Release v0.9.1

Fixed

• A continuation window no longer invents markup. parse_gemtext classified
  from the top of whatever string it was handed, but a window resumes
  mid-document, so a cut inside a ``` block read the block's contents as live
  markup: a fenced => url text line came back as a resolved link in
  document.links, the closing fence was read as an opening toggle, and every
  line after it was typed wrong. partial_line stays false on both windows, so
  nothing flagged it, and it needs no hostile capsule -- any page with a code
  block longer than the render cap does it. The tail of an over-long line had
  the same defect from the other side: 0.9.0 declines to parse the window that
  ends mid-line, but the window that starts mid-line was still parsed from
  the top, so a => just past the cap boundary arrived as a complete link to a
  target the capsule never offered. Both are one missing bit of state, and both
  contradicted what models.py and the API reference promise the model on every
  call.
• The install docs no longer deny that the container image exists. README,
  docs/installation.md and docs/troubleshooting.md all said registry
  publishing was still to come and that the release on PyPI predated
  --version; both stopped being true when 0.9.0 published the image to
  ghcr.io/cameronrye/gopher-mcp and shipped the flag. The PyPI long
  description is immutable per release, so 0.9.0's copy keeps the old text and
  only a new release replaces it.
• The changelog's own pre-0.3 tail. 0.2.1 and 0.2.2 carried January dates
  that were 8-11 months off their tags, 0.1.0 was dated 2025-01-XX, and the
  link definitions for 0.1.0 and 0.4.0 pointed at v0.1.0 and v0.4.1 tags
  in ways that 404'd or misnamed the release. Neither version was ever tagged,
  which is now stated rather than papered over.

Changed

• The container image is built for linux/arm64 as well as linux/amd64. The
  0.9.0 image was a single-platform amd64 manifest because the publish step
  declared no platforms at all, so every Apple Silicon and ARM-server user ran
  it under emulation.
• The release path runs the image before pushing it. publish-image built and
  pushed in one step and never started the container; the image it pushes is
  also not the one CI smoke-tested, because the Dockerfile resolves its
  dependency closure fresh from PyPI at build time. It now builds the host
  architecture, asserts --version matches the tag being released and that the
  registry annotation is present, and only then builds and pushes both
  platforms.
• server.json advertises the container as an oci package, so the MCP
  registry lists it alongside the PyPI one. This needed three things that were
  all missing: the image carries the io.modelcontextprotocol.server.name
  annotation the registry verifies ownership with, publish-registry now waits
  for publish-image instead of racing it, and scripts/prepare-release.py
  bumps the version inside the image tag -- the registry forbids a version
  key on an OCI package, so the tag is the version, and nothing was moving it.
• The container base image moves to python:3.14-slim, the top of the test
  matrix, now that the 3.14 legs have been green on CI.
• The minimum-versions CI job runs on Python 3.14 as well as 3.11. The declared
  floors are not one set of versions: mcp markers its pydantic requirement by
  interpreter, so pydantic==2.11.0 is not merely untested on 3.14, it is
  unsolvable there.
• Dependabot can propose a base-image minor bump again, reversing what 0.9.0
  said. The old rule ignored both major and minor updates for python, and the
  Dockerfile's tag has no patch component -- so no update of any kind could ever
  be proposed and the entry was structurally inert rather than conservative.
  Major stays ignored. What now protects the "base image must stay on an
  interpreter the matrix covers" rule is review rather than configuration, so
  check the matrix before merging one.
• pyproject.toml records what should end the mcp<2 hold instead of leaving
  it open-ended, and notes that the first condition has already fired:
  llama-index-tools-mcp requires mcp>=2 as of 0.5.0, so installing it beside
  this package is already unsatisfiable. It also corrects which advisory sets
  the mcp>=1.28.1 floor -- the DNS-rebinding one was fixed in 1.23.0 and is
  not it -- and stops citing the container as the deployment the SDK's rebinding
  default protects, since the container's own CMD disables that check.
• The release checklist no longer tells you to run a bare pre-commit autoupdate. Two hook revs deliberately track uv.lock, and a bare
  autoupdate bumps them out from under it; the checklist now scopes the
  update to the revs with no lockfile counterpart.
• docs/development/repository-setup.md describes the branch protection that
  actually exists -- the main ruleset lists the admin role as an
  always-bypass actor -- and carries the API calls to re-check every claim on
  the page.

---

Release Information:

• Version: 0.9.1
• Release Date: 2026-09-03
• Commit: e38ca01
• Workflow: 33804509908

Installation:

pip install gopher-mcp==0.9.1

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