Where the four drivers spent the last 12 hours
Each driver hands work to sub-agents (workers) and reviews what comes back. This page splits every minute of the drivers and their workers into phases, read from their session transcripts, and counts the pull requests they opened and merged.
What stands out
Throughput
"Open → merged" is the median minutes from a PR being opened to it landing, for PRs opened in the window. "Review rounds" counts reviewer verdicts delivered to the driver per merged PR; driver-2 does not receive review notices, so its rounds read near zero.
Where the working time went
Share of active time by phase. Parked time (a worker waiting to be woken) is left out of the bars and shown in the table above.
Flow across the day
Active hours per 15 minutes, drivers and workers combined, colored by phase. A height of 4 means four agents were busy for that whole quarter hour. Ticks under each chart mark PR merges.
What each driver was doing, minute by minute
The driver's own session: dark is working, the rest is stopped and waiting. Hover for times.
Pull requests by driver
How this was measured
Source: the four drivers' Claude Code transcripts and their sub-agent transcripts, plus gh pr view for PR open and merge times. Every turn ends with a stop record carrying its duration, so a driver's working time is the union of its turns (turns nest when notices arrive mid-turn, so they are merged rather than summed).
Phases come from the tool calls: file edits are writing code; reads, searches and plain shell reads are planning; test, lint and typecheck commands are tests; commit, push, open and merge are shipping; gh pr checks/view and review skills are reading review results; sleep and watch loops are waiting on its own background job (almost all of these watch a local push, typecheck or test run finishing, not GitHub); spawning, messaging and board updates are coordinating. Time the model spends writing a tool call counts toward that call. Time with no tool running is reasoning and writing reports, unless it runs longer than five minutes, which means the agent was parked.
These are heuristics from command text, so treat shares within a few points as equal. Driver idle gaps are labelled by what the driver's last message said it was waiting for.