[REGRESSION] v2.1.185: Reduced reasoning on routine actions breaks planning autonomy

Summary

Claude Code v2.1.185 (released 2026-06-20) introduced "improved auto mode classifier latency by reducing reasoning on routine actions" which appears to have systematically degraded planning and architecture reasoning across all complex engineering workflows.

Observed Behavior

• Before v2.1.185: Claude agents consistently discovered existing code patterns, read architecture before coding, performed data-layer-first planning
• After v2.1.185: Claude agents skip architecture discovery, invent duplicate solutions, skip data layer planning, trial-and-error without pausing for reflection
• Scope: Affects all sessions, all domains, all complex tasks

Metrics

Comparable tasks (TKP 7940 vs TKP 8011):

• Pattern reuse: 4 → 0
• Files modified (existing codebase): 6 → 1
• Commits spacing: days → 11 minutes
• Autonomy score: 0.85 → 0.15

Request

Revert or toggle the "reduced reasoning on routine actions" optimization in v2.1.185. Restore full reasoning on planning-phase actions (codebase exploration, architecture reading, reflection).

Session

• Version: 2.1.185
• Observed: 2026-06-19 to 2026-06-22
• Impact: 100% of sessions