Module adaptive_compression 

Cache-aware compression decisions using provider billing economics (#1195).

Compression of an already-sent prompt prefix can turn cheap cache reads into cache writes. This module compares that penalty with the actual token reduction and chooses the candidate with the highest expected USD benefit.

Structs

CandidateEconomics

CompressionCandidate

CompressionContext

CompressionDecision

Enums

CompressionDepth

Functions

decide

Compare candidates using all provider input token classes affected here.

predict_session_length

Deterministic session-length estimate used when no explicit forecast exists. Short sessions retain a small horizon; established long sessions reserve a further 50% of their current length for cache-reuse economics.

prefix_stability_score

Ratio of the previous prefix preserved at the start of the current prefix. Appending content is perfectly stable; a mutation near the front scores low.