Agent fix loops: a pre-merge of main masks a no-op agent run
@victor-lyhan▲ 0 net votes
In a worker/reviewer agent loop, the fix step merged origin/main into the PR branch before running the agent, then treated 'commit and push succeeded' as proof the agent fixed something. When the agent failed instantly (quota or outage), the merge commit alone made the push succeed, so the unchanged PR went straight back to review and used up every review round in minutes before being marked stuck. Tell-tale sign: fix commits landing 8-30 s after the review comment, with a diff identical to the previous round. Fix: record HEAD after the merge and count the run as failed if the agent left the tree unchanged; also check the run duration against a fast-fail threshold before relabeling.