Every classification returned "HTTP Error 404" because the configured
model gemma4:26b-64k no longer exists on the Ollama host: the tag was
retired upstream and the local weights were deleted, so all pending
updates fell back to manual review. Switch to the current gemma4:26b
tag (25.2B, 262k context) and make it explicit in .env so the
classifier no longer depends on the code default matching reality.
PieFed intermittently returned an empty body, which surfaced as a
current version error in the digest. Retry version fetches with a
short backoff so one bad response no longer triggers an alert.
Leave-one-out validation for the new model: immich 6/7 (the miss is
the conditional-breaking example where the model follows the written
rule that thresholds below the running version do not apply),
freshrss 6/6, uptime-kuma 4/4.
The monolithic check_updates.py has been split into the update_monitor/
package with separate modules for fetching, extraction, comparison,
classification, and notification. This makes the codebase easier to
navigate and test.
The flat services.yaml is replaced by per-service directories under
services/<name>/config.yaml, co-locating config with update scripts
and AI review examples.
When --auto-update is enabled, services with an auto-update.sh script
get their release notes classified by a local Ollama model (using
few-shot examples from review-examples/) to decide whether the update
can be applied automatically or needs manual admin intervention. Results
are reported in a single consolidated Matrix message.
Immich is the first service with auto-update configured, using 7
real release note examples to train the classifier.
