The forwarded log named the cause in one line: the prompt cache limit
defaults to 8192 MiB. llama-server saves the full KV state of every idle
slot it evicts, 112 kiB per token, so RSS climbed about 170MB per
distinct prompt until the deployed server held 7.9GB for a 1.1GB model.
Measured on homesrv today, uncapped versus `--cache-ram 512`: RSS
plateaus at 932MB from the fourth distinct prompt instead of climbing.
The task's leading guess was wrong. `-ngl 99` costs almost no RSS,
because RADV keeps device memory outside the process. Numbers and method
in docs/evals/2026-08-03-llama-prompt-cache.md.
`-c 4096` is untouched. The knob is `phraser.cache_ram_mib`, unset means
512, negative passes no flag for a llama-server too old to know it.
The deploy still runs the old image, so the box keeps its 8 GiB default
until mavend is rebuilt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
mavend scraped the child's stderr for the listen line and threw every
other line away, and never piped its stdout at all. Nothing about the
resident model's memory was diagnosable from a running box: no buffer
sizes, no KV-cache layout, no offload lines, no prompt-cache limit.
Both streams now share one pipe and every line lands in mavend's log
with a `llama:` prefix. The last 12 startup lines are also kept and go
into the error when the server dies before it listens, because bare
"EOF" never named which allocation it choked on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"что я говорил про бэкапы?" is his data by definition, and nothing outside the
box has ever heard him say anything. The boundary matched possession words only,
so the question walked past it into SearXNG and came back answered out of a Habr
article about somebody else's backups.
A speech-verb marker class was written first and dropped. Russian gives every
verb a dozen surface forms and the "как я говорил, ..." preamble list has no end,
so each form the lexicon missed was one more question reaching the world, and a
missing verb looks exactly like no bug.
The boundary now embeds two frozen seed sets and scores the turn's own query
vector, already computed upstream, against both. Nearest side wins. The
possession markers stay as the offline floor for a handler with no embedder.
19/19 held-out utterances correct against multilingual-e5-small; see
docs/evals/2026-08-03-personal-boundary.md. The live probe on the deployed box is
not done.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Automatic rather than a flag, unlike -reembed: only voice-tapped facts are in
this index, so it is tens of embeddings rather than thousands of notes. And
waiting for an operator to know the repair exists is the failure being fixed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every write-path fix leaves the rows already stored wrong, and a box in that
state looks fine: recall answers with the wrong text and nothing logs an error.
That is how the original poison survived four restarts.
RepairFactVectors resolves each fact vector against the fact it names,
re-embeds the ones whose text is stale, and deletes the voided, superseded and
orphaned ones. Marker-guarded and idempotent, so it runs once per box and a run
that dies partway is simply redone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
queryMemory returns a fact's stored text verbatim, so the text the write path
indexed is what he hears. It was the utterance, which made recall of any
voice-tapped fact answer with the sentence he said: go_version = 1.20 was
indexed as "какая последняя версия языка Go?", and that question came back.
FactRecallText renders the fact instead, and the utterance stays in meta as
provenance. Correcting a value now drops the key's vectors the way voiding one
does, since the superseded value was still answering.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
queryMemory and queryNotes both gate on score alone, so both needed it. The
eval keeps its own copy of bestRecall — package main is not importable — and a
fixture that measures a weaker gate than the daemon runs flatters it, so the copy
moves in step and its test pins the new rule.
Measured on the held-out recall fixture with the real embedder: 17/32 cases pass
→ 22/32, false recall 1/5 → 0/5, answered after gate 18/27 → 17/27. The one true
recall lost is en-hard-024, an English question against a Russian note, where no
lexical test can help.
The score gate cannot separate the right note from an unrelated one: the
held-out fixture puts the right note at 0.791-0.890 and the must-be-silent cases
at 0.795-0.835, so a note about his slow network answered 'почему небо синее?'.
RecallAllowed adds a topic veto, and applies it only to a question that mentions
nothing of his. That restriction is the whole design: demanding a shared word of
every recall silenced four true recalls on the fixture to kill one false one,
because recall exists to find the note whose words he no longer remembers. A
question about his own life keeps the embedder as its only judge.
IntentFact used to persist whatever the model invented for a question-shaped
utterance, at confidence 1.00, and index it for recall under the question's own
text. Two such rows then claimed seven unrelated world questions and silently
disabled world answering.
A question now goes down the query chain, which is what he asked for. The second
half is confidence: a value grounded in what he said stays 1.00, a value the model
supplied for words he never said drops to 0.60 and says so in the log. Same
reasoning as 'LLM output is not authorization' on the act path.
The predicate a fact write needs before it trusts a routing decision. Tokenized,
not substring: 'что' inside 'чтобы' is not a question. Capture verbs win over
every question signal, because 'запиши что я пил воду' contains an interrogative
and is still a capture.
Revert voided the fact row and left the vector, so recall kept serving the
voided fact's utterance and the documented repair reported success on a box that
stayed broken. There was no way to repair a poisoned box at all.
DeletePrefix covers every vector for the key, earlier rows included: their values
are superseded, and a superseded value has no business claiming a turn. It is
best-effort — the audit trail is already committed, and a fact that is voided but
still recallable beats a void that failed.
Owner's correction. It is the same host CLAUDE.md already calls workpc, and
two names for one machine read as two machines. The dated eval file keeps the
old name: a measurement is never edited after the day it was taken.
The offload inventory grows a column, because "seven callers of the resident
model" stopped being the useful fact. Which of them is offloaded, and under
which half of the rule, is. Three are resident-only on purpose and the table
now says why rather than leaving it to be rediscovered.
The three-outcome table is the part that was not obvious from the rule as
written. A configured-and-asleep workstation names the gap; a box with no
workstation block does not, because naming a gap requires a gap.
queryGeneral has nothing fetched to fall back on, so it is the sharp case:
with a workstation configured and asleep he is told that, rather than told
something false in a confident voice. The 1.7B answering a world question is
where "Война и мир" got Левитан as its author.
The sources that already hold a passage — a live search, a ZIM article, a
page he named — go through the world model too, but read the passage back
when it is not there instead of naming a gap. A real quote beats "не могу
сейчас", and nothing is invented on either path.
The Stub and every test double keep the Phraser interface they have.
PhraseWorld is reached by assertion, and a phraser without it is the
no-workstation case.
The middle outcome is the whole task: a workstation that is configured and
asleep produces a gap, and the resident model is never asked. The parity
test compares the bytes PhraseWorld sends the workstation against the bytes
PhraseQuery sends the resident model, so the fixtures and the daemon cannot
measure two different prompts.
The nudge tests cover the silent half from both sides, including the
temperature, which is how the workstation would otherwise change how she
sounds without anyone deciding to.
The naming half of the degradation rule in docs/offload.md. PhraseWorld has
three outcomes: no workstation configured means the resident model answers
exactly as today, a workstation that is taking work answers, and one that is
asleep returns ErrNoWorldModel so the caller can say so. Naming a gap
requires a gap — on a box that never had a second model, refusing every
world question would remove a capability he has now.
Both prompts move into knowledgePrompt and evidencePrompt, shared by
PhraseQuery and PhraseWorld, because prompt parity across two models stops
holding the moment there are two copies of a prompt.
The silent half comes with it: chatWithSystem and chatWithMessages prefer
the workstation when it will take work, at the same 0.7 the resident
transport samples at, and say nothing when it will not. That covers the
digestion worker's nudge and reminder phrasing without touching tick.go.
Only Available and CompleteRemote are in the Remote interface. Pair.Complete
has its own floor and the phraser already owns one; two floors under a
single call is one too many.
The phraser's own transport has always sampled at 0.7 and this client has
always been greedy. Routing a phrasing call through the client must not
change how it decodes, so Req carries the temperature and 0 — the zero
value, and what every existing caller wanted — is still greedy.
Both fixtures, run from homesrv across the LAN with the proxy env stripped.
Routing: 84.4% full / 93.5% intent-only at p50 329ms through the cascade, against
72.7% / 77.9% at p50 0.80-1.04s for Qwen3-1.7B. Talk: 25/27 against 20/27, with
knowledge 9/9. Nudges 15/15. Settles #485's first assumption by measurement.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NJYcaBiuny9UGSpFweQVQ1
modelSeam builds an llm.Pair when a workstation is configured and hands it to
the router and the replier. Both are the silent half of the degradation rule:
the big model is only better there, and he is never told which model answered.
No block, no probe, and the box behaves exactly as it did.
Owner's call, 02-08-2026. Without it the 12B spends the reply budget on
reasoning tokens and answers empty at low max_tokens. Verified on the box:
"Столица Франции?" now answers "Париж" with no reasoning_content.
make build now catches a broken supervisor on homesrv. deploy/mavgpud.json
carries the owner's gemma-4-12b line with the MTP draft model, passed to
llama-server untouched. The unit is a systemd user unit because sudo on the
workstation wants a password; lingering is the one command left to the owner.
The fixtures are the live numbers sampled from the box on 02-08-2026, where the
CPT run held 12.8GB of 16 as proc/478104/vram_35881.
The cases that matter are the ones where a mistake is silent: our own
llama-server counting as a contender, an unreadable card reading as free, and
/health hanging or proxying into a closed port instead of answering 503.
The lifecycle rule from Vikunja #488. Not on demand, because a 7-14B takes tens
of seconds to load and a world question would meet a gap every time the card
had been quiet. Not always on, because that is what holds the card.
/health is answered locally and always, so Maven's prober costs nothing and
works while the model is down. Everything else is reverse-proxied to
llama-server, which is what makes the idle window measurable at all.
Yielding is checked before starting, and both transitions are damped by a poll
streak so a short-lived rocm process cannot evict the model.
The workstation cannot keep a 7-14B resident: it would hold 16GB against the
owner's CPT runs, Correx and the manga-recap pipeline. So the process that
stays up costs no VRAM and the model comes and goes under it.
Contention is detected by presence on the KFD, not by a VRAM threshold. A ROCm
process registers under /sys/class/kfd/kfd/proc when it initialises HIP, well
before it allocates, so we see a contender during its startup instead of after
it has already lost an allocation race. rocm-smi is not installed on that box
and a per-second subprocess would get tuned down until useless, so this reads
sysfs and forks nothing.
Free VRAM is read only to decide whether to start. It is never a reason to
stop: by the time free VRAM has dropped, the other job has already failed.
The remote model cannot be a llama-server that is simply left running: a
resident 7-14B holds 16GB against the CPT runs the card is for. So what
is always up on the workstation is a supervisor, and llama-server is
loaded while the card is free.
Still not a scheduler. It arbitrates nothing between callers, and Maven
never asks it to start anything.
Seven cases. The load-bearing ones are the constraint from 483: an
unconfigured deploy never probes and always reaches the floor, a busy
card degrades silently with the remote untouched, and a remote that dies
between probes still completes the turn and corrects the cached answer on
its way out.
CompleteRemote is pinned not to fall back, because a named gap that
quietly became a 1.7B guess is the failure this whole split exists to
prevent. And 1000 Available calls are pinned to make zero probes.
Pair holds both models and decides which answers. A prober asks the
remote whether it will take work and caches the answer, so a request
reads an atomic bool rather than paying for a health check. Routing sits
at p50 825ms on the hot path and must never wait on a machine that may be
asleep.
The two methods are the two halves of the degradation rule in
docs/offload.md. Complete falls back silently, for routing, replies and
nudge phrasing, where the big model is only better. CompleteRemote
returns ErrRemoteUnavailable instead, for a world question, where the
1.7B does not answer worse but invents.
A nil remote is the unconfigured deploy: nothing probes, everything goes
to the floor, and the box behaves exactly as it does today.
483 is an umbrella and its children are the work, so what it owes them is
the shape they must all obey. docs/offload.md records it: the degradation
rule and where its line falls, admission control rather than a GPU
arbiter, the embedder staying on homesrv because it backs the classifier,
and the inventory of what runs a model on the box today.
CLAUDE.md gets a pointer, because an agent about to add a model caller or
touch a daemon seam needs to know this before it starts, not after.
It was untracked but not ignored, so one git add -A would have committed
MAVEN_AMBIENT_TOKEN. Same class as deploy/telegram.env, which is already
ignored.
Five hardcoded transports, three in internal/ipc and two in
internal/worker, all now go through the seam address. The unix perms
logic moved into netaddr, so the two copies of parentDir and the umask
dance are gone.
peerCaller already returned ok=false for a non-unix conn, so the
SO_PEERCRED path degrades correctly on tcp with no change.
Six cases. The load-bearing one is the first: every deploy in the tree
writes a bare path, and it must keep meaning a unix socket with no
handshake in front of the payload.
The rest cover the tcp seam: a good token round-trips, a wrong one comes
back ErrUnauthorized, a stranger that speaks HTTP at the port is dropped
while the listener stays up for the next peer, and a tokenless tcp bind
fails rather than serving his turns to anyone who connects.
internal/netaddr parses a daemon seam address and dials or binds it. A
scheme-less address is unix and behaves exactly as it does today: same
0700 parent dir, same 0600 socket, same bytes on the wire. tcp://host:port
is the new option, and it is what lets a module live on another host.
Over tcp the filesystem permission that authenticated the unix socket is
gone, and what crosses this seam is audio of the owner speaking. So a tcp
listener requires a shared token, checked in constant time before the
first protocol frame is read, and a peer that fails is dropped without
taking the listener down with it.
morning_routines, feeds, crawl.on_demand and netscan.enabled in
deploy/mavend.json; -ambient-token on mavweb, interpolated from a gitignored
/.env. All verified on the box: the dispatcher builds the morning plan,
/api/ambient answers 401/201, the crawler reads a named page, netscan finds 3
devices, and /events fills with scan:lan and ambient:notif.
Filed 482 (ambient reads a notification's wall clock as UTC). Corrected 479:
both capabilities work once configured, so it is not a routing defect.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NJYcaBiuny9UGSpFweQVQ1
claude
kami