Continues the decomposition PR #50 started. voice.go 542 -> 365:
quiet_toggle.go 144 resolveQuietToggle, quietInflections, quietStem,
quietTokens, quietPhrase, quietOn/OffPhrases,
classifyQuietToggle (quiet_toggle_test.go already
existed for these)
patterns.go +44 detectPattern, next to detectAndPropose which it calls
and which patterns.go's own header already pointed at
What is left in voice.go is the handler: reactiveHandler, HandlePushToTalk,
handleText, runTurn, applyAction, replySystem, chatHistory, reply.
Move-only: all 133 distinct non-blank lines removed from voice.go were
matched in the two destination files, zero lines added to voice.go. The only
non-move edits are import lists (log added to patterns.go, unicode and
internal/pattern dropped from voice.go) and two comments that pointed at
voice.go for code that is no longer there.
The template's wildcard `listen 80` with no ACL was fixed in 50cc17f, but it
still only covered nexus/praxis/hexis. mavweb — the one service in the set
that serves an RCE surface (POST /tools defines argv internal/tool executes)
— had no block at all, so anyone wiring it up wrote their own, which is how
the wildcard got there the first time.
Adds a maven.kvmx.ru server with the same wg+LAN bind and allow/deny,
proxying 127.0.0.1:9201, with the WebSocket upgrade /ws needs, a 32m body
limit for push-to-talk PCM, and a 300s read timeout because an LLM turn on
the iGPU is slow.
Also records in deploy/ecosystem/docker-compose.yml that the sibling
`build:` paths pin nothing and ship the sibling working tree, with the
command to check what is about to be deployed. The stale public DNS records
(item 2) are outside the repo.
Verified: nginx -t on the template inside a minimal http{} accepts it.
/api/chat reaches the router, the LLM and, through applyAction, the whole
act path, so it is the widest state-changing surface mavweb serves. It was
the only one with no gate. It now goes through stepUpOK like POST /tools,
POST /routines and POST /api/revert: unchanged in the default deploy
(WebAuthn unconfigured, fail-open behind wg+nginx), 403 under
-require-stepup or an unasserted passkey session.
The route table now carries an explicit enumeration of every state-changing
route and its gate, and the two startup SECURITY log lines name /routines
and /api/chat alongside /tools and /api/revert.
The loopback -addr default the task also asked for landed earlier in
d12de58; the compose already publishes mavweb on 127.0.0.1 only.
Pure move: actionFact, actionReminder, actionAct and actionNote each get
their own actions_<intent>.go. The two small ones (chat, system) and the
actionHandlers table stay in actions.go, which is now just the dispatch
layer and the notes about what does not belong in it. No behaviour
change — only the file a handler is read in.
The six answer sources were hand-unrolled inside one 127-line function.
The intent table is a closed set of 7, but this list is open-ended —
Kiwix (#286), RSS (#258), the crawler (#259) and email (#246) each add
one. Each is now a registry entry: a name plus a method on the handler,
walked in order until one claims the question.
Order is unchanged and still load-bearing (memory before the notes-only
pass, #373), the confidence gate keeps its position and semantics, and
every reply string, log line and best-effort failure is verbatim.
HandlePushToTalk and handleText hand-wrote the same eight-step turn
sequence twice, comments in the latter saying "same as HandlePushToTalk"
four times. Extract it into runTurn(ctx, text) string: the voice path
wraps it in stt/tts, the text path returns it directly.
The two had drifted. The text path was missing the quiet-hours toggle
check entirely, so "тихий режим" over IPC/telegram fell through to the
classifier; unifying gives it the check. It also logged the route result
and applyAction return where the voice path did not — both logs are kept
for both paths.
voice.go is still the biggest file in cmd/mavend and most of what is left
has nothing to do with the audio path. The ecosystem integration is one
such lump: it talks to Nexus, Praxis and Hexis over HTTP and only touches
the handler for its store and clock. Lifting it into ecosystem_acts.go
puts it next to ecosystem.go, where the clients it drives already live.
Move-only: handlePraxisAct, recordPraxisTrace (called from nowhere else),
handleHexisAct and execHexis verbatim, plus the two imports that became
unused in voice.go.
applyAction (cmd/mavend/voice.go) dispatched all 7 intents from one giant
switch. Extract each case body verbatim into its own actionXxx method in
new cmd/mavend/actions.go, dispatched from an actionHandlers table keyed by
router.Intent. applyAction itself is now just the dec.Clarify guard plus a
table lookup.
No behaviour change: same reply strings, same side-effect order, comments
moved verbatim. The destructive-act confirm gate and the enabled-tool
allowlist stay entirely inside actionAct, exactly where they lived in the
old switch's IntentAct case — they're act-specific, not cross-cutting, so
they don't move to a separate layer. dec.Clarify short-circuit, dialogue
bookkeeping and detectPattern stay outside the table since they run
regardless of intent.
voice.go: 1638 -> 1344 lines. New actions.go: 362 lines.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
dispatch() replaces the hand-written switch with a package-level
map[Method]handlerFunc built once at init. Each entry is one
withParams/withParamsVoid/withoutParams call closing only over the
CoreAPI method it invokes — adding a method is now one table line
instead of a new arm.
Check still runs once at the top before any unmarshal, unchanged. The
three non-CoreAPI methods (assert_stepup, store_encryption_key, unlock)
are special-cased before the table lookup since they drive Server
fields (StepUp/WrapKeyFn/UnlockFn), not store state. The current
CoreAPI is loaded once per dispatch and passed into the handler as an
argument, so SetAPI's runtime swap (the unlock transition) still takes
effect on the next request — the table itself never captures an api
value. No wire-format change; existing round-trip and unknown-method
tests in ipc_test.go pass unmodified.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Vikunja #281. The interruption policy promised four outcomes — deliver_now,
queue, digest, drop — but only three existed: a care candidate the restraint
gate suppressed for quiet hours / away / calendar-busy simply vanished in
loop.Tick's `continue`, with only the trace remembering why.
internal/morning turned out not to be the natural drain: it's a fixed
Item/FactKey checklist engine, not a generic message bundler, so gate-
suppressed nudge text has nowhere to plug into its evidence model. Built a
parallel (but small, reusing the outbox's shape) durable digest instead:
- internal/store: digest_entries table + EnqueueDigestEntry (dedupes by
rule+body, mirroring the delivery outbox's bodyHash), PendingDigestEntries,
ExpireStaleDigestEntries, DrainDigestEntries (mark, never delete — an
audit trail of what she actually said).
- internal/loop: DigestEligible(severity, blockedBy) is the pure boundary —
only genuine restraint blocks (quiet_hours/calendar_busy/presence) even
qualify (cooldown/snooze are not "suppression"); within care, Sev2 (break)
digests, Sev1 (water/meal — stale by the time anyone could resurface them)
drops. High severity never digests; alarms bypass the gate and deliver
unchanged, on purpose.
- cmd/mavend/tick.go: each tick scans ExplainTick's trace for eligible
blocked candidates, enqueues them, sweeps stale entries (24h expiry — the
care rules are daily-cadence, so anything older is describing a day
that's over), and drains the bundle only once the suppression reason has
actually cleared, capped at 3 spoken items plus a trailing count so a
digest can't turn into the exact nagging it was built to avoid.
Tests: store-level round-trip/restart-survival/dedupe/expiry/drain, loop-
level severity-boundary unit tests, and tick-level integration tests for
the drain-only-when-clear and never-digest-high-severity behavior.
detectPattern only ever fired as a side effect of a voice fact-write, so a
recurring pattern already sitting in history went unnoticed until he
happened to mention it again by voice — the opposite of proactive.
Split the pipeline: extraction (fact -> normalized event) stays where a fact
is written, in voice.go, since it's tied to that write regardless of who's
talking. Detection (events -> stable pattern -> proposed_routines row) moves
into shared code (patterns.go's detectAndPropose) that both the voice path
and the new tick.go:detectPatterns call. The tick runs it every cycle over
every action+object pair on record (store.DistinctEventPairs, added), so a
pattern gets noticed on the daemon's own schedule.
Idempotence and the dismiss-must-stick requirement turned out to already be
handled by the store, not something the tick needs to reinvent:
proposed_routines has UNIQUE(action, object) and CreateProposedRoutine does
ON CONFLICT DO NOTHING, and DismissProposedRoutine flips status in place
without deleting the row. So a pair already proposed, accepted, OR
dismissed is a silent no-op on every later tick — a dismissed pattern can
never resurface, and re-running the scan never spams the /routines page.
Kept the voice-path call (immediate spoken confirmation is a nice feature
UX-wise and is now redundant-but-harmless with the tick, since both paths
share the same guarded detectAndPropose).
Tick-side detection only ever writes a row; it does not notify, ring, or
speak, keeping Maven "not a nag, not autonomous" — the /routines page is
still the only place a proposal becomes visible, and only accepting it
starts producing nudges (fireAcceptedRoutines).
Also fixed the stale vikunja#46 reference in proposed_routines.go — the
TODO it named is what this commit does.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
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
PR #47 added two state-changing routes (POST /chat, POST /routines)
behind the -addr flag, which defaulted to ":9200" (all interfaces).
Default now binds 127.0.0.1:9200; anyone who wants LAN/wider exposure
still passes an explicit bind (as deploy/docker-compose.yml already
does with "-addr :9201" inside the container, unaffected by this
default change).
Vikunja #317.
The template said "drop into your nginx sites" but listened on the
wildcard `listen 80;` with no allow/deny ACL, unlike the actual deployed
hexis.kvmx.ru config which binds only to the WireGuard (10.42.0.1) and
LAN (192.168.1.104) addresses with allow/deny all. Anyone following the
template as written would expose these unauthenticated admin UIs to the
open internet.
Bind explicitly to those two addresses and add the matching ACL block,
mirroring cmd/mavweb/nginx.conf which already does this correctly.
Added a comment naming both addresses as host-specific so a deploy on a
different box swaps the IPs instead of reverting to `listen 80` when the
bind fails.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
CLAUDE.md's routing section said "llmrouter is wired nil" and called the
classifier cascade the committed default. That stopped being true when the
integration merge landed: voice.go:214 wires pickLLMRouter, DefaultLLMRouter is
on, and deploy/mavend.json sets llm_router true. It is the first thing anyone
reads before touching the router, so it was pointing the next reader at a
wiring job that is already done.
Rewritten to say the LLM router is the default, the classifier is the failure
floor and must not be deleted, and what the two actually measure — 36.8% at
p50 31ms against 67.5%/72.7% at p50 ~2.7s, a trade accepted on purpose. Names
the one thing still open on that path: Confidence is hardcoded 1.0 in
llmrouter.go, so the LLM never asks for clarification (#359).
Also in CLAUDE.md: the persona line pointed at a memory file that does not
exist, so the actual rule was nowhere in the repo. Written out instead —
feminine self-reference, informal singular address, pet names forbidden but his
name allowed — plus the three eval checks that enforce it.
MODEL-BAKEOFF: three claims had gone stale within hours of being written. There
IS a make eval-models target now; the routing numbers ARE the production path,
not a bench artifact waiting on a wiring change; and the truncated 293 MB gguf
is deleted. Struck through rather than removed, since the caveats are part of
how the evening read at the time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Wording fixes from the review of the clarify + phrasing PRs.
- "На когда напомнить?" → "Когда?". After she has just been asked something,
the long form is the phrasing of a form field, not of a person.
- A reminder now wants a subject as well as a time. "напомни в 11" had a time
and nothing to say at 11, and she asked nothing at all — she now asks
"О чём напомнить?". Subject first, since a reminder with no subject is not
worth setting.
- The expiry notice is five phrasings picked at random instead of one fixed
sentence. It is the line he hears every time he walks off mid-request, so it
is the line that repeats most.
- The nudge prompt's ban on "обращения" is now "ласковые обращения". It was
meant to forbid "милый"/"дорогой", not his name — "Ками, ноутбук на трёх
процентах" is how she talks, and the eval's cringe check already only flags
pet names.
- The nudge example no longer claims she plugged the laptop in. She has no
hands and no smart plug; an example where she acts teaches the model to
invent actions Maven never took.
- replySystem: "тепло" → "спокойно и без официальных формулировок". A one-word
mood instruction a 1.7B can't act on, replaced with the behaviour meant.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
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
kami
kami