The Windows build launched, printed help, drew the specimen card, and showed a
menu containing one item: Quit. Not a permissions problem and nothing to do with
Bluetooth — the tools were simply not in the bundle.
`meshterm/tools/__init__.py` finds every tool with `pkgutil.iter_modules` and
imports it so its `@register` decorator fires. Nothing static ever names a tool
module, so PyInstaller had no reason to include one, and the registry came up
empty: no menu entries, no CLI subcommands, an app that starts and does nothing.
`collect_submodules("meshterm.tools")` fixes it, which is the same treatment
`bleak.backends` already had for the same reason — the spec knew about other
people's dynamic imports and not about ours.
The worse half is that CI said this was fine. The smoke test ran `--help` and
`specimen`, and a build with no tools at all passes both. It now runs a
tool-provided subcommand, so an empty registry fails the pipeline instead of
reaching a release page. Green tests that cannot see the defect are worse than no
tests, because they are believed.
`MESHTERM_HOME` comes out of the same session. Every copy of MeshTerm reads and
writes `~/.meshterm`, so trying a downloaded build meant pointing it at the same
34MB history database as the checkout it was built from, with no way to keep them
apart. The variable moves the whole directory. An empty or whitespace value falls
back rather than resolving to the current directory, because a shell exporting it
blank is common and scattering someone's history into their downloads folder is a
poor way to find that out.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The distribution is named `mesh-term`, because `meshterm` is on PyPI's prohibited
list and a request to free it is 319 deep in a queue. The name after
`pip install` is the only thing that changes: the import package is still
`meshterm`, the console script is still `meshterm`, and nothing in the code reads
the distribution name, so this is one line rather than a rename. Proven rather
than assumed — the wheel was built, installed into a clean interpreter, and the
command run from outside the source tree.
The rest is what a stranger sees on a package page and what the app never shows
them: project URLs, trove classifiers for the console, the three operating
systems and the four Python versions, and the licence moved to its PEP 639
expression form. No `License ::` classifier, which that form forbids, and no
`Typing :: Typed` either — there is no `py.typed` marker, and a classifier that
claims something untrue is worse than an absent one.
`release.yml` publishes to PyPI and is built to be hard to fire. It has no
`push: tags` trigger, and that absence is the design: a workflow that published
on a `v*` tag would make tagging v0.9.0 on launch day claim a name on a public
index, permanently, as a side effect of an unrelated action — and PyPI has no
rename and no real delete. So a human runs it from the Actions tab and types the
version, the job compares that against the version the tree actually built and
refuses on a mismatch, and a `pypi` environment can hold it for a reviewer.
Trusted Publishing means there is no API token in the repository to leak. It
publishes whatever version is committed; there is no version argument, so it
cannot publish something other than what is there.
CI gains a packaging job that does the same build and clean-install check on every
push, publishing nothing. Packaging breaks quietly, and the alternative is finding
out on the one day that is worst for finding out.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`.github/` starts here: the three commands the contributing guide asks a
contributor to run, run for them on every push and pull request.
Two jobs rather than one. Style is a property of the source, so `lint` runs once
on Linux and a failure means one thing. Behaviour is a property of the machine,
so `test` runs across the systems the description now promises — Windows, macOS
and Linux — which until this file existed had only ever been checked on one of
them, whatever the description said.
The matrix is shaped for a private repository, where minutes are metered and
billed by a multiplier: Linux 1x, Windows 2x, macOS 10x. Linux therefore carries
the whole 3.10 to 3.13 range, and the expensive runners cover its ends, since a
version-dependent break shows up at an end rather than in the middle. Seven legs,
about fifty billed minutes a run. The comment says to widen it to the full
product once the repo is public and standard runners stop costing anything.
Two small decisions worth their lines. `fail-fast` is off: the point of a run is
the whole picture of what broke, and a first red leg cancelling its siblings
hides the rest until the next push. And `concurrency` cancels a superseded run,
because a second push makes the first run's answer obsolete and those are metered
minutes spent on a stale commit.
One line of `.gitignore` came with it. The catch-all `.*/` that hides `.venv`
and the caches was hiding `.github/` too — the same rule `.claude/` already needed
an exception to. It fails silently in the worst way: a workflow that is never
committed looks exactly like a workflow that never runs.
The version risk this was expected to surface is already gone: the suite was run
locally on 3.13 as well as the usual 3.10, and both report 1888 passing. What
remains genuinely unchecked is macOS.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>