Watching a run live

The HTML viewer can follow an evaluation while it runs. Start the viewer in one terminal before or after starting the evaluation:

inspect-robots view logs --serve --open

Then start the run in another terminal:

inspect-robots run --task cubepick-reach --policy agent --embodiment cubepick

The index and the running report refresh every two seconds. Running pages show each scene's instruction and, for scenes that carry a rubric, the rubric dropdown, live. The report groups the policy transcript into observation turns. Each turn keeps assistant prose, agent notes, readable tool argument chips, and operator or voice feedback next to the frames that provide their context. Open the collapsed Raw transcript section to inspect the complete raw exchange, including state dumps and tool results.

When frame storage is enabled, running pages show the most recent camera frames that fit the live frame budget. A camera player above each transcript can play or scrub those existing frames at about four frames per second. Its camera, position, and pause state survive the two-second browser refresh. Live and served pages always use this lightweight flipbook and never run ffmpeg.

After a run finishes, a plain inspect-robots view logs pass can replace the per-camera flipbook with one side-by-side composite MP4 per trial when ffmpeg is installed. The composite shares one playhead across every camera. A page first created by --serve stays eligible for that upgrade. Use --no-video to keep the flipbook on a plain rendering pass.

Scores appear only in the final log. Scoring happens after the live sink sees a trial end, so score cells remain pending while the run is active. When the run finishes, the browser returns to the index and the canonical final report replaces the transient running report.

Pass --no-live-log to run or eval-set to disable transient snapshots. The final JSON log is still written normally.

Stale running pages

A force-killed process can leave a *.live.json snapshot behind. Its index row continues to say running, and the report shows the last update time. This is a stale snapshot, not evidence that the process is still alive. Remove it after confirming that the evaluation process has stopped.

Higher-rate visualization

The browser view is designed for turn-by-turn policy transcripts. It rewrites a JSON snapshot at most once per second and refreshes the browser every two seconds. See Rerun visualization for high-rate control-timeline streams and their recording options.