Claude Code plugin

Idea to merged PR, through one gated pipeline.

harness-kit is a set of Claude Code agents that carry one idea through six gated steps, from spec to pull request. A script checks every step and a judge scores it, so you only decide twice: the direction, and the PR.

Install Pull an idea through it Open the brief builder

Open full screen

Why it is worth it

Without it	With harness-kit
The ticket lives in a chat thread	A PRD with the problem, the customers, and a numeric success metric
"Done" is whatever the agent decided	Acceptance criteria you can check one by one
The agent edits whatever it finds	A plan that names the files, and with graph engineering a gate that fails on any other file
Review is reading a diff and hoping	Every document scored against a rubric, retried until it passes 8.0
Quality depends on the day	The same six steps every time, with a score history you can watch

Install

/plugin marketplace add Pierry/harness-kit
/plugin install harness-kit@harness-kit
# restart Claude Code, open your repo, then:
/harness-kit:install

It asks two questions. Both defaults are fine, and you can change either later.

Question	Default	Other choices
Eval judge	`local`	`jev`
Graph engineering	`off`	`manifest`, `full`

Needs python3, git, and the gh CLI. To update: /plugin update harness-kit, restart, then /harness-kit:update.

Phase 1 of 3

Spec

Brief, PRD, PRP. Turn an idea into something you can check.

You do

Type a four-line brief after /golden-path, or fill the brief builder. Then read the PRD and approve the direction. This is your first decision, and the cheapest place to be wrong.

/golden-path

Squad: checkout
Problem: Returning guests abandon checkout when a card is declined once.
Hypothesis: If we add one-tap retry, completion rises 5 points.
Success metric: checkout completion, from 71% to 76% within 30 days

It does

The product manager agent writes the PRD: problem, customers, scope, and metrics with baselines. Then the PRP, an engineering spec with the files to change, patterns to follow, and acceptance criteria. Each must pass its structure check and score 8.0.

With Jev

Jev, a different model, scores each document. Claude decides only when Jev is unsure.

With graph

Each acceptance criterion gets a stable id like REQ-001, and trace/{feature}.yml is created.

Want fewer stops? /pipeline:run "<idea>" runs all phases and pauses only at your two decisions.

Phase 2 of 3

Build

Plan, dev, test. Every step runs the same gate.

write→structure check→score 8.0+→next step

Below 8.0 the agent rewrites only the failed checks, up to three times.

You do

Nothing. Watch it run, or run a step alone with /sse:plan, /sse:dev, /sse:test. hk status shows where it is.

It does

The staff engineer agent plans which files change and in what order, implements in small commits with your linters and .claude/conventions/, then runs your tests and reports failures by name.

With Jev

Jev scores the plan, the dev summary, and the test report instead of a Claude sub-agent.

With graph

The plan records each file it expects to touch, with evidence. A scope gate fails dev on any other file. Commits and tests are linked to the requirement they serve.

Phase 3 of 3

Ship

PR and merge. Your second decision.

You do

Read the PR description and approve. It opens as a draft on GitHub.

It does

Writes the title, summary, and test plan. A monitor watches the PR and clears the pipeline when it merges.

With Jev

Jev scores the PR description.

With graph

The PR carries a traceability table: each requirement, the code it touches, and the test that proves it. The next feature's plan marks these links VALIDATED or STALE.

Quality dashboard: each step's score across runs. Open it.

Cockpit: watch steps and approve gates from a terminal, npm i -g @pieerry/harness-kit then hk-tui.

Options

Four setups, same commands

Start with Plain. Each option is one command and can be switched back.

Setup	Best for	Adds	Cost
Plain (default)	trying it out, most features	the full gated pipeline	nothing extra
Jev judge	a judge that is not Claude grading Claude	Jev by TypeSafe AI scores with yes/no checks, Claude when unsure	about $0.04 per million tokens
Graph engineering	shared or regulated code you must audit	REQ ids, trace file, scope gate; `full` adds a local graph database, no Docker	free
Graph and Jev	a neutral judge and full traceability	both	about $0.04 per million tokens

/hk:eval local | jev
/hk:graph off | manifest | full
HK_STATUSLINE=1   # at install: show the pipeline step in your status line

Getting Started: every step, in all four setups.
Evals: how each document is scored.
Jev and System One: the Jev judge and when it hands off to Claude.
Graph Engineering: requirement ids, trace files, the scope gate.
Graph Theory: the model behind the graph.
Agents: product-manager (PRD, PRP), staff-software-engineer (plan to PR), system-architect (system design and review).