v1.4.0

Fixed

• Output tokens (and cost) were undercounted by ~20%. Claude Code writes an assistant
  message to its transcript once per streamed content block; the interim lines carry a
  placeholder output_tokens and only the final line reports the settled figure. The
  extension was keeping the first line of each message and discarding the rest, so it
  banked the placeholder and threw the real output count away. It now takes the final,
  settled usage. Your output-token and cost figures will step up after this release —
  the new numbers are the correct ones, not an increase in your actual usage.

Changed

• The 5-hour pace meter no longer cries wolf. It was a bare extrapolation
  (utilization ÷ elapsed), which is a knife-edge at exactly the moment it matters: being
  a point or two ahead of an even pace rendered as a confident → 104%, and one idle
  minute later as → 99%. Three changes:
  • It waits until the window is half over. The meter was shown from 10% elapsed, but
    the breach warning already refused to fire before 50% because the projection isn't
    reliable before then — so the extension was printing a number it was too suspicious of
    to warn on. Both now share one gate. Early in a window you see just 5h 15%.
  • New · on pace reading. Within ten points of 100% the projection is noise, so the
    meter now says 5h 53% · on pace instead of committing to a number that flips on the
    next refresh. Clearly under or over pace still shows the figure (5h 23% → 45%,
    5h 95% → 119%). The breach warning clears the same band, so a popup can no longer
    contradict an "on pace" meter.
  • It hides when you stop. The meter ignored whether you were still burning, so a
    heavy start followed by an hour of idling kept projecting a breach. It now disappears
    when you're idle, like the weekly ETA and the warnings already did.
• The burn readout says what it actually measures. Burn: 1.6M tok/min was ~95% cache
  reads — a long session re-reads its whole context every turn — so the headline tracked
  context size × turn count rather than anything you consume, and it sat under the "Claude
  plan limits" heading as though it drove the pace projection (it never did). It is now its
  own tooltip block, split into new work versus context re-reads, with the rate labeled as
  an API-equivalent estimate on a subscription, matching the dashboard:
  49.5k new tok/min · 849.1k/min context re-read · ≈ $0.80/min at API rates.