Release v0.10.1

Fixed

• refresh now invalidates the whole resource instead of one window of it. A
  windowed read caches each rendered window under its own key -- the resource's
  key plus the offset -- and refresh only ever skipped the cache read for
  the exact key it was passed. So refreshing at offset 1000 re-fetched that
  window and left every sibling in place: the next read of offset 0 or 2000 was
  served the bytes from before the refresh, marked cached: true. A caller who
  had asked for the current state was told about the state it was replacing,
  and told it in the one way that reads as merely old rather than superseded.
  The body held for continuations needed the same treatment. Declining to
  serve it under refresh -- which it already did -- is not dropping it: if
  the refreshed response has no next_offset, because the resource shrank or
  now fits in one window, nothing overwrites the slot and the pre-refresh body
  is still there for the next continuation that does not pass refresh. It is
  released only when it belongs to the resource being refreshed; releasing
  another one would cost a re-download nobody asked for.
  Both halves are 0.9.x bugs rather than fallout from 0.10.0's one-download
  change -- the per-offset key and the single-slot store both predate it.
  The refresh description shipped in every tools/list claimed the old
  behaviour re-cut every window, which was never true of any release. It now
  says what happens: the windows are discarded and re-downloaded when next
  read, which is still a reason to finish a walk before re-reading.

---

Release Information:

• Version: 0.10.1
• Release Date: 2026-09-06
• Commit: 444a392
• Workflow: 34004200936

Installation:

pip install gopher-mcp==0.10.1

Verification:
All packages are signed and can be verified using Sigstore.