Skip to main content
Page through a project’s runs, open one run’s recorded work, or analyze many runs at once. In the dashboard, the same results are on Read results and Run details. A run is one agent attempt on one task with one treatment. Each trial is its own run, so a trial ID is also a run_ ID.

Run scope flags

The run scope flags narrow a command to the runs that a set of records shaped. obss runs list, obss frictions list, and obss frictions get take all of them. obss experiments list takes all of them except --experiment-id, and lists the experiments of those runs. obss results analyze takes them too, but its --experiment-id, --task-id, and --friction-id are its own analysis flags. Each flag takes one ID, and a run counts only when it matches every flag you pass.
  • An ID with the wrong prefix is a usage error. The message says where the right ID comes from.
  • An ID that the project doesn’t have fails with scope_not_found instead of returning an empty list. The error names each ID it couldn’t find.
  • The runs and frictions results return scope: each record they counted, with its kind, id, current name, and dashboardUrl.

obss runs list

Page every run of the current project, newest first, with stable IDs, outcomes, and errors.
Plugin (MCP) tool: list_runs Flags The first twelve flags are the run scope flags. Also takes the global flags. Examples List the failed runs of one experiment:
List every run where a friction occurred:
List every run whose treatment gave the agent one MCP server:
Result fields Each run has its runId and trialId, the task, the treatment, and the agent with its model and effort, plus these fields. The result also has scope, hasMore, and nextCursor. Notes Continue with --cursor while hasMore is true. To fix a setup failure, see A run could not start because a setup command failed.

obss runs profile

Show where one run’s tool calls, time, and tokens went, and which skills, MCP servers, CLIs, and SDKs from its treatment the agent used.
Plugin (MCP) tool: get_run_profile Arguments Takes only the global flags. Examples
Notes This is the Run profile rail on run details: treatment use, tool calls, what stood out, and time. Counts and times are exact or absent, because Obss never estimates a missing count or time. A run that a setup command stopped has setupFailure, with the fields listed under obss runs list. The text output names the step and what to fix, prints the obss resources inspect command for the machine or tool that declared it, then shows the command and its output.

obss runs trajectory

Read one page of a run’s ordered steps, with token usage and measured tool durations.
Plugin (MCP) tool: get_run_trajectory Arguments Flags Also takes the global flags. Examples Read the first 500 steps, then the next page:
Result fields Notes For a summary of the whole run, use obss runs profile.

obss runs transcript

Read one byte-bounded page of a run’s original transcript, unchanged.
Plugin (MCP) tool: get_run_transcript Arguments Flags Also takes the global flags. Examples Read the first page, then the next:
Result fields Notes Use --output json to receive the exact bytes. It’s the same file that ⋯ → Download trajectory saves on run details.

obss runs output

Read what a run left behind: the file manifest, a diff of one file or the whole run, one file, the agent’s final message, or the full archive.
Plugin (MCP) tool: get_run_output Arguments Flags Also takes the global flags. Examples List the files a run changed, then read one file’s diff:
Result fields Notes Use --output json to receive the exact bytes. The dashboard shows the same files on the Output tab of run details.

obss results analyze

Run one structured analysis over the runs in a scope: summarize, compare treatments, find failures and frictions, or read one run’s evidence.
Plugin (MCP) tool: analyze_results Arguments Flags --schedule-id through --agent-model-id are run scope flags. Also takes the global flags.

Modes

compare_treatments compares mean costs only within one cost source, so it never mixes costs that the agent reported with costs that Obss estimated. Examples Summarize one experiment:
Print each treatment’s pass rate:
Find recurring failures for a task since July:
Result fields With --output json, the analysis is in data.result. The Baseline is the treatment that the experiment’s plan names as referenceTreatmentClientId. If none is marked, compare_treatments uses the first treatment. Notes A pass rate is passedRuns / (passedRuns + failedRuns), which is 1 - failureRate. Runs that are blocked, inconclusive, or whose Evals failed have no verdict, so they stay out of both counts. Wait until pendingRuns is 0, because the counts can still change until then. --limit is an upper bound, and the response size limit can return fewer items. The rows that page.completeData names can be finer than the ranked items, so check the hint before you compare counts. For how results become final, see What Evals read and What Insights read.

Next steps