v2.1.2

What's changed

• Preserve each discovered model's advertised reasoning-effort subset.
• Keep unsupported OpenClaw thinking levels disabled instead of falling back to canonical levels when effort_tiers is present.
• Normalize and deduplicate valid effort tiers while ignoring malformed or unknown values.
• Verify exact subset admission and OpenAI-compatible request construction.

This fixes models such as DeepSeek V4 Flash advertising off, low, medium, and xhigh without incorrectly exposing high.

Included pull requests

• #24
• #25

Full changelog: v2.1.1...v2.1.2