135 commits. make build produces all 8 binaries; make test exits 0 across 38 packages with no failures, no data races, gofmt and vet clean.
See PR #47 for what had to be fixed to make it build as a unit.
The file ran two sweeps and the second one changed the resident model, but
the lede still opened with "Recommendation: keep Qwen3.5-0.8B". Anyone
landing on the file read the wrong conclusion and had to scroll 100 lines
to find that it had been replaced — and it contradicted CLAUDE.md, which
already says the resident model is Qwen3-1.7B.
Both sweeps are accurate, so nothing is rewritten. The lede now states the
outcome and the first sweep's verdict is scoped to what it actually tested:
it rejects LFM2.5-1.2B, which still holds. It never was a case for keeping
0.8B as the resident model.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Numbers behind the resident-model change, plus the answer to "could a 230-350M
model do this instead" — no, and the reason is worth keeping: LFM2.5's published
instruction-following scores beat Qwen3.5-0.8B, and every one of those benchmarks
except Multi-IF is English. In Russian the 350M invents non-words and the 230M
answers in Spanish.
Also fills the row TALK-EVAL-31-07-2026.md had to void for contamination, and
corrects a wrong call I nearly made: the 1.7B's 16s p95 looked like the reasoning
trace, but the 0.8B sits at 17s in every run and the 1.7B beat it twice out of
three. The long tail is shared and is not the Thinking block.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Stock Qwen3-1.7B, not the CPT'd one — that training is still running. It won
on both fixtures we have, measured tonight on an otherwise idle box:
routing, 77 RU cases, intent-only: 67.5% vs 59.7% for Qwen3.5-0.8B
talk fixture, 27 cases: 20/27 vs 11-17/27
It also beat Qwen3.5-2B, which is 20% larger, on every routing column.
Two other things came with it:
n_ctx goes 2048 -> 4096. This is a Thinking variant, so reasoning tokens need
the room, and 4096 is the context every score above was measured at. Shipping
2048 would ship something nobody measured.
The doc now says not to bother with sub-500M models, because I checked and they
are not close. LFM2.5-350M routes at 5.2% — worse than guessing among 7 intents
— and answers "столица Франции?" with "Сторзит", which is not a word. The 230M
replies to Russian in Spanish. Their published IFEval and BFCL numbers are good
and they are all English.
Note the routing gain needs the LLM router actually wired on to show up. It is
still nil, so this commit buys the phrasing improvement today and the routing
improvement when that lands.
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
Same 9 cases as the retrieval eval, so the numbers compare directly:
hand-written keywords hit 8 of 8, this is what the model reaches on its
own. Reports the hand-written query next to the model's for every case,
because where the phrasing differs is the useful part.
Opt-in on MAVEN_KIWIX_URL + MAVEN_LLM_URL, like the other evals.
Result on Qwen3.5-0.8B: 3 of 8, identical on all three runs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Kiwix ranks by keyword, not meaning, so a translated question finds song
and TV titles. This asks the resident model for the TOPIC instead: a short
English noun phrase, like a Wikipedia article title.
Locked down three ways, because a wrong query is silently wrong:
- A GBNF grammar, same idea as routeGrammar and responseGrammar. The
reply must be {"query":"..."} with Latin words only. The JSON wrapper
matters: this model always thinks out loud and this llama-server build
ignores the thinking switch, so a bare word-list grammar just captured
"Let me analyze this request carefully" for every question.
- max_tokens 32, since the answer is a few words.
- CleanQuery, which throws away empty, Russian and prose replies rather
than passing them to Kiwix, and drops question words like "why" and
"how much" that a keyword ranker cannot use anyway.
Client side only. Nothing is wired into the daemon or any config.
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
Records all four configurations on the 27-case talk fixture, three runs
each: no grammar, plus grammar, plus Russian prompts, plus the truncation
fix. Composite, per-path and per-check, with the reproduce command.
The short version is that the plumbing got fixed and the score barely
moved. Grammar was the real win. Russian prompts helped a little and cut
latency by 5x. The truncation fix was necessary and bought nothing.
Also writes down three things that are easy to lose:
- The truncation cause was the grammar's 400-character bound, not the
token cap. Measured at three caps, same 400 characters every time.
- Then I set the bound to 1000 against a 768-token cap and made it worse.
The two limits have to agree.
- One run is contaminated and marked void: I ran an agent against the same
llama-server, and the report still claimed zero errors while a third of
the fixture silently answered "не знаю.". That is #397 and it is worse
than filed — a busy server is indistinguishable from bad phrasing.
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
Step one of letting Maven read instead of recall. No LLM yet.
internal/kiwix/client.go: search a local Kiwix server, parse the RSS
reply, hand back title + path + plain-text snippet + word count. The
snippet is the unit of context; a full article is ~100KB of HTML and
will not fit a 4096 token window.
internal/kiwix/retrieval_eval.go plus knowledge_v1.json: the 9 knowledge
questions from the phrasing fixture, each with hand-written English
keywords, scored on whether a wanted article comes back in the top 5.
Opt-in via MAVEN_KIWIX_URL, since CI has no Kiwix. No pass bar, the
number is the finding.
Result on the live mirror: 8/8 answerable questions hit, 7 of them at
rank 1. Retrieval works. Keywords are written by hand on purpose, since
Kiwix ranks by keyword and not by meaning, so a natural question fails.
A query-rewrite step is the next piece of work.
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 owner's call, 2026-07-31: a 0.8B model does not know enough about the
world to be useful without reading something. So she may now read external
sources to answer world questions.
What replaces the old rule, in all three docs:
- No telemetry, no cloud model, no third-party account. Unchanged.
- Local first: the Kiwix ZIMs on the box before anything on the network.
- External search is allowed but off unless configured, same as weather
and telegram.
- His notes and facts are never search input. Only the utterance goes out
— never the persona block, the history, or matched notes.
Docs only, no code.
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 block listed what she can do and ended with "nothing else". It sits
in front of the chat and general-knowledge prompts too, so that told her
to refuse the exact thing those prompts are for. Talking is now first in
the list, and the closing line limits ACTIONS rather than everything.
Also dropped the self-introduction from the knowledge prompt. It said
"Мавена, персональный ассистент" — a different name and a masculine
noun, right after the block says she is Maven and feminine. Identity
lives in the block now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The context block now lists her real capabilities: reminders, notes and
facts (write and recall), and the calendar — all three are code paths in
mavend today. Weather, telegram and shell acts are listed only when the
config actually has them, because offering something she cannot do is
worse than staying quiet about it.
Also drops the pronouns from the optional name/city line. The block's
own "ты" is Maven, so "тебя зовут" read as her name and "его" would have
shown her the third-person form she must never use about him. They are
plain labels now.
Vikunja #394.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
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
Tried rewriting the nudge prompt to lead with five on-topic examples instead
of rules. Three eval runs each side: before 12/13/14 of 15, after 11/12/11.
The loss is all in the address check — formal "вы" and plural imperatives came
back once the "говоришь на ты" rule stopped being its own sentence, and the
on-topic examples leaked their wording into the wrong cases.
Prompt reverted. Only the finding is committed.
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
"Nothing on the host" meant two different things and the script treated
them the same. If docker answers and names no running containers, Maven
really is down and the script should say so and exit 0. Only when docker
cannot be asked is the answer unknown, and that is the case that must
fail loudly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The migration copied rows with SELECT *, which matches columns by
position. It is correct today, but if the old table's order ever
differed it would shuffle every row instead of failing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
docker-compose.yml has no 'pid: host', so each container has its own PID
namespace and pkill on the host matches nothing inside them. The script
then printed "All services gracefully stopped" while mavend, its
llama-server and the rest were still running.
Now it checks for running compose containers first and stops them with
docker compose. If it cannot ask docker and finds nothing to kill on the
host, or anything survives the kill, it says so and exits non-zero
instead of claiming success. The bare-metal path is unchanged apart from
verifying the SIGKILL actually worked, and no longer risks killing the
shell it was launched from.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Dropping a sev1-2 care nudge while you're away is right and still happens.
But it was a bare `continue`: no row, no log, so "she dropped it", "the gate
suppressed it" and "the rule never fired" all looked identical afterwards.
Adds a 'dropped' delivery status (migration #12 widens the CHECK constraint;
sqlite can't do that in place, so the table is rebuilt) and records the drop
as one delivery_attempts row plus a log line.
No nudges row for a drop: that table feeds the ignored_rate signal, and a
nudge nobody could see must not count as ignored.
TestVoiceNoSessionFallthroughLeavesOutboxTrail expected exactly one row for
sev1-2 when voice had no session. It now expects the voice failure plus the
drop, which is the point of the change.
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 away sinks fell back to the whole Body when Summary was empty. ntfy and
telegram leave the box, and the 0.8B phraser drops fields regularly, so that
fallback could push full detail off the machine.
The dispatcher already strips detail from away sendables. This exports that
one rule as delivery.AwayMessage and has both sinks use it, so a sink can't
leak the body on its own either: empty Summary means a generic line plus the
rule name, never the body.
The two sink tests named TestSendFallsBackToBodyWhenSummaryEmpty asserted the
old, wrong behaviour, so they are rewritten to assert the generic line.
TestSendRejectsEmptyMessage is likewise replaced: an away message can no
longer be empty, so the sink has nothing left to reject.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The user only ever asks the time in his own zone, so answering other
cities was code kept in step with the weather city list for no gain.
Any named place now gets the honest "local time only" answer that was
already there for unknown cities.
Removes the 22-entry table, the lookup and the embedded tz database.
Closes Vikunja #389 — there is only one city list again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
replySystem had two arms that PR 30 made reachable, and both answered confidently wrong: the date arm keyword-matched "числ" and always answered today, so "какое число завтра" answered today; the clock arm ignored a named city and answered local time. The date arm now reads the day word through router.ParseCalendarDate (which grew послезавтра/вчера and now cuts the day boundary in the local zone instead of UTC). The clock arm answers the named zone when it resolves offline from the tz database embedded in the binary, and otherwise says plainly that she only knows local time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The prompt named seven intents but never said which one a clock or date
question belongs to, so the model guessed: system->query x4 in every eval
run. The rule now says the clock and the calendar date themselves are
system, what is written in the calendar or in memory stays query, and a
time named inside a request is just part of the request.
That split follows what the daemon can answer. Only replySystem owns the
clock and the date formatter, while the agenda is answered from
CalendarEvents inside the query branch.
Also adds one calendar-agenda fixture case so an over-broad system rule
cannot pass unnoticed, and writes up the before/after numbers. The
targeted confusion is gone; the headline accuracy did not move.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The embedder swap left every stored vector in the old model's space, so cosine against a new query vector is noise. Add the one-shot backfill: store.ReembedAll re-embeds every note and fact text with the currently configured embedder (the passage side, which is the side stored text was written with) and rewrites both places a vector lives — the notes table embedding column and the memory_vectors rows.
All of it plus the embedder marker happens in one transaction, so a failure partway changes nothing and writes no marker: re-run it. A run against a DB whose marker already names the current embedder does nothing.
Triggered explicitly with `mavend -reembed`, not automatically on mismatch: ONNX on the laptop CPU makes this minutes of work, and a silent multi-minute stall on boot would look like a hang. The mismatch warning now tells the user to run it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
kami
kami