wolven-harness (@wolven-tech/harness)

CLI tool that installs and maintains a set of tools for AI agents to find instructions, use recorded decisions, plan and implement, then check their work and maintain human-decided patterns.

Docs: wolventech.github.io/wolven-harness. See also Commands.

Install

pnpm add -D @wolven-tech/harness && pnpm exec wolven-harness setup

Run at the git top level. You need Node 22 or later, git, pnpm, and macOS or Linux.

Setup

setup creates only paths that are missing. It never creates or edits AGENTS.md, and it never commits. It also:

• adds missing harness:validate, harness:comments and harness:score scripts to an existing package.json, and writes a starter .harness-score.json;
• rewrites .wolven-harness.json on every run, with your setup choices and packageVersion.

To be asked questions, stdin and stdout must both be a TTY. Otherwise pass --git-host gh|bit and --runtimes (a comma list of claude, codex, cursor), unless .wolven-harness.json already records them.

Nine core skills are always installed. Two optional sets add more:

• ship: code-commit, code-pr, code-review, code-ci.
• discovery: create-prd, prototype, handoff, the-fool, the-jury.

Interactive runs preselect ship. Non-interactive runs without --skills install core and ship. Use --skills ship,discovery to pick sets, or --skills none for core only. Re-runs never remove an installed set.

What to do next

Ask your agent to run harness-init. On a fresh or thin-evidence repo, the lean path asks for your immediate goal in one sentence (unless you stated one), then may defer a few items after naming each to you; lean-path.md lists them. Your choices land under Deferred / skipped steps in the session note, and skill proposals still run. See harness-init. Then:

1. code-spec: agree on the change and how to verify it.
2. code-plan: approve the implementation steps.
3. code-execute: make the change and run the checks.
4. code-commit: commit the result when you are ready.

Commits are manual by default. For a one-file change, go straight to code-execute.

create-prd, prototype, handoff, the-fool, and the-jury need the discovery set: pnpm exec wolven-harness setup --skills discovery. code-pr, code-review, code-ci, and handoff are ask-only and never merge.

All 18 skills are listed on the skills page.

License

MIT © WolvenTech. See LICENSE. To contribute, see CONTRIBUTING.md.