claude-harness-engineering
Personal Harness Engineering plugin for Claude Code. Deterministic guardrails, fast Write/Edit feedback loops, Proto-first architecture, and multi-model PR review — for TypeScript, Go, and Protocol Buffers projects.
日本語版 / Japanese README
What is this?
An opinionated Claude Code plugin that applies Harness Engineering principles to everyday coding sessions:
model = CPU, context = RAM, harness = OS
The goal: make it structurally impossible for the agent to ship code that violates your architecture, silence your linters, or break proto wire compatibility — without you having to babysit it.
Philosophy
"Most of the value comes from the harness, not the model." — Mitchell Hashimoto, OpenAI harness engineering, 2026
Instead of hoping the LLM behaves, this plugin wires deterministic tools into the Claude Code session lifecycle. The agent gets fast, tight feedback loops — not vague instructions.
Features
1. Write/Edit → Lint Feedback Loop
Every file write triggers a language-aware lint + format pass. Violations are injected back as additionalContext on the next turn, so Claude self-corrects within the same flow — without losing context or requiring a second prompt.
| Language | Formatter | Linter | Arch check |
|---|
| TypeScript / JS | Biome | Oxlint | dependency-cruiser / sheriff |
| Go | gofumpt + goimports | golangci-lint v2 | go-arch-lint |
| Proto | buf format | buf lint + protolint | buf breaking |
All formatters complete in <100ms (Biome, Oxlint, gofumpt are Rust/Go binaries — no slow Node startup).
2. Config Protection
biome.json, .golangci.yml, .go-arch-lint.yml, buf.yaml, tsconfig.json, and similar harness files are write-protected for agents. When Claude hits a lint violation, it cannot "fix" it by disabling the rule — it must fix the code.
Protected file categories:
| Stack | Files |
|---|
| TypeScript | biome.json, biome.jsonc, .oxlintrc.json, tsconfig*.json, .eslintrc*, eslint.config.*, .prettierrc*, .dependency-cruiser.* |
| Go | .golangci.yml, .golangci.yaml, .go-arch-lint.yml, gosec.yaml |
| Proto | buf.yaml, buf.gen.yaml, .protolint.yaml |
| Cross-cutting | lefthook.yml, .pre-commit-config.yaml |
| Claude Code | .claude/settings.json, .claude/review-profile.yaml |
Blocked edits produce a clear exit 2 message: harness-engineering: BLOCKED edit on protected config file. Only a human can change these configs.
3. Bash Firewall
Dangerous shell commands are blocked at the PreToolUse hook before execution:
git commit --no-verify — bypasses pre-commit hooksrm -rf on protected directoriesgit reset --hard on protected branches (main, develop, master)git push --force to protected brancheschmod 777, sudo rm, and other destructive patterns
The firewall exits with exit 2 and explains why the command was blocked. Claude cannot proceed without explicitly working around the constraint — which requires justification visible in the conversation.
4. Proto First-Class Support
Every .proto file write runs three checks:
buf lint — validates proto style and correctnessbuf format --diff — enforces consistent formattingbuf breaking --against .git#branch=main — detects wire-breaking changes before they land
Wire-breaking changes (removing fields, renaming enum values, changing field numbers) are caught immediately — not at deploy time. The buf.gen.yaml template supports both Connect RPC and gRPC code generation.
5. Multi-perspective PR Review
The pr-reviewer-3a skill reviews PRs from three angles — Business, Development, and Testing — in a single Claude session. Four modes:
| Mode | Description |
|---|
standard | Sequential Business / Dev / Testing review |
mob | Broad scan first, then deep-dive on flagged files |
deep | 3-round progressive review (surface → logic → integration) |
multi | Parallel personas, each perspective runs independently |
Run via /harness:review-pr or /harness:review-pr --mode mob.
Skills
25 skills ship with this plugin. Each lives in harness-engineering/skills/<name>/SKILL.md.
Harness (bootstrap + guardrails)
| Skill | Description |
|---|
harness-setup | Interactive repo bootstrap: detects stack, copies CLAUDE.md, installs configs and lefthook |
config-protect | Write-protection policy for linter configs; what's protected and why |
edit-lint-feedback-loop | PostToolUse → lint → additionalContext injection pipeline details |
careful-mode | Production safety guard that blocks destructive commands for the session |
Language stacks
Jeffallan
greyhaven-ai
markus41
upstash
wshobson
nyldn