After Repeated Corrections, Rebuild the Task Context

The agent repeats the wrong approach after two corrections. My rule of thumb is to stop adding conversational patches and restate the task from the evidence. Two is a reminder, not a measured threshold.

Before clearing anything, save the failing case, the actual constraint, attempted approaches and current working-tree state. Distinguish a misunderstood instruction from a missing dependency or a wrong assumption in the brief.

Then start a focused session with that record: 'Approach X fails on this input. The constraint is Y. Verify the alternative against this test.' Do not quietly discard an unfinished change or a useful failure receipt.

Fresh context may help when conflicting directions accumulated. It will not repair an incorrect requirement or unavailable tool. The useful habit is diagnosing why the loop stalled, then changing the input that caused it.