Skip to main content
Live observability of every intent, mandate, and tool call — plus the inline correction loop that trains the per-org classifier.

Reading the feed

Each row is one event. The badge tells you the kind: PROMPT, TOOL, SHELL, ALLOW, DENY, APPROVAL, STEP-UP, MANDATE, CONFUSED DEPUTY. The correlation badge shows how confidently the event ties back to a captured prompt.

Forensic chain

Click any row to open the forensic chain panel — the full arc from prompt to dispense to action to outcome. Under it sits the cascade’s reasoning trail: which layers were consulted, what supporting signals fired, what alternatives were considered, and the cost + latency rolled across all layers.

Inline corrections

Every non-demo row has an ✎ correct affordance. Click it to record a correction: pick the intent the cascade should have chosen + an optional reason. The correction lands in classifier_corrections and feeds the per-org fine-tune trigger.

Real-time updates

The feed subscribes to Pusher channel private-org-{orgId}-dashboard. New events stream in within seconds; the existing rows are eligible for re-render when their state changes (approval granted, mandate refreshed, etc.).