Architecture

SWUST Code is organized around terminal interfaces, task orchestration, session runtime, tools, and project-state storage. A useful first pass is to read the system by responsibility rather than by exact source directory.

Runtime Layers

Interface Layer
  CLI / TUI / Web / Desktop / API server

Extension Layer
  Plugins / Skills / MCP / custom commands / integration surfaces

Coordination Layer
  Agent orchestration / subagents / workflow / goal / compose

Execution Layer
  Session runtime / tool registry / permissions / model routing / context rebuild

Foundation Layer
  Effect services / SQLite + Drizzle / config / filesystem / project state

This diagram describes responsibility boundaries. It does not map one-to-one to every source directory.

Monorepo Areas

Package group	Main responsibility
opencode	runtime, CLI, sessions, tools, config, server routes
app / desktop / web / console	user interfaces and interaction surfaces
sdk / ui / shared	shared types, components, and common utilities
plugin / extensions / function	extension and runtime integration surfaces
script / containers / identity / enterprise / slack / storybook	build, integration, deployment, and surrounding services

Critical Paths

Session Runtime

The session runtime assembles input, selects models, calls tools, handles permissions, and records messages. Long tasks also use memory, checkpoints, and context reconstruction.

Agent Orchestration

The primary agent, subagents, goal, compose, and workflows share task state and tool boundaries. Subagent results return to the primary task path, while workflows make multi-stage execution resumable.

Engineering Safeguards

Task Gate, Bash Safety, Write Guard, and Document Validation affect execution boundaries. They handle premature stopping, risky shell commands, write targets, and structured-document edits.

Reading Order

To continue into the code structure, start with:

1. Agent Modes
2. Persistent Memory
3. Workflow Engine
4. Security
5. Config Schema