How an AI coding agent works on GitHub so its replies and PRs are worth merging: skills, hooks and an installer for Claude Code, Codex and skills.sh agents
FAQ (answers come from our records)
How do I install it?
The repository provides this install command: npx mergeworthy. It comes from the repository itself; we cannot guarantee it runs as-is in your client.
Who is it for?
We filed it under these tags: 开源项目. The tags come from the entry itself, not a taxonomy we imposed.
Who maintains it?
repository nitedani/mergeworthy;first catalogued 2026-10-12;last seen 2026-10-12。There is no official certification or quality rating here — we record sources and dates only.
mergeworthy
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
Work with an AI coding agent through GitHub, the way you work with a coworker. You give it an issue or a request, then review its PR and answer its questions on the thread. You don't prompt it step by step or babysit a chat.
- **It ships merge-ready PRs.** It opens a draft with the first push, then verifies bugs, reviews and simplifies the change before marking it ready (converge). A UI or runtime change is also checked in the real app.
- **It keeps its threads moving.** It watches its PRs and issues, and answers every review comment, bot finding and red CI run. When a PR it depends on lands, it updates its own.
- **It writes for busy reviewers.** A post leads with its verdict, ends with one decision and its pick, and links the long material. An independent review checks every post before it goes out.
- **It asks only what isn’t its to decide.** The agent decides, acts and reports, asking for permission or an unrankable fork with its recommendation (always-on.md).
- **Hooks hold the line.** Hooks block an unreviewed post, a PR marked ready without its review, and a third comment in a row within three hours.
Install
The installer finds the coding agents on this machine and asks which to install into and how to set the options. It installs through each agent's own plugin system:
| Agent | Gets | Updates |
|---|---|---|
| Claude Code | the skills, the hooks, the bin/ commands and the always-on rules | automatically |
| Codex | the skills, and the always-on rules in ~/.codex/AGENTS.md | re-run npx mergeworthy |
| Others, through skills.sh | the skills | npx skills update |
Run npx mergeworthy again to change the options or add an agent; npx mergeworthy uninstall removes mergeworthy.
Options
| Option | Values (default first) | What changes |
|---|---|---|
| badge | on, off, auto | The icons of the agents that worked on a post, at its start; auto: only when posting from a human account. |
| merge | on-request-squash, reviewer | reviewer: the agent never merges. |
| watcher | on, off | off: no GitHub watcher. |
What's inside
always-on.md is loaded in every session: the rules that always apply, and when to open each skill.
| Skill | What it covers |
|---|---|
| task | the first skill of any task: triage, scope, the project file, the watcher, then the deliverable's skill |
| pull-request | any change, from an issue or not, to a merge-ready PR |
| reply | the watcher's events and every comment the agent answers |
| design | a design question, an API, protocol or module, a proposal |
| review | reviewing a PR it didn't write |
| open-issue | one finding a newcomer can find, reproduce and judge |
| docs | docs, READMEs and JSDoc that read as the project's own |
| delegate | briefing subagents and checking their work; skills/delegate/*.md holds one brief template per role |