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
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
15 root markdown files, ~4,900 lines against ~33,000 lines of Go, with at least
three pairs contradicting each other. When five documents describe the
architecture, the code becomes the only trustworthy one — which defeats the
point of having them. That drift is why the resident-model question had four
incompatible answers.
SPEC.md, maven.md and ROADMAP.md are deduped into DESIGN.md rather than
concatenated, with a "Superseded" section carrying eight retired decisions and
what replaced each: classifier-owns-the-route (the cascade is still the live
path, but as a stopgap, not a design to extend), faster-whisper/vosk/silero,
the small-model phrasing claim, sqlcipher, the Kotlin/Spring sketches,
obsidian->chroma, script deployment, and FloorEnrollment. Superseded material
is kept and marked rather than deleted, so it cannot read as current.
SESSION-05/06-07-2026.md and PLANS.md are removed outright — git history holds
them, and both were verified tracked before deletion.
Go doc comments citing the deleted files are repointed to the equivalent
DESIGN.md sections. Several asserted designs that were already retired, so the
claims are corrected and not just relinked: stt.go named faster-whisper as
production (it is whisper.cpp), tts.go named silero (it is piper), intent.go
still described the classifier as owning the route, and stale vosk/chroma
vocabulary is replaced. ECOSYSTEM-SPEC.md references are deliberately
untouched — that is a different document, and a naive grep for SPEC.md matches
it.
Root markdown drops from 4,880 to ~3,700 lines. The review's ~1,500 target is
not reachable while keeping the files it also said to keep — those alone are
2,553 lines — so trimming further needs a separate decision on
MAVEN_ECOSYSTEM_ARCHITECTURE.md and PROGRESS.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5JApcrCRVGmqrxnhynSik
Add a read-only /ecosystem page that consumes the sibling services'
JSON APIs (Nexus entities, Praxis attention, Hexis capabilities),
fetched concurrently with honest per-panel error states. Siblings stay
headless — mavweb is their human surface (arch §16). Wired via mavweb
-nexus/-praxis/-hexis flags; mavweb joins the ecosystem compose network.
Fix mobile horizontal overflow across all pages: .content is a flex
child with default min-width:auto, so it refused to shrink below the
tables' intrinsic width. min-width:0 lets wide tables pan inside .scroll
instead of dragging the page sideways. Verified via CDP geometry check
(scrollWidth === clientWidth at 430px).
Also includes in-progress Ethos UI redesign, ecosystem deploy compose,
and planning docs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Route contract is now a JSON array of action objects (one per ask) so
compound utterances route all their intents, not just the first. Grammar
root emits `[{intent...},...]`; parseActions tolerates a bare object.
Cascade still returns one Decision — full N-action dispatch lands with the
engine turn-on (marked in-code).
Router prompt rewritten shorter + decision-ordered (prompt-guy feedback),
fact redefined as "implicit update" not "trackable state", kept in Russian
to match the CPT base + phraser. "интент" → "намерение".
CLAUDE.md: routing-architecture section + refreshed open items.
docs/plans: route-data generation plan.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017GMrVfuYN3nE4L1vEiFYC9
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 LLMRouter: grammar-constrained LFM call for intent classification
after stage-0, before classifier cascade. Errors fall through gracefully.
- Add IntentChat: conversational intent with no store side-effect, routed
through LLM -> phraser chat endpoint.
- Extract slots for Chat: no structured slots, full utterance is payload.
- Extend stage-0 grammars to fire through Cyrillic wake-word spellings
(Мэйвен/Мейвен/Майвен/etc.) produced by Russian STT model.
- StripWakeToken helper strips leading wake in any script so time/date
grammars still match when wake is present.
- Add classifier examples for chat utterances (EN + RU).
- Wire LLMRouter into Router.Config; optional, nil-safe.
PythonDateParser shells out to python3 with the dateparser library for
full natural-language date/time extraction. Two-step approach:
1. search_dates() finds the date substring in surrounding text
2. parse() re-parses the substring for correct time resolution
Russian time qualifiers (утра/вечера/дня/ночи) are pre-processed to
AM/PM before parsing — dateparser drops them during substring extraction.
Falls back to StubDateTimeParser when python3 or dateparser isn't
available (graceful degradation, no hard runtime dependency).
Dockerfile updated: python3 + dateparser==1.4.1 in runtime stage.
five fixes spotted during routing investigation:
- gitignore: replace blanket models/ ignore with per-dir exceptions
(/models/embedder/, /models/stt/, /models/tts/) so the seed text
files under models/seeds/ are tracked in version control
- query.txt: fix merged line — 'сколько стоит свет в этом месяце' and
'найди заметку про сервер' were fused with no separator
- reminder.txt: add 11 pure-verb reminder seeds without time expressions
to shift centroid toward the reminding intent rather than time-lexicon
- StubDateTimeParser: add Russian 'через <N> <unit>'/'через час'/'через
полчаса', 'сегодня'/'завтра'/'послезавтра' with optional clock, and
'в <clock>' scan. Add Russian word numbers (один-десять) and unit
inflections (час/часа/часов, минута/минуты/минут, день/дня/дней,
неделя/недели/недель). Also adds missing English day/week units.
- replySystem: guard time branch against duration queries ('сколько
времени прошло') reaching it via the classifier path after the
stage-0 grammar's build filter rejects them. Mirrors stage0.go
duration keywords.
three bugs causing time queries to land on reminder or fact intent:
- seed collision: query.txt and system.txt shared identical time/date
seeds (который час, сколько времени), making system intent
indistinguishable from query intent in centroid space
- threshold (0.35) too low for ONNX embedder — cosine similarities
cluster 0.5-0.7 for related intents, so Clarify never fired
- reminder centroid contaminated by time-lexicon (every seed has a time
expression), pulling any time-word utterance toward reminder intent
fixes:
- remove 3 duplicate time/date seeds from query.txt (keep in system.txt)
- DefaultRouterThreshold 0.35 -> 0.55
- stage-0 grammar for напомни/remind me -> IntentReminder, bypasses
classifier (fixes 'напомни через час' being misrouted to fact)
- stage-0 grammars for time/date system queries (сколько времени,
который час, какой сегодня день) -> IntentSystem, with Build filter
to exclude elapsed/duration queries (сколько времени прошло)
- time parser fallback in applyAction for stage-0 reminder matches
(extractor doesn't run on stage-0 decisions)
- session.go: add History []Turn + Turn type for multi-turn context
- slots.go: add AnaphoraResolver with Resolve() for RU pronoun detection
(это/он/она/оно/тот/мой and inflected forms)
- followup.go: extend followUpMerge with cross-intent inheritance:
Query/Fact/Reminder after a Fact with anaphora inherits the key.
Same-intent path unchanged. Anaphora detection from utterance.
- voice.go: add fact-by-key lookup path in applyAction for IntentQuery
when dialogue resolved an anaphoric reference (calls LatestFact,
formats with formatTime helper). History tracked in Session.History
capped at 4 most recent turns.
- followup_test.go: 7 new test cases: anaphora query-after-fact,
no-inheritance-without-anaphora, three-turn break, anaphora in
reminder, anaphora in fact, explicit key wins, time inheritance.
make test green (303+, -race, all 29 packages).
- 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>
- Stub.PhraseQuery: empty notes → "не знаю." (was: "no notes")
- LLMPhraser.PhraseQuery: empty notes → general knowledge prompt to LLM
- Voice handler: on notes RAG failure, try phraser before giving up
- New router.KnowledgePrompt() pure function with test
Co-Authored-By: opencode <opencode@anthropic.com>
Store: CalendarEvents(ctx, from, to) — filters caldav facts by key date prefix.
IPC: full wiring through interface, server dispatch, and client.
Router: ParseCalendarDate (сегодня/завтра), CalendarEventFormatter (RU reply).
Voice: calendar detection before notes RAG in IntentQuery handler.
Tests: store integration test, date parser tests, formatter tests.
Co-Authored-By: opencode <opencode@anthropic.com>
kami