Module optimizer 

Description token-pruning.

Two layers:

1. HeuristicOptimizer — deterministic, offline, no model dependency. Targets the known failure modes of human-written MCP descriptions: politeness filler, self-reference (“This tool is a tool that…”), duplicate sentences, Markdown emphasis, whitespace runs.

2. DescriptionOptimizer trait — drop-in slot for an LLM-driven rewriter (e.g. Claude, GPT) implemented outside this crate so the core binary stays pure-CPU and offline.

A Jaccard guardrail prevents the heuristic pass from losing meaning.

Structs

HeuristicOptimizer

Default offline pruner.

OptimizationOutcome

Traits

DescriptionOptimizer

Pluggable optimizer interface. Out-of-tree backends (Claude/GPT/…) implement this.