Skip to content
/

Claude Conductor (lackeyjb)

claude-conductor-lackeyjb · lackeyjb/claude-conductor · ★ 14 · last commit 2025-12-30

Simplified Context-Driven Development plugin with 3-agent specialization (planner/implementer/reviewer) and lifecycle hooks for passive context management.

Best whenCommands should be thin dispatchers — all real work belongs in specialist agents. Never implement directly from a command.
Skip ifCommands implementing logic directly (must always delegate to specialist agents), Multi-phase runaway (must use phase-at-a-time by default)
vs seeds
bmad-method(persona-based agents), but lackeyjb's agents are dispatched dynamically rather than selected by the user.
Primitive shape 15 total
Commands 5 Skills 3 Subagents 3 Hooks 4
00

Summary

Claude Conductor (lackeyjb) — Summary

Claude Conductor by lackeyjb is a streamlined Claude Code plugin implementing Context-Driven Development, described as "Inspired by Conductor for Gemini CLI." It ships 5 slash-commands, 3 named sub-agents (planner, implementer, reviewer), 4 lifecycle hooks, and 3 skills as a clean, reduced-surface version of the rbarcante design. The core novelty versus the seed superbasicstudio/claude-conductor is the explicit 3-agent architecture: /conductor:new-track always delegates to a planner agent, /conductor:implement always delegates to an implementer agent, and phase checkpoints spawn a reviewer agent — establishing a clear separation of concerns through agent specialization. It also ships 4 hooks (SessionStart, PreToolUse:Write|Edit, PostToolUse:Write|Edit, Stop) that auto-load context and track changes without user intervention. Compared to rbarcante, it removes the Python CLI helper, Pattern Reference Layer, snippet library, ADR logging, and skill ecosystem — focusing on the core three-phase lifecycle with tight TDD enforcement. The MZWASHERE variant is a byte-for-byte copy of this repo, created on the same day, differing only in its README language.

01

Overview

Claude Conductor (lackeyjb) — Overview

Origin

Philosophy

Context-Driven Development — treating context as a managed artifact alongside code. The README tagline mirrors rbarcante: "control your code" by transforming the repository into a single source of truth.

Key beliefs:

  1. Three-agent separation: Planning, implementation, and review are distinct cognitive tasks that benefit from specialized agents.
  2. Parallel reads: Performance notes throughout prompt files remind the agent to run 3-4 Read calls simultaneously.
  3. Hooks for passive context: SessionStart loads conductor context automatically; Stop records session end — no user action required.
  4. TDD Red-Green-Refactor enforced: The implementer agent follows TDD cycle per task, with atomic git commits per task.
  5. Phase-at-a-time default: implement defaults to single-phase mode with interactive selection, preventing runaway multi-phase execution.

Relationship to rbarcante

lackeyjb is a simplification of the rbarcante approach. It removes:

  • Python CLI helper (no conductor_cli.py)
  • Pattern Reference Layer
  • Snippet Library
  • ADR logging
  • Skill ecosystem (skill-registry.json)
  • 4 of rbarcante's 9 commands (codeReview, patterns, skills, snippet are absent)
  • 6 sub-agents reduced to 3 specialized agents

It adds:

  • 4 explicit lifecycle hooks (not in rbarcante)
  • Cleaner 3-agent architecture with explicit Task tool delegation

Relationship to MZWASHERE

MZWASHERE/claude-conductor is an exact copy created on 2026-05-26 with only a different README (spam-style "download now" text). Operationally identical to lackeyjb.

02

Architecture

Claude Conductor (lackeyjb) — Architecture

Distribution

  • Type: Claude Code plugin
  • Install method:
    /plugin marketplace add lackeyjb/claude-conductor
    /plugin install conductor@claude-conductor
    
    or manual: ln -s /path/to/conductor ~/.claude/plugins/conductor
  • Required runtime: Node.js (for hook scripts), Claude Code

Directory structure

lackeyjb/claude-conductor/
├── .claude-plugin/           # Plugin manifest
├── commands/                 # 5 slash commands
│   ├── setup.md
│   ├── new-track.md
│   ├── implement.md
│   ├── status.md
│   └── revert.md
├── agents/                   # 3 sub-agents
│   ├── planner.md
│   ├── implementer.md
│   └── reviewer.md
├── skills/                   # 3 skills (subdirectories)
│   ├── code-styleguides/
│   ├── context-awareness/
│   └── tdd-workflow/
├── hooks/
│   ├── hooks.json            # Hook definitions
│   └── scripts/
│       ├── load-context.js   # SessionStart hook
│       ├── warn-config.js    # PreToolUse:Write|Edit hook
│       ├── track-changes.js  # PostToolUse:Write|Edit hook
│       └── session-end.js    # Stop hook
├── reference/                # Reference documentation
└── templates/                # Templates
    └── (workflow, styleguides, etc.)

Runtime directory (created at setup)

conductor/
├── product.md
├── tech-stack.md
├── workflow.md
├── tracks.md
└── tracks/<track_id>/
    ├── spec.md
    ├── plan.md
    └── metadata.json

Target AI tools

  • Primary: Claude Code (plugin installation)
  • Cross-tool note: Commands are Claude Code slash commands; no support for Gemini CLI or Codex listed

Languages / runtimes

  • Plugin prompt files: Markdown
  • Hook scripts: Node.js (JavaScript)
  • No Python required
03

Components

Claude Conductor (lackeyjb) — Components

Commands (5)

Command Purpose
/conductor:setup Initialize Conductor environment; detect brownfield vs greenfield; generate context files
/conductor:new-track Create track; delegate spec+plan generation to planner agent
/conductor:implement Execute tasks with TDD; delegate to implementer agent; spawn reviewer at phase boundaries
/conductor:status Display track progress
/conductor:revert Git-aware logical undo

Note: No codeReview, patterns, skills, or snippet commands (cf. rbarcante's 9 commands).

Sub-agents (3)

Agent Purpose
planner Specialist for spec.md + plan.md generation; dispatched by new-track
implementer Specialist for TDD task execution; dispatched by implement
reviewer Specialist for phase checkpoint review; dispatched by implement at phase boundaries

Defined as markdown files in agents/; spawned via Task tool.

Skills (3 subdirectories)

Skill Purpose
code-styleguides Language-specific style guide templates
context-awareness Skill for detecting and loading conductor context
tdd-workflow TDD Red-Green-Refactor cycle guidance

Skills are attached to agents (e.g., planner uses context-awareness) rather than auto-activating globally.

Hooks (4)

Event Matcher Script Purpose
SessionStart `startup resume` load-context.js
PreToolUse `Write Edit` warn-config.js
PostToolUse `Write Edit` track-changes.js
Stop * session-end.js Record session end

Templates

Templates directory contains workflow and styleguide templates.

Reference

Reference directory contains implementation notes and reference documentation.

05

Prompts

Claude Conductor (lackeyjb) — Prompts

Excerpt 1: commands/new-track.md — Parallel reads + Task tool delegation

---
description: Create a new feature or bug track with spec and plan
argument-hint: [description]
allowed-tools: Read, Write, Glob, Task
model: inherit
---

# Conductor New Track

Create new track (feature/bug/chore) with spec and plan.

## Pre-flight Checks

**Read these files in parallel (4 Read calls in one response):**
1. `conductor/tech-stack.md` - Verify exists, load tech context
2. `conductor/workflow.md` - Verify exists, load workflow context
3. `conductor/product.md` - Verify exists, load product context
4. `conductor/tracks.md` - Get existing track IDs (handle if missing)

If any of the first 3 files are missing: "Run `/conductor:setup` first."

...

## Delegate to Planner Agent

ALWAYS use Task tool:

Task tool:

  • subagent_type: 'conductor:planner'

  • prompt: | Create specification and implementation plan for:

    Track type: <feature|bug|chore>

    Context files:

    • conductor/product.md
    • conductor/tech-stack.md
    • conductor/workflow.md

    Generate:

    1. Specification (spec.md): Requirements, acceptance criteria, constraints
    2. Implementation plan (plan.md): Phased breakdown with TDD-oriented tasks
    3. Generate unique track_id: _

    Return track_id, file paths, and summary when complete.


Technique: Explicit parallel read instructions (performance optimization). Mandatory Task tool delegation — commands are thin dispatchers to specialist agents. Frontmatter allowed-tools: restricts each command to minimal tool access.


Excerpt 2: agents/planner.md — Specialist agent with skill attachment

---
name: planner
description: Specialist for generating specifications and implementation plans.
tools: Read, Write, Glob, Grep
model: inherit
skills: context-awareness
---

# Conductor Planning Agent

## Your Expertise

1. **Requirements Analysis**: Extract clear, testable requirements
2. **Specification Writing**: Create comprehensive spec.md
3. **Task Decomposition**: Break features into phases/tasks
4. **TDD Planning**: Structure tasks with test-first approach
5. **Risk Identification**: Spot blockers and dependencies

## Context Loading

**Read these files in parallel (3 Read calls in one response):**
- `conductor/product.md` - Product vision
- `conductor/tech-stack.md` - Technologies
- `conductor/workflow.md` - Methodology

## Generating plan.md

Structure:
```markdown
# Implementation Plan: <Track Title>

## Overview
Phase breakdown and approach

## Phase 1: <Name>
Goal: <What this phase achieves>

Tasks:

**Technique**: Agent specialist pattern with `skills:` frontmatter injecting skill content automatically. Agents have narrower `tools:` than commands. Performance-annotated parallel reads. Structured output format prescribed in the agent prompt.

---

## Prompting techniques observed

1. **Mandatory delegation** — commands always use Task tool; never implement directly
2. **Parallel read annotations** — performance notes explicitly specify "N Read calls in one response"
3. **Skill injection via frontmatter** — `skills: context-awareness` in agent header
4. **Type inference without asking** — track type inferred from description (add/create → feature; fix/broken → bug; update/refactor → chore)
5. **Frontmatter tool restriction** — `allowed-tools:` enforces least-privilege per command/agent
09

Uniqueness

Claude Conductor (lackeyjb) — Uniqueness

differs_from_seeds

Closest seed is superbasicstudio/claude-conductor (Archetype 4 — 0 commands, 0 skills, 0 hooks), but lackeyjb adds 5 commands, 3 specialist agents, 4 hooks, and 3 skills — transforming it from a passive markdown scaffold into a true active orchestration plugin. The mandatory agent delegation pattern (commands always use Task tool; never implement directly) is a clean architectural principle absent from the seed. Among seeds, spec-driver is the closest in spirit (skills-based structured spec → implement), but lackeyjb uses explicit slash commands rather than auto-activating skills. The 4-hook setup with passive context loading via SessionStart distinguishes it from all other CDD variants in this batch — only lackeyjb and MZWASHERE (its copy) have this.

Positioning

  • Versus rbarcante: lackeyjb is a simplified, cleaner derivation — fewer commands, no Python CLI, no Pattern Layer, but adds 4 hooks that rbarcante lacks
  • Versus MZWASHERE: identical codebase; MZWASHERE is a spam-forked copy
  • Versus MadAppGang conductor: MadAppGang has no hooks; lackeyjb hooks provide passive context management
  • Versus superbasicstudio seed: lackeyjb is the active plugin; the seed is the passive scaffold

Distinctive properties

  1. Mandatory Task tool delegation — commands are pure dispatchers; specialist agents do all work
  2. SessionStart hook — context auto-loaded at session start without user action
  3. 3-agent specialization — planner/implementer/reviewer separation of concerns
  4. Phase-at-a-time default — prevents runaway multi-phase execution by default
  5. skill attachment in frontmatter — agents declare skills: context-awareness to auto-inject skill content

Observable failure modes

  1. Node.js dependency: hook scripts require Node.js; silent failure if missing
  2. No code review automation: unlike rbarcante, no auto-triggered review on track completion
  3. No ADR logging: architectural decisions are not persisted
  4. 14 stars only: low adoption; uncertain long-term maintenance
  5. MZWASHERE copy dilutes discovery: the spam fork may confuse users
04

Workflow

Claude Conductor (lackeyjb) — Workflow

Phases

Phase Command Artifacts
1. Setup /conductor:setup conductor/product.md, conductor/tech-stack.md, conductor/workflow.md, conductor/tracks.md
2. Track Creation /conductor:new-track [description] conductor/tracks/<id>/spec.md, conductor/tracks/<id>/plan.md
3. Implementation /conductor:implement [track-name] [--all] Updated plan.md checkboxes; atomic commits per task
4. Status /conductor:status (read-only)
5. Revert (optional) /conductor:revert Git history modification

Approval gates

  1. Planner agent output review: new-track generates spec+plan via planner agent; user can review before implement
  2. Phase selection: implement defaults to single-phase interactive mode; user selects which phase to run
  3. Phase checkpoint: reviewer agent spawned at phase boundaries for quality check
  4. --all flag: to bypass phase-at-a-time gating user must explicitly pass --all

TDD cycle (per task)

  1. Red: create test file, write failing tests, confirm fail
  2. Green: write minimum code to pass tests, confirm pass
  3. Refactor: clean up while tests pass
  4. Commit: atomic git commit with task reference

Delegation model

/conductor:new-track
  └→ Task tool → planner agent
      (uses context-awareness skill)
      → returns spec.md + plan.md content

/conductor:implement
  └→ Task tool → implementer agent (per task)
      (uses tdd-workflow skill)
      └→ Task tool → reviewer agent (at phase boundaries)

Git commits

Atomic commits per task with git notes linking to track/phase/task.

Phase-at-a-time vs all-phases mode

/conductor:implement           # Single phase, interactive selection
/conductor:implement --all     # All remaining phases, no gating
06

Memory Context

Claude Conductor (lackeyjb) — Memory & Context

State storage

File-based in conductor/ directory:

File Contents
conductor/product.md Product vision and goals
conductor/tech-stack.md Technical preferences
conductor/workflow.md Development methodology
conductor/tracks.md Master track index
conductor/setup_state.json Setup resumption state
conductor/tracks/<id>/spec.md Feature specification
conductor/tracks/<id>/plan.md Implementation plan with checkboxes
conductor/tracks/<id>/metadata.json Track state

Passive context loading via hooks

The key difference from superbasicstudio's seed: context is loaded automatically at session start:

  • SessionStart hook runs load-context.js — auto-injects conductor context into the session
  • PostToolUse:Write|Edit hook runs track-changes.js — passively records what files change
  • Stop hook runs session-end.js — records session termination

This means the agent doesn't need to manually read all context files on each command invocation — the hooks maintain ambient context.

Cross-session handoff

  • setup_state.json enables setup resumption
  • tracks.md provides persistent work-in-progress state
  • metadata.json per track preserves status across sessions

Memory persistence scope

  • Project-scoped: all state in conductor/ within the repo
  • No global state

Context compaction

  • Parallel read instructions in agent prompts reduce sequential read overhead
  • No Python CLI consolidation (unlike rbarcante); relies on direct file reads
07

Orchestration

Claude Conductor (lackeyjb) — Orchestration

Multi-agent

Yes — 3 named specialists dispatched via Task tool.

Agent When dispatched By
planner Always by /conductor:new-track Task tool from new-track command
implementer Per task by /conductor:implement Task tool from implement command
reviewer At phase boundaries by /conductor:implement Task tool from implement command

Orchestration pattern

Hierarchical — commands are thin dispatchers; agents do all real work. The implement command spawns one implementer agent per task sequentially, then spawns a reviewer agent at phase boundaries. This is a sequential-with-checkpoint pattern, not true parallel.

Isolation mechanism

Git branches (inferred from workflow; explicit git notes per commit). No worktrees.

Multi-model

No — model: inherit throughout; uses whatever model Claude Code has configured.

Execution mode

Interactive loop — slash-command triggered per phase. Not a background daemon.

Prompt chaining

Yes — planner output (spec.md + plan.md) is consumed by implementer as its work queue.

Hooks as passive orchestration

The 4 hooks provide ambient orchestration:

  • SessionStart auto-loads context (replaces manual read)
  • PreToolUse:Write|Edit guards config file modifications
  • PostToolUse:Write|Edit passively records changes
  • Stop records session end

This is a form of event-driven side-channel coordination not present in rbarcante.

Crash recovery

Partial — setup_state.json for setup resumption. No explicit task-level crash recovery.

08

Ui Cli Surface

Claude Conductor (lackeyjb) — UI & CLI Surface

Dedicated CLI

No standalone binary. No Python CLI helper (removed from rbarcante design).

Local UI

None.

IDE integration

  • Claude Code plugin only
  • Installed via /plugin marketplace add lackeyjb/claude-conductor
  • Slash commands: /conductor:setup, /conductor:new-track, /conductor:implement, /conductor:status, /conductor:revert

Observability

  • tracks.md: master track index with status
  • metadata.json per track: machine-readable track state
  • PostToolUse hook: passively records file changes (track-changes.js)
  • session-end.js: records when sessions terminate
  • No ADR logging (unlike rbarcante)
  • No auto-generated code review report (unlike rbarcante)

Hook scripts (Node.js)

Script Trigger
hooks/scripts/load-context.js SessionStart — auto-load conductor context
hooks/scripts/warn-config.js PreToolUse:Write
hooks/scripts/track-changes.js PostToolUse:Write
hooks/scripts/session-end.js Stop — session termination record

Related frameworks

same archetype · same primary tool · same memory type

Claude-Flow / Ruflo ★ 55k

Eliminates single-agent context limits and sequential bottlenecks by orchestrating fault-tolerant swarms of specialized AI agents…

Hermes Agent (NousResearch) ★ 168k

Self-improving personal AI agent with closed learning loop, 7 terminal backends, and messaging gateway — not tied to any AI…

OpenCode ★ 165k

Terminal-first AI coding agent with multi-model routing, native desktop app, and a typed .opencode/ configuration system for…

OpenHands ★ 75k

Open-source AI software development platform (open-source Devin alternative) with Docker sandbox isolation, 77.6% SWE-bench…

DeerFlow ★ 70k

Long-horizon superagent that researches, codes, and creates by orchestrating parallel sub-agents with isolated contexts in Docker…

oh-my-openagent (omo) ★ 60k

Multi-provider AI agent orchestration for OpenCode: escape vendor lock-in by routing Sisyphus (Claude/Kimi/GLM) and Hephaestus…