Watching a run

A run reports twice, on its row and in its log.

The row

A running row carries its state, elapsed time, tokens spent, and a dollar figure when the runner’s own stream reports one.

A run adopted from a previous lerp reads like any other, with its age and totals read back from its log rather than the stretch since adoption.

The activity line

Under a running row, once the run has a log, a second line reads the last call the agent made, a shell command as $ go test ./... and anything else by tool and target, then a sparkline of recent activity. A run that has fallen quiet reads as a flat line, and an adopted run draws the flat line it earned rather than a fresh one.

These are numbers to read, not a timeout. Lerp sets no threshold and never acts on one, and ejecting a stalled run stays your call.

The log

enter on a running ticket opens a live tail in the main pane, with scrollback that survives the run’s exit, and esc gives the screen back. pgup and pgdn scroll, end resumes following, and r toggles the raw output the runner wrote.

The tail reads as activity rather than bytes, one line per tool call, prose as prose, thinking collapsed to a line with its token count. Output lerp does not recognize is shown exactly as written. The log on disk is untouched either way.

The loop’s own log

Agent output goes to each run’s log, which is what the pane tails. The loop’s diagnostics, provision, dispose, adopt and reap, go to .lerp/loop.log, appended across sessions. Look there when a run never started, rather than when a run is going badly.