links: site | repo | project page
Memcrate v0.4.2 shipped with a lesson I keep relearning: a test that has never failed has not earned much trust yet.
The headline change was support for Claude Desktop, but the surprise was that there was no new installer target to build. Claude Desktop local agent mode reads the same Claude Code skills directory, so the right move was to correct the docs and explain the support clearly: Claude Code, Claude Desktop, and Codex all work with the same vault pattern.
The real work was live-agent testing. Static checks can prove that the files install. They cannot prove that an actual agent finds the vault, reads the right note, and uses the context instead of guessing. So the new test creates a scratch home, seeds a temporary vault with a canary phrase, installs the current skills, runs the agent, and fails unless the answer proves it read the file.
That test failed immediately. The first failure caught a vault-discovery issue where the instructions depended too much on tool identity instead of the capabilities the agent actually had. The second failure was in the test harness itself. Both were useful because they looked like the kind of bug a real user would hit: the skill appears installed, but the agent politely misses the context and asks for a path.
v0.4.2 also cleaned up docs drift on memcrate.dev and moved the visible site version to one place so the header, footer, and hero stop disagreeing.
This is the shape I want for Memcrate releases now: install checks for the binary, visible-site checks for the docs, and a live agent test for the part that actually matters.