Skip to main content
Run details show one run in full. You usually open a run from a friction or a failed Matrix cell when you read results. A run has three tabs, Traces, Output, and Evals, with the Run profile in the side rail of each. To review two or three runs side by side, see Compare runs side by side.

Traces

Traces shows what the agent did: the Timeline first, when the run recorded times, then the transcript.
Run Traces tab with the Timeline, transcript filters, and the Run profile rail.Run Traces tab with the Timeline, transcript filters, and the Run profile rail.
Above the Timeline, a label names the Insights model that wrote the review you see, A note there says when Insights didn’t read the whole run (see What Insights read).

Timeline

Timeline shows when tool calls happened. It appears when the run has at least two model turns, or two steps if it recorded no turns. Each column is a turn, a short range of turns, or a step. Select a column to jump to it in the transcript. When a run has no Timeline, start at the transcript. Each row is a tool the agent called, in this order:
  1. The tools in the treatment
  2. Other launched skills, MCP servers, or CLIs
  3. Other tools
A run whose tool calls aren’t split by tool has one Tool calls row. A treatment tool with no calls to show gets no row. Instead, a line under the Timeline counts those tools by state, for example “2 Not used, 16 Can’t tell”, so Can’t tell never reads as Not used. Treatment use lists each tool with its state.

Transcript

The transcript lists the same work by model turn. Each step starts as one line. Open a step to read the agent’s reasoning, the tool call, and the result, or use Expand all and Collapse all. Each friction shows at the first step it cites. One that cites no step, or a friction on a run that recorded no steps, shows above the Timeline as No specific step. After Obss groups a friction, its name opens the friction’s page. While the run is still going, Traces says Steps appear when the run finishes. If the run recorded nothing, it says No steps were recorded for this run.

Output

Output shows what the agent delivered:
  • Answer shows the agent’s final message. Select Copy answer to copy it.
  • Files shows the changed files that the run recorded, each with its diff. Use Search files to find a file, and Download file to save one.
Run Output tab with the Answer view of the agent's final message, and the Run profile rail.Run Output tab with the Answer view of the agent's final message, and the Run profile rail.
If the run ended without a final message, Output says so and points you to Files for any saved work.

Download the trajectory

Open the run, then choose ⋯ → Download trajectory. The file is usually a .jsonl, .json, or .txt file, depending on how the trajectory is stored. The same menu has Download output files and Download changes (diff), each shown only when the run has that file, and Cancel run while the run can still stop. A trace larger than 32 MB isn’t saved, so that run has no trajectory to download, though it keeps its outcome.

From the CLI and the plugin

Next steps