledger
• Settings – role registry, skill index, MCP configuration

Features

• Real-time updates via SSE (no polling)
• DAG rendered as an interactive graph
• Channel messages displayed as a chat timeline
• Progress bars driven by progress.json from each room
• Error highlighting with escalation status

Subsystem 3: Bot (TypeScript)

The bot subsystem provides chat platform integrations for human interaction with OSTwin.

Platform	Adapter	Status
Discord	bot/discord/	Production
Telegram	bot/telegram/	Production
Slack	bot/slack/	Beta

The bot enables users to:

• Start plans from chat messages
• Monitor room progress in real-time
• Receive alerts on failures and escalations
• Override lifecycle states manually
• Query the memory ledger via natural language

Subsystem 4: MCP Servers (Python)

Four Python-based MCP servers provide tool interfaces for agents:

Server	Module	Key Tools
Memory	.agents/mcp/global-memory-server.py	publish, query, search, get_context, list_memories
War-Room	.agents/mcp/warroom-server.py	post_message, read_messages, get_latest, update_status, report_progress, list_artifacts
Channel	.agents/mcp/channel-server.py	post_message, read_messages, get_latest
Knowledge	.agents/mcp/global-knowledge-server.py	search_knowledge, read_knowledge

Each server runs as a separate process, launched by roles/_base/Invoke-Agent.ps1 with room-specific arguments. Servers communicate with agents over stdio using the MCP protocol.

Process Isolation

Each MCP server process is isolated. A crash in the memory server doesn’t affect the war-room server. The engine detects crashed servers and can restart them without interrupting the agent session.

Filesystem as Coordination Layer

The filesystem is OSTwin’s coordination backbone. All state is stored as files:

File/Directory	Purpose	Format
.agents/plans/*/PLAN.md	Plan definitions	Markdown
.agents/plans/*/DAG.json	Dependency graphs	JSON
.agents/war-rooms/*/config.json	Room contracts	JSON
.agents/war-rooms/*/channel.jsonl	Message channels	JSON Lines
.agents/war-rooms/*/status.txt	Current lifecycle state	Plain text
.agents/war-rooms/*/progress.json	Completion tracking	JSON
.agents/war-rooms/*/lifecycle.json	State machine definition	JSON
.agents/ledger.jsonl	Shared memory	JSON Lines
.agents/roles/*/role.json	Role configurations	JSON
.agents/skills/*/SKILL.md	Skill definitions	Markdown
.agents/roles/registry.json	Role catalog	JSON
.agents/vault.json	Secrets (gitignored)	JSON

Why Filesystem?

Design Decision

OSTwin chose the filesystem over a database or message broker for five reasons:

1. Inspectability – debug with ls, cat, grep, not SQL queries
2. Portability – copy a .agents/ directory to reproduce any state
3. Version control – git tracks every state change automatically
4. Simplicity – no database migrations, no broker configuration
5. Resilience – no single point of failure; files survive process crashes

The tradeoff is that concurrent writes require file locking (handled by fcntl.LOCK_EX in MCP servers) and cross-machine coordination requires shared filesystem access or synchronization.

Concurrency Model

Component	Concurrency	Mechanism
War-rooms	Up to 50 parallel	Wave-based scheduling from DAG
Agents within a room	Sequential	Manager orchestrates turn order
MCP servers	1 per agent session	Process-per-session isolation
Channel writes	Serialized	fcntl.LOCK_EX file locking
Ledger writes	Serialized	fcntl.LOCK_EX file locking
Dashboard reads	Concurrent	Read-only polling, no locks needed
Bot commands	Queued	Event loop with sequential dispatch

Sequential Agent Turns

Agents within a single war-room do not run concurrently. The manager invokes engineer, waits for completion, then invokes QA. This is intentional – concurrent agents in the same room would create race conditions on shared files.

Data Flow

User Input
    │
    ▼
PLAN.md ──► Build-PlanningDAG.ps1 ──► DAG.json
                                  │
                                  ▼
                           Start-Plan.ps1
                                  │
                                  ▼
                           New-WarRoom.ps1 ──► room directories
                                  │
                                  ▼
                           Invoke-Agent.ps1
                            │         │
                     ┌──────┘         └──────┐
                     ▼                       ▼
              LLM Session              MCP Servers
                     │                       │
                     ├── channel.jsonl ◄──────┤
                     ├── artifacts/ ──────────┤
                     ├── progress.json ◄──────┤
                     └── ledger.jsonl ◄───────┘
                                  │
                                  ▼
                           Dashboard (SSE)
                                  │
                                  ▼
                           Next.js Frontend

Key Source Locations

Directory	Language	Purpose
.agents/	PowerShell	Orchestration scripts
.agents/mcp/	Python	MCP server implementations
dashboard/api/	Python (FastAPI)	REST + SSE backend
dashboard/web/	TypeScript (Next.js)	Frontend application
bot/	TypeScript	Chat platform adapters
.agents/	Config files	Runtime state and configuration
.agents/roles/	JSON + Markdown	Role definitions
.agents/skills/	Markdown	Skill documents
.agents/plans/	Mixed	Plan files and DAGs
.agents/war-rooms/	Mixed	Room state and artifacts

Getting Started

To understand the system, start with .agents/roles/_base/Invoke-Agent.ps1 – it is the single entry point that ties all four subsystems together. Every agent invocation flows through this script, making it the best place to trace the complete execution path.