What problem does it solve? When a Synapse coding-agent session shows unexpectedly high token costs or low prompt cache hit rates, it is hard to tell whether the misses are normal incremental misses or an abnormal full cache eviction, and even harder to find the root cause of an eviction. ## Core Features & Use Cases - Overall and per-turn hit-rate computation: Aggregates cache_read and input_tokens from checkpoint SQLite messages to compute overall, per-turn, and per-call cache hit rates. - Miss classification: Distinguishes normal incremental misses (large new tool results) from full cache eviction by tracking the absolute cache_read value rather than the percentage. - Eviction root-cause diagnosis: Compares wire fingerprints in model_request_compression_events to identify whether system prompt changes, tool set changes, summarization, truncation, or upstream provider eviction caused the cache bust. - Use Case: A user notices one turn of a session suddenly billed tens of thousands of uncached input tokens; this Skill locates the exact call, confirms it was a full eviction, and shows the tool schema set changed between calls. ## Quick Start Analyze the prompt cache hit rate of my Synapse session and tell me why the hit rate collapsed on turn 12.
