Andon plugin
A plugin that catches an agent false "done": a Stop hook asks for evidence and every defect becomes a lasting safeguard
- 168 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The Stop, UserPromptSubmit and PreToolUse hooks run Python scripts on every agent step and can block a reply
- Verification logs and project memory live in files that reach the repository unless excluded
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/andonDetects 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.
- The Stop, UserPromptSubmit and PreToolUse hooks run Python scripts on every agent step and can block a reply
- Verification logs and project memory live in files that reach the repository unless excluded
This is third-party code. Review the repository files before installing.
What it does
Andon brings lean manufacturing practice to work with an AI agent. The name comes from the cord on a Toyota line that you pull to stop production. The plugin installs into Claude Code and Codex and turns on three hooks. The Stop hook, the claim check, finds completion phrases such as "the migration is complete and the tests are fixed" and compares them with verification entries. In warn mode it asks for evidence, in block mode it returns a decision to stop. The UserPromptSubmit hook reads the project `memory/MEMORY.md` at session start. The PreToolUse hook guards against test pipes that hide the exit status and is off by default. The set includes the commands `/andon:wrap`, `/andon:orient` and `/andon:log-mistake` and five agents: researcher, auditor, memory-steward, builder and chief-of-staff. The command `andon-log-claim "what I claim" "how I verified it"` records a verified claim that the hook then reads. The project runs in four layers: a memory file, a wrap and orient loop, the Stop cord and lane discipline. In `DEFECT_LEDGER.md` each defect is logged as a chain of cause, countermeasure and result with a review date. An overdue countermeasure fails the tests.
Who it is for. People who work in an agent every day and are tired of repeated mistakes and "done" on red tests.
Good fit when
- The agent declares work finished without running tests, and the Stop hook has to demand evidence
- A project resumes after three months and the agent learns the stack, rules and decisions straight from `memory/MEMORY.md`
- A repeated mistake is logged with `/andon:log-mistake` and closed with a safeguard such as a hook or a test
- At session end `/andon:wrap` saves the outcome so the next one starts with orientation through `/andon:orient`
Not a fit when
- You only want memory across sessions: the author suggests a lighter memory repository
- You need a hook in Cursor, Cline, Windsurf or Aider: templates are copied by hand there and the hooks were not tested
- You need proof of real execution: the agent writes its own verification entry, and word overlap can accept unrelated work
Example request
Install Andon, keep the claim check in warn mode and show how the hook reacts to the phrase "the migration is complete"Limitations
The hook matches words between the claim and the verification entry, so by words it can accept unrelated work or a fabricated entry. The author advises staying in warn mode until you have checked its behavior on your work and only then enabling block. In Codex the hooks work after you trust them in `/hooks`, the warning text does not show in `codex exec`. The `tool` and `both` evidence modes read only the Claude Code transcript. The plugin and a manual hook install must not run together. Python 3 and a project directory with `memory/` are needed.
How to disable. In Claude Code run `/plugin uninstall andon@primefold-tools`, in Codex disable the plugin and the hooks in `/hooks`. The `memory/` folder and logs stay in the project as ordinary files.
FAQ
How do I install Andon in Claude Code?
Run `/plugin marketplace add PrimeFoldTools/andon`, then `/plugin install andon@primefold-tools`. The Stop hook starts in warn mode, the orient hook is also on, and the guard against hidden test statuses stays off.
How do I install Andon in Codex?
Run `codex plugin marketplace add PrimeFoldTools/andon` and `codex plugin add andon@primefold-tools`, then type `/hooks` in `codex` and trust the three hooks. Codex does not run hooks without trust.
What does the claim check hook do?
It finds completion phrases in the agent reply and compares them with entries made by `andon-log-claim`. In warn mode it asks for evidence, in block mode it returns a decision to stop.
How is Andon different from memory repositories?
Memory stores the lesson, while Andon closes a defect with a safeguard that makes the mistake harder to repeat, such as a hook or a test. Memory is only the first of four layers.
Can I try the hook without installing?
Yes. Clone the repository and run the command from the "Reproduce it in 30 seconds" section: `claim_check_hook.py` in `warn` mode takes a completion phrase and returns `continue: true` with a request for evidence.
Discussion
No questions or reviews yet.
Related
PluginsProcoder is a single Go binary with a commit gate, quality controllers and a lessons ledger so agents cannot ship unfinished work
Shared Claude Code settings for a team: hooks against dangerous commands, a check before finishing and skills from spec to review
A 33-skill plugin with gates: question the request, verify facts against sources, test first, review by a second model
A Claude Code plugin that hands a task to a local no_human and reviews a branch with a reviewer that never saw the coder's session
A plugin makes the agent explain its code and, before a push, ask questions so you understand what you are merging.
lessloc uses hooks to keep an agent to the least code: it refuses placeholders and junk and sends unfinished work back
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Andon (this card) | no | 168 | |
| HOL Guard | yes | 1,094 | checks commands, secret reads and MCP calls before they run, while this catches false completion and keeps a defect ledger |
| Superpowers | no | 316,679 | a development process skill set, while this adds evidence hooks and safeguards against repeated mistakes |
README in short
The README sets the goal of not repeating the same AI mistake twice and gives a no-code quick start with a memory file. A 30-second hook demo, installation as a plugin for Claude Code and Codex, a table of supported apps, the defect ledger and the four system layers follow. The repository contents close the page.