memcrate sat untouched all summer. picked it back up today and the first thing i wanted was a from-scratch install test, which i’d somehow never done. that test is what killed the old design.
setup was four commands: init to make the vault, setup to fill in your
profile, install claude-code for the skills, and the binary install
before all of it. each one made sense on its own. running them in order on
a clean machine felt like homework. so v0.4.0 collapses the whole thing:
memcrateit asks one question, where the vault should live, then creates it and
installs the skills. --vault and --yes cover scripts, and --yes is
implied when there’s no terminal so it never hangs in CI. the identity
questions moved to memcrate profile, optional, because folding them into
the main flow turned one question into five and “what do you do?” is a
terrible prompt.
that’s a breaking change. i did it now specifically because npm support landed in the same release, and breaking a handful of early users beats breaking a larger group later.
codex, and one skill source for both
codex uses ~/.codex/skills/<name>/SKILL.md, the exact format claude code
uses. so instead of maintaining two near-identical copies of ~440 lines to
change about six, there’s one canonical set that installs to both. the
only genuinely tool-specific part is how each one finds your vault, and
that names the concrete call per tool inline.
also added --force protection that i needed personally: my own load,
save, and pin skills live at exactly the path memcrate installs to.
installed skills now carry a marker, and memcrate refuses to replace
anything it didn’t put there.
the part that actually stung
after shipping, i started clicking around my own site instead of reading source. four things were broken:
memcrate.dev/install.sh, the headline command on the homepage, returned a 404 page. anyone copying it pipedNOT_FOUNDintosh. i’d added a_redirectsfile, which is netlify and cloudflare syntax. the site runs on vercel, which ignored it silently.- the install snippet’s windows, npm, and cargo tabs were styled buttons with no javascript behind them. purely decorative.
- both install scripts ended by telling you to run
memcrate init, a command that release had just deleted. - codex invokes skills with
$load, not/load. every doc taught the slash form, so codex users were handed commands that do nothing.
i’d “audited” the docs twice before finding these, by grepping for strings
like memcrate init. that never matched, because the page renders it as
<span>memcrate</span> init. the audit that works renders the site, strips
the tags, and checks the visible text. same blind spot bit me on file
extensions: filtering for .md and .rs meant .gitignore and the
license file were never scanned, and both had problems.
v0.4.1 shipped the codex fix. install it however you like:
curl -fsSL https://memcrate.dev/install.sh | sh
npm install -g memcrate
cargo install memcratestill untested: the windows installer, and running /load in a real
session rather than just verifying the files land correctly. the codex bug
slipped past every automated check i had, which is a decent argument that
the live session is the test that counts.