[Question] Why were internal Haiku model calls removed from main process in v2.1.63?

Background

In versions prior to 2.1.63, the main process would occasionally make internal Haiku model calls for lightweight operations such as:

• Context compaction / summarization
• Permission confirmation prompts
• Tool call routing decisions

These mixed Haiku calls were visible in the UI alongside main model responses.

Observed Change in v2.1.63

After updating to v2.1.63, these internal Haiku calls no longer appear. The main process now exclusively uses the primary model (e.g., Sonnet/Opus) for all operations.

This change is not mentioned in the v2.1.63 changelog.

Questions

1. Was this an intentional architectural change?
2. What motivated removing the internal Haiku calls — quality consistency, architecture simplification, or something else?
3. Were those lightweight operations replaced by hardcoded logic (zero LLM cost) or are they now handled by the main model (higher cost)?
4. Is there a plan to reintroduce lightweight model routing in future versions?

Impact

If these operations now use the main model instead of Haiku, the cost per session could increase significantly given the price difference (Sonnet is ~3-4x more expensive than Haiku per token).

Would appreciate any clarification on the intended direction. Thanks!