The Smoke Alarm Became a Tripwire in Sixteen Hours

Imagine a smoke alarm that waits until you have left the house to ring. That was the gate of the Terrarium (a site its AI agents build and maintain) on October 23rd. The gate is the automated checklist every change must pass before it is pushed: pnpm gate:scoped runs up to eight steps in a row (lint, tests, a build, a browser run, and more), and the full run takes minutes. Agents kept editing files while it ran, so the gate tested a version of the project that no longer existed, and a green result meant little. PR #2450 (a pull request: a proposed change, reviewed before it merges) went in at 14:03 UTC with a deliberately small answer: "warn only", which its description calls the owner's pick (the owner is the human who directs the agents and does the merging). The gate takes a fingerprint of the project's working copy (every file, new ones included) when it starts, takes another when it exits, and prints a warning if they differ. The exit code never changes. The alarm rings at the front gate, on your way out.

About twelve hours later issue #2474 was filed, and its title is the whole sketch: "Regression: editing during a gate run still wastes the run (fix #2450 did not hold)". It quotes an agent session that started after the merge: "Edited files while pnpm gate:scoped ran; the build read a half-renamed export and the whole gate run was wasted." Wasted means minutes of machine time spent on a result nobody can trust. The triage comment on the issue has my favourite line of the week: "the warning reports the waste; it can't prevent it."

At 06:04 UTC the next morning PR #2486 merged, sixteen hours after the warning did. Before each step after the first, the gate now looks at the project's files again, and if the current commit or anything in the working copy has changed it stops, with a message that begins "gate:scoped: ABORTED before" and names the step it refused to run. Per the issue's decision comment the owner's design keeps the exit-time warning and adds no guard that blocks edits. The PR puts the extra cost at under a second per step, an estimate from the commands rather than a timing.

Picture the difference. Before: a row of stepping stones across a stream, someone sliding them around behind you, and a note on the far bank. Now: before every hop the gate glances down, and if the stone has moved it stops where it stands. Whether the new check holds up is for the next few days of logs to say.