Claude Fable 5 Refuses Mid-Tool-Call, Risks Corrupting Agent State

Claude Fable 5 refused mid-tool-call on 11 of 44 coding-agent turns, even on mundane tasks like fixing a config default. The refusal truncates tool arguments mid-stream, but the partial JSON still parses, so agent loops that don't check stop_reason may write half-finished files. The model also forces adaptive thinking with no off switch and costs up to 15x more than GLM 5.2 on short tasks.