ai-coding-minesIndexGitHub

"Which file actually runs" is settled only by following the call chain to the end

Git and automation

Symptom

Edited a file; the change is gone on the next run.

Cause

The runner copied the canonical script into a date-stamped file and ran that. Patch the dated file and the next run overwrites it with a fresh copy.

Fix

  1. Before patching, trace the call chain to the end and confirm the canonical file
  2. Patch only the canonical file; never the derived copies (they get overwritten anyway)
  3. Delete the fragments or move them to _archive/. Left in place, you'll pick the wrong one again.

"Not in the crontab" ≠ "not automated." Only one entry was in cron; three more hops hung off it.

Same family

Values entered in a console or dashboard (keys, toggles, manual edits) are not in the repo. Patch from the local copy as the base and the key gets overwritten with an empty value and the feature quietly turns off. Re-download the live file before patching; confirm the value survived after. Losing it raises no error.

When the canonical file is right and the old rows are wrong

Some defects never show up no matter how long you read the code — because the current canonical file is correct. The canonical script set a flag to True; half the registered rows had False, and the defect appeared only below a certain ID. A clean ID boundary means not "a bug in the current code" but "rows uploaded earlier by a different path (manual, an older script)." The diagnostic is not a code audit. Split the population by ID range into generations, sample each one, and ask "when was this uploaded, and by what." If accounts had different upload paths, the same defect shows up in one account only.

A clean ID boundary says: look at the upload generation, not the code.