I had deliberately simplified the main instructions my agents loaded. I wanted a small set of operating expectations, with detailed policies available separately when the task needed them.
When the behavior still didn’t match those instructions, the first explanation was that the file was too long. I asked the agent to check the actual configuration. The generated version had drifted away from the draft I intended to use, and extra rules had accumulated around it.
Adding another paragraph about following instructions would have left that source problem in place. I needed to know which file was authoritative and what could change it.
Make the intended configuration checkable
The repair restored my draft as the source for the main file. Changes to that draft required approval, and a line limit kept the generated result from growing indefinitely. The additional policies moved into an indexed directory.
That kept the main instructions readable while preserving the detailed rules. It also gave the deployment process something concrete to check. If the protected draft changed or the output exceeded the agreed limit, the configuration update should stop.
Some operating rules also moved into hooks. Git commands were a useful example because several prohibited options could be detected before the command ran. That gave the written policy an enforcement mechanism, although the coverage still depended on which agent environment and shell were being used.
I didn’t have complete coverage everywhere at the end of the session. The report separated the tested paths from the environments that still needed work. That distinction belongs in the handoff; a guard existing in one place doesn’t establish that every agent is using it.
I also asked Sol to review the broader audit. Several claims from the first pass needed correcting because they came from searches that were too narrow. Having a second model challenge the findings helped expose assumptions before they became more changes to the system.
My instruction to the repair work was to take one problem at a time: investigate it, implement the fix, test it, then continue. The configuration had enough moving parts already. I wanted each change small enough to see what had actually been repaired and what remained open.