Curated state copied to every node is state that can disagree, so the index,
the model and the shared memory stay on the owner and each node reaches them
over the SSH trust onboarding already builds. Chats stay on the node that had
them; memory and bug reports stay the owner's to write.
Both checks already answered their question on demand and both needed
somebody to press a button on the right row. One host here has returned
nothing but 5xx for months.
Filed as findings, which is the existing answer to a condition that
persists while nobody is looking. Grouped by cause rather than by
hostname: a default policy of bypass produced twenty-two findings that
were one sentence repeated, and they have one fix between them.
vv_is_ai_host() was `=== 'host1'`, which made a physical fact — that is where the
GPU is — look like a rule. AI_OWNER_HOST declares it, so the card can move to a
rebuilt host3 or a friend's spare.
The gate was also doing two jobs. Assistant docks and findings strips now ask
whether a model is reachable, so a node without a GPU gets them by borrowing;
the AI tab asks whether this is the owner, because that page carries the bug
reports, the index and the model configuration — the surface where the
vocabulary assumes you built the mesh.
Resolution is local, then owner, then anyone else declaring a model, pinned once
it answers. Pinned rather than re-derived per call: a mesh that re-decides every
request eventually decides differently mid-conversation, and a chat whose second
turn lands on another machine has no history there. Cleared only on a transport
failure, and only when there is somewhere else to go — a single-node mesh whose
model is down should say so, not report AI as unconfigured.
Identity now includes the host: "Bazarr is on the skip list" is a different fact
on each machine that says it, and without the host in the hash the second node
to report a subject would land on the first node's record — one machine's
dismissal silencing another's live fault.
Nine dismissed findings existed when this changed, so the write path migrates a
legacy id once, at the only moment it matters. Verified: a dismissal survives the
rename rather than resurfacing as new.
acted_by is stamped in apply_action, not set_state — the sweep calls set_state to
move a candidate to needs_operator, and crediting a machine for classifying is
not the same as recording who decided.
Declared in PHP rather than in each bash header, against scriptinfo's usual rule
that the header next to the code is authoritative: it is authoritative about what
the script does, and an enhancement is something else reading its output, gated
by a flag the script has never heard of. A description of PHP inside a file that
cannot enforce it would drift the first time either changed.
Shown with its switch, so "there is an enhancement" and "it is running" are never
the same claim. Discovery is listed even though nothing acts on its output — it
takes the first accessible root folder with no regard for content, and the
classification scan is what notices that night. That relationship explains where
misfiled series come from and was written down nowhere.
The triage could say Narcos is filed under anime and offer nothing to do about
it. Findings that recorded a destination now carry a Move action, which the AI
tab renders as a button on its own — actions are read from the server, so no UI
knew this existed. Armed like Fix and Dismiss, because it is the only action in
that card that relocates files rather than settings and the undo is another move.
Offered only where the triage said misfiled, never for uncertain: three
narrowings before a file is touched. Re-reads the series from Sonarr first, so a
finding gone stale cannot undo a move made by hand in the meantime.
The scan's own header calls reverse-anime leaks "genuine judgment calls" and
reports them as a count. Seventeen of them concealed two live-action crime
dramas filed under anime for as long as nobody read the list. The scan now
persists that bucket as its own verdict, and the sweep asks the model to sort
only those eleven — not the eleven hundred it already classifies correctly.
Files findings for misfiled and uncertain only; donghua and anime-adjacent are
answers, not problems, and filing them would rebuild the undifferentiated list
this exists to break up. Moves nothing — --move stays a flag a human types.
A fourth candidate source beside the arrs, the system log and container logs:
skip-listed containers, strikes past their limit, repeated restarts, unattended
reboots and sustained pressure become findings. The kind is deliberately not
conf-bound, so it cannot autofix by construction rather than by a switch — and
WATCHDOG_SCAN_IGNORE suppresses it, so a knowingly broken container stays quiet.
First consumer AI_ASSIST_WATCHDOG has ever had.
The evidence of a syslog or container finding carries a line count and the timestamp of the
first matching line, so it differs on every pass even when nothing about the fault has changed.
Both the acknowledgement and the announcement were pinned to it, which meant acking one expired
fifteen minutes later and the notification repeated every sweep, forever — worst exactly where
it matters least, on a fault that has been there since the machine was built. Those kinds now
pin to their identity. A fault that genuinely worsens has a different class and so is a
different finding, and still announces on its own.
Triage could see what Varaverk wrote about itself and what the arrs said about themselves, and
nothing else — so a disk throwing I/O errors, a filesystem going read-only or a PCIe link
retraining every two minutes was invisible to the thing whose job is noticing. Container state
was already watched; no line any container actually wrote ever was.
Container patterns match the environment rather than the application. Fifty containers are fifty
programs with no shared vocabulary for malfunctioning, but an exact shared one for a full disk or
a corrupt SQLite file, because those strings come from libc and SQLite rather than from the app.
Both halves are checked by Tools/ai_log_check.sh, which is two tests because the failure modes
are opposite: fixtures for recall on faults this host has never had, and a replay of its real
logs for precision — 74,519 syslog lines and 79,193 container lines, matching only the PCIe
errors it genuinely has.
Severity was being graded from a hand-picked three fields, so every one of these would have been
filed as a warning however bad it was, and notified as one.
A finding nobody is told about is a finding nobody has, and the card added earlier only shows
them to someone who opens the tab. Only needs_operator is announced — an open finding may still
be repaired by the next pass — one notification for all of them, and each is announced once and
stays quiet until the fault changes or gets worse.
vv_notify() hands the message to common.sh's notify() rather than reimplementing the channels,
and calls detect_hosts() explicitly because load_config.sh deliberately does not: without it the
Unraid notification arrives and Discord silently never does. It also reports false when no
channel is switched on at all, since notify() exits 0 either way and a caller believing that
would mark a finding as told and never mention it again.
Notification text is folded to ASCII. Unraid's notifier dropped an em dash outright and left the
double space behind, which was found by sending one and reading what arrived.
Several scripts delete inside a conf path — the orphan cleaner runs rm -rf under a download dir
and rsync runs --delete against a destination — so /mnt/user/Movies becoming /mnt/user is the
edit that turns a cleanup into a sweep. Depth cannot be the test, because /tv and /movies are
real container-internal values here; direction can. Clearing a path, making it relative and
'..' segments go with it, and autofix additionally requires a proposed path to exist, since
every probe it has is a network probe and proves nothing about a directory. Refusals now reach
the caller: a save whose only change was refused answered ok with no explanation.
An ack is supposed to hold only while the thing acknowledged is still true, but a finding with
no key compared '' against '' and so was silenced permanently — dismiss wearing ack's label,
which is every arr health finding on this host. Those pin to the shape of the fault instead.
Dismiss and reopen become real answers rather than states only reachable by editing JSON, and
apply_action now refuses anything the finding did not offer.
Conf writing is granted to this profile alone and it is not offered as a button — making it
selectable would put a write one click from any question. Its prompt leads on what it does
not decide, because the resolver picks the key and the probe picks the value.
Learned alias, then the literal key, then the keys and script ids whose own words contain
every word said. No similarity scoring anywhere: a near-match fails silently and confidently,
and here it would write to a key nobody named. Several matches is a question, not a ranking.
A term alias decides what gets written to conf and has to be earned by repetition; a spelling
decides how a sentence is read, and the same typo is rarely made three times.
A term that resolved the same way three times stops being an inference and becomes a lookup;
the corrections are the rows worth having, since a term that meant two things is still a
guess and is never promoted.
They already publish what they believe is wrong, structured and graded, so there is nothing
to parse and no severity to second-guess. Findings widen from naming a conf key to naming
anything specific — Radarr's import lists being down is actionable in Radarr, not here.
Nothing in Varaverk fires when a job ends, so this picks up run records that completed
since the last pass rather than adding a hook to forty scripts. Ahead of stability in the
cycle on purpose: a wrong port is not fixed by rebooting the machine.
The reply to "shall I fix it" ends in a conf write, so the three actions are matched by
pattern with an ambiguity refusal instead of being inferred — and a phrase that supports
two readings asks again rather than picking one.
A model can be confident a port should be 8686; a probe can report that 8686 answered, and
only the second is a fact. Detecting and repairing are also separate things to trust, so
the feature runs read-only until the write switch is turned on under it.
A bug stays open until the code changes and nothing here can close it; a finding closes
itself when the probe that proved the fault starts passing, so the two cannot share a
store. Acknowledging one is scoped to the value it was acknowledged at.
Gmer4Lfe