Browse traces and the query bar
Narrow an Environment’s traces with the query bar, facets and time range; read quality chips and tri-state scores on the list; jump to a trace by id.
After this page you can get from thousands of production traces to the handful worth reading without learning a query language, share that filtered list as a URL, and open any trace someone pasted into a ticket. Everything on the page belongs to the Environment in the switcher; changing it re-reads the list for the other Environment with the same filters.
The page
Sub-views along the top: Traces · Sessions · Users · Errors · Logs · Services · Topics · Agents. Below them the query bar, the time range (default 24 h) with the Live toggle, a facet rail, the table, a Saved views bar and Export. Table columns: time · operation · service · status · latency · tokens · ≈cost · score · chips; a column picker hides or shows them. Ticking rows opens a selection bar with Evaluate, Add to dataset, Add to review queue, Annotate, Export and Copy ids.
The query bar
The bar is the whole query language: typed tokens with autocomplete from the facets, serialised into the URL as you type (debounced), so a link reproduces the list.
| Token | Matches | Notes |
|---|---|---|
status: | the root span status, exact (ERROR, OK) | also a facet |
model: provider: service: | substring on the trace’s model, provider or service name | also facets |
session: | substring on the session id | a typed session:<id> opens the same rows the Sessions view groups |
user: | substring on the user id | a typed user:<id> is lifted into a removable chip; the Users view row actions produce the same filter |
method: | HTTP verbs, comma separated | prefix of the root operation |
latency> tokens> spans> | numeric with >, >=, <, <=, = | latency in ms; tokens = in + out; spans = span count |
| free text | operation, attributes, prompt, completion, model, provider, trace id | a bare 32-character hex id opens that trace |
The API also filters by exact operation, appVersion, environment and agentVersion; the Services, release-comparison and Agents views set those for you. Sorting is by latency, tokens or status. Facet counts are self-excluding: choosing one status still shows the other statuses’ counts. Facets cover status, method, provider, model, service, app version, environment and agent version.
Chips and the score column
Each row carries flags the trace store computes from the whole span tree: error (any span ended in ERROR), empty completion (a model returned nothing without erroring), truncated (a provider stopped at its token limit), model swapped (more than one model answered inside the trace), tool loop (the same tool called with identical arguments three or more times) and subagents (how many coding-agent subagents ran). The chips need no click; they are why a row is worth opening.
The score cell is tri-state — a number, no verdict (dashed, neutral) or error (never drawn as zero) — and comes from the trace’s evaluations, which live in the control plane rather than the trace store. The list joins them in the browser for the rows on the page, so sorting or filtering by score covers the loaded page only.
Jump to a trace by id
Paste a 32-character hex id into the query bar or into the command palette (⌘K anywhere in the project). The palette tries the selected Environment first and the project’s other Environments in parallel, says which one it found the trace in, and opens it with the URL carrying that Environment — a trace id is unique only inside an Environment. The palette also finds runs, datasets, evaluators and pages by name over what the browser has loaded, and offers commands such as New run, New dataset and Invite member. There is no server-side cross-object search yet.
Limitations
- A filtered list has no cursor: paging past the first result set works on the unfiltered recent view only. Narrow the time range instead.
- Scores are joined in the browser for the loaded page, so score sort and score filter do not reach traces that are not loaded.
- The browser reads traces with the Environment key; there is no control-plane proxy for trace reads yet.
- Redacted traces (content capture off) are marked per span but cannot be filtered or counted: the capture marker has no facet.