Skip to main content
Observe → LLM Routing (/observe/llm-routing) is the layout for a per-dispatch model-tier audit: which tier an agent declared, which tier the router resolved, why it changed, and which model served the call. The page is not connected to dispatch data. Its data source returns no rows, so the summary cards and the table are always empty, whatever your permissions or filter. Use the surfaces below for model and dispatch evidence.
The LLM Routing sidebar entry ships disabled. Its feature flag, nav.observe.observe-llm-routing, is off by default, so the entry is shown grayed out with the tooltip “Feature not enabled” and does not open from the sidebar. The page itself still opens at /observe/llm-routing, empty as described here.

What you can do

What the page shows

  • Summary cards — Dispatches (24h), Tier-adjusted, Adjustment rate and Top reason. All four are empty.
  • Filter by reason — All adjustments, Skill floor, Tenant cap, Platform cap, Context overflow, Provider unavailable, Latency budget or Budget block. Every choice returns the same empty result.
  • The table — columns When, Agent, Declared, Final, Adjustments and Model. With no rows it shows No tier-audit events yet.

Key state to watch

  • An empty table is not evidence. It does not show that no model calls occurred, or that no tier was adjusted. Check the owning run and the model surfaces.
  • Models — models, providers and fallback chains.
  • Runs — the evidence for a single run.
  • LLM routing — the model-tier architecture and selection rules.