Auto-remove thinking blocks from context

[Tarquinen]

Problem

Thinking blocks (<thinking>...</thinking>) consume significant tokens (often 1-3K+ per response) but are never referenced again after generation.

Proposed Solution

Add deterministic removal of thinking blocks (no LLM analysis needed):

• Pattern match thinking block formats (XML tags, message part types)
• Run alongside deduplication in Phase 1
• Config: removeThinkingBlocks: boolean (default: true)

Could save 5K-30K tokens in typical sessions with extended thinking enabled.

[dehidehidehi]

FYI : Some providers return <think> and </think>, such as Perplexity.