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
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 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
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
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
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 memory pass ran only after the notes-only gate had already rejected
the same note at the same score. Notes and facts share one vector index,
so a note that failed there failed again — the branch could only ever
return a fact.
Now the memory pass runs first: one search over everything Maven
remembers, one gate, and the memory that clearly matches best answers
(a note gets phrased, a fact is read back). The notes-only pass stays
behind it for notes the vector index does not hold. No threshold moved,
so the set of questions answered is unchanged — only which memory
answers them.
Fixture gained two mixed note+fact cases, so the answerable count goes
25 -> 27: hash recall@1 36.0% -> 37.0% (ratchet 0.32 unchanged, comment
updated), e5 recall@1 72.0% -> 70.4%, false recall still 1/5.
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
Vikunja #363. The follow-up session was a plain in-memory map, so any
mavend restart dropped the thread. It now mirrors to a small TTL-pruned
sqlite table and is loaded on startup; expired sessions are deleted on
load, not revived. Clarify's pending question is untouched.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
MaxAttempts was 1, justified as "not a nag". Wrong reading: "not a nag" is about
interrupting unprompted, and a clarifying question is part of a conversation he
started. Now three, configurable via voice.clarify_max_attempts (default 3).
Three, because after that the likely problem is she misheard the whole request,
not one slot.
Answer used to keep the parked value, so "в три" then "нет, в пять" threw the
five away. Now a value the answer carries wins for the slot she asked about.
Only for the clarify answer — a correction in a fresh turn is followUpMerge.
The eight-field chained assertion in the Answer test is one DeepEqual now, so a
new field in Slots is covered without touching the test.
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
The e5 embedder puts every cosine in one narrow band (0.79-0.89), so the
absolute query_min_score gate cannot tell a real hit from a made-up
question: any value under the band answers everything, any value above it
answers nothing. False recall was 5/5.
New gate asks whether one note is clearly the best instead: top1 - top2 >
delta. New query_min_margin config knob, default 0.008, read off the sweep
in the recall harness. The absolute floor stays as a second check.
On the recall fixture with e5: answered 72% -> 68%, false recall 5/5 -> 1/5.
Vikunja #359
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The two things that made this unsafe are fixed: the router can now
refuse, and slot extraction runs on its decisions.
On the held-out fixture it gets 63.2% of intents right against the
classifier's 50.0%, with no route errors. It costs about a second a
turn instead of 30ms.
The flag is a pointer now, so leaving it out of the config means on
and only writing false turns it off.
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
Two merge fixes on top of the branch:
- migrations: keep both new steps, snooze stays #8, the routine columns
become #9. Both agents had numbered theirs #8.
- accepting no longer takes a reminder id, on the web surface too. The
web accept path had the same one-shot-reminder bug the voice path did,
so both now just flip the status and let the tick loop schedule.
The test that asserted "accept creates a reminder and links it" asserted
the bug. It now asserts that accepting creates no reminder.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The tick loop now reads accepted routines from the store and nudges when
their interval has passed; accepting no longer builds a one-shot reminder.
Look at routine.DueAccepted for the schedule rule (no catch-up backlog) and
at fireAcceptedRoutines for the restraint gate — routines do not bypass it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Two agents wrote the same three test helpers and names for the same two
bugs. Kept the real assertions from dispatcher_test.go and removed the
skipped placeholders they replace. panicSink stays in durability_test.go
since both files use it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
kami
kami