How It Works

The three planes of state

1. Global user state — ~/.codymaster/kanban.json holds projects, tasks, deployments, changelog, chain executions (src/data.ts).
2. Per-project agent state — .cm/ holds continuity, memory JSON (legacy path), SQLite DB, bus, sprint artifacts, config (src/continuity.ts, src/context-db.ts, src/context-bus.ts).
3. Runtime processes — dashboard HTTP server, optional browse daemon, MCP stdio server (src/dashboard.ts, src/browse-server.ts, src/mcp-context-server.ts).

How a typical day looks

flowchart LR
  subgraph Human["You"]
    H[Plan / review]
  end
  subgraph CM["CodyMaster CLI"]
    C[cm commands]
  end
  subgraph Store["Persistence"]
    G[kanban.json]
    P[.cm memory]
  end
  subgraph Agents["AI tools"]
    A[MCP client]
  end
  H --> C
  C --> G
  C --> P
  A -->|MCP tools| P

Text fallback: you use cm for ops; global kanban and .cm store continuity; MCP clients read/write the same project memory.

Skills vs commands

• Commands (cm …) are implemented in TypeScript under src/cli/commands/.
• Skills are documentation + procedure for agents, living in skills/<name>/SKILL.md. They tell how to think and act; the CLI wires where data goes.

Progressive loading (L0 / L1 / L2)

CodyMaster can resolve cm:// URIs at different depths so agents do not load huge context at once (src/uri-resolver.ts, MCP tool cm_resolve in src/mcp-context-server.ts).

See also

• CodyMaster Brain
• Data flow
• Using skills