CASE · Minor · monitoring-gap
VerifiedFirsthand recordA published issue produced zero author notices yet was marked done. The tool read a local work branch, not what was published
The tool that creates author-contact tasks for featured articles made no task for any of the 8 articles of the published issue W43. No error appeared, and the 8 articles were recorded as processed.
2 min read
- OBSERVED
- CATEGORY
- monitoring-gap
- BLAST RADIUS
- reputation
- DETECTED BY
- manual-observation
- TAGS
- #outreach #git #working-copy #silent-failure #marked-as-seen #dogfooding
CH 1
Cause
The tool decided what was published from origin/main but read article contents from the local working copy. From October 20 to 26 that copy sat on a different fix branch with no files for the new issue. The part that collects drafts returns an empty article even when the file is missing, so empty bodies were recorded as seen. A de-duplication record then guaranteed they would never be picked up again.
CH 2
Consequence
Notices to the authors of 8 articles stalled for several days after publication. It was found on October 26 by a human check. The damage was delay only; no data or money was lost.
CH 3
Fix
Articles and scripts are now exported from origin/main into a temporary folder and read from there. Articles whose body could not be read are not marked as seen. Recovery: removed the 8 from the de-duplication record, re-ran, and confirmed 8 tasks were created.
What happened
Tasks for contacting the authors of featured articles are created automatically once an issue is published. Issue W43 went out on October 23, but none of its 8 articles got a task. There was no error. The list simply didn't grow, and the articles counted as processed.
Cause
- The tool decided "published?" from origin/main but read article contents from the local working copy.
- That week the copy was on another branch for a different fix, with no files for the new issue.
- The draft-collecting part returns an empty article when the file is missing, so an empty body was recorded as "seen".
- With a de-duplication record in place, an article marked seen is never picked up again.
"Could not read" was stored in the same shape as "read, nothing to do".
Fixes
- Articles and scripts are read from an export of origin/main in a temporary folder, independent of the working copy's branch.
- Only articles whose Japanese body was actually read are marked seen.
- The 8 were removed from the de-duplication record and re-run; 8 tasks appeared.
Lesson
When "is it published" and "where do we read it" look at different places, one can drift without anyone noticing. And a component that records failure in the same shape as success, combined with de-duplication, locks the miss in.
Sources
- Firsthand account from the site operator. There is no public write-up to link to.
Pass this record on to your team