Traces

Traces is where every other screen ends up. A chart, a cost total and a model comparison all exist to get you to the one run that explains something.

#The trace list

The list shows runs newest first for the selected environment and time range. Each row carries the run name and when it started, its status, duration, span and tool counts, tokens, recorded cost, environment, scores and the end user or session behind it — the narrower the screen, the fewer of those columns it keeps.

#Search and filters

Search supports quoted phrases, OR, and exclusion with -. Failed only and Status are independent: the first asks whether anything failed anywhere in the run, the second matches the run's own final status.

Filters live in the URL, so a filtered list is a link you can send to a colleague. Live counts and row refreshes use the same predicates as the list itself, so the count never disagrees with what you can see. Invalid input is reported rather than silently dropped, and there is a clear-filters recovery when nothing matches.

Keyboard: j and k move through the focused list, Enter opens the row, g t jumps here from anywhere, and ? shows the shortcut sheet.

#Saved views

Save an investigation you keep repeating — "failed production runs over 10 seconds" — and give it a name.

• Private to you, or shared with everyone in the project.
• Rolling ranges, fixed dates and explicit environments are all preserved.
• Only the creator can rename, replace the filters, change visibility or delete it. Viewers open shared views read-only.
• The command palette opens saved views by name and remembers recent ones.

If a saved view names an environment that no longer exists, the view says so rather than showing an empty list.

#Reading a run

Open a run to get the waterfall, the span details panel, and four tabs.

Trace

The waterfall of spans and the selected span's detail

Timeline

Events in the order they happened

Logs

Log events recorded on the run's spans

Metadata

Tags, metadata, agent, release, session and user

A failed run opens on the span that broke, not on the root. That is the point of the product principle "show the failure": you should not have to hunt for it.

#The waterfall

Each bar is a span, nested under its parent, sized by duration and coloured by kind — agent, generation, tool, retrieval, embedding, guardrail, custom. Rows are a keyboard-navigable tree.

Selecting a span opens its details:

• Generations — model, provider, request settings, reported token usage (including cached and reasoning tokens), finish reason, time to first token, provider error type, cost, and the prompt and output when bodies were captured.
• Tools — the call id, arguments, output, error flag, attempt events and the time between the model asking and the result coming back.
• Any span — attributes, events and its own status and timing.

#Comparing two runs

Select two runs in the list, or start from a run's detail, to put them side by side: run metrics on both sides, with independent span inspection. Both run ids, the selected spans and the filters you came from stay in the URL, so swapping sides or drilling into a full trace does not lose the comparison. Narrow layouts stack the two runs instead of shrinking them.

Prompt diffs, automatic span alignment and release regression analysis are not built yet.

#Feedback and comments

Rate a run or a span with thumbs up or down and an optional comment, on any plan. Leave a private note on one span for a teammate Pro. See Feedback, comments and sharing.

#Limits worth knowing

• A trace detail loads the first 1,000 span summaries; the rest page in.
• Search covers finalised runs. A run still in flight is pending until it ends.
• Captured bodies may be redacted, truncated or absent — see Privacy and data handling.
• Runs expire with your plan's retention window. Aggregates and usage counters survive; span detail does not. See Plans, quotas and retention.
• Manual trace deletion is not available yet.