oj

OpenJunto for Claude Code

A Claude Code configuration that transforms the AI into a coordinated team of 16 expert sub-agents with structured peer review.

Quickstart

OpenJunto ships as a Claude Code plugin — no build step. Install it from the marketplace inside a Claude Code session:

/plugin marketplace add openjunto/oj-claude
/plugin install oj@openjunto

Then start a new Claude Code session (restart, or /clear) to load the plugin. To iterate on the plugin locally without installing it, load the working tree directly:

git clone https://github.com/openjunto/oj-claude.git
claude --plugin-dir ./oj-claude

On the next session start, you should see a banner on stderr:

OpenJunto v<version> active — OpenJunto coordination system

The banner is printed by the SessionStart hook, so it appears on session start (startup, resume, /clear, compaction) — not on /reload-plugins or /plugin reload. After a reload, start a new session to see it. <version> comes from this repo's VERSION file.

Usage

The manager protocol loads at session start. Put OpenJunto to work by invoking a coordinated cycle with /oj:cycle <what you want done> (or /oj:run-task for a backlog item):

• /oj:cycle Review this pull request for security issues.
• /oj:cycle Fix the flaky test in auth_service_test.go.
• /oj:cycle Evaluate whether we should migrate from REST to gRPC for internal services.

The Manager handles triage, expert selection, peer review, and quality gates. You don't invoke experts by name — the Manager selects them.

What's Included

• 16 expert agents — full + compact profiles in agents/
• Templates — technical analysis, ADR, retrospective, session state, comms playbook (templates/)
• Complexity triage — Simple / Moderate / Complex tiers with process weight proportional to risk
• Peer review workflow — adversarial review by a different domain on all Moderate/Complex work
• Circuit breakers — auto-escalate after 3 revision cycles or 2 hours without progress

How It Works

1. Triage — Incoming requests scored against 4 criteria (multi-domain, regulatory, production risk, resource commitment)
2. Delegation — Manager spawns domain experts via the Task tool
3. Review — A different expert adversarially reviews the work, testing failure modes
4. Synthesis — Manager consolidates findings, surfaces dissenting views, hands back to you

Advanced: Backlog Sprint

For projects with a .claude/BACKLOG.md, run:

claude '/run-task'

The Manager picks the next item, delegates to experts, enforces peer review, and marks it done.

Directory Structure

The plugin tree (loaded by the plugin host from this repo's root):

.claude-plugin/
└── plugin.json              # Plugin manifest

CONDUCTOR.md                 # Manager coordination protocol (injected at SessionStart)
agents/                      # 16 full + 16 compact expert profiles
skills/                      # /run-task, /show-backlog, /save-session, /cycle
templates/                   # Structured output formats
reference/                   # Tier-loaded background (workflow, stakeholders, examples)
hooks/
└── hooks.json               # SessionStart + SubagentStart wiring
bin/
└── oj-helper                # Hook dispatcher + backlog helpers

Per-project, OpenJunto reads and writes a local .claude/ directory:

.claude/
├── CLAUDE.md                # Project-local protocol overrides (optional)
├── BACKLOG.md               # Project backlog (consumed by /run-task)
├── state/                   # Session state, carry-over notes
└── artifacts/               # Generated deliverables (ADRs, analyses, retrospectives)

Documentation

• WHY.md — the problem OpenJunto solves, concrete examples, honest tradeoffs
• docs/onboarding.md — your first 10 minutes after installation
• CONDUCTOR.md — the full Manager protocol
• agents/index.md — expert roster and engagement triggers