devagent-architecture

name	DevAgent Architecture
description	Explains DevAgent's actor model, runtime state store, and TUI overlay system so the agent can answer questions about its own architecture accurately.
tags	["architecture","runtime","tui","devagent","actors","overlays"]
version	0.1.0

DevAgent Architecture

DevAgent models every subsystem as an always-alive actor (conversation, planner, executor, tasks, git, logs, memory, models, mcp, skills). Actors never stop; only what is observed changes when a view is focused. Views map to a frozen set of numbered keys 1-8 (see docs/SPEC.md §7-8); skills and a few other subsystems are overlay-only and never occupy a numbered view slot.

State flows one way: actors publish events onto an EventBus; a single reducer (src/runtime/store.ts) folds events into RuntimeState; the TUI renders RuntimeState and never mutates it directly.

The layout is frozen: Header, Active View, Activity Strip, Prompt, Context Strip — five zones, always present, at every terminal size. Only content density changes as the terminal shrinks or grows; the zones themselves never rearrange, and no zone is ever removed.

Overlays (command palette, model switcher, search everywhere, skills) are ephemeral: they render over the Active View, never mutate runtime state directly, and always close with Esc. Every searchable overlay reuses one component, UniversalPicker, rather than reimplementing list UI per overlay.