The chachamaru127/claude-code-harness repository provides a structured development harness designed specifically for Claude Code, aiming to improve the quality and reliability of agent-driven software development. Its core purpose is to transform the often ad-hoc, chat-based workflow of AI coding agents into a disciplined, repeatable operating loop that emphasizes planning, verification, independent review, and evidence-based release. By enforcing a Plan→Work→Review→Ship cycle, the harness ensures that every development step is documented, validated, and independently checked before code is shipped.
The harness introduces a set of five core verb skills—plan, work, review, sync, and release—that encapsulate the main stages of the development process. Users begin by describing their desired outcome in plain language. The harness then generates drafts for specification and planning files (spec.md and Plans.md), which include scope, acceptance criteria, dependencies, unknowns, and stop conditions. These drafts are subject to user approval or correction, ensuring that the plan is clear and agreed upon before any work begins. Non-trivial planning is validated through team or sub-agent perspectives, checking for alignment, memory reuse, product fit, security, and practical viability.
Once the plan is approved, the harness executes the work phase, implementing only the approved slice of the plan. Test-driven development (TDD) is enforced when required, and verification is performed to ensure the work meets the specified criteria. The review phase is conducted independently from implementation, treating major findings as blockers that must be resolved before proceeding. Finally, the release phase packages only verified evidence, checks readiness, and ensures that all boundaries for changelog and tagging are respected.
The harness is compatible with Claude Code v2.1+, Codex CLI, Codex app, OpenCode, Cursor, GitHub Copilot CLI, and Antigravity CLI, though support tiers vary. Installation is streamlined, with quick setup commands and onboarding guides tailored to both new and existing users. Migration tools are provided to help users transition from older setups, inventorying plugin caches, skills, symlinks, and memory states without deleting data.
Advanced features include team-style execution (Breezing), Codex companion review for schema-backed second opinions, OpenCode bootstrap for guidance mirroring, and optional project-scoped memory recall via harness-mem. The harness does not inherit support claims from other projects and maintains strict boundaries for what is considered supported or proven within its own framework.
Comprehensive documentation is available, covering onboarding, installation routes, migration checks, skill trigger gates, capability matrices, compatibility notes, integration boundaries, distribution scope, runtime safety, and evidence packaging. The repository encourages contributions and acknowledges influences from hierarchical skill design and test tampering prevention patterns. Licensed under MIT, the harness aims to provide a robust, Go-native guardrail engine for agent-driven development, minimizing drift, enforcing discipline, and ensuring that every release is backed by verifiable evidence.
