harness

Vimeflow

An experiment in Harness-Engineered, AI-Native Development — where autonomous agent loops build the application from specification to implementation, governed by layered rules and specialized agents.

Vimeflow is a coding agent conversation manager built as a Tauri desktop app (Rust + React/TypeScript). But the product itself is secondary to the process: this repository is a testbed for exploring how far autonomous development harnesses can go when given well-structured specifications, safety guardrails, and progressive disclosure documentation.

What Makes This AI-Native

Traditional projects have humans write code and AI assist. Vimeflow inverts this:

1. Humans write specs — product requirements, design system, development rules
2. An autonomous harness builds features — a two-agent loop (Initializer + Coder) decomposes specs into feature lists and implements them incrementally
3. Specialized agents review the work — 10 AI agents handle planning, TDD, code review, security, and documentation
4. Rules govern everything — a hierarchical rule system (common + language-specific) ensures consistency without human intervention per commit

The CI/CD infrastructure, linter configuration, and git hooks already in this repository were all built by the harness from an app_spec.md specification.

Repository Structure

CLAUDE.md           <- AI navigation hub (start here for agents)
README.md           <- You are here (for humans)
DEVELOPMENT.md      <- Commands, tech stack, code style
ARCHITECT.md        <- Architecture decisions, Tauri patterns
DESIGN.md           <- UI design system (Obsidian Lens / Catppuccin Mocha)

agents/             <- 10 specialized AI agent definitions
rules/              <- Hierarchical development standards (common + TS + Rust)
harness/            <- Autonomous development loop (Claude Code SDK, Python)
docs/design/        <- Screen mockups, Stitch HTML/CSS, design spec

The Harness

The autonomous development harness (harness/) is the engine of this experiment. It is a Python-based loop built on the Claude Code SDK:

• Initializer agent reads app_spec.md and decomposes it into a phased feature_list.json
• Coder agent picks the next pending feature, implements it, marks it done, and auto-continues
• Safety layers include a bash command allowlist, sandboxed execution, and write-protection on the feature list

cd harness && pip install -r requirements.txt
python autonomous_agent_demo.py                    # Unlimited iterations
python autonomous_agent_demo.py --max-iterations 5 # Capped

See harness/CLAUDE.md for the full details.

Current Status

Phase: Foundation / Pre-implementation

• Development rules, agent specs, and CI/CD tooling are established
• Design system (5 screens) specified via Google Stitch
• Application code (Tauri scaffolding, src/, src-tauri/) does not exist yet
• Next: harness builds the app from the design spec

Tech Stack

Tauri 2 (Rust backend + web frontend) | React 19 + TypeScript | Vitest + Playwright | ESLint + Prettier | Husky + commitlint | GitHub Actions CI/CD