Repository Harness skill
Repository Harness installs an AGENTS.md entrypoint, a document map and explicitly invoked onboarding skills into a repo and updates them safely
- 1,251 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Executes a binary downloaded from a GitHub release; a checksum and the release identity are verified before it runs
- Writes and updates files in the repository: `AGENTS.md`, `docs`, `.agents/skills`, `.harness-core`, and `--override` replaces existing ones
- On your command the `$encode-invariant` skill adds checks to the project and `$improve-harness` changes agent instructions
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/repository-harnessDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Executes a binary downloaded from a GitHub release; a checksum and the release identity are verified before it runs
- Writes and updates files in the repository: `AGENTS.md`, `docs`, `.agents/skills`, `.harness-core`, and `--override` replaces existing ones
- On your command the `$encode-invariant` skill adds checks to the project and `$improve-harness` changes agent instructions
This is third-party code. Review the repository files before installing.
What it does
Repository Harness prepares a repository for coding agents such as Claude Code, Codex and Cursor. From the target repository you run the `install-harness.sh` script (`install-harness.ps1` exists for PowerShell). A `harness` binary then carries out the installation. The repo gains a compact `AGENTS.md`, a document map in `docs/README.md`, places for product notes, decisions and execution plans, templates, a reference on turning rules into checks, and five skills in `.agents/skills`. The skills run only on an explicit command, installation does not activate them. `$onboard-repository` studies an unfamiliar repo, traces one working path and proposes evidence-backed improvements, and its first pass is read-only. `$audit-onboarding-proposal` independently audits such a proposal. `$encode-invariant` turns an accepted rule into a check with positive and negative proof. `$improve-harness` tests one improvement against a baseline and a fresh rerun. `$engineering-wisdom` gives engineering advice and installs separately with `--with-engineering-wisdom`. Afterwards `scripts/bin/harness` provides `status`, `doctor` and `update`. An update keeps the exact upstream base in `.harness-core/`, runs a three-way merge and backs up changed files. If your edits and upstream edits overlap, no managed file changes: BASE, LOCAL, UPSTREAM and RESOLVED copies stay, a human decides, then `update --continue` or `update --abort` follows.
Who it is for. For teams and solo developers who hand a repository to coding agents and want one entrypoint with a document map and reproducible checks.
Good fit when
- Key decisions live only in chat and the agent has nowhere to read which documents are authoritative
- You need a compact `AGENTS.md` and a document map that can be updated without losing your edits
- An accepted rule, such as a ban on a certain dependency, should become an automatic check through `$encode-invariant`
- You point an agent at an unfamiliar repo and first want a map marking what is known and what is missing, without writing to files
- You want to confirm that a change to the agent's instructions actually helped on a fresh rerun
Not a fit when
- You need a task tracker, an agent orchestrator or an application runtime: the author explicitly rules those out
- You cannot download and run a release binary: installation and updates go through it
- You need ready build and test commands or credentials for your app: Harness does not install them
- You only need a project description for Claude Code in one file: maintaining `AGENTS.md` by hand is simpler, or take the Caliber CLI from the catalog
Example request
Prepare this repository for agents with Repository Harness, show me what will change first, then run onboard-repositoryLimitations
- Installation pipes a `curl` download into `bash` and fetches a binary. Preview what it will do with `--dry-run` before any files are written.
- Harness does not install application architecture, product policy, validation commands, credentials, a database or background processes. Operating the whole application is left to the repository owner.
- The former SQLite `harness-cli` and protocol v1 ended support on August 10, 2026. The last compatible release is `harness-cli-v0.1.22`; the current repo no longer builds it and does not delete old databases.
- `--merge` and `--override` decide the fate of existing files: the first keeps them and adds what is missing, the second replaces them. Use the second deliberately.
- The skills are written in English. The repository carries a license, see the `LICENSE` file for the other terms.
How to disable. Check the installed file list in `scripts/harness-install-files.txt` and delete those files from the repo, including `.agents/skills` and the `.harness-core` folder. The sources do not include an uninstall script.
FAQ
How do I prepare a repository for an AI agent and add AGENTS.md?
Run `install-harness.sh` from the repository folder, ideally with `--dry-run` first. It adds a compact `AGENTS.md`, the `docs/README.md` document map, plan and decision templates, and skills in `.agents/skills`. `--merge` preserves files you already have.
Do the skills switch on by themselves after installation?
No. Each skill runs only on a direct user command such as `$onboard-repository`. Neither a matching request topic nor installation activates a skill; the author states this for all five.
What happens if my edits conflict with a Harness update?
`scripts/bin/harness update` runs a three-way merge and, when edits overlap, changes no managed file. BASE, LOCAL, UPSTREAM and RESOLVED copies are kept. After your decision run `update --continue`; `update --abort` discards only the staged resolution.
Does `$onboard-repository` change repository files?
Its first pass only reads and produces a proposal. It records the initial Git root, revision, branch and status, never prints secret contents and preserves your uncommitted changes. Only exactly approved items within the requested scope get applied.
How do I check the installation is healthy?
`scripts/bin/harness status` shows the installation state and `scripts/bin/harness doctor` checks its integrity. `update --dry-run` is useful before an update.
How do I add the engineering advice skill?
It is a separate optional part. Run `scripts/install-harness.sh --with-engineering-wisdom --yes /path/to/project`, then invoke `$engineering-wisdom` by hand; it gives advice and does not install repository policy.
Discussion
No questions or reviews yet.
Related
SkillsThe author's personal set: shared agent rules and dozens of skills for GitHub, the browser, macOS, messaging and releases, plus a sync script
A coding agent for the terminal and desktop: it does not count "done" without tests or a diff and works with DeepSeek, GLM, Kimi, Claude, Codex and Gemini
Builder Skills
Skills for agents
A set of small skills from Builder.io for agent work habits: visual plans and diff recaps, docs-first checks, cross-agent review and clean status updates
Two skills for coding agents: wtf audits a project before shipping, better-codex keeps the agent disciplined while working on code
A shared skill and tooling infrastructure for 1C development, wired in as a git submodule and laid out across several agents at once
An autonomous iteration skill for agents: modify, verify, revert failed changes, and repeat the loop toward a target metric
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| Repository Harness (this entry) | no | 1,251 | |
| rpa-init | no | 1 | a single onboarding skill: the agent reads code and tests, sets up the environment and writes a report, with no AGENTS.md install or updates |
| Caliber | no | 1,309 | a CLI that creates and maintains context files for several agents: CLAUDE.md, Cursor rules, AGENTS.md |
README in short
The repository contains install scripts for Bash and PowerShell, the Rust binary sources, five skills, and `docs` and `tests` directories. The front page lists the problems agents hit (intent left in chat, authoritative documents unnamed, completion claimed without proof), what gets installed, maintenance commands, three release-evidence boundaries and the date protocol v1 support ended. The development section describes the pre-merge checks: Rust formatting and tests, Clippy, installer and documentation checks.