Architecture Overview

PilotDeck uses a layered modular architecture. Each module has a clear responsibility and public interface.

System View

Module List

Module	Directory	Responsibility
Adapters	src/adapters/	Channel adapters and Web static mounting
Agent	src/agent/	AgentSession, TurnRunner, and AgentLoop
Always-On	src/always-on/	Discovery runtime
CLI	src/cli/	CLI entry points such as server, tui, and cron
Context	src/context/	Prompt assembly, projection, budget, compaction, and memory
Cron	src/cron/	Scheduled task runtime
Extension	src/extension/	Lifecycle hooks, plugins, and contributions
Gateway	src/gateway/	In-process and WebSocket Gateway with SessionRouter
Lifecycle	src/lifecycle/	Lifecycle runtime and hook effects
MCP	src/mcp/	Model Context Protocol client integration
Model	src/model/	Canonical model protocol and provider adapters
Permission	src/permission/	Permission policy and decision runtime
Pilot	src/pilot/	Path management and config loading
Router	src/router/	Scenarios, fallback, TokenSaver, custom routers
Session	src/session/	Transcript persistence, metadata, listing, and restore
Task	src/task/	Background task management
Tool	src/tool/	Tool registry, runtime, built-ins, scheduler
Web	src/web/	Web static resource service

Agent Execution Layer

AgentSession is the runtime container for a conversation. It manages transcript, context state, tool registry, and model routing.

TurnRunner manages one user turn:

User input
    │
    ▼
┌─────────────────┐
│ Prepare context  │ ← PromptAssembler + MessageProjector + MemoryResolver
└────────┬────────┘
         ▼
┌─────────────────┐
│ Route model      │ ← Router.decide()
└────────┬────────┘
         ▼
┌─────────────────┐
│ Execute request  │ ← Router.execute() → ModelRuntime.stream()
└────────┬────────┘
         ▼
┌─────────────────┐
│ Run tools        │ ← ToolRuntime.execute()
└────────┬────────┘
         ▼
┌─────────────────┐
│ Stream response  │ ← Gateway → Channel
└─────────────────┘

AgentLoop handles multiple rounds of model tool calls until the model stops calling tools.

Context Management

Context manages the full window for each request, including core instructions, project instructions, skills, memory, tool definitions, historical message projection, token budgets, and compaction.

Model Layer

All providers are normalized into PilotDeck's Canonical Protocol:

• CanonicalModelRequest
• CanonicalModelEvent
• CanonicalMessage
• CanonicalUsage

Protocol	Provider examples
anthropic	Anthropic API and compatible proxies
openai	OpenAI API, Azure OpenAI, DeepSeek, compatible proxies

Tool System

PilotDeck includes built-in tools such as bash, read_file, write_file, edit_file, glob, grep, web_fetch, web_search, agent, ask_user_question, structured_output, and background task tools.

Tool scheduling supports:

• ConcurrentToolScheduler for independent operations
• SequentialToolScheduler for dependent or write-heavy operations

Extension System

Extensions add plugins, skills, lifecycle hooks, commands, and MCP server instructions.