Investigate an issue

Open an issue from the Feed list and you land on its detail page: a header, a metadata sidebar, and a tab bar with Overview, Traces, Trends, and Fix, all views onto the same cluster of traces that failed the same way. Fix has its own guide; this one covers the other three.

This guide works the detail page in the order that actually finds a problem: orient in the header, read the pattern on Overview, evidence the divergence with the trace evidence reel and split compare, drop into Traces only if the pattern doesn’t hold up, and check Trends for how urgent it is.

flowchart TD
accTitle: The order to read an issue's tabs in
accDescr: Start on the Overview tab reading the pattern-summary cards, then decide whether one consistent failure mode explains the issue. If it does, evidence it with the trace evidence reel and the split compare. If it does not, or you need one specific run, go to the Traces tab. Either path ends on the Trends tab to judge urgency.
A["Overview<br/>read the pattern-summary cards"] --> B{"One consistent<br/>failure mode?"}
B -->|"yes"| C["Evidence it<br/>evidence reel + split compare"]
B -->|"no, or need one run"| D["Traces tab<br/>find the specific run"]
C --> E["Trends tab<br/>how urgent is it?"]
D --> E

Header and sidebar

The header and the right-hand sidebar stay fixed across every tab. The header carries:

• A breadcrumb and error-type chip
• The issue title
• Status and severity badges
• A trace-count chip

The sidebar holds status, severity, and assignee as editable controls. See Triage issues for how to use them.

Use Copy cluster ID to paste the cluster identifier into a ticket or message, and Share for a direct link to this issue. See Issue fields & filters for what everything on the header and sidebar means.

Overview tab: is this one failure or several?

Overview opens by default, and it’s where every investigation starts: work out whether the cluster is one clean failure or several tangled together before you dig into individual traces.

The Overview tab: pattern-summary cards, the events-and-users chart, and the trace evidence reel

Read the pattern

The pattern-summary cards describe what’s common across the whole cluster, not just one trace. Read them first: if they point at one consistent failure mode, you’re likely looking at a single clean cluster; if they point in different directions, the cluster may be mixing more than one failure mode and needs a closer, trace-by-trace look.

A chart below the cards plots events and users for the cluster, so you can see whether it’s a steady trickle or a recent spike.

Open the evidence reel

The trace evidence reel is on the Overview tab, with a switcher above it for its three view modes:

• Breadcrumb: a linear read of what happened, the one to reach for first
• Agent Graph: every step the agent could take, useful for seeing whether the failure sits on one path among several or is the agent’s only option
• Agent Path: the sequence this particular run actually took, useful for tracing exactly where this one run went sideways

Within the reel, two tabs separate the evidence: Failing shows one failing trace at a time from those backing the pattern, and Working shows the nearest trace that succeeded.

Split-compare to find the divergence

Toggle Split with working to line the open failing trace up against that nearest working trace (toggle Single view to go back to one trace at a time). This pairing is matched ahead of time by Error Feed, not a random working trace picked on the spot, so it’s built to show exactly where the two runs diverge.

Not every cluster has a working trace to pair against. If none was found, split compare has nothing to show; work from the Traces tab instead.

Traces tab: find the specific run

If Overview’s pattern doesn’t hold up under a closer look, or you need one specific run rather than the aggregate, drop into the Traces tab: it lists the cluster’s traces, one row each.

Five aggregate cards sit at the top: Total traces, Avg score, Avg turns, P50 latency, P95 latency. The grid below carries a column for each: Trace ID, Input, Start Time, Duration, Tokens, Cost, Score. Click any row to open it in the trace drawer for the full detail.

Trends tab: is this urgent?

Trends is a single chart: errors and traffic plotted on two axes over time. Read the two lines as a pair, not separately. If the error line climbs while traffic barely moves, something got worse in the system itself. If both climb together, you’re most likely looking at more volume, not a rising failure rate, which is often enough on its own to tell you whether an issue is urgent or just a side effect of growth.

Dive deeper