Not a bug, but now it's about 4 months that a change I did not review is paying back - within that four months the whole other code is built on the wrong premise and assumptions. It's way too much to git back, so the Claude workflow agents are struggling and always fall back to that incorrect premises. It can't get out of the loop. Now I sitting for weeks and babysitting it. No chance. 3 runs of dynamic workflows, debugging mcp and introspection abilities into the live execution, burned 30m+ tokens, ran in the sum for about 50h now.. still no change, no fix, no anything. I am slowly thinking to start from new.