loop

/harness-plugin:loop — Autonomous Development Harness

Launch the VIBM autonomous development harness. Gathers feature requirements, brainstorms the spec, generates app_spec.md, and starts the agent loop.

Step 0: Worktree & Environment (MANDATORY — do this FIRST)

The harness creates commits and pushes code. It MUST run inside a git worktree, never on main.

0a. Create or enter a worktree

Check the current branch:

git branch --show-current

If it says main, you MUST create a worktree before proceeding. Use the EnterWorktree tool:

EnterWorktree(name="harness-<feature-name>")

Or create one manually:

git worktree add .claude/worktrees/harness-<feature-name> -b feat/<feature-name>
cd .claude/worktrees/harness-<feature-name>
npm install

DO NOT skip this step. If you are already in a non-main branch/worktree, you may proceed.

0b. Source .env from the source repo

Git worktrees do NOT include untracked files like .env. The API keys live in the original project root, not in the worktree. Find the source repo root and source from there:

SOURCE_ROOT=$(git worktree list --porcelain | head -1 | sed 's/^worktree //')
set -a && source "$SOURCE_ROOT/.env" && set +a

Verify the key is set:

echo "ANTHROPIC_API_KEY is ${ANTHROPIC_API_KEY:+set}"

If it prints "set", proceed. If not, STOP — the harness will fail without API keys.

Why this matters: Without this step, the harness dry-run hangs, the pre-bash hook blocks on missing ANTHROPIC_API_KEY, and agents waste iterations regenerating app_spec.md from scratch.

Step 1: Gather Requirements

Use the AskUserQuestion tool to ask the user TWO questions:

Question 1 — "What feature(s) do you want to build? Describe the functionality, screens, and user flows."

• Header: "Feature"
• Options:
  • "Full VIBM app" — Build the complete conversation manager (all screens, data model, IPC layer)
  • "Specific feature" — Build one feature or module (user will describe in the text input)
• multiSelect: false

Question 2 — "How many harness iterations should we run?"

• Header: "Iterations"
• Options:
  • "3 (quick test)" — Good for validating the harness works
  • "10 (one session)" — Enough for scaffolding + a few features
  • "Unlimited" — Run until all features are done or you stop it
• multiSelect: false

Step 2: Brainstorm the Spec

Invoke the superpowers:brainstorming skill using the Skill tool:

skill: "superpowers:brainstorming"

When brainstorming, focus on producing a product specification for the harness, NOT an implementation plan. The brainstorming output should cover:

• Overview — What the feature/app does in 2-3 sentences
• Tech Stack — Tauri 2, React 19 + TypeScript, Rust, Vitest, Testing Library, ESLint, Prettier, CSpell, Storybook
• Core Features — Bulleted list of every feature with brief description
• Data Model — TypeScript interfaces and matching Rust structs for all entities
• IPC Commands — List of #[tauri::command] handlers with argument/return types
• UI Screens — Description of each screen/view and its components
• User Flows — Step-by-step flows for key interactions

IMPORTANT: After brainstorming, do NOT invoke writing-plans. Instead, proceed to Step 3.

Step 3: Generate app_spec.md

Take the brainstormed spec and write it to harness/prompts/app_spec.md using the Write tool. The file should be a clean markdown document that the harness initializer agent can read to generate a feature_list.json.

Structure the file as:

# VIBM — App Specification

## Overview

[from brainstorming]

## Tech Stack

- Framework: Tauri 2
- Frontend: React 19 + TypeScript (arrow-function components)
- Backend: Rust (Tauri commands, managed state)
- Testing: Vitest + Testing Library (frontend), cargo test (backend)
- Linting: ESLint flat config, cargo clippy
- Formatting: Prettier, cargo fmt

## Core Features

[from brainstorming — bulleted list]

## Data Model

[from brainstorming — TypeScript interfaces + Rust structs]

## IPC Commands

[from brainstorming — command table with args/returns]

## UI Screens

[from brainstorming — screen descriptions]

## User Flows

[from brainstorming — step-by-step flows]

Step 4: Launch the Harness

Pre-flight check: Confirm you are NOT on main and that ANTHROPIC_API_KEY is set (both from Step 0). If either is missing, go back to Step 0.

Run the harness using Bash:

cd harness && pip install -r requirements.txt 2>/dev/null && python autonomous_agent_demo.py --max-iterations <N>

Where <N> is the iteration count from Step 1. If "Unlimited", omit the --max-iterations flag entirely.

Notes:

• On WSL2 only, add --no-sandbox (the OS-level sandbox is unreliable on WSL2). On macOS/Linux, omit it to keep sandbox isolation enabled.
• The env vars from Step 0b are inherited by the subprocess automatically

IMPORTANT: This command will run for a long time. Use run_in_background: true on the Bash tool so the user isn't blocked. Tell the user the harness is running and they can check progress with:

• cat feature_list.json | grep '"passes": true' | wc -l — count completed features
• cat claude-progress.txt — read the latest progress notes
• git log --oneline -10 — see recent commits from the agent