Problem

prune_tools fires an LLM call on every agent loop iteration where tool count exceeds min_tools_to_prune. A multi-turn conversation with 5 tool-use steps would make 5 extra LLM calls per user message, adding ~500ms latency at 100ms/call.

Solution

Cache the pruned tool set per user message, keyed on (message_content_hash, tool_list_hash). Reset the cache when a new user message arrives or when the MCP tool list changes.

Implementation note

Cache should live in the agent loop state (not in prune_tools itself, which is a stateless free function).

Priority

Must be addressed before the wiring PR for acceptable UX.

Component

zeph-core (agent loop), zeph-mcp (pruning.rs)