v0.2.0 — active time metric + benchmark

Added

• agentxray benchmark [dir] — runs agentxray across every transcript and prints aggregate findings, so the numbers in the README are checkable rather than trusted. Aggregates only: no file paths, prompts or tool inputs leave your machine.
• SessionStats.activeMs — time the agent actually spent working (sum of gaps under 5 minutes).

Fixed

• stats reported first-event-to-last-event wall clock as session duration. On resumed transcripts this was wildly wrong — across the benchmark set it overstated agent time by 41× (1,455 h reported vs 35.6 h actual). stats now leads with active time and shows the span only when the two differ:

  session  2156 events, 1291 assistant turns, 7h 27m active (spanning 246h 28m)
  

  Found by writing the benchmark, not by a bug report.

• Durations over an hour rendered as inflated minute counts (14751m 29s); they now render as 246h 28m.

New benchmark

56 sessions, 9 projects, 20 days: 1 in 7 tool calls was a failure or a repeat. Full method and caveats: docs/BENCHMARK.md.

Reproduce on your own machine:

npx agentxray@latest benchmark