Investigate

Investigate a run

Two layers of issue analysis. Every ingested run gets an automatic verdict: healthy or failed, with every confirmed issue listed and its spans cited. When you want to go deeper, the trace investigator is a chat grounded in the trace: ask “what went wrong?” and get a cited answer.

Verdicts & issues

When a run is ingested, Kalmia's judge reads the trace and issues a verdict for it: a status (healthy, failed, or needs attention), a short summary of what the run did, and every issue it confirmed, each with a severity and the exact spans that carry it. It does not single out one root cause: a run that completed with three issues shows all three, and a run with none shows an empty list. Where a fix is clear, the issue carries a concrete suggested fix. The verdict appears on the trace. When the judge proposes a behavior worth watching for, Add as rule on the verdict saves it as a rule in one click.

The judge is grounded in your domain context and rules, so verdicts reflect your team's standards, not just the agent's own prompt.

Every confirmed issue also lands in the Issues tab, grouped with the same failure from every other run, alongside the span-level signals Kalmia detects in code for every completed run.

Trace investigator

The verdict is one shot. The investigator is a conversation. From a trace, click Investigator to open a side panel grounded in that run, then ask anything in natural language. The answer streams in, with:

  • Cited spans — chips that scroll to and highlight the referenced span in the timeline.
  • Cross-run links — mention another run and it opens in a new tab so you can compare.
  • A saved thread — the conversation persists and reopens where you left off.

It can also propose labeling runs or adding a behavior; those show up as a card you confirm or decline — nothing is applied without you.

No setup required

The investigator works on any captured trace out of the box. Viewers get a read-only view.

Finding runs

The Traces view is one list of your runs, newest first. Each row carries its status, and its variant when that differs from the agent name. The toolbar narrows the list:

SearchOne box, three scopes: Runs (input text, PR, trace id, or the name of a step, tool, or error), Errors (text of tool-call errors), and Semantic (describe what you are looking for)
FilterVariant, status (Failed, Warnings, Healthy, or Unreviewed, meaning not judged yet), issue tag, behavior label, and customer feedback (thumbs up, thumbs down, or any)
TimeIts own control beside Filter: any time, the last 24 hours, 3, 7, 30, or 90 days, or a custom range
Metric rangesMinimum and maximum for turns, tool calls, files read, files changed, tool errors, subagents, backtracks, and retries, plus has errors or no errors

The list loads 100 runs at a time and fetches older ones as you scroll toward the end. The count in the toolbar says how many are loaded, and out of how many when the list is the whole workspace. To jump further back, set the time range: the list then loads that window from the workspace, newest first, and scrolls through it the same way. The range is by when Kalmia received the trace, which is the order of the list; a run sent long after it started is found by when it arrived. The Variant filter works the same way: tick one or more variants and it loads their runs from the whole workspace. It offers every variant in the workspace, not only the ones among the loaded rows. A long list of options gets a search box: type part of a name, and Enter picks the only match.

Search, the other filters, and the metric ranges narrow the runs that are loaded. While one of them is on, the list stops fetching on its own and says how many loaded runs it searched; Load older traces at the end of the list extends the search further back.