Spoken capture was dead. "добавь в задачи купить молоко" routed act, so the
gate found no allowlisted fn and asked "Что сделать?", and the list stayed
empty. Capture rides the note intent by design (#130, no eighth intent), and
nothing under actionNote was reached any more. The model also rewrote the
payload on the way — "купить молоко" came back as "сделать покупку молока",
and a task must read as the words he said.
TaskCaptureGrammar answers it at stage 0, the same place the agenda rules
went. It matches any utterance and lets ParseTaskCapture refuse, so the
marker list stays data. Three phrasings he used are added to that list:
"запиши в список дел" and the two next to it were missing.
The other deterministic matchers were checked for the same exposure. They
are all question-shaped — money, habit, feed, day plan, task list, calendar —
and a question lands on query, which is where they already sit. Capture was
the only imperative among them, which is why only it was taken.
ru-note-006 is the fixture case. The classifier alone cannot pass it, and the
hash baseline drops by that one case; the daemon answers it at stage 0.
"какие планы на сегодня" worked and "какие планы на завтра" answered
"пока не умею": the agenda rule needs "у меня" or a calendar noun, and
that phrasing carries neither. "когда планёрка?" had the same shape.
Two rules. One takes a plan noun aimed at a named day, one takes a closed
list of event nouns after "когда"/"во сколько". Both route intent only,
so the query chain still decides which source answers.
classifier+onnx over the fixture: 55/79, 69.6% full, with the two new
cases passing and no case moving the other way.
Slots.Text was the raw utterance for every intent, so a reminder could not
have an empty subject. StillMissing never reported SlotText, the question
"О чём напомнить?" was unaskable, and the branch in PendingQuestion.Answer
that fills a text slot could only overwrite the whole request.
The LLM path now keeps the model's own text, empty included, and the gate
turns a subjectless reminder into a question. The classifier path is
unchanged: it has no subject parser, so the utterance is the only signal it
has.
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.
#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>
"что у меня сегодня" and "что у меня в календаре сегодня" both routed
IntentSystem on the deployed daemon, and replySystem has no agenda arm,
so both answered "пока не умею". The calendar source that can answer
them lives in the query chain and was never reached. The fixture has
said query since ru-query-019 was written; the daemon disagreed with the
fixture and the daemon was wrong.
AgendaQueryGrammars routes them at stage 0, after the clock rules so
"какой сегодня день" keeps reaching replySystem. Intent only — which
source claims the turn stays the query chain's decision.
This is what made the follow-up continuation look like it only worked
for "what day is it". It did: the query half inherited an intent whose
handler could not answer, so both halves came back "пока не умею".
Measured on the 77-case RU fixture: full accuracy 70.1% → 72.7%,
intent-only 75.3% → 77.9%, calendar 0/2 → 2/2, clarify counts unchanged.
The eval harness wires the new grammars too, or the fixture would stop
being a measurement of the daemon.
Go's \b is ASCII-only and never fires after a Cyrillic letter, which the
first version of the pattern learned the hard way.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Deployed check, second round: "привет" after "какой сегодня день" answered
with the date. followUpMerge fills an empty Text from the previous
same-intent turn, so the topic-widening added a minute earlier was reading
an inherited topic on turns that had nothing to do with it.
Decision gains Continued, set only by continuation.go and never by the
router. replySystem widens on that and nothing else, so an inherited Text
is back to being invisible.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The clarify gate thinned any one-word utterance to 0.3 confidence, which
trips the stage-3 gate and comes back as "не совсем поняла". That is an
English intuition. Russian packs subject, tense and gender into one word,
so "поужинал" is a complete report and "привет" a complete greeting, and
both got clarified.
thinSingleToken keeps the rule for bare nominals, where it is real ("вода"
is a fact-or-query coin flip), and spares two classes: a closed lexicon of
social and control singles, and any token carrying a verb ending. Both
tests are offline.
Fixture: false clarifies 3 → 2, intent-only 74.0% → 75.3%, full accuracy
unchanged at 70.1%, missed clarify still 1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
"что нового?" is a greeting, and the feed matcher claimed it: "нового" was a
feed noun and "что" an ask. With no feeds block, which is what ships, the answer
to hello was "я пока не читаю ленты — они не настроены". A newness word now
needs a named topic or a real feed noun beside it. The topic prepositions lose
"о" for the same class of reason: one rune of filler produced a category of
whatever followed it, and then "по этой теме в лентах пока ничего".
An undated feed was re-noted in full on every boot. Dated items are deduped
against the durable mark, undated ones against a map that dies with the process,
so five items became five more on the next start, stamped now, at the top of the
recent-notes window. A crash loop made that a flood. The mark is now set for an
undated feed too, and its existence marks the first poll after a restart as a
resync: those items are recorded as seen rather than written.
A burst larger than max_items lost its middle. The poll walked the feed
newest-first, stopped at the cap, and marked the newest item written, which put
everything below the cap behind the mark forever. The cap now applies to the
oldest candidates and the mark follows what was written, so max_items paces
instead of dropping.
The category tag was read out loud: "Заголовок [технологии]" went through piper
brackets and all, because the answer path took the whole first line. The tag is
parsed off for reading and is now the only thing a topic is matched against.
Matching the whole note meant "что нового про погоду" hit any tech headline
whose link contained "pogod".
Also: the charset comment on dec.Strict described something Strict does not do,
and a skipped feed is named in the log.
Found in review of #66.
The day total rolls over at midnight and the poller had nothing to write until
the first spend of the new day, so at 09:00 the latest money_today fact was
yesterday's spending and looked perfectly fresh. The value now carries the
first instant of the window it covers, and a today question that the stored
window does not cover is refused rather than answered with yesterday's number.
Staleness was measured off the fact timestamp, which only moved when the figure
moved, so a quiet month was reported as data from three days ago while being
current. The value now carries when it was last read and the poller writes on
every read.
Amounts in an instrument the window diff never named were spoken with a numeric
instrument id as the currency. Instruments are resolved from one cursor-zero
diff, cached for the process, and an amount still unnamed is dropped from
speech rather than recited wrongly. "сколько я потратил вчера" was answered
with the month total, a real number to a different question, and is now
refused by naming the two windows she keeps. Income questions led with the
spending.
Found in review of #62.
A task extracted from mail deduped on the live-norm index only, so once he
finished it the row left the live set and the next poll of the same immutable
message re-extracted it as a fresh candidate. mavmaild is a read-only reader
and marks nothing read, so that repeats forever. Derived rows now carry an
ext_id built from the message uid and the extracted span, unique across every
status, while voice keeps live-only norm dedupe because saying an errand again
is the recurrence signal. A derived source can no longer capture straight to
open, and saying a task out loud that Maven had only proposed promotes the
candidate instead of answering that it is already in the list.
SetTaskStatus was classified AuthRead. Resolving a task is not additive, it
erases work off his list, so it is a write, and the row now records the caller
that moved it. ListTasks was unbounded. The list-query matcher claimed any
utterance with "что мне делать", including "с чем мне помочь", and the urgency
stripper matched inside words.
Found in review of #60.
IsDayPlanQuery only rejected the сегодня family, so "какие планы на
понедельник?" carried no other-day token, did carry "планы", and the
plan claimed it ahead of the calendar listing and recited today under
today's date. Weekday names, week, weekend and month join the refusal
list. This is a refusal and not a feature: it stands until the plan can
build a day other than the clock's own.
isRestOfDayQuery also lived in cmd/mavend and matched by substring while
IsDayPlanQuery tokenized, so the two predicates deciding one utterance
could disagree, and "проверь nextcloud" read as a request for the rest
of the day. It moves to the router and tokenizes.
Found in review of #58.
IsDayPlanQuery fires on the token "планы" and its other-day list does not know
weekday names, so "какие у меня обычно планы по вторникам?" was claimed by the
day plan, which answered today's calendar stamped with today's date. The habit
source never ran. The matcher now declines any utterance ParseHabitQuery
claims, which keeps the decision out of the source table's ordering.
Two gaps in the same matcher. Sunday had only its dative plural listed, so "в
воскресенье" found no weekday. "по выходным" named days that no weekday word
matches, so it was answered with the whole-week profile. Both are recognised
now, and the weekend is read back as two days rather than pooled.
Found in review of #59.
Seven fixes, each answering a line comment on the stack.
**Weather no longer invents Moscow** (PR 50). extractWeatherLocation returned
the string "Moscow" when he named no city and voice.weather.default_location
was unset — a made-up answer presented as fact, which is the one thing maven
must never do. It returns "" now and the query path says it does not know.
**Digest statuses are a defined type** (PR 50). DigestStatus string plus the
three constants, so a rule name cannot reach the status column.
**Quiet-mode negation is not adjacency** (PR 53). The OFF list carried
{"не","тих"}, an adjacency pattern, so "не надо тихий режим" missed OFF, hit
the ON pattern {"тих","режим"}, and asking for quiet mode to stop turned it
on. Negators are scanned over the whole utterance now, with the two ON phrases
that are themselves built on "не" excluded. "тихий режим выключи" works too,
which it did not before.
**Pattern stability uses a median band** (PR 54). max/min over the extremes
asked whether every gap resembles every other gap, so 7,7,7,7,20 — four clean
weeks and one holiday — was thrown away at a ratio of 2.9. Each interval is
now tested against the median and 70% must be in band, and the reported
interval is the median of the in-band ones, so a holiday no longer drags a
weekly habit to "every 9.6 days". The reviewer's 5,8,10,3 is still rejected.
**The weekday profile stops reciting everyday habits** (PR 59). "What do I do
on Saturdays?" answered "you drink water" — true, and useless, because it is
equally true of every other day. Activities that are habits on six or more
weekdays move to Profile.Everyday and are read back as daily habits instead of
as an answer about that day.
**Russian phrase tables move out of Go** (PR 59, PR 61). The behaviour glosses
and weekday names, and the task capture/urgency/list vocabulary, are now
behavior_ru.json and task_phrases.json, embedded with go:embed. Single-binary
deploy is unchanged; wording edits are no longer source diffs.
**nginx template stops taking nginx down** (PR 52). Two host-side failure
modes, both plausible causes of today's crash. The $connection_upgrade map is
fatal when duplicated, so it moved to its own nginx-upgrade-map.conf with a
grep-first note. And `listen 10.42.0.1:80` fails with EADDRNOTAVAIL when wg0
is not up yet, so nginx exits on a reboot that beats WireGuard — the header
now documents net.ipv4.ip_nonlocal_bind.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The network fallback behind the local sources, off unless configured.
internal/crawl is pure: a stdlib robots.txt parser (group specificity,
wildcards, Crawl-delay, cached per host), HTML-to-plaintext extraction, and a
watcher that notes a watched page only when its text changed. It has no store
access and no net/http; cmd/mavend/crawls.go is the impure half.
Every limit is code and tested: the guarded fetcher from #258 enforces the host
allowlist/denylist, refuses private addresses in the dialer Control hook (so DNS
rebinding and each redirect hop are covered), caps size and redirects, times out,
and spaces requests per host. A robots.txt Disallow is refused with no override.
On demand, reading is a query source placed last in the chain, after his memory,
his notes, and the local Kiwix ZIMs once those are wired: no URL in the
utterance means no fetch, and only the URL ever leaves the box. Scheduled
watches write notes and announce nothing.
The vendored tree has no x/net/html, goquery or temoto/robotstxt, so the parsers
are stdlib. No new dependency.
internal/rss parses RSS 2.0 and Atom, and polls each configured feed on its own
interval; internal/webfetch is the one door either of them uses to touch the
network. The poller writes items as notes with source "rss:<feed>" and nothing
else: the answer path reads them back when he asks "что нового в лентах?", and
nothing is announced on arrival. A feed that dispatched would be a nag, which is
why the plan's breaking-news rule was left out rather than built.
webfetch is where the limits live, as code rather than a paragraph: http(s)
only, an allowlist (the configured feeds' hosts) and a denylist, a 2 MiB body
cap, a 3-redirect cap, one request per host per second, and a refusal to connect
to any private address — checked in the dialer's Control hook so it holds for
every resolved address and every redirect hop, not just for a literal IP.
Off unless configured: no "feeds" block, no poller, no outbound request. How far
a feed was read is a config fact (rss:latest:<name>), so a restart does not
re-note yesterday's headlines.
The trust boundary is zenmoney, not maven — they already hold his bank
sessions. So the poller reads /v8/diff/ and writes totals as
facts(kind=env, source=poll:zenmoney); core reads those back when he asks
and never sees the token.
internal/zenmoney sums transactions per currency over a window, skipping
tombstoned rows and transfers between his own accounts, and refuses to
encode a summary built from zero transactions. That refusal is the whole
design: a failed or empty read writes nothing and leaves the last good
total alone, because a zero recited as fact is worse than silence. No
currency conversion either — a figure he can check against his bank beats
one he cannot.
Off unless configured, and the token is read from a FILE rather than a
flag so it never lands in `ps`, in docker-compose.yml, or in shell
history. Nothing about the money is search input, no tick rule reads the
keys, and the log lines name keys, never figures.
The live-credential half is BLOCKED: there is no zenmoney account or token
here, so everything is verified against a recorded diff fixture.
Ordering is computed, not generated. Asking a 1.7B which of his tasks
matters most produces a fluent opinion with no basis in anything, and a
confidently wrong priority is worse than none — same posture as the
behaviour profile in internal/memory, which counts instead of summarising.
internal/tasks is a pure package (no ipc, no store, no cgo) holding the
score, the order and the Russian rendering, so the spoken list and the
/tasks page cannot drift. Four signals, all of them things he stated:
deadline (overdue > today > tomorrow > this week), stated urgency, age
with a cap so nothing rots at the bottom, and confirmed work always
ahead of mail-derived candidates. A task with no due date and no weight
scores nothing and carries no reason string — inventing a "потому что"
about a priority he never set is the failure mode this avoids.
Capture now picks up urgency he says out loud ("добавь в задачи срочно
оплатить интернет"), stripping the marker from the task text, and the web
add form offers the same three rungs. Ranking is a read: it sorts and
renders, never writes, schedules or announces.
A task is not a fact and not a note. A fact is a claim about the world that a
correction supersedes; a note is something to recall by meaning. A task is work
with a lifecycle, and the read that matters is "everything outstanding right
now" — which over an append-only log would mean replaying history on every
question. So: a tasks table, migration #14, statuses candidate/open/done/dropped
that each move forward exactly once.
Dedupe is on normalised text among LIVE rows only, via a partial unique index.
That is the property the mail side needs: an extractor may call CaptureTask for
every message it reads, as often as it likes, without growing the list — while a
weekly errand is still capturable again once the last one is done.
Three ways in, one seam. ipc.CaptureTaskReq is it: the voice path
(router.ParseTaskCapture on an explicit marker — "добавь в задачи …", never
"надо бы поспать"), the /tasks form, and the email reader from #246 when it
exists. Mail-derived items set Source "email:<account>", Status "candidate" and
Evidence to whatever makes the row reviewable; a candidate is inert until he
confirms it on /tasks, and Maven names it as unconfirmed when she recites the
list rather than putting words in his mouth.
No new intent — the router enum is a contract with the relabelling prompt, so
capture rides the note intent and the list rides a query source, both matched
deterministically like the calendar and plan matchers already are.
Nothing here speaks. No tick rule reads tasks; the list is answered when asked
about, which is why /tasks POST is not step-up gated the way /tools and
/routines are — a task write moves no boundary.
Vikunja #130
Behavioural memory, narrowed on purpose. internal/memory/behavior.go builds a
profile out of self-facts — distinct days per weekday, median time of day — and
reads it back in RU; router.ParseHabitQuery finds the weekday deterministically;
a `habits` query source answers the question.
Three things the plan doc asks for are deliberately absent, and the doc now
records why:
- The profile is COUNTED, not LLM-generated. A 1.7B asked to summarise a year of
habits writes fluent claims about the owner's life that no row supports, and a
wrong claim about him is the most expensive kind of wrong maven can be.
- No cached profile fact, so no "update on fact write" machinery. It is
recomputed on the question; a cache that can disagree with its own rows is two
truths.
- No proactive daily plan nudge. A dispatcher proposal at 08:00 every day is the
definition of a nag. The path from "she noticed a pattern" to "she acts on it"
already exists in internal/pattern with the proposal queue on /routines, and it
goes through him.
A one-off is not a habit: an activity needs two distinct days before she will
call it usual, and until then she says she does not know yet. Only self-facts
count — env rows are the world, config rows are her own tuning state. The typical
time is a median so one 03:00 outlier cannot move a morning habit into the night.
An unrecognised fact key is read back verbatim rather than glossed into something
she made up.
The source sits before "calendar" in querySources, and its matcher requires a
habit marker, so "что я делаю в среду?" still reaches the calendar — answering a
question about this coming Wednesday with a statistical average would be
answering a different question.
Verified: make build and make test both exit 0.
The plan answers "какие планы на сегодня?" by putting one day in order:
calendar events (with #126's ambient provenance carried through and hedged),
pending reminders, and one line per morning routine that still has items
outstanding. "что дальше?" trims what has already passed.
It lives in internal/morning, not in a parallel system, because it is the same
question the checklist asks at a different scale — the routine knows what is
missing from a window, the plan knows what the whole day holds, and both read
the same facts and the same idea of "today". BuildPlan is pure; tickLoop.dayPlan
is the impure half that reads the store.
It is not a nag. Nothing here fires, schedules or announces: the plan is built
only when asked, over IPC (day_plan) or on the existing /morning page.
Unprompted delivery stays with the morning nudge and the dispatcher's policy.
The query source sits before "calendar" in querySources because both match
"…на сегодня" and the plan's matcher is the more specific one; IsDayPlanQuery
matches whole words so "планёрка" (a meeting) is not read as a request for the
plan, and refuses any utterance naming another day, since the plan is built for
the clock's own day only.
Verified: make build and make test both exit 0; new tests cover plan ordering,
the checklist-only-what-is-left rule, other-day rejection, the RU rendering
against the persona checks, rest-of-day trimming, the source ordering, and the
matcher's refusals.
Maven does not get a work credential. A corp mail or calendar session living on
the homelab ties the box's blast radius to the employer's data, which is the
thing this task exists to refuse. What she reads instead is the signal: an
Android notification-listener on the phone relays meeting notifications over
wg/LAN to POST /api/ambient, and the ones that clearly describe a meeting become
calendar events at source=ambient:notif, confidence 0.6.
The provenance is the point. A notification is evidence about a meeting, not a
reading of a calendar, so it is never indistinguishable from one: it is stored
below full confidence, store.CalendarEvents keeps the source and confidence on
every row it returns, and the query path hedges — "похоже, Планёрка @ 14:00" for
a relayed event, plain text for a CalDAV read.
The parse is deliberately conservative (internal/calendar/ambient.go). It needs
a real clock reading and a summary that is not just that clock reading;
otherwise it stores nothing at all. A bare hour is not a time, an unread count
is not a time, and "срок 2026.08.15" does not offer 08:15 as a meeting — loose
digits in a notification are far more often a badge or a date, and a mailbox of
noise rendered as invented meetings is worse than a gap.
The ingest is off unless configured: no -ambient-token, no route registered. The
token is a shared secret compared in constant time, because the poster is a
background Android service and WebAuthn has no answer for one. The endpoint is
write-only, accepts one shape of write, and cannot read anything back out.
Reposts of the same notification dedupe against the latest fact for that
key+source, the same append-only discipline cmd/mavcaldav follows.
Not shipped: the Android relay app itself, which is a separate artifact and a
device, not Go in this repo.
Confidence was hardcoded to 1.0 for every LLM decision, and the LLM branch
in Router.Route returned straight from fillSlots without ever touching the
stage-3 threshold gate — so the LLM path could not produce a Clarify no
matter what confidence a model reported. That is why all 6 want_clarify
cases in the 77-case RU fixture were missed by every model in the bake-off.
Fix reads structural signal instead of changing the (parity-locked) router
prompt: a single-token utterance ("вода", "бэкап") is flagged thin evidence
in llmrouter.go; a fact left keyless or an act that never resolves to an
allowlisted fn, checked after fillSlots so the deterministic parsers get
first crack, is flagged in router.go's new gateLLMDecision. Anything below
config.DefaultRouterThreshold (0.55) now sets Clarify=true through the same
path the classifier already uses.
Added unit tests with a stubbed Completer proving both directions: thin
cases clarify, clean multi-word/resolved-slot cases stay confident. The
77-case fixture re-run against a live llama-server is still needed to
confirm the 6/6 moves — not done here, no llama-server on this box.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
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 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
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 moved from paraphrase-multilingual-MiniLM-L12-v2 to
multilingual-e5-small. Both are 384-dimensional, so nothing in the code
noticed: cosine between an old stored vector and a new query vector is
noise, and recall degrades silently.
So the DB now records the embedder that wrote its vectors. One value for
the whole DB (migration #11, a small `meta` key/value table) rather than a
column on every vector row: the backfill re-embeds every note and fact in
one pass, so a per-row marker would hold the same string in every row and
cost a column on two tables for nothing.
The identity comes from the embedder itself via a new optional ID() method
("multilingual-e5-small@384", model file name plus dimension), so pointing
the config at another model changes the string without anyone editing a
constant. mavend logs a loud WARNING at startup naming both the stored and
the configured embedder when they differ.
Detection only — recall behaviour is unchanged. TODO(#378) in
store.CheckEmbedder marks where the backfill will hook in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The 67.1% "thinking off" column in ROUTING-EVAL-31-07-2026.md was an
artefact. It came from a hand-rolled HTTP client in the eval test that
did not send repeat_penalty, so it differed from the reference run on two
axes and the penalty was the one that mattered.
Re-scored back to back on an idle box with everything else held equal:
thinking off is identical to thinking on, case for case, same confusion
matrix, same three unparseable replies. A direct probe of the running
llama-server shows enable_thinking, thinking and reasoning_budget are all
ignored for this model on this build, so there was nothing to turn off.
No defaults changed. The misleading third configuration is removed from
internal/router/eval so its table cannot be quoted again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
An LLM-routed reminder came back with no parsed time and an act with no
fn, because only the classifier path ran the extractor. Now the router
runs the same extraction after an LLM decision and fills only the empty
slots. No time in the utterance still means no time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The old model was a symmetric paraphrase model, so it scored "do these
look alike" instead of "does this note answer this question". Also fixes
the file mismatch: the Makefile, the deploy config and both evals now all
name the same quantized file, and the quantized one is what gets measured.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Note recall is asymmetric: a short question goes in, a longer note comes
out. Adds EmbedQuery/EmbedPassage helpers and the e5 prefixes, and points
the note/fact write path at the passage side and the query path at the
query side. Reviewers: the three call sites in voice.go.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Chose an 8th enum value over a confidence number: the model already picks
one enum token, so it costs nothing in the grammar, while a score from a
0.8B model would be uncalibrated noise. A refusal returns "no decision"
with no error, which is the fall-through the caller already uses for a
bad parse, so the classifier and its clarify gate take the turn.
Reviewers: the prompt's counter-examples matter most — a small model will
over-use any easy escape hatch. The training workspace copy of the prompt
still needs the same edit (Vikunja #362).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The bake-off in #278/#250 needs two models' scores side by side, and the
report names only carried the config, so the rows were indistinguishable.
ModelID reads /v1/models instead of taking a string that goes stale.
New target: make eval-models MAVEN_LLM_URL=...
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Route now sets RepeatPenalty on the request, and the grammar's string rule is
capped at 120 characters. Two of 76 fixture cases looped one sentence inside
the text field until MaxTokens, which cut the JSON in half.
Reviewers: the new constant and the grammar string rule.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The router prompt tested "reports current state -> fact" before "wants
information -> query", so a question naming a fact key was written as a fact.
Query now comes first, plus an explicit question test.
Reviewers: the prompt block in llmrouter.go, and the note about the
training-side copy of the prompt that needs the same edit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Completes #319's comparison. Three configurations, because "the LLM router"
was ambiguous: the model alone, the cascade #320 would actually ship (stage-0
grammar → model → classifier floor), and a thinking-off diagnostic.
intent-only full RU missed-clarify p50
classifier+onnx 36.8% 36.8% 25/61 5/6 31ms
llm-only (0.8B) 48.7% 23.7% 13/61 6/6 850ms
cascade+llm (0.8B) 50.0% 32.9% 18/61 6/6 825ms
On the question asked — does the resident model route better? — yes, 50.0%
vs 36.8% intent accuracy. REARCH.md's premise holds. It costs 27x the
latency (p50 825ms vs 31ms, max 3.1s), on the same llama-server the phraser
needs, so it is a trade rather than a free win.
Three things the numbers surface that the headline hides:
query→fact x15 is the dominant failure, four times the classifier's x4 on
the same axis. routeSystem's decision order puts "сообщает или обновляет
состояние" (rule 3) above "хочет получить информацию" (rule 4), so any
utterance naming a fact key matches the earlier rule and a question about
past state reads as an assertion of it. A prompt fix, not a model limit.
The LLM router cannot clarify: llmrouter.go hardcodes Confidence 1.0, so
stage 3's gate can never fire on its decisions — 6/6 missed. With #359's
finding that the classifier's gate is miscalibrated under ONNX, neither path
currently refuses. Flipping #320 as-is removes the refusal lane.
The gap between 50.0% intent and 32.9% full accuracy is entirely slots: the
LLM path fills neither Fn nor Time (it returns Slots.Text for acts, and
Extract never runs on an LLM decision).
Also settles a hypothesis rather than leaving it in the air: thinking mode is
a non-issue under a grammar (identical score), and the grammar's unbounded
("," ws action)* repetition that ran away in an isolated smoke test does not
reproduce under the real prompt — 2 errors in 76, not 76. internal/llm
deliberately does not grow a chat_template_kwargs field.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5JApcrCRVGmqrxnhynSik
The onnxruntime .so was already vendored at deps/onnxruntime-linux-x64-1.26.0
— nothing to download. make eval-router now defaults MAVEN_ONNX_LIB there, so
both baselines run by default and only a fresh clone without deps/ falls back
to the hash ratchet alone.
Prod-representative result, deployed 0.55 gate: 28/76 (36.8%), RU 25/61,
EN 3/15, hard 0/11 → 4/11, p50 31ms / p95 71ms. Versus the hash floor's
13/76 at p50 9µs.
The finding is not the accuracy, it's the refusal lane: missed clarifies went
0 → 5 of 6. Better embeddings raise cosine everywhere, so the 0.55 threshold
that used to hold ambiguous utterances back stops holding — "сделай это"
routes to act at 0.847, "бэкап" to chat at 0.755. The gate was implicitly
tuned to the hash floor's low similarities. That is an argument about the
threshold, not about the embedder, and it lands before #320 rather than after.
Also fixes a fixture-model mismatch: ReminderGrammar deliberately skips the
extractor at stage 0 and the daemon's applyAction parses the time downstream
(stage0.go says so). Charging the router for that slot made 4 exact-match wins
read as misses; they are now counted as SlotsDeferred instead. Hash baseline
moves 13/76, ratchet to 0.15.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5JApcrCRVGmqrxnhynSik
#319 asks for a measurement before #320 flips the route decider from the
classifier cascade to the resident model. There was nothing to measure
against: the only routing tests assert single utterances, and the
classifier's seed corpus is its own training set — scoring it there
measures memorisation of frozen centroids, which is the illusion that hid
the weak RU query handling in the first place.
internal/router/eval is a separate package so both paths can be scored
from outside router (including cmd/mavend, where the real llama-server
client lives). The fixture is embedded; the scorer takes a Router
interface, so *router.Router and a bare LLM stage both go through the same
76 cases.
The fixture is a CONTRACT, not a snapshot: cases the cascade fails today
stay in the file and fail loudly. TestFixtureIsHeldOut enforces that no
utterance appears verbatim in models/seeds/*.txt.
Baseline, hash embedder at the deployed 0.55 gate: 9/76 (11.8%), 63 false
clarifies, 0 missed clarifies, p50 9µs. Almost everything falls to the
confidence gate — the documented floor behaviour, not a new bug. The
number worth comparing is TestONNXBaseline's (skipped without
MAVEN_ONNX_LIB); the assertions here are a regression ratchet plus a tight
bound on the dangerous direction: ambiguous utterances must not start
being routed confidently.
Seeding is order-fixed on purpose — a few phrases appear under two intents
and map iteration handed them to a different centroid each run, which made
the score jitter between 9 and 10.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5JApcrCRVGmqrxnhynSik
claude
kami
kami