Harness

This is an opinionated framework for running agent tasks that span multiple context windows or sessions. It uses two files (an append-only log and a structured JSON state) to track progress, plus git-based checkpointing so an agent can recover exactly where it left off after a crash or context reset. The lock mechanism uses portable mkdir atomics, and everything is designed to be grep-able and idempotent. The core insight is that premature completion is the main failure mode for long-running agents, so it enforces explicit task dependencies and validation commands. If you're building workflows that need to survive interruptions or run distributed work across sessions, this gives you the scaffolding.

skills.sh installs: 126
GitHub stars: 2.7K
last updated: Jun 3, 2026

View on GitHub

Install

$ npx -y skills add cexll/myclaude --skill harness --agent claude-code

Installs into .claude/skills of the current project.

Files

Harness — Long-Running Agent Framework

Executable protocol enabling any agent task to run continuously across multiple sessions with automatic progress recovery, task dependency resolution, failure rollback, and standardized error handling.

Design Principles

Design for the agent, not the human — Test output, docs, and task structure are the agent's primary interface
Progress files ARE the context — When context window resets, progress files + git history = full recovery
Premature completion is the #1 failure mode — Structured task lists with explicit completion criteria prevent declaring victory early
Standardize everything grep-able — ERROR on same line, structured timestamps, consistent prefixes
Fast feedback loops — Pre-compute stats, run smoke tests before full validation
Idempotent everything — Init scripts, task execution, environment setup must all be safe to re-run
Fail safe, not fail silent — Every failure must have an explicit recovery strategy