System Health
Activity events / minute (last 2h)
Retrieval method breakdown last graph window
loading…
Latency distribution bm25-memory hook
Utility rate share of injected items Claude actually referenced ⓘ how
Quality notices rate-based, informational
Other signals
Knowledge graph decisions · docs · prompts — coral = selected prompt, edges = retrieval + topic + temporal
Function proof — real samples recent prompts → actual hook outputs (eye-check relevance)
Per-prompt utility pill:
≥50% CTX surfaced items Claude actually used
≥25% partially used
<25% injected but not referenced — honest signal, not a bug
— reply not yet persisted (Stop hook fires ~30s after turn)