claude-code-harness
SHA-256A disciplined plan-work-review cycle for Claude Code that prevents agent drift by turning raw AI coding into structured, auditable delivery with spec, tests, and evidence packaging.
Smart Download
Download Download Version
v4.12.5 · 11.4 MB
A disciplined Plan-Work-Review-Release loop for Claude Code, turning raw agent sessions into structured, auditable delivery with spec, tests, and evidence.
Core Features
- Enforces Plan → Work → Review → Release loop with spec.md and Plans.md as source of truth
- TDD required for approved tasks; review and implementation are separated
- Automatically generates evidence pack for PRs and release preflight checks
- Zero Node.js dependency, built with Go native for minimal overhead
- Supports multi-host: Claude Code, Codex CLI, OpenCode, Cursor (candidate)
What It Can't Do
- •Works reliably only with Claude Code v2.1+; other hosts (Codex CLI, OpenCode, Cursor) are at varying support levels — don't expect full parity
- •/harness-work all should only be used when the plan is stable and repo baseline is known; misuse may cause chaos
- •harness-mem is optional and must be separately configured and healthy; absence won't break core functionality
- •Avoid using raw alternatives (e.g., raw codex exec) instead of the Harness command surface; you lose review and evidence packaging
Use Cases
- Teams that need predictable, auditable AI-assisted coding with clear gates and quality checks
- Individual developers who want to avoid 'drift' when using Claude Code for complex tasks
Detailed Introduction
Claude Code Harness is a lightweight, Go-native development harness that enforces a repeatable Plan → Work → Review → Release loop around Claude Code (and compatible hosts). Unlike raw agent usage where plans vanish in chat and testing becomes optional, Harness keeps spec.md and Plans.md as source-of-truth files, requires TDD for approved tasks, separates review from implementation, and packages only verified evidence for PRs or releases. It supports multi-host deployment (Claude Code, Codex CLI, OpenCode) via plugins or scripts, and provides commands like /harness-plan, /harness-work, /harness-review, and /harness-release. Compared to heavier tools like Superpowers or Hermes Agent, Harness is intentionally minimal (5 core verbs), has zero Node.js dependency, and follows a “not_observed != absent” principle to avoid data invention. It is ideal for teams or individuals who want predictability and auditability in AI-assisted coding without sacrificing flexibility.
Tags
Getting Started
Install the software
Double-click the downloaded installer and follow the prompts
Run /plugin marketplace add Chachamaru127/claude-code-harness in Claude Code
Run /plugin install claude-code-harness@claude-code-harness-marketplace
Run /harness-setup to initialize the harness
- Run /plugin marketplace add Chachamaru127/claude-code-harness in Claude Code
- Run /plugin install claude-code-harness@claude-code-harness-marketplace
- Run /harness-setup to initialize the harness
Latest Release Notes
## 🎯 What's Changed for You
**Breezing benchmark dependency alerts are closed, and the same lockfile is now guarded in CI.**
### Before → After
| Before | After |
|--------|-------|
| GitHub showed 10 open Dependabot alerts for the tracked Breezing benchmark `agent-eval` lockfile. | Open Dependabot alerts are now `0`, with patched dependency ranges locked and verified. |
| Benchmark dry-run scripts referenced stale `task-01`-`task-10` fixtures. | Dry-run scripts now target the tracked `task-11`-`task-20` eval fixtures and pass locally/CI. |
| The benchmark lockfile was not covered by scoped Dependabot npm updates or the main validation audit gate. | Dependabot and `validate-plugin` now check the benchmark lockfile path explicitly. |
## Fixed
- Closed the remaining Breezing benchmark `agent-eval` lockfile alerts by updating `@vercel/agent-eval`, using scoped npm overrides for patched `undici`, `minimatch`, and `uuid` ranges, and aligning benchmark dry-run task references.
## Security
- Added a scoped Dependabot npm update entry and CI audit gate for `benchmarks/breezing-bench/agent-eval`.
---
🤖 Generated with [Claude Code](https://claude.com/claude-code)
SHA-256 checksum verified
Checksum extracted from GitHub official Release page
SHA256 Checksum
f938a7f58cb7b10ff902d2969b9d8b198811e8e5f87ead178e39d758adef810eThis checksum is extracted from the GitHub Release page. Verify file integrity after download.
All SHA-256 checksums on this platform are extracted from the project's official GitHub Release page, without any modification. You can independently verify them on the GitHub Releases page.
Open Source Transparency
View GitHub SourceUninstall Info
Remove the plugin via /plugin remove claude-code-harness in Claude Code. Optionally delete generated files (.harness/, spec.md, Plans.md) from your project root. If harness-mem was used, run bin/harness doctor --migration-report first to check for residues.
No Extra Dependencies
Ready to use after download. No additional runtime required.
Similar Projects
Visual Studio Code (Code - OSS)
A free, open-source code editor by Microsoft with rich extensions, debugging, and Git integration. The most popular developer tool worldwide.
scrcpy
Display and control your Android device from your computer via USB or TCP/IP. No root required, high performance, low latency, supports audio and video mirroring.
Hoppscotch
Open source API development ecosystem with a lightweight, fast, and feature-rich interface for testing HTTP, WebSocket, GraphQL, and more.