A review bot should earn every message it sends
The findings were fine, the volume was the problem. Cutting our review bot down to one card per thread took two moves: edit in place, and ignore events that were never requests.
What changed this week
- Each thread now has one card, taken over and updated in place every hour instead of posted again.
- A re-review in a thread costs four fewer messages.
- The bot stops after two rounds and names what actually breaks.
How to do it
- Count the messages one review produces, not just the findings it reports. Our re-review in a thread was spending four messages more than the work needed.
- Give each thread exactly one card. On the next run, take over the card that is already there and update it in place, rather than appending a second one.
- Decide what counts as a request before the bot is allowed to speak. Another reviewer’s verdict is not a request for yours, a merge needs no notice, and a roundup of open merge requests is not a review request.
- Collapse the repeats. One message for a request that covers several merge requests, one notice per duplicate pair, and a conflict said once instead of on every pass.
- Cap the argument. After two rounds the bot stops and states what breaks, so a disagreement with the author does not grow into a thread nobody finishes reading.
- Print nothing when there is nothing. A row of zeros is noise, and it teaches the team to scroll past the card that will matter next week.
Here is my review bot's code and the log of one real run. List every message it sent for a single review, in order, and the event that triggered each one. Then tell me: which of those messages could have been an edit of an earlier message instead of a new one, and which triggers were not actually a request for the bot to act. Finish with the smallest change that gets me to one card per thread, updated in place, and a hard stop after two rounds of back and forth.