Four lines he hears out loud lived as string literals in three Go files, so
rewording one meant a rebuild. They move to fallbacks_ru_v1.json on the shape
nudges_ru_v1.json already uses: embedded, schema-versioned, several variants,
never the same one twice running.
The gap phrase is marked fixed, because it names one specific missing model and
must not drift into a general "I do not know". Every accessor falls back to the
literal it replaced, including on a nil receiver: these strings exist because
something already failed, so a broken template file must not take her last
words away.
Review of #108: "поговорили." reads as a summary of a conversation that did
not happen. One exported constant now, so the Stub, the LLMPhraser fallback
and the daemon all say the same thing.
internal/voice/replier.go keeps its own copy — that is the separate replier
seam, not this one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
It sat in world.go, which is about the workstation model; it is a phrasing
error and belongs in llmphraser.go. Also trims the PhraseQuery doc.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Call sites take the fallback text and log the error instead of treating a
canned string as success. phraseSource drops the text entirely — its callers
hold the passage and read it back better than "вот что я нашла: <passage>".
The talk scorer's before-and-after model probe (the #395 workaround) goes;
the run now fails only when every case errored, which is the honest
"nothing was measured" condition. TalkFixture gets its own schema version so
the two fixtures can be versioned apart.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
PhraseChat and PhraseQuery returned canned text with a nil error, so a dead
or OOM-killed server was indistinguishable from bad phrasing — "не знаю." is
also a legitimate answer.
Both now return the fallback text AND the error. The daemon keeps using the
text, so the turn still survives; a measuring caller counts a real failure.
An empty response is its own error: the model is up and said nothing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Nine reply cases and a fourth column in the talk report. The reply path is a
separate object from the phraser in the daemon, so Pair joins a Talker and a
Confirmer for a run that covers everything Maven says.
Cases carry intent/key/value because the replier is phrased from the decision the
router resolved, not from the raw utterance. Three of them are baits the other
paths cannot produce: a masculine verb about himself that she must not copy onto
herself, a polite plural input that must still come back на ты, and an unresolved
note that invites a question a confirmation is not allowed to ask.
Not scored against a model here — this box has no llama-server, and the baseline
test is opt-in on MAVEN_LLM_URL.
llmReplier lived in cmd/mavend, so the confirmation he hears after every fact,
note and reminder was the one phrasing path nothing could import or score.
Replier owns the prompt, the call and the parsing, and returns its errors instead
of hiding them — a dead model shows up as an error rather than as bad phrasing.
It has no stub fallback of its own; the daemon keeps that. StripThink is exported
for the daemon's own model callers.
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>
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.
Every phraser test built the phraser with NewLLMPhraserAt, which starts no
process, so NewLLMPhraser, spawnLlamaServer, startLlamaProc, llamaProc.Close
and extractPort sat at 0% while the package headline read 65.3%.
These drive the real spawn code against a fake llama-server script: the port
scrape, the three reachable startup-race arms (start failure, stderr EOF,
context cancel), and Close actually reaping the child. The orphan test
re-execs the test binary as the daemon, SIGKILLs it, and asserts Pdeathsig
killed the grandchild. The last test rebuilds the production command line and
checks kill-maven.sh's pattern still matches it — that pattern has gone stale
twice and leaked orphans both times.
Package coverage 65.3% -> 76.9%. The 60s timeout arm stays untested; it needs
an injectable clock in production code.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012YQGVXu5J1iCMCff5J4S1R
#400 rewrote the chat and query prompts in Russian and left three pieces
of English prose behind.
PhraseReminder's user prompt was fully English. It is Russian now, and it
no longer restates the JSON contract or the persona rules: the call goes
through chat(), so nudgeSystem already states both, and a second copy of a
contract is one more thing that can drift out of step with the first.
querySystemPrompt and router.KnowledgePrompt both closed with the English
"Respond ONLY with valid JSON:". That sentence is prose instruction, not
wire format — the JSON skeleton after it is the wire format, and it is
unchanged. Kept rather than deleted: the GBNF grammar makes it close to
redundant, but the grammar is switchable off (phraser NoGrammar), and the
sentence is the floor when it is.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018CotYKycuio1GwLbYh9jfc
Seventeen markdown files at the repo root, twelve of them dated one-shot
reports sitting next to CLAUDE.md. That is why stale docs read as
current: nothing in the path said which was which.
Root now keeps CLAUDE.md and AGENTS.md. Living docs move under docs/
and carry a Last verified line. Dated measurements move to docs/evals/
ISO-prefixed, and are never edited after the day, so a newer number is
a new file. The senior review moves to docs/archive/.
Every reference was rewritten across markdown, Go comments, the Makefile
and the recall fixture. The touched Go packages still build.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The evidence branch of PhraseQuery framed every source as "твои заметки" and
joined them into one quoted run-on. A live search snippet is not his note, and
a run-on gives a 1.7B one blurred claim to merge rather than sources to answer
from. That is the shape that named Левитан as the author of Война и мир.
Sources now arrive numbered, one per line, and the system prompt says three
ways that the answer comes out of them: only from the sources, say plainly
when they do not answer, add nothing of your own.
Blank sources take the knowledge branch. One empty string used to reach the
evidence branch and ask the model to answer from an empty list, which is the
one prompt guaranteed to make it fill the gap from memory.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The replier and the meeting summariser were the two call sites without a
GBNF. Both are exactly the shape that makes a Thinking variant answer with
its reasoning as prose, and neither had anything downstream that could
remove it.
The replier already parses {"response","mood"}, so it now sends the phraser's
grammar for that contract, exported once as phraser.ResponseGrammar so the
two definitions cannot drift.
The summariser stays text-in/text-out. The JSON wrapper is attached and
unwrapped in the daemon's Completer, so internal/capture is unchanged and a
Completer without a grammar still works.
The simulator told routing from phrasing by "has a grammar", which stopped
being true here; it now looks for the intent enum.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
wireVoice runs before the tick loop exists, so it could only be handed
the bare store adapter — and that adapter answers DayPlan with "not
available via direct store API", because a day plan is assembled by the
tick loop and is not a table to read. So queryDayPlan, which the query
chain reaches for "какие у меня планы на сегодня", failed for every
caller on the deployed daemon.
main already back-patches the other direction (daemonAPI.chatFn =
handler.handleText). This is the same seam in reverse, at both wiring
sites. No recursion risk: nothing in the voice path calls api.Chat.
With the plan reachable, it recited its reminders as literal JSON. The
payload unwrapper existed but was private to the phraser, so the day
plan had its own non-unwrapping copy. One owner now, store.ReminderText,
with the phraser delegating to it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Two things that were each half-done.
The prompts handed the model copyable examples. chatSystemPrompt lost its
openers this morning and the "Я подумала, что" tic went with them, but "не
забыл ли я" appeared in its place: the removed example had been suppressing
the masculine self-reference by accident. Two predicatives are not enough
signal, so the rule is now stated as morphology (-ла) rather than as a pair of
words — a suffix rule generalises where an example only gets copied.
querySystemPrompt had the same defect and gets the same treatment; its "вот что
я нашла: " opener is deliberate and stays.
internal/kiwix had no caller. actions_query.go said "once internal/kiwix is
wired into this chain" and that never happened. It is wired now, between the
notes pass and the web source: everything of his answers first, and only what
is left over is looked up. Off unless a `kiwix` block names a server and a book.
Reading the search snippet does not work. Kiwix builds it from wherever the
keyword matched, which on Wikipedia is the navigation box at the foot of the
page — the first version of this answered "что такое фотосинтез?" by reciting
"Ecological economics Ecological footprint Ecological forecasting …". Client
grows an Article method; the head of the article is the lead paragraph, which
is the definition the snippet was meant to be. Verified on the box: the same
question now answers correctly off the ZIM.
Only the rewritten query leaves the process. A test asserts it: a turn carrying
a stored note must not put that note in the search string.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The grammar examples in chatSystemPrompt were full clauses — ("я подумала",
"я рада") for her, ("ты сказал", "ты забыл") for him. A 1.7B copies those
instead of generalising from them.
Observed on homesrv 2026-08-01, in one session: all three chat replies
opened with "Я подумала, что ...", and one ended "...немного тревожусь.
ты сказал" — the second example pasted onto a finished sentence, which
reads as a truncation and is not one.
Contrastive pairs replace the openers, so the rule reads as a correction
rather than a template. The him-examples are dropped; the "ты" instruction
carries that on its own, and those two produced the worst output. A closing
line tells her not to echo the instructions, because a small model treats a
quoted string as licence to reuse it.
Verified after rebuild: three chat turns, no "Я подумала" opener, no
dangling example, feminine forms intact.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Two fix branches independently added a Gate to internal/llm. One is priority
between a voice turn and a background job, the other is admission control while
the resident model is swapped. They are orthogonal and both are needed, so the
swap one is now SwapGate, with SetSwapGate to install it.
Complete takes the priority gate first and the drain second. A background
request can wait a long time on priority, and counting it as in flight against
the drain that whole time would stall a swap on a request that has not started.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The drain counted only the phrasing paths in internal/phraser. The router, the
replier, the mail extractor and the memory evaluator reach llama-server through
llm.Client, so quiesce could report zero requests in flight while the router was
mid-generation, and the old server was killed under it. The turn then finished
on the new model, which is the split turn the swap exists to prevent. llm.Client
now enters an optional Gate before every completion and LLMPhraser implements
it, so one counter covers every holder of the base URL.
A total failure also reported itself as a rollback. Swap set RolledBack on the
path where the rollback failed too, so the page rendered "rolled back to — she
is still answering, with the old model" over an empty model name and a daemon
with no model at all. The total failure has its own flag now, LiveModel stops
naming a gguf that is not loaded, and the log says another attempt can recover
without a restart, which is true.
The swap also ran on the connection every other page shares. ipc.Client holds
its mutex for a whole roundtrip with no read deadline on either side, so a load
froze /dash, /history and /notifications for minutes. mavweb dials a second
connection for /models alone. POST /models joins the route table, and the load
settings no longer come off a form that renders no input for them.
Found in review of #68.
Loading a different gguf was a one-line edit to phraser.model_path plus a
restart. It is now an owner-triggered IPC call, off unless configured.
internal/phraser/swap.go holds the safety properties as code:
- Never two models resident. The old llama-server is killed and reaped
before the new one is launched. One 1.7B fits the Vega iGPU; a
blue/green overlap would OOM the box, so it is not offered.
- Atomic from a turn's point of view. Swap drains the in-flight turns
(they finish on the old model), then refuses arrivals with ErrSwapping
until the new server has answered /v1/models. No turn ever sees half a
swap; refused turns fall back to the classifier cascade.
- A failed load rolls back. If the new model does not start or does not
probe, the previous one is reloaded and the call returns RolledBack
with the error. If the rollback also fails the daemon says so and
degrades to the classifier rather than pretending to serve.
Holders of the completion client are re-pointed, not rebuilt: llm.Client
guards its base URL and LLMPhraser.OnSwap re-points it, so the router, the
replier, the mail extractor and the memory evaluator follow the new port
without knowing a swap happened.
Reach is deliberately narrow. phraser.swap_models is an exact-match
allowlist of absolute paths a human wrote, rejected at startup otherwise,
so "swap the model" can never mean "load any file on my disk"; the running
model is always swappable back to. MethodSwapModel is AuthStepUp, the same
rung as mutating the tool allowlist, and /models gates POST through the
same stepUpOK the tools page uses. Nothing calls Swap on a timer and no
act, intent or utterance reaches it.
Vikunja #250
Wording fixes from the review of the clarify + phrasing PRs.
- "На когда напомнить?" → "Когда?". After she has just been asked something,
the long form is the phrasing of a form field, not of a person.
- A reminder now wants a subject as well as a time. "напомни в 11" had a time
and nothing to say at 11, and she asked nothing at all — she now asks
"О чём напомнить?". Subject first, since a reminder with no subject is not
worth setting.
- The expiry notice is five phrasings picked at random instead of one fixed
sentence. It is the line he hears every time he walks off mid-request, so it
is the line that repeats most.
- The nudge prompt's ban on "обращения" is now "ласковые обращения". It was
meant to forbid "милый"/"дорогой", not his name — "Ками, ноутбук на трёх
процентах" is how she talks, and the eval's cringe check already only flags
pet names.
- The nudge example no longer claims she plugged the laptop in. She has no
hands and no smart plug; an example where she acts teaches the model to
invent actions Maven never took.
- replySystem: "тепло" → "спокойно и без официальных формулировок". A one-word
mood instruction a 1.7B can't act on, replaced with the behaviour meant.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
DEPRECATION, flagged not asked: LLM-phrased nudges are no longer the default.
LLMPhraser.PhraseNudge now returns a hand-written Russian template. The model
still phrases chat, queries and reminders — only nudges moved.
Why: measured over many runs, Qwen3.5-0.8B wrote formal "вы" and plural
imperatives, used masculine self-reference, and invented facts and units
(90-95 seconds to boil an egg). A nudge is five words of known content, so
generation buys nothing and risks the persona every time. Templates score
15/15 on the nudge fixture, the model 11-13/15.
Nothing is deleted: the prompt, the fallbacks and the whole LLM nudge path
stay. Set phraser.llm_nudges = true in deploy/mavend.json to get them back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Nudge wording as data instead of generation. The wording lives in
internal/phraser/nudges_ru_v1.json (embedded), about 10 variants per rule:
water, meal, break, service_down, netdata_critical, routine:, morning:, plus
a contentless default. That JSON is long because it is data — the owner can
edit any line of Russian without touching Go.
The picker:
- random, but never the same variant twice in a row for the same rule
- deterministic when seeded (math/rand with an injectable source)
- fills {since} / {service} / {what} from the candidate, and skips any variant
whose value is missing, so no raw placeholder can reach the piper voice
- {since} is spelled out in words ("полтора часа", "семь часов"), because
"3 ч" is wrong in a Russian voice
Scores 15/15 on the existing nudge fixture, on every seed swept. Nothing is
wired yet — that is the next commit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
It scored 27/27 on a run where two replies were "{" and "{\n \"". It only
tested that the string was not blank, so punctuation counted as content and
the worst replies of the run passed the first check.
Now a reply needs at least one letter, Cyrillic or Latin. Latin counts
because answers about ssd or vpn are legitimately part English.
Digits alone fail too. The same run answered "сколько варить яйцо
вкрутую?" with "15-16" — no unit, no words, and the wrong number as well.
That is not something she said.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Two bugs, one symptom. A run of the talk eval produced replies that were
literally "{" and "{\n \"" — those strings went out as things Maven said.
First bug: the parser could not tell "the model answered in plain prose"
from "the model started a JSON object and got cut off". Both came back as
empty, and every caller then shipped the raw text. Now an unfinished object
returns an error and each caller uses its own fallback instead. Bare prose
with no JSON in it still passes through, because small models do sometimes
answer that way and the reply is fine.
Second bug, and the actual cause: the grammar capped the response field at
400 characters. I measured it against Qwen3.5-0.8B at three different token
caps — 256, 768 and 2048 — and the reply came back exactly 400 characters
every time, cut mid-word. So the token limit was never what stopped it.
The bound is 1000 now, about six Russian sentences, still low enough to cut
off a repetition loop.
Token caps go from 256 to 768 on the chat and query paths so 1000
characters of Russian actually fits. The nudge path keeps its own cap; a
nudge is meant to be one sentence.
Note: cmd/mavend/replier_llm.go has its own copy of this parser with the
same bug. Left alone here so this commit stays small — that duplicate is
Vikunja #396.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The reply has to be Russian, but two of the phrasing prompts told her
what to do in English. Both are Russian now, in the same style as the
nudge prompt that already works better.
Also dropped the "you are maven, a self-hosted personal assistant"
line from both. The persona block right above it already says who she
is, so it was said twice.
The JSON part is unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The 0.8B answered about one chat turn in three with open reasoning as plain text, so no JSON ever closed and the fallback shipped "Thinking Process:" to the user. A grammar makes that output impossible.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
From a real reply in a nudge eval run: "Смотрите на его потребление
воды" is a plural imperative AND third person about him. Only the plural
printed.
Two separate faults. The check returned on its first hit, so the second
break stayed invisible and the failure read as milder than it was; it now
joins them. And "его" was not detected at all — looksVerb knows the
-й/-йте imperative but not the -те plural, so "смотрите" counted as the
person being talked about, which is what an antecedent means here.
pluralVerb already knows that form, so the antecedent test uses it too.
Third time a verb form has blinded this check. A fourth means it wants a
morphology table rather than another suffix.
The phrasing fixture was 15 nudge cases, so every prompt change we
measured only told us about nudges. But the shared context block sits in
front of five prompts, and three of them — chat, note query, general
knowledge — had no scorer at all. Those are the long free-form replies,
where a persona break is most likely and where nothing could see one.
27 cases, nine per path. Nine rather than five because the nudge fixture
already cannot resolve a change smaller than about three cases, and a
per-path score off five would be worse.
Reuses the persona checks instead of copying them. Length, mood and
"no questions" are left out on purpose: these paths return no mood, and
a follow-up question is a feature in chat, not a fault.
The run refuses to score unless the model answers before and after it.
PhraseChat and PhraseQuery swallow model errors and return a canned
string, so without that guard a dead server produces a full report with
zero errors and a bad score — which reads as bad phrasing rather than as
nothing measured. Vikunja #397 is the real fix.
The "address him as ты" rule had only reached two of the five system
prompts. Instead of pasting it into the other three (five copies drift —
that is how this happened), there is now one block, in internal/persona,
prepended to all five: nudges, action replies, chat, note queries and
general knowledge.
The block says who he is and how to address him (a man, always "ты",
never "вы", never "он" about him; Maven stays feminine), plus the
current local date and time. It is rendered fresh each turn because the
time changes, and it is correct with an empty config — the address and
gender rules are defaults in code. Config only adds optional facts:
owner_name, city, and the existing free-text `persona` string, which is
now the static half of the block.
Russian even in front of the English prompts: the rules are Russian
grammar, so they read best stated in Russian, and there is one copy.
Vikunja #394.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The third-person check asks whether anyone else was named before "он".
A nudge is mostly verbs, and they were counted as possible people, so
"попробуй встать и отдохнуть — у него есть перерыв" passed. Infinitives
and imperatives now join past tense as words that cannot be a person.
A plain noun before the pronoun still blinds it. That needs a parser,
and the comment says so.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The prompts stated the feminine self-reference rule but never said whom she is
speaking to, so the model produced formal plural ("Жду вас") and talked about
him in third person ("Он не ел 11 дней"). Adds the address rule right next to
the feminine one, in the nudge prompt and the confirmation prompt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The check asks whether anyone else was named before "он". Time words
were not stoplisted, so "сегодня он не ел" read "сегодня" as the person
being talked about and passed — which is the recorded break with a word
in front of it, and nudges open with those words constantly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The phrasing eval printed "llm (0.8B, ...)" no matter which gguf
llama-server had loaded, so two runs of two different models came out
named the same and were easy to mix up when comparing.
It now asks llama-server over /v1/models, same as the router eval
already did. The helper moved to internal/llm so both share it, and it
now errors instead of returning a blank name when the id field is
missing — an unreachable server gets labelled "unknown-model", never a
plausible-looking guess.
Both eval paths stay opt-in behind MAVEN_LLM_URL; no server needed for
go test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The phrasing run produced two persona breaks that scored clean:
"Приходите… Жду вас" (formal plural) and "Он не ел 11 дней" (talks
about him instead of to him). She is feminine, he is male, and she
speaks to him informally, one to one.
The new `address` check flags the "вы" family, plural imperative
endings, and a third-person "он" with no other subject named earlier in
the message. Like `hisgender` it is a keyword/suffix heuristic, not a
parser, and it prints the word it tripped on so a false alarm is easy to
dismiss. Limits are written out in the comment.
Both recorded strings are pinned as unit tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The system prompt showed the JSON contract as {"response": "..."} and the
user prompt repeated it. A 0.8B copies whatever sits in the response slot, so
7 of 15 nudges came back as literally "...".
Changes, all prompt-side — the {"response","mood"} contract is unchanged:
- nudge system prompt is Russian, feminine self-reference, with filled-in
examples on topics that never appear as rules, so copying them is visible
- rule names get a Russian gloss and a required keyword, named last in the
prompt where a small model weights it hardest
- durations render in Russian, not English
- the no-parse fallback says something Russian instead of "water — care",
which was going straight to a Russian piper voice
- same "..." placeholder removed from replier_llm.go
Scored on internal/phraser/eval: 0/15 -> 13/15.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Review internal/phraser/eval/checks.go -- it IS the measurement. Each check
names in a comment which DESIGN.md line it defends: length, feminine
self-reference (windowed around "я" so the operator's own masculine
second-person forms are not flagged), the cringe list (pet names, emoji,
"!!", fake concern, apology, emotional support, asking how he feels,
praise), on-topic, mood enum. No send/veto signal anywhere, per
DESIGN.md § "Rules decide, LLM phrases".
Fixture (158 lines) and tests (252) do not count toward the diff ceiling;
the scorer itself is still ~650. Splitting eval.go from checks.go would
give two commits neither of which measures anything.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The most load-bearing decision in the project was stated four incompatible
ways: the docs said Qwen3-1.7B, deploy/mavend.json said Qwen3.5-2B, the repo's
models/llm/ held an LFM2.5-1.2B gguf, and five code comments still said LFM.
Answering "which model is deployed" meant re-deriving it from scratch every
time.
Two facts the review missed, found while resolving it:
- /mnt/hdd1/llms is bind-mounted over /opt/maven/models/llm, which shadows the
repo's models/llm/. The LFM2.5 gguf sitting there was never loaded by
anything, so it was not evidence of the deployed model at all.
- That library holds Qwen3.5-0.8B, -2B and -4B, and no Qwen3-1.7B. The config
pointed at a file that does exist; the docs' Qwen3-1.7B was the stale claim,
the reverse of the assumed direction. Qwen3-1.7B is the CPT target, and that
training is still in flight (Vikunja #122), so no such gguf exists yet.
phraser.model_path moves to Qwen3.5-0.8B (Q4_K_M) — the smallest checkpoint on
disk, chosen for latency, and relevant to whether the LLM router is affordable
on this box. Docs and comments now say the same thing in one voice: 0.8B
resident now, CPT'd Qwen3-1.7B as the target, and the bind-mount shadowing
written down so the next reader does not mistake models/llm/ for ground truth.
Comments name the model, never a filename, so a swap stays a one-line config
change.
n_gpu_layers: 99 is correct and stays — compose passes /dev/dri and the render
gid for Vulkan offload to the Vega iGPU. CLAUDE.md's "CPU-only" was the stale
half of that contradiction and is corrected.
phraser.go also dropped a wrong "sub-1b, prompted not trained" size claim: the
target is trained end-to-end (RU CPT + joint persona/router SFT).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5JApcrCRVGmqrxnhynSik
Daemon side of Decision B: parse {"response","mood"} across the 4 consumers
(replier, nudges, reminders, chat), fall back to legacy formats. Drop the
LLM router — the classifier handles routing; replier/phraser share one
llm.Client (timeout 20s->60s). llm.Client reads reasoning_content when
content is empty (thinking models).
Docs: TTS piper-student plan (OmniVoice teacher -> piper student, from
scratch, phoneme-first). CLAUDE.md training guide.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Add llmphraser: LFM-based phraser implementing Phraser interface with
PhraseChat, PhraseNudge, PhraseReactive, and PhraseReminder methods.
- Add shared internal/llm/client: llama-server completion client used by
both the phraser (talking back) and router (routing), sharing one model.
- Add LLMReplier in mavend: replaces StubReplier for chat/nudge/reactive
replies, falls back to stub on model errors.
- Update Phraser interface: add PhraseChat method, update stub to match.
- Wire LLM phaser into mavend voice init, plumb LLM config from JSON.
- VoiceConfig: add QueryMinScore (default 0.55) + Persona config fields
- voice.go: remove queryMinScore const, wire from cfg.Voice.QueryMinScore
as reactiveHandler field
- llmphraser.go: add Persona to Config, prepend to system prompts in
chat and query paths (systemPrompt/querySystemPrompt methods)
- main.go: pass personaFromCfg into both phraser config blocks
- Makefile: add download-embedder target (Xenova/paraphrase-multilingual-
MiniLM-L12-v2, ~90MB ONNX)
- AGENTS.md: document embedder model download + libonnxruntime setup
- server.go: fix pre-existing wg.Add vs wg.Wait data race using accept
mutex. make test green, zero races across all 29 packages.
Task 4 created and tested router.KnowledgePrompt() but the live path in
LLMPhraser.PhraseQuery used a separate hardcoded copy of the same RU
anti-hallucination prompt, leaving KnowledgePrompt() as dead code and two
strings that could drift. Point the phraser at the tested helper so there
is a single source of truth.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Add CalendarEvents method to recordingAPI in auth_test.go
- Add CalendarEvents method to fakeCore in handlers_test.go
Co-Authored-By: opencode <opencode@anthropic.com>
claude
kami
kami