Module cache_alignment 

KV-Cache alignment for commercial LLM prompt caching.

Claude’s prompt caching stores KV-tensors for byte-exact prefix matches. GPT models have similar mechanisms. This module ensures lean-ctx outputs are structured to maximize cache hit rates.

Key strategies:

1. Stable prefix: invariant content (instructions, tool defs) comes first
2. Cache-block alignment: content segmented to match provider breakpoints
3. Delta-only after cached prefix: only send changes, rest stays in KV-cache
4. Deterministic ordering: same inputs always produce byte-identical output

Structs

CacheAlignedOutput

CacheBlock

DeltaResult

Functions

cache_order_code

Order file contents for maximum cache reuse across tool calls. Stable elements (imports, type defs) first, then variable elements (function bodies).

compute_delta

Generate a delta between two versions of content for cache-efficient updates. Returns only the changed portions, prefixed with stable context identifiers.