Victoriametrics

2026.08.30.1

system-overview now counts each physical disk once.

node_exporter exports every layer of the block stack. On Unraid an encrypted array slot is dm-N -> mdXp1 -> one physical sdX, so the disk view listed the same spindle two or three times and a dm-*/sd* pair reporting the same utilisation looked like two independent disks corroborating each other. On Tower, 2026-08-30, dm-3 and sdl were flagged together as a disk-saturation watch item: dm-3 = md4p1 = disk4 = /dev/sdl, one 14.6TB drive, and neither layer was near saturation (7d mean 20%, peak 52%).

• DISK_IO_QUERY (new export) adds {device!~"(dm-|md|loop|sr|zram|ram|nbd|drbd|zd).*"} so VictoriaMetrics drops the virtual layers at query time.
• isPhysicalDiskDevice() (new export) re-checks each returned series client-side, so a server that ignores the matcher still cannot double-count.
• Both derive from one VIRTUAL_DISK_PREFIXES list, so the PromQL string and the client-side check cannot drift apart. Tests import DISK_IO_QUERY instead of restating it.
• Unlabelled series still normalise to "unknown" and are kept — an unattributable series must not be silently dropped.

No resource schema change; identity upgrades[] entry appended.

2026.08.02.1

Real fixes for all 11 latent bugs tracked in the LOCAL victoriametrics-latent-bugs issue-lifecycle model (NEVER filed to the swamp.club Lab — see CLAUDE.md's anti-bypass rule). model.version and manifest.yaml both bump 2026.07.16.2 -> 2026.08.02.1, with an identity upgrades[] entry appended (no resource schema change).

• VM2 (HIGH, multi-series collapse): extractValues now flatMaps EVERY series in result.data.result (result.data.result.flatMap((s) => s.values ?? []).map(...)) instead of reading only result.data.result[0]. system-overview's CPU/memory/load stats now characterize the FULL set of series a range query returns, not just the first one. Single-series results are byte-identical (flatMap of one series equals that series). The flatten's inter-series timestamp boundary is a backward jump by construction and never trips the >600s forward-gap "Metric gap" reboot detector.
• VM1 (MED, health has no notion of expected targets): health gains a new backward-compatible method argument, expectedTargets (default []). Any name in that list not seen in the up vector is appended with status:"unknown" instead of being silently omitted. An empty/default list is a no-op — existing callers see no change.
• VM3 (MED, absent metric indistinguishable from idle 0%): system-overview now pushes a distinct "<CPU|Memory|Load> metric absent (no series returned)" anomaly whenever a metric's range query returns zero series. The numeric fields still report {0,0,0,0} (unchanged) — only the anomalies array gains the flag.
• VM4 (MED, boot-time epoch fallback): an empty node_boot_time_seconds result no longer falls back to bootTs=0 (the 1970 epoch, with an enormous computed uptime). uptime.bootTime is now the string "unknown", uptime.uptimeMinutes is 0, and a "Boot time unavailable" anomaly is pushed. A real boot time is completely unaffected.
• VM5 (MED, scalar query garbage-mapping): query now dispatches on data.resultType: a "scalar" result ([ts, "<value>"]) maps to a single [{metric:{}, value:<parsed number>}] row instead of the old two-row [{metric:undefined,value:null},{metric:undefined,value:null}] garbage. Vector results are unaffected (same r.value ? parseFloat(r.value[1]) : null mapping as before); a matrix body fed into query still maps every series to value:null (matrix support was never in scope for query, and stays a documented, intentional gap).
• VM6 + VM11 (MED, uninspected {status:"error"} responses): a new shared vmData(result) helper is called by the three direct single-query methods (query, query-range, health) right after the HTTP round-trip. It throws Error("VM query error: <result.error | 'response missing data'>") when result.status==="error" or result.data.result isn't an array, instead of letting a partial/hostile 200 body crash several frames down with an uncaught TypeError. system-overview and container-memory deliberately do NOT go through vmData — they keep their existing lenient degrade-to-zero behavior on missing/empty data.
• VM7 (MED, query-range vector-shaped-body crash): query-range's series mapper now does (r.values ?? []).map(...) — a series missing values (e.g. a vector body mistakenly fed to query-range) degrades to values:[] instead of throwing.
• VM8 (MED, unguarded inner .values/.value access): the disk loop, container-memory's series loop, and health's per-target status now all guard their inner access instead of crashing on a partial/hostile response: disk pre-filters series via Array.isArray(r.values) && r.values.length>0 before mapping; container-memory uses (r.values ?? []) and r.metric?.name; health uses r.value ? (...) : "unknown".
• VM9 (MED, disk NaN/-Infinity): folds into the VM8 disk pre-filter — a series with values:[] (present, empty) is now dropped BEFORE the Math.max/reduce computation runs, instead of computing NaN/-Infinity and relying on the >10 threshold filter to silently absorb it. The observable result (disk:[]) is unchanged.
• VM10 (LOW, negative topN): container-memory's topN is now clamped via Math.max(0, args.topN) before .slice(0, ...). A negative topN now yields [] instead of slice(0, -n) silently dropping from the END of the already-desc-sorted array. topN:0 is unaffected (Math.max(0,0)===0).
• Tests: all 11 latent-bug pins in victoriametrics_adversarial_test.ts flip from characterizing the bug (title prefixed pin:) to asserting the fix (title prefixed fix (VMn):), reusing the existing fixtures plus a handful of new inline hostile bodies. New adversarial coverage: two multi-series regressions for VM2 (a LOAD spike

Release 2026.07.16.2 — align model versions with manifests

Maintenance release across the @magistr extensions. For most packages this carries no functional change: the only edit is the model's version: field, brought back in line with its manifest version so the published model type version and the package version no longer drift.

Functional changes in this release are limited to:

• anime-cron: normalizeTitle now strips a ": subtitle" suffix and a trailing parenthesized year before comparison, fixing dedup false-misses where the torrent title carries a subtitle or year that the AniList romaji does not.

• arckit: first publish. Standalone ArcKit port — a 12-phase architecture governance state machine with 65 bundled templates, driven by a bundled skill.

Also tracks three extensions (kaiten, observability-agent, music-library) that previously existed only as untracked working-tree directories, recovered from stashes.

Maintenance version bump. No functional changes.

Merge pull request #5 from umag/extensions/jscad-stl-pair

extensions: add 15 more @magistr extensions + auto-discover CI