Observability

The Operate category of the admin console carries the pages that answer "what is this workspace doing?". This page covers each of them: what it shows, where the data comes from, and which question it is the right page for.

For deployment-level monitoring and wiring metrics into an external stack, see Monitoring.

Right now

Active Chats

#/manage/operate/active-chats

One card per chat turn or agent run that is in flight at this moment: the stage it reports, the tools it has running and the ones it has finished, the sub-agents nested inside it, how long it has been going, and how close it is to being cancelled for having gone quiet. Each card links into Inspect Chat for the conversation behind it.

The page holds one streaming connection and the server pushes a whole snapshot only when a turn actually moves, so an idle workspace costs one held connection rather than a poll every few seconds.

Nothing here is persisted

A turn is visible only while it runs. This is the page for "is that request stuck?"; for what a finished turn did, use Inspect Chat.

The cards carry names, counts and timings, not conversation content.

Query Metrics

#/manage/operate/queries

The graph queries executing right now, with the option to hide short ones so the list is the queries worth looking at. It also carries the switch for query debugging, which records more about each query at some cost.

This is the page for "the workspace feels slow": a query that is always in this list is the one to look at.

Live Activity

#/manage/operate/live-activity

A rolling stream of workspace-wide user activity, searches, chats and endpoint calls, as it happens. Useful for confirming that traffic is arriving at all, and for watching the effect of a change during a rollout. The same widget can appear on a dashboard in compact form.

File Processing Queue

#/manage/operate/file-processing-queue

Files waiting to be extracted and indexed. A queue that grows and never drains is the usual first symptom of an extraction problem or of an ingestion rate above what the workspace can process; a queue that is empty while documents are missing from search points at the connector instead.

See reindexing and re-embedding for the operations that fill it deliberately.

Deletion Queue

#/manage/operate/deletion-queue

Nodes queued for delayed deletion, with the instant each becomes eligible. Read-only: entries fall off as the background sweeper processes them. See the deletion queue for why deletion is delayed at all.

Over time

Usage Metrics

#/manage/operate/usage

Workspace activity aggregated daily, weekly or monthly, over periods from the last month to the last ten years, with a strip of headline numbers (users online now among them) above the charts.

This is the page for "is the workspace being used, and by how many people?".

User Analytics

#/manage/operate/user-analytics

The same question per person: chats, messages, searches, sessions and cost for every user over the selected period, ranked by message count and searchable. A session is one streaming connection the user opened.

Cost columns only carry a value when LLM cost calculation is on and the models have prices, which is configured on the LLM Usage page below.

LLM Usage

#/manage/operate/llm-usage

Everything about external model consumption, in three tabs:

Dashboard — tokens, requests and cost as headline numbers and as charts over time, broken down by category.
Records and Categories — the per-category aggregate, plus every recorded call with category and search filters.
Models and Pricing — the model catalog, the cost-tracking switch, and the per-model price editor. Reference prices are seeded into the catalog automatically.

This is the page for "what is the AI costing us, and which feature is spending it?". Pair it with quotas to cap per-user spend.

Agent Metrics

#/manage/operate/agent-metrics

Calls, errors and average latency for the workspace's agents: a rollup for the last hour on top and a card per configured agent below, with errors carried in the card's colour so a failing agent is visible without reading the numbers.

Every agent run is counted, including runs started as sub-agents from another agent's tool call.

Uptime

#/manage/operate/uptime

Per-service availability captured by the in-process monitor: a 90-day status strip and per-month calendars. It also holds the maintenance-announcement template, whose scheduled window is substituted into the banner and rendered in each viewer's local time.

Monitoring

#/manage/operate/monitoring

Performance and health charts, driven by the range the X axis is on: the toolbar, the wheel and a drag all move the axis, and whatever range it lands on is what gets loaded. Covered in Monitoring, along with the metrics endpoints for scraping into an external system.

Logs

Workspace Logs

#/manage/operate/logs

The server's own log, live. Lines written in the same millisecond as a poll are not dropped and genuine repeats are still shown, so a burst reads correctly rather than as a single line.

The same log pane appears inside the endpoint editor and on connector runs, so a log line can be read next to the thing that wrote it.

Integrations Logs

#/manage/operate/api-integrations

Two tabs:

Connector runs — each run of an API-based connector, with its log.
WebSocket — live diagnostics for a library connector's socket connection: which connections are open (remote address, user agent, when they connected and authenticated, per-connection counters) and the last few that were lost with the reason; the last ten frames sent and received with times, sizes and a truncated preview, interleaved with connects and losses on one timeline; and round-trip time measured from real traffic, as last, best, worst and average.

A Ping button probes a connector that has not been asked anything recently. It sends a reserved frame the client library answers directly, without the connector's own message handler seeing it, so an unanswered ping means the connector is built against an older client library rather than that the socket is broken.

Everything is held in memory and capped per connector, so it costs the same after a month as after a minute, and it starts empty after a restart.

Diagnostics

Developer tools

#/manage/operate/developer-tools

Inspect HTTP requests and how they were routed, which is the page for a read-only replica deployment where the question is usually "which node served this?".

Profiling

#/manage/operate/profiling

Memory dumps, CPU profiles, runtime traces and allocation reports of the running server, with a choice of profile, output format, duration and buffer size. Allocation captures are capped at a shorter duration than the others because they are far more verbose, and the pass that attributes allocations to where they came from is off by default because it takes noticeably longer.

See Profiling for reading the output.

Which page for which question

Question	Page
Is a request stuck right now?	Active Chats, Query Metrics
Is anything reaching the workspace?	Live Activity
Why is search missing a document?	File Processing Queue, then Integrations Logs
Why did the assistant answer that?	Inspect Chat (under AI)
What is the AI costing?	LLM Usage, then User Analytics
Which agent is failing?	Agent Metrics, then Inspect Chat
Was the workspace down last Tuesday?	Uptime
Is the server itself unhealthy?	Monitoring, then Profiling
Who changed the configuration?	Audit log (under Access)