JSPM

  • ESM via JSPM
  • ES Module Entrypoint
  • Export Map
  • Keywords
  • License
  • Repository URL
  • TypeScript Types
  • README
  • Created
  • Published
  • Downloads 53
  • Score
    100M100P100Q59172F
  • License MIT

Community harness kit for Claude Code — multi-agent orchestration with 7 patterns, 8 teams, built-in operations tooling

Package Exports

    This package does not declare an exports field, so the exports above have been automatically detected and optimized by JSPM instead. If any package subpath is missing, it is recommended to post an issue to the original package (harnesses) to support the "exports" field. If that is not possible, create a JSPM override to customize the exports field for this package.

    Readme

    harnesses

    npm License: MIT

    Community harness kit for Claude Code — multi-agent orchestration with 7 patterns and 8 specialist teams.

    한국어 문서

    Install

    npm install -g harnesses

    Note: npx harnesses does not work on npm 10+ due to changed binary resolution. Use global install.

    Claude Code (default)

    harnesses                  # install all teams to ~/.claude/
    harnesses --claude         # same, explicit
    harnesses be-team fe-team  # specific teams only

    Agents and skills are installed to ~/.claude/agents/ and ~/.claude/commands/.

    Codex / OpenAI

    harnesses --codex          # installs codex-harnesses + Codex auto-routing plugin

    Requires Python 3.11+ and uv (recommended) or pip3. After install:

    codex-harnesses route "Redis vs Memcached 결정해줘" --json
    codex-harnesses run "Review this PR for correctness and security"
    codex-harnesses "PostgreSQL vs MongoDB?" --option-a PostgreSQL --option-b MongoDB

    The Codex package ships native routing for debate, explore, review, and research teams. The installed Codex plugin teaches Codex to auto-route natural-language requests such as Redis vs Memcached 결정해줘 to the right harness, while /debate remains available as an explicit slash command.

    Teams

    Team Pattern Agents Skills Use when
    dev-team Pipeline 5 1 Building a full feature (FE + BE + review + QA)
    review-team Fan-out/Fan-in 6 2 Auditing existing code, a PR, or full codebase
    fe-team Expert Pool + Reflection 6 5 React, Next.js, Tailwind, a11y, perf
    be-team Pipeline + Expert Pool + Reflection 8 5 Hono/Express API, LLM integration, MCP
    explore-team Hierarchical Delegation 4 3 Investigating codebase, root cause, arch review
    research-team Blackboard 4 3 Web research, library comparison, fact-gathering
    debate-team Adversarial Debate 4 2 Architecture decisions, trade-off analysis
    ops-team Skills + Hooks 0 3 Release, CI watch, orphan process cleanup

    Patterns

    # Pattern Description
    1 Pipeline Sequential stages — each stage's output feeds the next
    2 Fan-out / Fan-in Parallel workers, results aggregated by a moderator
    3 Expert Pool Route to domain specialist agent
    4 Hybrid Pipeline + Expert Pool Sequential stages, each stage is an expert
    5 Hierarchical Delegation Scout orchestrates, delegates to specialists
    6 Adversarial Debate Advocates argue positions, judge decides (max 5 rounds)
    7 Blackboard Shared state file; agents read/write with no explicit handoff

    Cross-cutting: Reflection Loop, Circuit Breaker, Escalation, Consensus Voting.

    See docs/PATTERNS.md for full details.

    Repository layers

    Layer Path Purpose
    Core core/ Shared orchestration pattern specs, routing schema, run state schema, blackboard schema, and trace schema
    Claude Code claudecode/plugins/ Claude Code teams, agents, skills, hooks, and harness docs
    Codex codex/ Python package that orchestrates Codex CLI workers

    The npm CLI in bin/install.mjs is the distribution wrapper: by default it installs the Claude Code layer to ~/.claude/; with --codex it installs the Codex Python package, slash prompts, and auto-routing plugin.

    Layout note: Since 1.0.3, Claude Code plugins live under claudecode/plugins/ instead of the old root-level plugins/ directory. Existing installed teams still use the same ~/.claude/ destination paths; only the repository layout changed.

    Usage

    Skills (slash commands)

    # Frontend work
    /fe-component "UserCard with avatar and role badge"
    /fe-page "Dashboard with stats and recent activity"
    
    # Backend work
    /be-api "POST /invoices with Zod validation"
    /be-mcp-server "GitHub integration MCP server"
    
    # Review
    /review-code src/auth/middleware.ts
    /review-codebase src/ --screeners 5    # full codebase, 5 parallel screeners
    
    # Research
    /research-web "Compare Zustand vs Jotai for Next.js 15"
    
    # Debate
    /debate-tradeoff "SSR vs SPA for internal dashboard"
    
    # Operations
    /zombie-collector     # clean orphan processes
    /release              # release workflow
    /ci-watch             # monitor CI run

    Routing

    Claude Code routes your request to the right team automatically based on triggers in CLAUDE.md. You can also be explicit:

    "explore로 이 버그 조사해줘"
    "review-team으로 PR #42 리뷰해"
    "debate-team: Redis vs Memcached 결정해줘"

    Examples

    See examples/ for 5 real-world walkthroughs:

    1. Simple CRUD feature — dev-team
    2. Code review — review-team (SARIF output)
    3. Research task — research-team (Blackboard flow)
    4. Architecture decision — debate-team
    5. Debugging + cleanup — explore-team → ops-team

    Uninstall

    harnesses --uninstall           # Claude Code: remove from ~/.claude/
    uv tool uninstall codex-harnesses  # Codex: remove Python package

    Creating custom teams

    See docs/CREATING_TEAMS.md.

    License

    MIT — justn-hyeok