0.5.1 · · Release on GitHub
Kernl 0.5.1
Upgrading from 0.5.0
- An install whose move to the built-in graph failed tries again on its
own. On a busy install the 0.5.0 copy from Neo4j failed and stayed on
Neo4j (
Neo4j import incompletein the log): nothing was lost, and the first start of 0.5.1 copies it again. Nothing to do by hand. - A deb or rpm install still needs the command once, for this update. The updater arrives with this package, so the install that is upgraded to it runs the old way; every update after it goes through the button.
Added
- A deb or rpm install updates from the button without a password,
servers included. The kernel runs as an ordinary account, and the only way
it had to ask for root was the desktop's password prompt (pkexec), so on a
machine without a desktop the button downloaded the package and handed back
a
sudo apt-get install …to run by hand. The packages now ship/opt/kernl/bin/kernl-system-updateand one sudoers line that lets any account run that script, and only it, as root. It copies the file somewhere root owns, checks it against the release'sSHA256SUMSfetched by root from the official repository, refuses anything that is not the kernl package or not newer than the one installed, and installs it with the distribution's own tool. To turn it off, create/etc/kernl/system-update.disabled; the button then falls back to pkexec and the printed command, as before.
Fixed
- The copy from Neo4j works while the kernel writes to it. It checked the built-in graph against the counts taken when it began, but the kernel keeps writing to Neo4j during the copy (the agents' experience graph alone syncs tens of thousands of runs on the first start), so every copy on a busy install failed and was retried, and failed, on every start. It now checks the graph against what it wrote, skips a relationship to a node created during the copy, and catches up after the switch. On a copy of a real install: 62k nodes and 136k relationships in 34 seconds while the graph grew, every label matching Neo4j five minutes after the switch.
- Everything is projected again after the live switch. Passes still running on Neo4j when the switch closed it never finished and held up every later sync and backfill (about 38k agent runs never reached the new graph). After the switch the backfills run again and the experience and decision graphs re-project from SQLite.
- A failed copy says why. The reason recorded in
kernl_upgradeswas three lines of stack; it is now the error itself. - The Docker image builds on ARM.
bun installhung in node-gyp on arm64; the Node headers are now fetched once during the build. AGENTS_SCHEDULER=offalso stops relaunching agents with answers waiting on start, like the scheduler and the inbox sweep.- Backups now include the built-in graph. Since 0.5.0 the Relate store
(
data/graph) is the only copy of chat long-term memory, the links between contacts and the news archive, andbackup.shleft it out. It now copies the store while the kernel keeps running (a read-only snapshot written out as a fresh store), reopens the copy and checks its node and relationship counts, and fails if they differ.restore.shputs it back for--dockerand--native, keeping the store it replaces asgraph.before-restore-<time>. A Docker restore also hands the restored files to the kernel's user:docker compose cpwrote them as root. On Docker the snapshot is verified inside the kernel container, so a host without bun (most servers) can runbackup.sh: it stopped at "bun not found" before.