Skip to content

ExaDev/mob-boss

Repository files navigation

mob-boss

Claude Code plugin — AI agent team orchestration with tiered incremental review, metrics-driven evolution, and per-package knowledge accumulation.

What it does

Mob-boss orchestrates a team of specialised AI agents to deliver features end-to-end:

  • Architect — designs implementation plans, verifies claims against actual code, enforces vertical-first phase shape
  • Designer — produces HTML/CSS/JS prototypes as contracts, with component manifests for 1:1 conversion
  • Developer — implements features with TDD discipline, emits tiered review signals (unit → composite → feature)
  • Reviewer — tier-aware code review with stack-specific checklists (React web supported)
  • Project Expert — accumulates package-specific knowledge across dispatches, orients every agent, offers opinions on HOW

The mob-boss skill itself is the top-level orchestrator — it dispatches the team, monitors metrics, and evolves agent definitions through patient A/B testing (minimum 5 tasks, 3 pattern occurrences before any modification).

Installation

# Add the marketplace
/plugin marketplace add exadev/mob-boss

# Install the plugin
/plugin install mob-boss@exadev-mob-boss

Usage

# Verify setup (run before first dispatch)
/mob-boss:init

# Start a new task
/mob-boss:dispatch Implement user authentication with OAuth2

# Check in-progress dispatch status
/mob-boss:status

# Resume an interrupted dispatch
/mob-boss:resume

# Close out an in-progress dispatch without further work
/mob-boss:close-out

Requirements

Platform file-watcher (optional)

The orchestration loop uses file-watching for real-time signal coordination. A native watcher is recommended for instant event delivery, but not required — mob-boss falls back to polling at 2s intervals using only standard POSIX tools (find, touch).

  • macOS: brew install fswatch (recommended)
  • Linux: sudo apt install inotify-tools (recommended)
  • Other platforms: polling fallback works automatically

No external dependencies are required to run mob-boss.

Supported stacks

Stack-specific guidance is provided for:

Stack Architect Developer Designer Reviewer
React on web (Next.js, Vite, CRA)
React Native
Other Falls back to base rules Falls back to base rules Falls back to base rules Falls back to base rules

How it works

Per-dispatch flow

  1. Phase 1 — Architect and designer collaborate to produce a unified design (plan + prototype)
  2. Phase 2 — Developer implements in chunked slices with tiered review signals
  3. Phase 3 — Close-out: archive, metrics, project-expert curation, user report

Tiered review

Every chunk emits a review signal at one of three tiers:

Tier Scope Focus
Unit (~50–150 lines) One testable piece Correctness, conventions, local bugs
Composite A cohesive slice of units Coherence, boundary tests, duplication
Feature The whole feature Architecture, plan compliance, security, integration

Agent evolution

Mob-boss tracks metrics across dispatches and evolves agent profiles when thresholds are met:

  • Minimum 5 completed tasks before any agent modification
  • Minimum 3 occurrences of a metric pattern before it's considered systemic
  • All changes are logged in the changelog with evidence

Project expert

Each package accumulates knowledge:

  • @expert: tags in agent reports surface facts worth remembering
  • Close-out curation investigates, cross-references, and records into .mob-boss/expert/knowledge/
  • The orientation snippet is injected into every agent spawn in future dispatches

State layout

Location Purpose
${CLAUDE_SKILL_DIR}/ (plugin) Read-only: canonical agent profiles, templates, preamble
~/.mob-boss/ (global) Mutable: evolved agents, experiment variants, metrics, changelog
<package>/.mob-boss/ (per-package) Mutable: signals, feedback, archive, project-expert knowledge

Structure

plugins/mob-boss/
├── .claude-plugin/plugin.json
└── skills/
    ├── dispatch/                     # Start new dispatches
    │   ├── SKILL.md                   # Full orchestrator protocol
    │   ├── preamble.sh                # Runtime environment setup + global state seeding
    │   ├── agents/main/               # Canonical agent profiles
    │   │   ├── architect.md + guidance/
    │   │   ├── developer.md + guidance/
    │   │   ├── designer.md + guidance/
    │   │   ├── reviewer.md + guidance/
    │   │   └── project-expert.md
    │   └── templates/                  # Copied into each new package on init
    ├── status/SKILL.md                # Read-only dispatch status check
    ├── resume/SKILL.md                # Resume interrupted dispatch
    ├── close-out/SKILL.md             # Close out without further work
    ├── init/SKILL.md                  # Verify setup / first-time seed
    └── team-manager/
        ├── SKILL.md                   # Dispatch coordinator
        └── context.sh                 # Agent profile loading

Licence

This plugin is the property of ExaDev Ltd. See the licence template in templates/LICENSE.md for terms applied to generated orchestration artefacts.

About

AI agent team orchestration with tiered incremental review, metrics-driven evolution, and per-package knowledge accumulation.

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors