ui

Plan. Work. Review. Ship.
Turn Claude Code into a disciplined development partner.

English | 日本語

v4.2 Update — Claude Code 2.1.99-110 + Opus 4.7

Hokage line carries forward. Full Claude Code 2.1.99-2.1.110 + Opus 4.7 integration. Plugin manifest now matches the official plugins-reference schema.

Anthropic shipped Opus 4.7 with literal-instruction-following semantics, plus Claude Code 2.1.105 added the PreCompact hook and the monitors/monitors.json manifest. v4.2 makes Harness fit those changes precisely:

What you'll notice:

• Long-running tasks no longer get cut off mid-flight by automatic compaction
• claude plugin validate runs clean on Harness for the first time since monitors was added
• harness sync stops mysteriously deleting your monitors/agents block
• Reviewer agent gives sharper feedback (xhigh effort) when running on Opus 4.7

Update with the usual flow:

/plugin update claude-code-harness

For detailed Before/After in Japanese, see CHANGELOG.md [4.2.0] entry.

v4.0 "Hokage" — What's New

Go-native engine. 25x faster hooks. Zero Node.js dependency.

Every tool call Claude makes passes through Harness hooks. In v3, each pass cost 40-60ms of bash + Node.js overhead — a subtle drag you feel across hundreds of calls per session. v4 replaces the entire stack with a single Go binary:

What you'll notice:

• The micro-pauses between tool calls disappear — Claude feels more responsive
• No more npm install or Node.js version issues on setup
• Session startup is instant (10-30ms vs almost a second)
• Optional harness-mem integration: sessions remember what you worked on last time

Just update the plugin — no configuration changes needed:

/plugin update claude-code-harness

Why Harness?

Claude Code is powerful. Harness turns that raw capability into a delivery loop that is easier to trust and harder to derail.

The 5 verb skills keep setup, plan, work, review, and release on one path. The Go-native guardrail engine protects execution with sub-10ms response, and validation can be rerun when you need proof.

Compared With Popular Claude Code Harnesses

What matters here is not the theoretical ceiling of Claude Code. It is what becomes the default operating model once you install a harness — and what happens when you pair it with harness-mem.

With harness-mem, Claude Harness is the only setup where the default path is planned, guarded, reviewed, rerunnable, and remembered across sessions.

Full notes and source links: docs/github-harness-plugin-benchmark.md

Supported baseline and latest verified snapshot: see Claude Code Compatibility.

Requirements

• Claude Code v2.1+ (Install Guide)
  • v2.1.105+ recommended (PreCompact hook + monitors manifest)
  • v2.1.111+ for xhigh effort and Opus 4.7 support
• Opus 4.7 (claude-opus-4-7) recommended for full v4.2 benefit (literal instruction following, vision 2576px, xhigh effort)
• No Node.js required (v4.0 Hokage uses Go-native engine)

Who Is This For?

Install in 30 Seconds

# Start Claude Code in your project
claude

# Add the marketplace & install
/plugin marketplace add Chachamaru127/claude-code-harness
/plugin install claude-code-harness@claude-code-harness-marketplace

# Initialize your project
/harness-setup

That's it. Start with /harness-plan.

Language

Harness now ships with English as the default setup language. Japanese remains
fully supported as an explicit opt-in:

i18n:
  language: ja

You can also start a temporary Japanese setup with
CLAUDE_CODE_HARNESS_LANG=ja claude. The Japanese README is
README_ja.md.

🪄 TL;DR: Verified Work All

Don't want to read all this? Just type:

/harness-work all

One command runs the full loop after plan approval. Plan → Parallel Implementation → Review → Commit.

⚠️ Experimental workflow: Once you approve the plan, Claude runs to completion. Validate the success/failure contract in docs/evidence/work-all.md before depending on it in production.

The 5 Verb Workflow

0. Setup

/harness-setup

Bootstraps project files, rules, and command surfaces so the rest of the loop runs against the same conventions.

1. Plan

/harness-plan

"I want a login form with email validation"

Harness creates Plans.md with clear acceptance criteria.

2. Work

/harness-work              # Auto-detect parallelism
/harness-work --parallel 5 # 5 workers simultaneously

Each worker implements, runs a preflight self-check, and waits for an independent review verdict before completion.

Advisor Strategy

Harness is starting to adopt an advisor-style execution flow.
In plain terms, the executor keeps moving, and only asks for help when the decision is genuinely hard.

• High-risk tasks can trigger one preflight consultation
• Repeated failures can trigger a corrective consultation
• Plateau detection can trigger one last consultation before escalation
• Final approval still belongs to the reviewer, not the advisor

The first rollout is in harness-loop, because long-running execution benefits most from "ask only when needed."
Details: docs/advisor-strategy.md

3. Review

/harness-review

4. Release

/harness-release

Packages the verified result into CHANGELOG, tag, and release handoff steps after implementation and review are complete.

Safety First

Harness v4 protects your codebase with a Go-native guardrail engine (go/internal/guardrail/) — 13 declarative rules (R01–R13), single binary with sub-10ms response:

Runtime differences between Claude Code hooks and Codex CLI gates are documented in docs/hardening-parity.md.

5 Verb Skills, Zero Config

v4 unifies 42 skills into 5 verb skills. Start with the verbs first, then add Breezing, Codex, or 2-agent flows only when you need them.

Key Commands

Architecture

claude-code-harness/
├── go/                # Go-native guardrail + hookhandler engine
│   ├── cmd/harness/   #   CLI entry point (sync, doctor, validate)
│   ├── internal/      #   guardrail / hookhandler / state / lifecycle / breezing
│   └── pkg/           #   config / hookproto (public API)
├── bin/               # Compiled harness binaries (darwin-arm64/amd64, linux-amd64)
├── skills/            # 5 verb skills + extension skills (plan/execute/review/release/setup, etc.)
├── agents/            # 3 agents (worker / reviewer / scaffolder)
├── hooks/             # CC hook configuration (hooks.json)
├── scripts/           # Auxiliary shell scripts
└── templates/         # Generation templates

Advanced Features

Breezing (Agent Teams)

Run entire task lists with autonomous agent teams:

/harness-work breezing all                    # Plan review + parallel implementation
/harness-work breezing --no-discuss all       # Skip plan review, go straight to coding

Phase 0 (Planning Discussion) runs by default — Planner analyzes task quality, Critic challenges the plan, then you approve before coding starts. 8+ tasks auto-split into manageable batches.

Session Memory (harness-mem)

When harness-mem is running, Harness automatically records session events — what you worked on, what tools were used, and how the session ended. Next time you start a session, that context is available for retrieval.

• Without harness-mem: events are logged locally to .claude/state/memory-bridge-events.jsonl (no external dependency)
• With harness-mem: events are also sent to the memory server for cross-session search and retrieval

No configuration needed — Harness detects harness-mem automatically.

/harness-work --codex implement these 5 API endpoints
/harness-review --codex  # 4 perspectives + Codex second opinion

/harness-release handoff  # Report to Cursor PM

/generate-slide   # 3 visual patterns, quality scoring, auto-export
/generate-video   # JSON Schema-driven pipeline with Remotion

Why Harness vs Skill-Pack Only?

Skill packs can teach a prompt. Harness also enforces behavior at runtime.

• Guardrail engine blocks destructive writes, secret exposure, and force-push patterns on the actual execution path.
• Hooks + review flow keep quality checks close to the tools that edit your repo.
• Validation scripts + evidence pack give you a rerunnable way to confirm docs, packaging, and /harness-work all behavior.

Troubleshooting

For more help, open an issue.

Uninstall

/plugin uninstall claude-code-harness

Project files (Plans.md, SSOT files) remain unchanged.

Claude Code Feature Highlights

Harness leverages the latest Claude Code features automatically. Here are the ones you'll notice:

Full technical list (19 features): docs/CLAUDE-feature-table.md

Documentation

Contributing

Issues and PRs welcome. See CONTRIBUTING.md.

Acknowledgments

• AI Masao — Hierarchical skill design
• Beagle — Test tampering prevention patterns

License

MIT License — Free to use, modify, commercialize.

English | 日本語