Harness — Long-Running Agent Framework

What it does

A comprehensive protocol for running agent workloads that span multiple sessions. The harness provides file-based progress persistence, atomic locking, checkpointing, dependency-aware task selection, and recovery rules for interrupted work. It defines lifecycle commands (/harness init, run, status, add) and explicit validation/rollback behaviors.

When to use it

Use this skill when a job requires progress to survive context resets, distributed workers, or when tasks have dependencies and need reliable retries and rollbacks (e.g., long CI jobs, multi-step migrations, batch automation across sessions). Triggers on harness-related commands or when multi-session orchestration is requested.

What's included

• Scripts: none detected in the skill directory, but the repo includes hooks and tests directories that support the framework.
• References: detailed protocol and example file formats (harness-progress.txt, harness-tasks.json) with sample JSON and log formats.
• Instructions: procedural guidance for locking, claiming tasks, validation, recovery protocol, and error categories.

Compatible agents

Best used by agents that can read/write files, run shell/git commands, and manage isolated working directories (CLI-capable agents like Copilot/Codex/Gemini CLI).