GUPPI

General Unit Primary Peripheral Interface.

A nod to Dennis E. Taylor's Bobiverse: GUPPI is Bob's non-sentient on-board assistant — it runs the ship, handles the peripherals, and frees Bob to think. Here, GUPPI is the non-sentient OS / orchestrator layer that sits underneath the Agentic Developer (Claude Code + the Agentheim plugin) and gives it hands, eyes, ears, and a voice. The wake word is, naturally, "Bob".

GUPPI does not think about the work. It runs the body so the agentic developer can run the mind.

What GUPPI is

A personal, single-machine mission-control surface for managing multiple Agentheim projects at once. Concretely, GUPPI provides:

• A canvas — a Miro-like infinite surface where every Agentheim project on disk shows up as a tile, with its bounded contexts as connected child nodes. One ambient view across all projects, instead of N terminals.
• A claude runner — spawns claude processes inside each target project's folder so each project's own CLAUDE.md, skills, hooks, and settings apply. GUPPI is a controller, never a host.
• Agent-awareness — per-bounded-context indicators for running / idle / blocked-on-question, surfaced on the canvas at the BC's location.
• A voice channel — wake-word "Bob", Whisperheim for STT, Utterheim for TTS. Voice is a first-class input modality, not a gimmick.
• Project discovery and creation — find every folder containing .agentheim/, count tasks per state, and kick off new projects with brainstorm already running inside them.

What GUPPI is not

• Not sentient. No agency, no judgment, no opinions about your code. The agentic developer is upstairs; GUPPI is downstairs.
• Not a host for claude. GUPPI never runs claude in its own folder to act on another project. Always spawns into the target project's cwd.
• Not a replacement for Claude Code, for Agentheim, or for the terminal. The terminal stays as an escape hatch.
• Not a team tool. Single user (Marco), single machine, local-only.
• Not a code editor or a generic project manager. It only knows the Agentheim shape: vision -> bounded contexts -> tasks.

Where the actual specs live

• .agentheim/vision.md — vision, problem, v1 scope, non-goals, ubiquitous language.
• .agentheim/context-map.md — bounded contexts and their relationships.
• .agentheim/contexts/ — each BC's README.md, backlog, and todo list.

Status

Pre-v1. The domain has been modeled; foundation tasks (desktop runtime, canvas rendering, claude PTY, persistence, project discovery, walking skeleton) are in infrastructure/todo/. No application code exists yet.

v1 is deliberately tiny: a read-only canvas that shows every project and its BCs with task counts. Voice, command execution, and live agent observation come after.