cc-harness

🌐 Language / 语言
English • 简体中文 • 日本語 • 한국어

cc-harness

Multi-agent harness framework for long-running application development with Claude Code.

Based on Anthropic's research: Harness Design for Long-Running Application Development

Why This Exists

Single-agent approaches to complex coding tasks produce superficially impressive but often broken results. Two persistent problems:

1. Self-evaluation bias — Models confidently praise their own mediocre work. Separating creation from evaluation is the strongest lever for quality.
2. Context deterioration — Models lose coherence over long sessions. Structured handoffs and context management keep multi-hour builds on track.

This plugin implements a generator-evaluator architecture (inspired by GANs) that drives real quality through iterative feedback loops.

Architecture

User Prompt (1-4 sentences)
        │
        ▼
┌──────────────┐
│   PLANNER    │──→ SPEC.md (ambitious product spec)
│  (read-only) │
└──────────────┘
        │
        ▼
┌──────────────┐         file-based          ┌──────────────┐
│  GENERATOR   │◄──────communication────────►│  EVALUATOR   │
│  (builds)    │                              │  (tests)     │
└──────────────┘                              └──────────────┘
        │                                            │
   git commit                                  tests live app
        ▲                                     via Playwright/
        └────────iterate if FAIL──────────────browser tools

Quick Start

Installation

# Add the marketplace
/plugin marketplace add Jinsong-Zhou/cc-harness

# Install the plugin
/plugin install cc-harness@cc-harness-marketplace

Then restart Claude Code.

Usage

# Start a harnessed development session
/harness Build a browser-based DAW using the Web Audio API

# Evaluate current work at any point
/evaluate
/evaluate the login flow
/evaluate frontend design

# Check session progress
/harness-status

What's Inside

cc-harness/
├── .claude-plugin/
│   ├── plugin.json                              # Plugin identity & metadata
│   └── marketplace.json                         # Marketplace distribution config
│
├── agents/
│   ├── planner.md                               # Expands prompts → ambitious product specs
│   ├── generator.md                             # Builds features one at a time, commits to git
│   └── evaluator.md                             # Skeptical QA — tests live apps, grades against criteria
│
├── skills/
│   ├── harness-loop/
│   │   ├── SKILL.md                             # Core generator-evaluator iteration loop orchestration
│   │   └── references/
│   │       ├── sprint-contract-examples.md      # Worked examples of sprint contracts
│   │       └── evaluation-examples.md           # Calibrated QA feedback examples from real runs
│   ├── context-management/
│   │   └── SKILL.md                             # Compaction vs reset strategies for long sessions
│   └── harness-tuning/
│       ├── SKILL.md                             # Evaluator calibration & harness simplification
│       └── references/
│           └── audit-template.md                # Template for auditing harness component necessity
│
├── commands/
│   ├── harness.md                               # /harness — start a harnessed development session
│   ├── evaluate.md                              # /evaluate — trigger standalone QA evaluation
│   └── harness-status.md                        # /harness-status — show session progress & scores
│
├── rules/
│   └── common/
│       ├── harness-workflow.md                  # Mandatory pipeline: plan → contract → build → evaluate
│       ├── evaluator-discipline.md              # Evaluator mindset rules — skepticism over politeness
│       ├── context-strategy.md                  # When to compact vs reset, model-specific guidance
│       └── file-communication.md                # File-based agent communication protocol
│
├── hooks/
│   └── hooks.json                               # Lifecycle hooks: SessionStart, PreCompact, Stop
│
├── scripts/
│   └── hooks/
│       ├── run-with-flags.js                    # Profile-based hook runner (minimal/standard/strict)
│       └── track-iteration.js                   # Iteration counter, state persistence, session summaries
│
├── examples/
│   └── todo-app-harness/                        # Complete worked example of a harness session
│       ├── SPEC.md                              # Example planner output