Morse v0.26.0

The usage panel says what the cache saved and how fast the model answered, and Morse finds pi however it was
installed.

New

• Cache split and generation speed in the usage panel. The session totals now answer four questions they
  could not before: how many prompt tokens the cache served versus missed, and how fast the model actually
  generated — average, peak and slowest. Cached reads pi's cacheReadTokens and Miss its inputTokens (pi's
  promptTokens = input + cacheRead + cacheWrite, so input is exactly the part the cache did not cover).
  Speed is different: pi keeps no clock, so the host times each assistant message (message_start →
  message_end) and persists the totals per session, so a resumed session continues its history instead of
  starting from zero. A message under 25 output tokens is refused — at that size the span measures request
  latency, not speed — and all three figures share one counted set, so the average can never sit outside the
  range the extremes show.

Fixed

• Fixed the "Morse could not read the built-in slash commands" warning, and the missing builtins behind it,
  for most ways of installing pi. Morse read the pi package by walking up from the pi binary, which only
  works when pi is a symlink into node_modules — true of exactly one install method. A managed install
  (curl pi.dev/install.sh, the PowerShell install.ps1) has no manifest anywhere above its launcher; pnpm
  writes a copied shell shim into its store; Windows .cmd/.ps1 shims are scripts outside the package; and
  bun on Windows writes a compiled .exe beside a .bunx sidecar. The package is now resolved through whichever
  of those shapes is present — the launcher's installer marker, the node_modules path named inside a shim, or
  the sidecar's target — so the palette gets its builtins and the "unreadable install" banner is gone. The
  "a newer pi is out" notice is restored with it, since the version reads the same way.
• Fixed the usage panel sitting on stale numbers until a run ended. Every figure came from a single pull
  that only ran at attach, compaction and settle, so during a run the panel showed attach-time totals — exactly
  when the reader wanted to watch the context grow. It now re-reads per completed message, and the in-flight
  response's cumulative usage is shown as it streams (guarded so a provider that reports zeros mid-message
  cannot wipe the panel, and so the same figure is not rebroadcast per token).