Long Running Harness

Long-Running Harness

A Claude Code and Codex skill for maintaining continuity across multiple context windows in long-running software projects.

Based on Anthropic's research: [Effective Harnesses for Long-Running Agents](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents)

Problem

When AI agents work on complex projects spanning multiple sessions, each new session starts with no memory of what came before. This leads to:

• One-shotting: Attempting to build everything at once, exhausting context mid-implementation
• Premature completion: Declaring the project done when features are still incomplete
• Lost progress: Having to guess what happened in previous sessions

Solution

This skill implements a planner-led development approach:

Phase 1: Initializer

Sets up the project environment on the first run:

• long_running//plan.md - Planner-owned scope and acceptance criteria
• long_running//feature_list.json - Comprehensive feature requirements in JSON format
• long_running//progress.txt - Session work log for tracking progress
• long_running//init.sh - Development environment startup script
• long_running//state.json - Current harness phase and selected feature
• long_running//handoffs/ - Worker/evaluator handoff files
• long_running//prompts/ - Reusable prompt files
• Git repository with initial commit

Phase 2: Planner-Led Agent Team

Every subsequent session follows this workflow:

1. Orient - Read long_running//progress.txt and git log
2. Plan - Choose ONE incomplete feature and update long_running//plan.md
3. Implement - Delegate focused changes to a worker subagent when available
4. Verify - Delegate skeptical end-to-end checks to an evaluator subagent
5. Document - Commit changes, update feature_list.json, and append progress.txt

Agent