Structured Thinking must fire BEFORE solution generation on system failures.
Structured Thinking must fire BEFORE solution generation on system failures. When 913-tool MCP server failed with Prompt is too big, jumped straight to dispatcher recommendation (additive, ordinary) without ST. The user gut reaction (split into two pure servers) was objectively superior. Root cause: existing ST trigger choosing between approaches fires too late. Fix: Added 4th trigger: system failure or architectural constraint violated then generate_hypotheses BEFORE any recommendation. Self-check: Have I generated 3+ alternatives? If no, STOP. Guard against complexity bias (preferring additive solutions when subtractive is correct).