fallback.sh starts covered containers with docker start and never creates them, so a coverage list
the partner has never been sent is a promise nothing can keep — all twelve were missing. Adds the
push and remove paths, a readiness card that checks rather than infers, and the fallback state the
assistant needs to answer for it.
The card described the machine and said nothing about the thing whose dashboard
it is. Cache size matters most: VV_CACHE_ROOT is under /tmp, which on Unraid is
RAM, so it is shown against the rootfs percentage it counts against. Jobs are
locks with a live process — four lock files were sitting there and one job was
actually running.
Twenty-four fetch chains ended in an empty catch, which is not error handling
but error deletion: the request fails, nothing renders, nothing is logged, and
the surface sits on Loading forever. That is how the mesh chat's ReferenceError
read as a slow load for as long as it did.
_vvNcStyle referenced a const declared inside vvNcRender, so with the
formatting toggle on every mesh transcript threw ReferenceError and drew
nothing. The plain-render early return skipped the line, which is why it looked
like a slow load rather than a broken one.
Every fetch ended in an empty catch, so a failed channel list was
indistinguishable from a slow one: nothing rendered, nothing was logged, and
the next attempt was a backoff tick away. Errors now land in the card's status
line and the console, a failure retries in two seconds, and opening the pane
before the list has arrived asks for it rather than waiting.
`unraid-api apikey` costs 1.98s and ran once per render just to answer whether
a key exists. Cached for a minute — still read from the key store, so a
deletion still shows, a minute later rather than instantly. Partner probes are
multiplexed at the same time.
The fold saved a scroll and cost a click on every long answer, and what it hid
was usually the part worth reading. Follow already handles holding your place
while an answer arrives.
Local stays the default and the cheap path; mesh asks each partner for its own
sessions over SSH, live rather than cached, because a stream is true for
minutes and a cached one would be confidently wrong about the only thing the
card exists to say.
MagicDNS does not work on this mesh, so `tailscale ip -4` missed and blocked on
a system DNS lookup until it timed out: 5.04s per call against 0.010s for the
status parse, paid by every mesh operation on both hosts. Exact match first, so
no precision is traded for the speed.
HOST2 has UHD Graphics 730 and reported no GPU at all. An iGPU has no VRAM,
no sensor of its own and no encode/decode split, so it is drawn from the
per-engine busy figures Intel actually publishes and the fields it lacks are
null rather than a confident zero.
lm-sensors prints high= and crit= on the same line as the reading, so scraping
the line for its largest number reported a flat 100C on Intel while the shell's
last-field parse got a bare ")" and never fired at all. One parse in the
adapter now, preferring Tdie over Tctl since Tctl carries a +27C offset.
mdNumDisabled counts a slot that was never populated, so HOST2 reported
"Emulating 1 disk" for a parity2 slot with no disk in it. Emulation is a
data-disk state, and an assigned identity is what separates a slot that lost
its disk from one that never had one.
A mirror was probing an Ollama it does not have and an index it never builds,
so the card read "unreachable / not built" about the wrong machine while the
model it actually uses was up. The card now names whose figures it is showing.
Every turn runs on the AI owner now, so the ledger's per-host column had
collapsed to a single name and the fleet breakdown said nothing. The asking
node travels with the request and is validated before it is written, which
also retires ai_token_sync.sh — a partner ledger is empty by construction.
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.
The image is wider than the transcript is tall, so as a background it overflowed and percentage positioning ran backwards — nudging it down visibly raised it.
Almost every message here is one line, so the common action takes the bare key. Shift+Enter also breaks a line, because that is the reflex people arrive with.
Gmer4Lfe