Chachamaru127/claude-code-harness

What it solves

Claude Code Harness provides a disciplined delivery loop for AI coding agents (such as Claude Code, Cursor, and Grok) to prevent "agent drift," where plans disappear into chat history and testing or reviews are skipped under pressure. It replaces open-ended requests with a structured, repeatable process: writing a specification, implementing approved slices of work, verifying, and reviewing independently.

How it works

The project implements a structured workflow through five core "verb" skills: plan, work, review, sync, and release.

• Planning: Turns intent into a spec.md and Plans.md for user approval.
• Implementation: Executes approved tasks, enforcing TDD (Test-Driven Development) where required.
• Review: Conducts a separate review of the result, independent of the implementation phase.
• Synchronization: Compares the actual implementation against the original plan to report drift.
• Release: Packages verified evidence into a changelog and release tag.

To ensure safety, a Go-native engine adjudicates every tool call before it runs. This includes a "Runtime Floor" that denies high-risk actions (like secret reads or production deploys) and configurable "Guardrails" for actions like direct pushes to main.

Who it’s for

Developers using AI coding agents who want to move from "asking the agent to code" to a professional software delivery pipeline with strict boundaries and verification gates.

Highlights

• Pre-execution Safety: A Go-native engine that blocks or warns about risky operations before they occur, rather than reviewing them after the fact.
• Cros-Session Communication: A roster and messaging system that allows multiple agent sessions on one repository to communicate and share reports.
• Non-Engineer Views: HTML surfaces for Plan Briefs, Progress, and Acceptance that allow non-technical sponsors to judge progress without reading code.
• Multi-Tool Support: Integrated support for Claude Code, Codex CLI, Cursor, and Grok.