On this page
Coding agents fail in a few familiar ways. They forget what you decided last week. They repeat a mistake you already corrected. And they say "done" when the work isn't. The last one costs the most, because you believe it and move on.
Where the name comes from
On a Toyota line, the andon is the cord any worker can pull to stop production when something is wrong. The fix that follows is a countermeasure, a change that makes the same defect harder to repeat. I ran a manufacturing floor before I worked with agents. Andon is that discipline, moved into AI work.
What it checks
The core is a Stop hook. When the agent's final message claims the work is done, the hook looks for a verification record from the same session that matches the claim. If there is none, it warns you. Switch it to block mode and the turn cannot end until the proof exists. An opt-in mode goes further and checks the claim against the tool-use record the harness writes, which the agent cannot edit.
The newest piece is a pipe guard, off by default. It stops a test command piped into something like tail without pipefail. In that case the shell reports tail's exit status, so a failing run looks like a pass. While I was drafting this post, the guard stopped my own agent's test run for exactly that reason. The agent reran it with pipefail and all 747 tests passed.
The ledger behind it
Every check in Andon came from a defect I hit. The Defect Ledger in the repo records 19 of them: what went wrong, the root cause, the countermeasure, and whether it held. Three of the entries: "Fixed five, claimed fourteen", "The gate that wasn't running" and "The test that passed by luck".
Who it is for
Anyone who works inside Claude Code or Codex every day and has been told "done" when it wasn't. It is MIT licensed and free. Install it in Claude Code with /plugin marketplace add PrimeFoldTools/Andon, then /plugin install andon@primefold-tools. The README at github.com/PrimeFoldTools/Andon has the 5-minute start.
