返回资讯中心
外部精选
软件工程
#开发者实时流

HAR 1.0: The open harness for AI coding workflows you can trust and verify

HAR is an open-source agent harness for AI coding workflows you can trust and verify. It runs every coding agent in its own isolated, consistent worktree, with deterministic validation, verifiable proof, and full observ…

Hacker News 最新karimtr2026/10/20 · 26 天前12 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

HAR is an open-source agent harness for AI coding workflows you can trust and verify. It runs every coding agent in its own isolated, consistent worktree, with deterministic validation, verifiable proof, and full observability. Run one agent or a whole fleet, with the same quality and control. It wraps the tools, tests, and CI you already use, and stays yours to customize.

Works with Claude Code · Cursor · Codex · any MCP agent.

HAR. The core harness, available as both a CLI and an MCP server. It turns any repository into isolated worktrees with deterministic launch, verify, and teardown stages for coding agents to work in.

Mission Control. HAR's open-source local dashboard. It gives you one place to keep track of every repository, worktree, run, validation, and artifact across your projects.

Plugins. A growing ecosystem of open source plugins that further expand HAR's customizability and functionality.

Install

npm install -g @osfactory/har

Get started

cd my-app har onboard # scaffold .har/ and print a prompt for your coding agent har env launch 1 # isolated worktree + running stack for agent slot 1 har env verify 1 --full # run the project's real checks, record what passed

Full walkthrough: Quickstart.

Running a coding agent is easy. Trusting what it produced, without re-checking every change by hand, is the hard part, and it only gets harder as tasks grow more complex and you run more agents at once. HAR was built to close those gaps:

No standard way to run or verify a repo. That knowledge is scattered across a README, a CLAUDE.md, Cursor rules, and CI yaml today, drifting out of sync with each other and the actual codebase. HAR replaces all of that with one machine-readable contract (.har/) that Claude Code, Cursor, Codex, or any MCP agent reads the same way.

Multiple agents on one repo collide. Shared dev server, shared database, shared ports, conflicting git state. HAR gives each agent its own worktree, ports, and database per slot, so a fleet can genuinely run concurrently.

Trusting an agent's change means re-verifying it yourself. Every task runs the same deterministic verify step and leaves an evidence trail, logs, artifacts, a validated tree hash, so a reviewer can check proof of what ran instead of relying on the agent's self-report.

One platform's sandbox locks you in. If the contract lives inside a vendor's hosted dashboard, switching coding agents later means rebuilding the whole verification setup. HAR's contract is an open standard living in the repo itself, portable across whichever agent or tool you adopt.

Hand-rolled scripts rot as the stack changes. A new dependency, a new service, a new env var, and nobody updates the script until an agent's run fails for a confusing reason. har env maintain diffs your installed harness against current templates and flags drift before it causes a silent failure.

HAR coordinates the work around the model, so agents can focus on the code and reviewers can trust the result.

flowchart LR D["Discover<br/>reads the harness contract"] --> I["Isolate<br/>one slot per agent"]

I --> A1["Agent 1<br/>own worktree, branch, ports, DB"] --> B1["Build"] --> V1["Verify"] I --> A2["Agent 2<br/>own worktree, branch, ports, DB"] --> B2["Build"] --> V2["Verify"]

V1 --> H["Hand off<br/>branch + evidence"] V2 --> H

Loading

Discover. The agent asks the harness what this project looks like, including its stack, its scripts, and what checks are available.

Isolate. Every task gets its own slot. That means a fresh copy of the repo on its own branch, with its own ports and, where the project needs it, its own database. Nothing is shared with the main checkout or with any other agent's slot.

Build. The agent edits and tests its work entirely inside that isolated copy. The main checkout stays untouched the whole time.

Verify. The project's own checks, whatever they are, run through the same pipeline every time and produce a consistent result. A full verification goes further and captures the state of the entire codebase at that moment, so a pass is tied to the exact code that was checked.

Hand off. Once verification passes, the session is torn down, but the branch and the proof of what ran are kept. A reviewer gets the code plus the evidence that it was checked.

Everything beyond install and first commands lives at harproject.dev.

See CONTRIBUTING.md for local setup, the dogfood harness loop, and architecture. Coding agents working on this repo should start with AGENTS.md. Maintainer release process: RELEASING.md.

HAR is sponsored by Kerno, runtime code and security tester for coding agents.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →