Coding Agent Harness
A document-native framework that turned AI coding agents' tasks, decisions and verification evidence into repository files, so a new agent could resume from durable facts instead of chat history.
Coding Agent Harness added working agreements, task plans, execution evidence, regression results, dashboards and closeout records directly into a repository as files, so that Codex, Claude Code, Gemini CLI and other coding agents stayed transparent and reviewable during long-running work. The core loop had a human state the goal, the agent read the repository's own harness files first, then followed a fixed sequence of diagnosing, deciding, scaffolding, configuring, verifying and delivering the work, with a command-line interface and dashboard exposing status, risk and review evidence.
The project's stated problem was not generating code, but keeping an agent oriented across days of work, preventing parallel agents from overwriting each other, and letting a new agent continue from facts recorded in the repository instead of previous chat memory. It required Node.js 24 or newer and ran as ordinary Markdown files, templates and check scripts with no required background service or database.
This repository is no longer maintained; its listed successor is a separate project.