DiOffice

Overview

DiOffice is an experiment in durable AI teamwork: a virtual software company where an Owner manages permanent AI employees. Employee identity and work history are persistent; the runtime — OpenCode first — is a replaceable tool behind a gateway. The v0.1 target is one honest end-to-end flow: task drafts, explicit start, isolated workspace, real repository work, checks and preview evidence bound to a commit SHA, PR review, approval, and merge — with a pixel-office UI driven only by real backend state.

Problem

Most AI-coding tools show output, not work — there is no persistent employee identity, no auditable event trail, and UI progress is often simulated rather than backed by real state.

Solution

An event-sourced office: every employee action becomes a normalized durable event, task state machines are explicit (DRAFT → BACKLOG/READY → DONE, with merge as a separate Owner action), and the pixel-office UI renders only real backend state — never simulated progress.

Responsibilities

• Authored the product spec — state machines, event schema, execution manifest, user flows
• Built the Go API — session/CSRF auth, task and event endpoints, explicit migrations
• Built the React + Vite shell with a Phaser pixel-office layer
• Defined the agent-gateway boundary (Node.js) and the first runtime adapter (OpenCode)
• Specified execution isolation — Docker + git worktree, Temporal for workflow durability

Key Features

• Persistent AI employees with identity and work history
• Explicit task lifecycle — Owner gates every transition
• Normalized event stream with durable SSE replay cursor
• Evidence-bound reviews — checks and previews tied to the tested commit SHA
• Pixel-office UI rendered from real backend state
• Swappable agent runtimes — OpenCode first

Technology

• React + TypeScript + Vite
• Phaser (pixel office)
• Tailwind CSS + shadcn/ui
• Go API
• PostgreSQL
• Node.js agent gateway
• OpenCode runtime
• Temporal workflows
• SSE realtime
• Docker + git worktree
• Playwright
• S3-compatible storage

Technical Decisions

• No simulated work: Character animation, monitors, diffs, tests, and progress must come from real task/runtime events — nothing fake.
• Persistent identity, swappable runtime: Employees keep identity and history; runtimes like OpenCode are replaceable tools behind the gateway.
• Owner-in-the-loop: Start, approve, and merge are explicit Owner actions — the AI never merges its own work.

Impact

Screenshots

Links

• GitHub repository