Claude Code over-investigates solo instead of taking the one cheap measurement that ends the ambiguity

Working a small, bounded numeric-display defect, the agent repeatedly reasoned in long internal chains across multiple turns ('solo thinking') — proposing and arguing competing theories — when a single cheap measurement available the whole time would have settled it in one shot.

Concretely:

• It debated whether a value was off by a factor of 2 for a long time before simply computing the quantity two ways (which proved the values were identical and the theory false).
• It debated which rounding rule the reference tool uses, instead of first capturing the reference tool's own intermediate table — one capture that contained the exact numbers needed and immediately revealed the correct rule.
• Each new session restarted the same reasoning rather than resuming a persisted diagnosis.

The cost is paid by the user in time and tokens while the agent talks itself in circles.

Request:

1. When a question can be settled by a measurement/observation, the agent should take that measurement FIRST and reason after — not produce multiple rounds of competing hypotheses on missing data.
2. Cap solo theorizing: after one unverified hypothesis, the next action should be a measurement, not a second hypothesis.
3. Persist the diagnosis so later sessions resume instead of re-deriving.
4. Effort should scale to defect size; a one-cell display issue should not generate hours of internal deliberation.