Content modelled on a real session (identifiers changed): goal "Complete PAY-142 and merge it to main then clean up the worktree". The judge ran 196 checks in 25 minutes (one every ~7 s). Every check after the third said the same thing in new words: merged, but 2 worktrees remain because killing their processes was denied. Another session ran 2,318 checks over 7.8 hours. Reasons average ~480 characters and are almost never word-for-word identical, so plain de-duplication will not work.
What others do: nobody shows the full list. t3code shows only a check count; Claude Code shows the latest reason and puts each verdict as a collapsed line in the transcript; Codex shows status with pause/resume/clear. Constraint: the judge is Claude Code's built-in /goal evaluator; conduit reads its verdicts and can pause the goal, but cannot change what the judge writes or how often it runs.