Skip to content

v0.4.3: Listed in the MCP Registry, so VS Code's own gallery can find it

Choose a tag to compare

@github-actions github-actions released this 13 Sep 21:39
· 332 commits to main since this release

Added

  • qikly has a listing for the MCP Registry, server.json. That registry
    is what VS Code's own gallery reads: searching @mcp in the Extensions view
    shows servers from it, and only servers listed there can be added that way.
    The listing carries the icon in four forms, PNG and SVG for a light and a
    dark background, served from the landing page, because clients are only
    required to support PNG and merely encouraged to support SVG.

    The install block carries the [mcp] extra, and that is the part worth
    writing down.
    A bare uvx qikly would install without the extra and
    produce exactly the server-that-will-not-start of 0.4.1 and 0.4.2, for every
    stranger who clicked install, and the console script is named qikly-mcp
    while the package is named qikly. Both facts live in the arguments:
    --from qikly[mcp]==<version> as a runtime argument, and qikly-mcp as a
    positional, both inside runtimeArguments. Verified with uv 0.12.13, and it
    is the shape the live ai.gateco/gateco listing uses, read from the registry
    on 2026-09-12. The placement is not cosmetic: with the script name in
    packageArguments instead, a client composing runtime arguments, then the
    package, then package arguments produces uvx --from qikly[mcp]==X qikly qikly-mcp, which fails with "unrecognized arguments: qikly-mcp". The
    version therefore appears three times in server.json, and a test holds
    them together.

    It was nearly built the hard way, and that is the useful part. A first
    pass sampled four pages of the registry, concluded that no entry can express
    an extra, and added a second, code-free distribution whose only job was to
    depend on qikly[mcp]. It worked. It also cost a second PyPI project, a
    second pending publisher, the version pinned in a fourth file, a workflow
    check to hold the pair together, and a shared console script that
    pip uninstall could take away from the package still needing it. Reading
    the whole registry rather than four pages of it found 1,513 package entries
    using runtime arguments, out of 21,240, by full pagination on 2026-09-12. So
    the extra had been expressible all along. The measurement was wrong before
    the design was, and the registry keeps growing, so that count is a reading
    on a date rather than a fact about the ecosystem.

    The other alternative, making mcp a hard dependency, was rejected on
    measurement too: 18 extra packages and about 19.5 MB on a 48 MB base, some
    40% more, paid by every command line and every CI run that never starts a
    server. Measured 2026-09-12 as the on-disk size of site-packages,
    excluding __pycache__, for qikly==0.4.2 against qikly[mcp]==0.4.2 in
    two fresh venvs, CPython 3.10 on Windows, where pywin32 is the largest
    single item and installs on Windows only. An earlier figure of 26 MB in a
    draft of this entry counted __pycache__, which is bigger in the bigger
    venv and inflated the difference. The extra resolves through mcp>=2.0,<3,
    so this number drifts with its dependencies and is a measurement on a date,
    not a property of the package.

    Publishing needs the package on PyPI first, because the registry proves
    ownership by finding mcp-name: io.github.gal-a/qikly in the README that
    PyPI serves. That marker ships here as an HTML comment, so the listing can
    only be published after this release, never before.

  • A run records which code produced it, not just which release. Every run
    summary already carried qikly_version, which is exact for an install from
    PyPI and misleading from a source checkout: the version only moves on
    release day, so 420 runs in this project's own history are stamped 0.1.0
    across weeks of changing code. When qikly runs out of its own git checkout,
    the provenance now also carries the commit, whether the tree had
    uncommitted changes, whether that commit is a tag, and the Python version.
    The same line opens the run in the console and appears in both HTML report
    headers, because a console capture or a saved report is often all anyone
    keeps.

    The guard is the part worth reading. A virtual environment commonly sits
    inside the user's own repository, so asking git about the package directory
    would record their commit as qikly's: not a gap but a confident wrong
    answer. A git root is believed only when it actually contains src/qikly.

Fixed

  • A third test stopped writing into the live project. tests/test_runs.py
    had a project fixture that redirects to a temp directory, and tests that
    did not ask for it wrote to the real one: a CALC_TAX run summary and
    transaction log from the file's fixed timestamp were sitting in the live
    outputs tree, describing a run that never happened, the log saying every
    test passed while the summary said it failed. Redirecting is now automatic
    for every test in the file rather than something each test has to remember.

  • Four warnings pointed at a heading that was never written. cli.py and
    orchestrator.py sent anyone running without acceptance criteria to
    README.md#auto-generating-acceptance-criteria, an anchor that has never
    existed in any version of that file, so the one message a person sees at the
    moment they need the explanation led nowhere. The section now exists, in
    docs/USING_YOUR_OWN_DATA.md, and the four messages point at it.

  • A docstring escaped nothing and warned about it. _demo_root documents
    cd C:\Temp\scratch in a plain string, where \T is not a valid escape, so
    Python raised a DeprecationWarning whenever cli.py was compiled fresh. The
    docstring is now raw. Invisible from a warm __pycache__, which is why it
    survived this long.

Changed

  • The MCP tool qikly_check_criteria is now qikly_validate. It always
    ran the offline check that qikly --validate runs, while the command line
    has a separate --check-criteria that asks a model whether the
    specification contradicts itself. Same name, different check, one of them
    paid. It misled the first host that used it: asked to check CALC_TAX's
    criteria, GitHub Copilot read cli.py, found --check-criteria, and tried
    to run the paid checker in a terminal; and when finally pointed at the tool,
    it reported "no contradictions found", which the tool never looks for. The
    new name says what it does, and the description now says in as many words
    that it is not the contradiction check. The old name is not kept as an
    alias: the server shipped a day earlier, and a fifth tool listed under the
    misleading name would keep misleading.

  • The README is 751 lines instead of 1,099, and nothing was deleted. It had
    become two documents interleaved: the case for reading further, and the
    reference you consult once you have. The reference half moved out whole, to
    docs/USING_YOUR_OWN_DATA.md (the task file field by field, criteria from a
    ticket, seeding your own code or tests, where files are read from, fixture
    proposals) and docs/CONFIGURATION.md (settings, environment variables,
    provider setup and determinism). Both are linked from where they used to sit.
    The prose was moved as text rather than rewritten, so the two pages inherited
    the documentation guards automatically: they are matched by the same
    docs/*.md globs, and the suite grew by six cases without a test being
    written.

  • The comparison table on the landing page stacks on a phone. Two prose
    columns at 600px left a couple of words per line; each row is now its own
    card, carrying the column name that the dropped header row used to supply.
    The table also described its two sides as qikly's "unit stage" and
    "integration and system stages", a word neither document defines until well
    below it. They are tests, and now say so.