Before the alarm got a vote, I ran it silently over my last 50 merges. It cried wolf three times, so I fixed the alarm first. Then it went off on its own merge, which is exactly why it still only logs.
AI agents now take tickets off my board, build them on a branch and land them on main with one human press of Approve. The merge script checks a lot before the push: tests green, branch clean, main not behind. Nothing looked at what actually landed. So I asked for a smoke alarm: a fixed catalog of danger signs and a judge, written test-first, no model anywhere in it. Rule from my own research page: deterministic checks may one day recall a merge, model opinions may only flag. For now it does neither. It logs.
An alarm you have only tested on fixtures is a rumour. So the runner got a --replay 50 mode: rebuild every merge on main from git, judge it, and print a fire-rate table. It cried wolf three times. A --no-verify inside a test fixture (tests push to scratch repos), a new npm script key read as a new dependency, and a data refresh counted toward diff size. Each one got a failing test first, then the fix.
Tonight’s replay window includes the tripwire’s own merge. It fired three hard checks on itself. One is fair: the alarm’s own files are a protected path, so only a human merge may change them, and that is what happened. The other two are the alarm reading its own rulebook as evidence: the fake API keys in its tests look like secrets, and the line in the catalog that defines the --no-verify pattern contains --no-verify. Nothing was blocked, because nothing can be. That is the whole point of shadow.
The fixed alarm fires on 21 of 50 merges, 7 of them hard. Most hard fires are honest (a merge touched scripts/tix-merge.ps1, which only a human may change). The noise is soft: size 13 times, traceability 4 times (data chores with no ticket id). A fire rate like that is not boring yet, and an alarm that is not boring does not get to block anything.
The tripwire runs as the last step of the one-touch merge, after the dashboard restarts, so it can probe the live routes too. It writes one line and one log row. A runner crash becomes a warning, never a failed merge. Next to it, a smaller fix from the same evening: three Approve presses in a row used to merge in random order. Now they queue in press order. Tap any image to enlarge it and read the exact prompt that drew it.




The tripwire now watches every merge my agents land, and it is allowed to do exactly one thing: write it down. It gets promoted from shadow to gate only when its fire rate is boring. Not before. Today that rate is 21 of 50, and the loudest thing it ever flagged was itself.
Pictures only: the catalog, the replay, the three false fires, and the alarm flagging its own merge.
For anyone putting a guardrail on AI agents that ship code: a deterministic catalog, a pure judge, a replay over your own history, and a shadow step that logs and never blocks. Everything in this episode is free and open — clone it, run it, make it yours.
# repo soon. The shape: catalog of plain patterns → pure judge (no I/O, no model) → --replay 50 over main → one test per false fire → shadow step after the merge, log only → gate only when the fire rate is boring
No GitHub? Comment TRIPWIRE on the post and the bot DMs you the link.
A read-only scan of the NAS found where the space went: Synology Drive version history, quietly keeping every old copy. Next: the space-to-free report, and what I let an agent touch.