Inspect a run

Updated 4 hours ago • October 18, 2026

A run is a workflow execution in your application. Cloud and self-hosted Studio use the same inspection views. Studio observes execution; it does not replace your application runtime.

Find a run

Open Runs, choose the project and environment, and narrow the time range. Search for a run, session, workflow, user, or error. Open a row to inspect its overview, execution, trace, events, model inputs/outputs, and feedback where available.

If the list is empty, clear filters and verify the application sent telemetry to this project's endpoint with its SDK write key. See Connect Your Project and Search and Filters.

Status and completion

Status describes execution, not answer quality. A failed run has execution failure evidence. An interrupted run has paused for human input; Waiting for input and Input expired describe its associated interrupt.

Good to know: A response can finish while background work continues. Check the final run status and trace before deciding the run is done. Handled tool denials or faults do not automatically fail the workflow.

Child workflows

Runs defaults to root executions. Enable Include child workflows in Filters to inspect individual calls. Opening a child selects it within its parent execution. Session run counts continue to count roots.

A child returns to its caller. Waiting, resumption, return, failure, and cached reuse are distinct states. Forked branches can inherit calls from their source; inherited evidence is not newly executed work. See Sessions.

Trace and events

Use Trace for nested spans, model/provider requests, tool attempts, and human waits. Nested spans overlap: do not add all their durations to estimate total time.

Events shows captured lifecycle facts and their payloads. Occurred is the emitter timestamp; Received is ingestion time. Ingest delay can reflect buffering, network delay, or clock differences. A missing terminal event leaves incomplete evidence rather than proving success.

Captured content

Prompt, input, and output content is excluded by default. Your application controls capture; enabling it in server configuration affects future telemetry. Not captured means evidence is unavailable, not that the original value was empty. Large content can be omitted with a capture-limit marker.

Use payload controls to inspect clean/raw representations, wrap long values, or copy authorized captured data. See capture configuration.

Feedback

Ratings supplied by your application and human correctness reviews are distinct from execution status and automated eval grades. A negative rating does not mark a run failed. Connect feedback through your authenticated backend; see Collect response feedback.

Next: Understand metrics, Inspect sessions, or Inspect interrupts.