Agent Harness Plugin

A lightweight Claude Code plugin implementing Anthropic's engineering best practices for effective long-running agent workflows.

For advanced context management with FIC (Flow-Information-Context) system, see UltraHarness

Overview

Long-running AI agents struggle across multiple context windows because each new session begins without memory of prior work. This plugin solves that problem by providing:

• Progress Tracking - Persistent log file (claude-progress.txt) that records accomplishments
• Feature Checklists - JSON file (claude-features.json) tracking feature status
• Git Checkpoints - Encourages frequent commits as safe recovery points
• Session Startup Routine - Automatically reads context at session start

Based on: Effective Harnesses for Long-Running Agents

Installation

Install this plugin globally to enable it for all your Claude Code projects:

claude plugins:add praneethpuligundla/harness

Or install from URL:

claude plugins:add https://github.com/praneethpuligundla/harness

The plugin is installed at user scope and applies to all Claude Code projects.

Usage

Initialize a Project

/harness:init

This creates:

• claude-progress.txt - Progress log
• claude-features.json - Feature checklist
• .claude/.claude-harness-initialized - Marker file
• Optionally: init.sh - Startup script

Check Status

/harness:status

Shows git status, recent progress, and feature summary.

Manage Features

/harness:feature add [name] - [description]
/harness:feature start [id]
/harness:feature pass [id]
/harness:feature list
/harness:feature next

Log Progress

/harness:log started [task]
/harness:log completed [task]
/harness:log blocker [issue]
/harness:log note [observation]

Create Checkpoints

/harness:checkpoint [message]

How It Works

Session Start Hook

When a Claude Code session starts in an initialized project:

1. Reads git log for recent commits
2. Reads progress file for context
3. Summarizes feature checklist status
4. Injects this context into the session

Session Stop Hook

When Claude stops responding:

1. Reminds to update progress file
2. Suggests committing work as checkpoint
3. Encourages merge-ready state

Best Practices

1. Initialize early - Set up harness at project start
2. List all features - Comprehensive checklist prevents premature completion
3. Work incrementally - One feature at a time
4. Commit often - Each commit is a recovery point
5. Log everything - Future sessions depend on this context

File Structure

project/
├── claude-progress.txt      # Progress log
├── claude-features.json     # Feature checklist
├── init.sh                  # Optional startup script
└── .claude/
    └── .claude-harness-initialized  # Marker file

Plugin Structure

harness/
├── .claude-plugin/
│   └── plugin.json
├── hooks/
│   ├── hooks.json
│   ├── session_start.py
│   └── stop.py
├── commands/
│   ├── init.md
│   ├── status.md
│   ├── log.md
│   ├── feature.md
│   └── checkpoint.md
├── skills/
│   └── harness-workflow.md
├── core/
│   ├── progress.py
│   └── features.py
└── README.md