The new guard stopped the step that finished an ordinary job.
It had been turned on to block dangerous commands. But it also blocked sending approved work to the shared copy on GitHub. Every small planned task ended with that step. Now the safety check stood across the way out.
How do you fix a guard that catches the work it is meant to protect?
At First AI Movers, the guard had been copied from a public library on Friday. It was a small program that ran before Claude Code, an AI developer, could run a command. Think of a gate: if the check said no, the command could not pass.
Copying the files did not turn it on. That happened on Saturday. Tests showed that it blocked dangerous commands and let harmless ones through.
But its rule put every push on the dangerous side. A push sends saved changes from one computer to the shared copy. The imported test even called for an ordinary push to be blocked. By that test, closing the gate on routine work was success.
Then the newly active guard caught the push that closed a planned task.
Fifteen minutes after activation, a fix was recorded. Ordinary approved pushes were allowed. Commands that could rewrite or delete the saved history were still blocked. The test now described ordinary publication as allowed.
The guard had not been removed. Its line had been moved.
The revised rules also made a point that mattered beyond this one catch: when a valid command was blocked, that was a defect to fix. Keeping the check strict did not mean keeping every refusal.
Soon, another check was added to stop saved changes that contained text that looked like secrets. Then came a guard for commands that could rent servers or delete data. Those actions required a consent word in the same command—like a written yes attached to the request.
This time, the gate caught words about the work.
The deploy guard stopped its own saved-change message. That message listed the exact text patterns that made the guard refuse a command. The description of the new check had set off the check itself.
The message was rewritten around those patterns. The record does not name who rewrote it or show the first wording. There was no new rule allowing that message through. The way out was to change the words.
On Sunday, the trap returned in a larger form.
A scanner was added to read the project files and flag risky text: private links, text that looked like secrets, and dangerous commands. It was more like a search through every page than a gate before one action.
Some files already had written exceptions at the top. The license, for example, needed to name its copyright holder. An allow line gave the scanner a reason to accept that name.
But the scanner’s own manual and policy listed the patterns it was built to catch. Without allow lines, those instructions tripped the scanner too.
The problem was found after the change had been merged. The trial run had used an earlier file state, before the new files were added to the saved change. It had missed the instructions that would set off the scan.
Ninety-five seconds after the scanner was added, a fix was recorded. Allow lines were placed at the top of its manual and policy. Each stated what kind of text the file could contain and why.
The next rules document arrived later that evening. Its allow lines were already near the top, before the rules began.
This time, the explanation came with its permission slip.
What any founder can take from this
- Test the step that finishes real work, including sending approved changes to the shared copy.
- Let a safety check read its own instructions. They may contain the very words it blocks.
- When a valid task is stopped, fix the rule or give a clear reason for an exception.
Next time: “The Assistant That Passed by Refusing.”
How this article was made: Based on the supplied facts sheet for the guards added from May 1 through May 3, 2026, drawn from saved-change messages, manuals, tests, and planning notes. Gates and permission slips are explanatory images. No scenes, dialogue, or feelings are dramatized. Technical terms are rendered in plain English.