LongHorizon-Harness CLI
A CLI that keeps an existing agent on a long task: plan a step, run it, verify the result in the real environment and checkpoint progress
- 1,870 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- It drives an agent that runs commands and edits files for hours on end
- It controls the desktop and terminal with real actions
- It acts in the launch directory, that is your real project
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/longhorizon-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.
- It drives an agent that runs commands and edits files for hours on end
- It controls the desktop and terminal with real actions
- It acts in the launch directory, that is your real project
This is third-party code. Review the repository files before installing.
What it does
The tool wraps an existing agent (Claude Code, Codex, OpenCode, DeepSeek Harness) in a loop for long computer-use tasks. Each round it recovers the original goal and verified state, picks the next bounded step, runs it in a fresh context, checks the actual result against files, UI, logs and tests, then checkpoints progress or records failure evidence and continues. The roles are split: a manager holds state and the next step, an executor performs the step, an auditor verifies independently. It covers desktop and terminal work, a command-line run and a web dashboard, and each role can use its own model and backend.
Who it is for. For developers and devops engineers who need to drive an agent through long multi-hour tasks without losing state.
Good fit when
- A task does not fit one agent pass and needs many verified steps
- You need progress to survive context resets and failures
- You want independent verification of the result rather than trusting the executor's report
Not a fit when
- The task is solved in a single agent pass without a long loop
- You have no access to a supported agent backend and its model
Example request
Carry this multi-hour task to completion: plan each step, verify the result in the real environment and checkpoint progress between roundsLimitations
The tool does not train or replace a model; it builds a loop around an existing agent, so a configured backend (Claude Code, Codex, OpenCode or DeepSeek Harness) and access to its model are required. Python 3.10 or later is needed. DeepSeek Harness in phase one is CLI-only, with desktop control and MCP promised later. MIT licensed.
How to disable. Remove the package: uv tool uninstall lh-harness or pip uninstall lh-harness. The ./.lh-harness working folder can be deleted by hand.
FAQ
Is this a separate model or a new agent?
No. It is a loop around an existing agent: it coordinates roles, verified state and cross-round progress, while your agent does the actual work.
Can I assign different models to the roles?
Yes. The manager, executor and auditor can each use their own model and backend to balance quality, speed and cost.
Discussion
No questions or reviews yet.
Related
CLI toolsThe Harness MCP server: one agent starts other Claude Code and Codex sessions on your computer, waits for them, reads their answers and closes them
rsc-harness is a CLI for configuring agent skills and project rules through a plan the user accepts before files change.
A Codex plugin that builds a team of subagents and skills for a project: an orchestrator splits tasks and checks the result.
The /better-harness command reviews how an AI agent's work is set up in a project: goals, checks, delivery and memory, and returns a prioritized fix report
A meta-skill for Claude Code that turns a domain description into an agent team and the skills it uses, using six team-architecture patterns
One hundred ready agent teams for Claude Code, each with specialists, an orchestrator skill and domain skills, in English and Korean
README in short
The README presents what the authors call loop engineering: the agent gets a goal once, and the tool repeatedly turns the remaining work into a bounded step, runs it on the right surface and verifies the actual result. Only what passes independent verification becomes trusted state; the rest stays failure evidence. It supports Claude Code, Codex, OpenCode and DeepSeek Harness, desktop and terminal work, a CLI run and a web dashboard. Installation is via uv tool or pip, with init, run, web and doctor commands.