Playground
Overview
A development workshop for building and integrating agentic workflow tools. Evolves HumanLayer, OpenSpec, and RepoMirror into a cohesive pipeline where the solo developer acts as Architect/Product Manager while AI agents act as the Engineering Team. Production-ready tools get promoted to the main organization.
Core Value
The integrated pipeline: GTD to OpenSpec to RepoMirror to HumanLayer working as one cohesive system that lets a solo developer scale themselves while maintaining strict control over intent and quality.
Components
- HumanLayer — Human-in-the-loop AI agent approval system (Go daemon + TypeScript CLI + React/Tauri UI)
- OpenSpec — Spec management CLI tool for spec-driven development
- RepoMirror — Repository mirroring tool with context boundaries
- github.gg — GitHub code analysis platform (Next.js)
- Jan — Open-source AI assistant for local LLM inference
Key Architectural Decisions
| Decision | Rationale |
|---|---|
| Spec-first, human-gated philosophy | Prevents drift — agents get stuck in loops without clear specs and approval gates |
| Strict context boundaries | Agents only see active spec + relevant src + their scratchpad |
| Agent loop budgets (10-20) | Prevents “forever refactor” cycles |
| Hands-off-keyboard rule | No manual edits while agent running — prevents merge conflicts |
Constraints
- Build on existing: Must use HumanLayer/OpenSpec/RepoMirror — don’t reinvent
- Tech stack: TypeScript 5.9.x, Go 1.24.0, Bun 1.2.23, Turbo 2.5.8
- Architecture: Event-driven with EventBus, repository pattern, service layer abstraction
- Solo workflow: No multi-user/team support. Local-first, no cloud infrastructure for v1.
Related
- Knowledge Vault
- Research Lab - Formal research environment
- agentic-pipeline-orchestration - The pipeline this playground produces
- polyglot-monorepo-architecture - Architecture pattern of this workspace
- human-in-the-loop-approval-gates - Approval gate taxonomy developed here
- agent-scratchpad-workspace - Agent workspace pattern developed here