Better Harness plugin
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
- 2,413 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Reads records of the agent's past sessions in the project, which can hold prompts, file paths and command output, and with permission also the home folder and memory
- Starts three extra agents for analysis, which spends requests to your model, and writes reports into the project folder
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/better-harnessDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Reads records of the agent's past sessions in the project, which can hold prompts, file paths and command output, and with permission also the home folder and memory
- Starts three extra agents for analysis, which spends requests to your model, and writes reports into the project folder
This is third-party code. Review the repository files before installing.
What it does
Better Harness is an open-source plugin that evaluates the harness around the agent: rules, specs, skills, checks and the order of delivery. It runs through your coding agent with the `/better-harness` command, or `$better-harness:better-harness` in Codex. The result is a report on five dimensions of the agent work loop: task understanding, controlled execution, change validation, reliable delivery and learning capture. Each dimension rests on what is found in the project: `AGENTS.md`, rules, skills, commands, MCP, hooks, tests, CI. The command assembles a data bundle from three independent sources: logs of the agent's past sessions in this project, the repository itself and the agent's configured assets. Then three separate read-only agents run, one per source, and a lead agent merges their conclusions. Missing data is marked explicitly in the report and gets no score. Each finding carries an impact, an expected output, a repair boundary and acceptance checks, and the repair action drafts a fix plan for your review. The report is saved as a self-contained `report.html` with `report.md` and `findings.json`, in Claude Code by default to the repository's `.claude/better-harness` folder. In Qoder and Cursor it opens as a native canvas. A separate `better-harness` command-line program can check the install state with `plugin status` and `doctor`. Claude Code, Codex, Qoder, Cursor, GitHub Copilot CLI, Qwen Code, Pi, Kimi Code, WorkBuddy and Grok are supported.
Who it is for. For teams where AI agents already write a lot of code and for team leads who want to understand why an agent solves the wrong problem, skips checks or repeats the same mistakes.
Good fit when
- An agent confidently solves the wrong problem or cannot tell what done means, and you need to check rules, specs and acceptance criteria
- The team has adopted skills, hooks and MCP but it is unclear whether agents really use them in live sessions
- Before bringing AI agents into the delivery process you need to check that they do not bypass review, tests and approvals
- After a few weeks you want to compare reports on the same dimensions and see how the harness changed
Not a fit when
- You need to check the code itself or find a bug in a diff: the report evaluates the process around the agent, not the final lines of code
- The project has no sessions and no configured rules yet: a report on empty data shows little and the missing parts stay unscored
- You need continuous real-time monitoring: the plugin runs on command and does not watch the agent all the time
- You are not ready to let the agent read logs of past sessions: for analysis it reads the local session records of this project
Example request
/better-harness analyze how the AI coding workflow is set up in this project and prepare a report with findings and checksLimitations
- The `@qoder-ai/better-harness` package carries version 0.7.0-alpha2, and the command runs only by hand: the skill description says it must be invoked with the slash command.
- There is no single entry point for all harnesses: install, verification and launch differ for Claude Code, Codex, Qoder, Cursor and Copilot. Cursor is not published to the plugin marketplace, and WorkBuddy has no managed plugin lifecycle.
- Session analysis relies on what the harness saves. Copilot CLI records no per-response token data and VS Code chat has no durable transcript, so that data stays unscored in the report.
- Reading the user's home folder, memory bodies and raw sessions needs a separate explicit permission. The three analysis agents spend requests to your model.
- The report shows observed trends and does not prove the cause of an improvement. Running from source needs Node.js 22.20 to 24. License MIT.
How to disable. Remove the plugin with your harness's own tools: through the `/plugin` menu in Claude Code, through Settings and Plugins in Codex Desktop. The `better-harness plugin plan` command builds an install, update and removal plan for the chosen harness. Delete the `.claude/better-harness` reports folder in the project by hand.
FAQ
What are the five dimensions of the report?
Task understanding, controlled execution, change validation, reliable delivery and learning capture. The first rests on `AGENTS.md` and specs, the second on skills and sandbox boundaries, the third on tests and hooks, the fourth on review and CI, the fifth on memory and reusable skills.
How does the plugin get data for the report?
The command assembles a bundle from three independent sources: agent session records, project files and the agent's configured assets. Three read-only agents each analyze their own source and do not see each other's conclusions. A lead agent merges the result and marks missing data explicitly.
What does the report do with missing data?
It keeps the gaps explicit. Behavior that could not be observed does not turn into a rating or an unsupported claim. Findings are ordered by priority, and each has an impact, an expected output, a repair boundary and acceptance checks. Report history can be compared on the same five dimensions.
What does the repair action in a finding do?
It drafts a scoped fix plan for your review and does not apply the change silently. A separate direct-fix mode starts with the words `fix` or `repair` at the beginning of the request. You review the result before any change goes into the project.
What is Harness Inspector?
An interactive viewer that links product intent, agent activity, sessions, files and commits on one timeline. It opens read-only. The public sample on the project site is built from fictional data and does not read your workspace.
How do I check the installation state?
The command-line program shows local traces of the install with `better-harness plugin status --host all` and `better-harness doctor --platform all`. It does not query registries or change settings. `plugin plan install` only prints the sequence of steps for the chosen harness.
Discussion
No questions or reviews yet.
Related
PluginsTurns a GitHub repo into a package with its own AI agent: an npx command, an MCP server, memory and a signed release
The Harness MCP server: one agent starts other Claude Code and Codex sessions on your computer, waits for them, reads their answers and closes them
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
A Rust app for macOS, Linux and Windows: skills, MCP servers, plugins, hooks and configs of 17 AI agents in one window, with a security audit
A curated, ranked list of agent harnesses with an MCP server and ready subagents to pick a harness for a task
One hundred ready agent teams for Claude Code, each with specialists, an orchestrator skill and domain skills, in English and Korean
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| Better Harness (this entry) | no | 2,413 | |
| Harness | no | 9,283 | a meta-skill that assembles a team of agents and skills from a domain description |
| AutoHarness | no | 10,482 | a self-learning skill layer that derives skills from your sessions and updates them |
README in short
The repository holds the main skill with references for the three data sources, extra skills (architecture model, change traceability review, spec cross-check), the evaluation model of the agent work loop, Node.js data collectors and report renderers, plugin manifests for several harnesses, the Harness Studio desktop app and docs. The front page opens with a choice of harness and a sample report, then explains the five dimensions, the architecture and installation.