When the live site breaks

The preview working says little about the live site: it has its own data, its own secrets and its own history since the last publish. Look at what the live site really has, in this order:

1. service_status with env production. When the live site answers errors it says when they began, the requests just before the first one, the likely trigger, live files changed outside a publish, and what the project's record shows in that window (a publish and the migrations it ran, a rollback, a restore, secrets set or removed on the live site, crashes, restarts and exit codes, a stop at the memory limit, a full disk, outbound connections the plan refused).
2. Read the full recent live log: logs_tail with env production and no grep (an error is often explained by the lines before it, which a search for "error" skips).
3. Compare the live site with the preview: fs_diff for the folder or file in question, and fs_read or fs_list with env production for what the live site has. Check the live database with db_read env production (a row or setting the code depends on), and the live secrets with secrets list env production (a name the code reads that is missing).
4. Say plainly what you found, with the evidence, and what you propose; ask before changing anything live. If nothing explains it, say so and say what you checked: never guess a cause, and never say something waits for an approval unless deploy_list says so (it says "Nothing is waiting for your approval" when nothing is).