v1.6.3

This release is about reviews that should have been posted and weren't. Nothing about how you install or use siGit Code Review changes.

#Room to answer

Some of the models behind siGit Code Review think before they write, and that thinking counts against the same output limit as the review itself. On a large enough diff, a model could spend the whole limit thinking and hand back nothing. The review then failed, and a failed review posts nothing on your pull request.

Reviews now get four times the output room. The answer itself is a few thousand tokens at most, so the extra space only matters to a model that needs it. Prices are set by the size of your diff, not by the model that served it, so nothing costs more.

When a reply does come back empty, our logs now say so plainly instead of reporting a parse error, and the review moves on to the next model in its pool as before.

#Steering reviews without a restart

Pausing reviews, pinning a model, or taking a misbehaving model out of rotation used to mean editing the server's environment and restarting it. Those switches now live on an admin page and apply to the next review. The same page shows the last day of reviews, how many completed and why the rest did not, so when something goes wrong we find out from the page instead of from you.

Release masters: @setoelkahfi and @sigit.