The console¶
Six views, in the order you use them.
| View | Answers |
|---|---|
| Overview | Is the fleet healthy right now, and how is it failing? |
| Servers | Which server, which version, which environment? |
| Tools / Prompts / Resources | Which capability is slow or failing? |
| Protocol | Is tools/list slow? It runs on every client connect. |
| Traces | What happened on this specific call? |
| Errors | Only the failures, newest first. |
Reading the Overview¶
- Tool calls — volume in the selected window.
- Error rate — failures over calls, using the single failure definition.
- p95 latency — over eligible spans, with a caveat when the host clock cannot support it.
- Failure breakdown — the categories, sorted. Click one to see those traces.
Banners appear above the cards when something qualifies the numbers: an unreliable clock, or a share of failures that are only coarsely classified.
Live vs paused¶
Overview and Servers refresh every 30 seconds. Traces and Errors do not — investigation surfaces should not move while you read them. The sidebar always shows how old the data is, and says why refreshing is paused when it is.
The trace drawer¶
Clicking a trace opens it beside the list rather than replacing it. Debugging is
comparison — you bounce between traces, and navigating away loses your place.
Esc closes it. The URL carries the trace, so it is a link you can send.