Maven/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Maven is a self-hosted, privacy-first voice assistant (Russian + English). Go daemons talking over unix sockets; one resident small model for routing + phrasing; whisper.cpp STT, piper TTS. Deploy target is a Ryzen laptop (homesrv) with Vulkan offload to the Vega iGPU (n_gpu_layers: 99, compose passes /dev/dri + the render gid) — the resident model stays ≤1.7B either way.

Resident model: currently Qwen3-1.7B (UD-Q4_K_XL), stock — not yet the CPT'd one. It replaced Qwen3.5-0.8B on 2026-07-31 because it measured better on both fixtures we have: 67.5% vs 59.7% intent-only on the 77-case RU routing fixture, and 20/27 vs 11-17/27 on the talk fixture. See docs/evals/2026-07-31-model-bakeoff.md. It is a Thinking variant, so n_ctx is 4096 — reasoning tokens need the room, and 4096 is what the scores above were measured at.

The target is still the locally CPT'd Qwen3-1.7B (Vikunja #122, training in flight). Stock already speaks good Russian; what it gets wrong is the persona — it writes я рад, masculine, where Maven needs рада. That is what the CPT is for.

Do not bother with sub-500M models. LFM2.5-230M and 350M were measured on 2026-07-31 and both are unusable in Russian: the 350M routes at 5.2% (worse than guessing) and answers "столица Франции?" with the invented non-word "Сторзит"; the 230M replies to Russian in Spanish. Their strong published IFEval/BFCL numbers are English-only. Model files live in /mnt/hdd1/llms, bind-mounted to /opt/maven/models/llm — which shadows the repo's models/llm/, so the LFM2.5 gguf sitting there is not loaded by anything. Swapping the resident model is a one-line change to phraser.model_path in deploy/mavend.json.

See docs/rearchitecture.md for the target architecture, docs/design.md for the folded design spec, and AGENTS.md for local-preview + model-download recipes.

Model work is moving to the workstation (owner's call, 2026-08-02). homesrv cannot grow a GPU and the workstation has 16GB of VRAM. So the resident model, STT and TTS become preferred remotes with a floor on homesrv. The workstation is never assumed up. Fall back silently when it would only do the job better. Name the gap when the 1.7B cannot do it at all. The embedder stays on homesrv permanently, because it backs that floor. It is multilingual-e5-small, quantized and asymmetric — EmbedQuery and EmbedPassage apply the query:/passage: prefixes it was trained with, and calling plain Embed on a note is a bug. It replaced MiniLM and bought ten points of recall@1 and 2.5× the speed; see docs/evals/2026-08-04-recall-e5-small.md. Read docs/offload.md before touching a daemon seam or adding a model caller. Vikunja #483 is the umbrella, #484 to #487 are the work.

Both halves are wired as of 2026-08-03. Routing and replies prefer the workstation silently through modelSeam; nudge and reminder phrasing prefer it silently inside the phraser. A world question goes through LLMPhraser.PhraseWorld and names the gap when the card is not free — worldGap in cmd/mavend/worldmodel.go, which the owner hears instead of an invented answer. A box with no workstation block behaves exactly as it did before the seam: naming a gap requires a gap. The offload table in docs/offload.md says which caller is which.

Speech-to-text moved on 2026-08-09 (V-486). sttSeam in cmd/mavend/voicewire.go builds an stt.Pair beside modelSeam, preferring CrisperWhisper 2.0 turbo on workpc with mavsttd as the floor. It takes only the silent half of the rule. A worse transcript is still a turn, so stt.Pair has no TranscribeRemote. The fallback is never spoken. CW2 turbo scores 10.4% WER in Russian against 27.5% for the ggml-small.bin mavsttd loads, over 200 Golos clips (docs/evals/2026-08-09-crisperwhisper2-russian-wer.md). It runs in Intended mode, not Verbatim, though that corpus cannot separate the two. whisper.cpp cannot load CW2 at all. It reads its language count off the vocabulary size, and CW2's 51897 tokens shift seven special token ids. So it is not a second endpoint on mavgpud. It is its own transformers service on port 8081 (deploy/cw2/serve.py), which Maven reaches directly. stt.HTTPTranscriber posts raw PCM to it with a bearer token, because audio is the most sensitive thing that crosses this seam. The switch is workstation.stt in deploy/mavend.json, and deleting the block sends every utterance to mavsttd. mavgpud runs that service as a second child. That is not an optimisation. CW2 is a ROCm process on the same card, so it registers on the KFD like any contender. Under its own systemd unit it made mavgpud evict llama-server every few seconds. That took the gemma-4-12b arm down for eight minutes on 2026-08-09 before anyone noticed. The card needs one owner. Any GPU service added beside this daemon has the same defect, so add it to cmd/mavgpud and not to systemd. CW2 is on the yield clock and not the idle one. At 1.6GB it denies the card to nobody, and unloading it would only send the next voice turn to the homesrv floor. Text-to-speech has not moved and piper on homesrv is still the only synthesizer.

Build & test

CGO daemons (mavend, mavsttd, mavttsd, mavenclient) need the vendored toolchain and libs wired through the Makefile — do not call go build on them bare, use make:

make build          # all 11 binaries
make build-web      # single daemon (pure-Go ones: web/waked/poll/caldav build without CGO)
make test           # go test -race across ./internal/... ./cmd/... with CGO env set

Run one package or one test with make t. Do not hand-write the CGO preamble. Past sessions pasted it about 390 times. That is where the shell-quoting failures came from. This box runs zsh, so an unquoted -run Test* or --include=*.go dies on "no matches found" before go is ever reached.

make t PKG=./internal/router/
make t PKG=./cmd/mavend/ RUN=TestSimulator
make t PKG=./internal/router/eval/ RUN='TestONNX' V=1   # V=1 for -v, RACE=0 to drop -race

t carries -race, so a green make t cannot turn red under make test. It carries -count=1, so a cached PASS from before your edit is never mistaken for a result.

It also sets MAVEN_ONNX_LIB, which the hand-written recipe did not. The four TestONNX* measurements self-skip when that variable is unset. The run still prints ok. So every targeted eval done the old way reported the hash ratchet while reading as a real embedder score.

Pure-Go packages (router, memory, mavweb, …) also run under a plain go test ./pkg/, but make t works everywhere and is one thing to remember.

The daemons (cmd/)

Binary	Role
mavend	Core. Router, phraser, memory, reminders, digestion tick. Owns the DB + IPC socket.
mavweb	HTTP UI + PWA (/dash, /history, /trace, /notifications, /tools); WebAuthn auth. Connects to mavend's socket.
mavsttd	Speech-to-text (whisper.cpp, CGO).
mavttsd	Text-to-speech (piper subprocess).
mavwaked	Wake-word / VAD gate. Not on homesrv — see below.
mavenclient	Voice loop client (mic → stt → core → tts). Not on homesrv — see below.
mavpoll	Environment poller: netdata alarms, uptime-kuma, zenmoney, wireguard presence. Writes facts, sends nothing. Telegram is internal/delivery/telegramsink, not this.
mavcaldav	CalDAV calendar sync.
mavmaild	Mail reader (IMAP, read-only). Holds the IMAP password; core never sees it.
mavgpud	GPU supervisor. Runs on workpc, not homesrv — own unit, deploy/mavgpud.service. Keeps llama-server loaded while the card is free (V-488). Maven never asks it for anything, it reads /health through llm.Pair.
mavupdate	Not a daemon. Operator CLI a human runs on the box to deploy a new build.

Two more binaries have no Makefile target and are built with go run or go build when they are needed. Neither is deployed.

Binary	Role
mavseal	Recovery tool. Encrypts a live tmpfs working copy back to the ciphertext file when mavend was killed before defer st.Close() sealed it.
labelgen	Runs the stage 0 grammars over utterances and prints JSONL, the training data for the routing heads (V-546).

Daemons are wired socket-to-socket, not linked. internal/ipc is the client/server wire protocol; the config in deploy/mavend.json (with ${VAR} env expansion from gitignored deploy/telegram.env) sets socket paths, model paths, and the phraser/embedder blocks.

docker-compose.yml runs five: mavend, mavsttd, mavttsd, mavweb, mavpoll. Count against compose, not against the table. Four of the nine daemons are absent, and each absence has a different reason.

mavmaild and mavcaldav are commented out in compose, each with the reason written beside it: the first needs a mail account, the second a CalDAV account, and this box has neither. mavcaldav used to appear nowhere at all, which was an oversight; it became a recorded decision on 07-08-2026 (V-644). Two things ride on that absence and the block names them. Agenda questions route to IntentQuery at stage 0 (V-498) and the calendar query source then reads a table nobody writes. And loop.State.CalendarBusy is fed by the same facts, so the gate's "do not nag mid-meeting" is permanently false. Its password is read from a file (-pass-file, and -render-pass-file for the render collection), never taken as a flag value, which is the rule mavpoll and mavmaild follow too.

mavwaked and mavenclient are absent by decision, not oversight (Vikunja #463, docs/plans/17-where-the-voice-loop-runs.md). homesrv has a microphone — it is a laptop — but it is in the wrong room, so a wake-word daemon there listens to nobody. They belong on a client machine where the owner is standing.

That machine is workpc (owner's correction, 2026-08-05). This section used to say no such machine existed, which was written when the workstation was only a model host. It is where he sits most of the day and it has the microphone. ipc.Dial already takes tcp://host:port?token=... through the netaddr seam, so the two daemons need deploying, not building. V-515 is that deployment.

Until they are deployed, the wake word and the VAD gate are covered by unit tests and by nothing else, and push-to-talk through /dash is what QA actually covers. Note that deploying them does not by itself prove a wake word: mavwaked gates on energy and has no keyword model (V-487), so the loop runs open until that lands.

The ecosystem: Nexus, Praxis, Hexis

Maven is one of four services. It owns conversation and personal memory. It does not own identity, operational state, or execution. Full contract in docs/ecosystem.md.

Nexus identifies.  Praxis observes.  Hexis acts.  Maven understands and coordinates.

Service	Owns	Maven's client	Configured at
Nexus	Canonical entity ids, names, aliases, relationships. Projects, services, devices, people, pets, places.	nexusClient in cmd/mavend/ecosystem.go, POST /api/v1/resolve	nexus.url (http://nexus:9740)
Praxis	Operational attention and item lifecycle. What needs looking at, what changed, what is still unresolved.	praxisClient, the HTTP tools API under /api/v1/tools/	praxis.url (http://praxis:8989)
Hexis	The capability registry and the only path to executing anything.	vendored github.com/kami/hexis/pkg/client	hexis.url (http://hexis:9741)

All three are nil unless configured, and every one of them degrades on its own. An outage means a named gap in the answer, never a broken turn and never a guess.

Rules that are not negotiable:

• No component reads another component's database. Praxis attention comes over HTTP, never from its SQLite file.
• Identity lives in Nexus. Do not invent a local fact key for something Nexus resolves. actionFact already sets Subject, and cmd/mavend/factenrichment.go resolves it in the background against Nexus.
• Free text never reaches a mutating Hexis call. Resolve to a canonical entity id first. Ambiguous resolution asks the owner, it does not pick.
• LLM output is not authorization. Confirmation binds capability id, target entity, arguments, requester and expiry. See cmd/mavend/confirm.go.
• Praxis lifecycle words mean different things. Surfaced is not acknowledged, acknowledged is not resolved, execution success is not recovery. Reading an item aloud calls Surface, never Acknowledge.
• No automatic attention-to-action path. Digestion may summarise Praxis. It may not call Hexis.

Every cross-service call carries a correlation id minted once per action (withCorrelationID), a contract version header, and X-Requested-By: maven.

Routing — read this before touching the router

internal/router/ has TWO layered engines. The LLM router is now the default and it is on in deploy — this section used to say it was wired nil, which stopped being true on 2026-07-31.

• LLM router (the intended design, docs/rearchitecture.md): the resident Qwen3-1.7B (llmrouter.go) emits GBNF-constrained structured JSON, and the SAME model phrases replies. Embedder is demoted from a routing gate to a RAG hint. Wired at voice.go:214 via pickLLMRouter(cfg.Voice.UseLLMRouter(), llmClient); the flag is voice.llm_router (config.go), DefaultLLMRouter is on, and deploy/mavend.json sets it true.
• Classifier cascade (the failure floor, not dead code): classifier.go + embedder.go nearest-neighbour over frozen seed phrases. It runs when the LLM router is off, when there is no llama-server to talk to (pickLLMRouter logs that and degrades), and on any per-turn LLM error. Do not delete it — routing by seed similarity is the known cause of weak RU query handling, but a turn must never break on the model.

Cascade order: stage0.go exact-match fast-path → LLM router (when non-nil) → classifier fallback. Any LLM error falls through to the classifier so a turn never breaks on the model.

A stage-0 decision is slot-extracted too, since 06-08-2026 (V-572). fillMatchedSlots in router.go runs the stage-2 extractor over whatever a grammar built and fills only the slots it left empty — a matched value always wins, because the rule read a literal pattern and the extractor guesses. It did not run before, so ReminderGrammar handed the daemon HasTime: false for "напомни в 11:00 позвонить маме" and missingFor read the silence as absence and asked "Когда?". It applies to every grammar and is inert for all but the reminder: Extract fills Time, Fn and Key and nothing else, and the query, clock, agenda, feed, list, task and narrative rules all emit intents with no such slot. Benchmarked at 20000x, a stage-0 query costs 3.7µs against 3.9µs before. Slots.Text is deliberately not filled — a grammar that left it empty meant it, and agendaQueryBuild hands the query chain the utterance itself. Fixture unchanged at 64/91, with "slots deferred to daemon" 6 → 0.

Measured on the 77-case RU fixture. Re-measured 2026-08-02: the classifier scores 68.8% full accuracy at p50 16.6µs, not the 36.8% at p50 31ms that stood here from docs/evals/2026-07-31-model-bakeoff.md. That older figure predates the stage 0 rules and the seed additions, both of which now score inside the classifier baseline. Qwen3-1.7B scores 77.9% intent-only / 72.7% through the cascade. So the router buys about 4 points of accuracy, not a doubling, and the trade is worth re-arguing rather than assuming. The ≈2.7s figure that stood here until 2026-08-02 was contention, not the model. See docs/evals/2026-07-31-routing.md line 61, which measures the LLM router at p50 825ms / p95 1.2s / max 3.0s and the full cascade at p50 0.80-1.04s. Do not plan latency work off the bakeoff table.

Re-measured 2026-08-05 on the fixture as it now stands, 91 cases (V-320 item 2, docs/evals/2026-08-05-routing-resident-model.md): cascade + resident model scores 75.8% full / 80.2% intent-only at p50 1.19s / p95 1.65s. That is a new baseline and not a movement, because 14 cases were added since the 77-case number above. The model alone scores 37.4% full against 61.5% intent-only, and the gap is slots rather than routing: it routes reminder and leaves the time to the daemon, which is what the contract asks. To re-run it, start a second llama-server on a fixed host port — the resident one binds --port 0 inside the container and no host process can reach it.

The numbers above are the homesrv floor, not the ceiling. With the workstation up, routing completes through llm.Pair against the model mavgpud holds, which is better than the resident model and about 2.5× faster. gemma-4-12b scored 84.4% full / 93.5% intent-only at p50 329ms (docs/evals/2026-08-02-workstation-gemma4-12b.md, Vikunja #485). The workstation is never assumed up, so both sets of numbers are live. Judge a routing change against the classifier and the resident model, since those are what always answer.

The workstation runs gemma-4-E4B since 2026-08-09 (owner's call), and it is a step down measured the same day (docs/evals/2026-08-09-e4b-vs-12b-routing.md). Against a same-session 12B control it scores 83.3% full / 89.6% intent-only, destination 19/33 against 23/33, at p50 294ms against 344ms. So it costs four destination cases and buys 50ms. Read destination as the finding: it names nothing where the 12B names recall or calendar, which is safe but walks the whole chain. It also has no MTP and cannot be given any here. The only gemma4-assistant draft on disk is trained against the 12B's hidden states.

The intended third engine is not a generative model (owner's call, 05-08-2026, V-546, docs/plans/18-routing-heads-on-e5-small.md). Routing has a bounded output space, so it is classification, and the 118M multilingual-e5-small is already resident. Three heads on one forward pass: intent, mood, and BIO slot tags. Roughly 5e15 FLOPs to train, so 10 to 30 minutes on the workstation. A 100M decoder from scratch is 10 to 20 GPU hours. Two things it buys that a decoder cannot. No grammar is needed, because a softmax cannot emit a value that does not exist. And max softmax is a calibratable confidence, where Confidence: 1.0 was a hardcode. Fine-tune a copy of the weights. The resident embedder backs memory recall. Training it in place couples routing accuracy to recall@1, with nothing in the suite to name the trade.

Two of those heads are trained as of 08-08-2026, and they are not the three above (V-661, docs/evals/2026-08-08-routing-heads-two-head.md). Intent and destination share one masked mean pool. Destination scores a mean 80.8% over three seeds, best 29/33 (87.9%). The classifier cascade scores 12/33 and the cascade with gemma-4-12b scores 24/33, so a 118M encoder beats the 12B teacher it was distilled from. Read the best run as one seed and not a headline, because one case is 3 points on a fixture this small. Recall is 15/15 and world is 5/5. Intent is 93.6% mean over three seeds. That is not comparable to the 76.0% and 84.4% those two arms scored: a softmax has no clarify class, so the head's fixture is the 88 cases carrying an intent.

A fourth head asks instead of guessing, same day (V-661, docs/evals/2026-08-08-clarify-head-four-head.md). Clarify is not a value of intent, so a softmax cannot emit it. It is a second question over the same pooled vector: can Maven act on this at all. That is why the head's fixture was 88 cases and not 96. Over three seeds it catches 7.0 of the 8 want_clarify cases and produces 2.3 false clarifies of 88. The cascade today misses 1 and produces 2, so this is parity with no rules in front of it. Accuracy is the wrong number here and a head that never asks scores 91.7%. Confidence is the other half. Max softmax over the intent head reads 0.851 where it is right against 0.604 where it is wrong, ranking right above wrong in 83.4% of pairs. Confidence: 1.0 was a hardcode, and this replaces it with a signal. The two are not the same signal: one says which intent is unclear, the other says the utterance carries too little to act on. The fourth head is not free the way the third was. Intent, destination and slot F1 each move down one to four points, inside the seed spread. поужинал is a false clarify on every seed, which is the same defect thinSingleToken was narrowed for on 2026-08-01.

The corpus for it is generated, because every existing row is answerable by construction. The router-prompt agreement filter cannot work here, since routeGrammar has no clarify value and a generated line always agrees with itself. A gemma judge replaces it. The first judge called 24 of 40 answerable rows underspecified, because it judged against a generic assistant rather than against Maven's contract.

Mood is cut, not deferred. The enum describes her own reply state, not the speaker's emotion, and no dataset maps onto it.

A third head landed the same day (docs/evals/2026-08-08-slot-head-three-head.md). BIO slot tags had no Maven-domain corpus, which was true of found corpora and false of made ones. label_slots.py distils spans out of gemma-4-12b under a GBNF closed over Maven's own five slots. A span survives only when it is a literal substring of the utterance, so the agreement filter costs no second call. 2178 spans over 1702 rows, 37 dropped, nothing unparsed. Three heads score intent 92.8%, destination 82.8% and slot span F1 72.4% over three seeds. The slot head is free: both other numbers move less than their own seed spread. Epoch selection reads the intent dev slice alone. Slot F1 is still climbing when it stops, which costs about 4 points.

The MASSIVE warm-start of step 2 is worth nothing here. Stock e5-small ties it on intent and leads by a third of a case on destination. Nothing argues for keeping that step.

The floor was a corpus defect and it is fixed. The first 120 floor rows carried one sentence shape, so the head named a destination where the fixture says walk the chain. Rotating six shapes took the floor 3/7 to 6/7 and destination 75.8% to 80.8%. What is left is calendar at 3/6 on every seed, which training cannot move: the possessive agenda rules claim those cases at stage 0 and name nothing, so no label reaches the head. That is the same trade V-660 flagged and it wants the owner's call.

The heads run in Go and route every turn, since 08-08-2026 (V-664, docs/evals/2026-08-08-routing-heads-in-go.md). This section used to say nothing of it ran. RouterHeads in internal/router/heads.go loads router_heads.onnx and reads intent, destination and clarify off one forward pass. It is stage 0b: after the grammars, before the resident model, and the classifier is still behind both. Through the cascade it scores intent 96.9% and destination 75.8% at p50 27.9ms. That beats the gemma-4-12b cascade, 84.4% and 72.7%, at a twelfth of its 329ms. The workstation stays the better phraser and is no longer the better router.

Three rules around it. The clarify head decides first, before the intent threshold. It answers a different question. A thin utterance scores low intent by construction, so gating it cost 6 of 8 ambiguous cases. The destination head is read on IntentQuery only, since no other intent reaches queryWalk. And headsThreshold is 0.6, the measured knee: every value to 0.85 drops right answers and keeps the same two wrong ones.

voice.embedder.heads_path is the whole switch. Empty, missing or unloadable means the heads are nil and the cascade is byte-for-byte what shipped before them. It must never be pointed at model_path. The resident e5-small must not be replaced by the fine-tuned copy. Recall depends on that file scoring what it scored.

The hand-written tokenizer read every long word backwards until this task (encodeWord, onnxembedder.go). It cost recall@1 7.4 points and recall@3 11.1. Nothing caught it because seeds and queries were mangled the same way, so cosine survived. The heads found it. They are trained through transformers and read through this. The embedder id now carries a tokenizer revision (@384/tok2), so fixing the tokenizer triggers ReembedAll the way swapping the model file does. Bump tokenizerRev on any change to what it emits.

Confidence: 1.0 used to be hardcoded in llmrouter.go, so the LLM path could never ask for clarification (6/6 refusal cases missed on the fixture) — Vikunja #359. Fixed 31-07-2026 with structural signal (single-token utterance, keyless fact, act with no allowlisted fn) feeding the same stage-3 gate the classifier path already had — see gateLLMDecision in router.go. Note the second half of that bug: the LLM branch never consulted r.threshold at all, so a correct low confidence would have been discarded anyway.

Re-measured on the fixture after the fix: missed clarify 6/6 → 1, at the cost of 3 false clarifies and 2.6pt of full accuracy (72.7% → 70.1%, intent-only 67.5% → 74.0%). Two of the three false clarifies are acts the model mis-routed and the gate caught — asking beats wrongly executing, so the fixture and the daemon disagree about what is correct there. The third, "поужинал", was a real defect: the single-token rule was an English intuition and does not transfer to Russian, where one word is routinely a whole sentence.

Narrowed 01-08-2026. thinSingleToken (internal/router/singletoken.go) still thins a bare one-word nominal — "вода", "бэкап" — but spares two classes: a closed lexicon of social and control singles ("привет", "спасибо", "стоп", "yes"), and any token carrying a Russian verb ending (past tense, 2nd person, reflexive), because a verb already contains its subject. Both tests are offline and cost nothing. Re-measured: false clarifies 3 → 2, intent-only 74.0% → 75.3%, full accuracy unchanged at 70.1%, missed clarify still 1. The two remaining false clarifies are the act-with-no-allowlisted-fn arm of the gate, not this rule.

Agenda questions taken off the model, 01-08-2026. AgendaQueryGrammars (stage0.go, wired after the clock rules in buildRouter) routes "что у меня сегодня", "во сколько у меня встреча" and anything naming a calendar to IntentQuery at stage 0. They were going to IntentSystem, where replySystem has no agenda arm and answered "пока не умею" — the fixture had said query since ru-query-019 was written. Measured: full accuracy 70.1% → 72.7%, intent-only 75.3% → 77.9%, calendar 0/2 → 2/2, clarify counts unchanged. Note that Go's \b is ASCII-only and never fires after a Cyrillic letter; the pattern needs an explicit (\s|[?!.]|$).

Two more shapes taken off the model, 04-08-2026 (V-498). rest-of-day-query inside AgendaQueryGrammars claims "что дальше?" / "what's next", and NarrativeQueryGrammar (stage0.go, wired last in buildRouter, after the capture marker) claims "расскажи про X", "объясни X", "опиши X". Neither carries a question mark or an interrogative, so the model called both IntentFact; the write was caught downstream by IsQuestionShaped, so this was a latency and fixture defect, not a correctness one. The narrative rule reads the same narrativeRequests lexicon IsQuestionShaped reads, and declines chatNarrativeTopics — a joke, a bedtime story, herself — because the query chain has no source that answers those. New fixture cases ru-query-024 and ru-query-025. Classifier + ONNX baseline 56/80 (70.0%) → 58/82 (70.7%), no case regressed, no new false clarify. The LLM arm was not measured (no llama-server in that run), so judge it again before quoting a cascade number.

Praxis taken off the model, 05-08-2026 (V-516). PraxisGrammars() (internal/router/praxis.go, wired in buildRouter before the capture marker because "отметь" is a capture verb) fills Slots.Fn with a Praxis capability name. These grammars are the only path to Praxis, not a faster one. Measured 2026-08-05 with the resident model as router (V-517, docs/evals/2026-08-05-reach-llm-router.md): the model alone reaches Praxis 0/12, the same as the classifier alone, because nothing in the router prompt names a Praxis capability and there is no string for it to write. Through the cascade it is 11/12. Deleting these rules costs every point. Praxis reach was 0/12 and structurally so: handlePraxisAct compares Slots.Fn to a capability alias, and that slot is filled from the deployment's enabled tool names, which no Praxis alias is on. Measured 16/30 → 27/30 overall, praxis 0/12 → 11/12, lifecycle 0/5 → 5/5 (docs/evals/2026-08-05-praxis-reach.md). Two rules to know before editing: a stative lifecycle word ("готово", "принято") needs an item named beside it, while a bare imperative ("закрывай") may ask which one. The bare arm additionally requires that the sentence name no object of its own, or "закрой шторы в комнате" goes to Praxis instead of the house. A demonstrative ("отметь это как сделанное") resolves against h.surfacedItems only when exactly one item was spoken. Otherwise the turn goes back to the cascade rather than transitioning the wrong item.

Who claimed a turn is now recorded, and so is who did not (V-564, umbrella V-558). Arbitration between the claimants on the utterance stream is order, hardcoded in the pre-route resolver ladder, in buildRouter and in querySources. internal/decision records one Record per turn: every claimant, what it would have made the turn, the score it reported, and whether it won, declined, lost on score, was thinned by a gate or was never asked. The record rides the context, the same seam querysource.go uses, so a claim site cannot change a route and a context with no record costs nothing. It is installed in runTurn, so the mic, telegram and the web all leave the same trail. Storage is a 25-turn in-memory ring on the handler (decision.Ring), read over ipc.TurnDecisions and rendered as the second table on /trace. It also persists, since 06-08-2026, and that reverses what this section used to say (V-629, docs/plans/21-persisting-the-routing-trace.md). The old rule was that nothing persists, because a turn record is read minutes later or never. The owner reversed it: the routing heads (V-546) cannot be fitted or calibrated without real utterances, and 9 of the 31 modes in internal/modes have no seed example at all. The ring did not move. It is still what /trace reads and still what a test with no store gets. cmd/mavend/routingtrace.go is a second sink beside it, writing routing_traces (migration #23). The utterance is stored in clear, because a 384-dimension vector of a short sentence is substantially recoverable and storing vectors instead would be a privacy claim we cannot support. What makes it safe is the same thing that makes the fact store safe. Retention is 14 days, enforced on write and again on start, so a box that goes quiet does not keep every row. Nothing reads it outward, and the rule that his notes and facts are never search input covers this table. Store.Wipe deletes it with everything else. A correction (V-630) is promoted out into a seed-shaped row in routing_labels (migration #24) and kept, because a label is not a transcript. The transcript still expires. The gesture that writes one is two buttons beside the reply on /chat, reached over ipc.CorrectTurn and the trace id that now rides back on ipc.ChatReply. A turn marked wrong with no target is a usable negative, so naming the intent is never required. The target is one of the seven intents and never free text. All three reaches offer it as of 06-08-2026, and this section used to say only /chat did. Voice is the repair rung, which has read spoken corrections since V-455 and now writes the durable label beside the classifier seed it always wrote; a spoken negative with no target is its own rung, repair-negative (V-636, docs/plans/22-correcting-a-turn.md). Telegram is an inline keyboard under the reply, and it needed the chat to become readable first — telegram is no longer outbound only (V-637, docs/plans/23-inbound-telegram.md). The poller is dark unless the telegram block says intake, it long-polls because the box takes no inbound connections, it accepts chat_id and no other sender, and it drops whatever queued while the daemon was down. It reaches the daemon through ipc.CoreAPI alone, so a chat turn takes the path POST /api/chat takes. Note that the turn source is still tap:text for both, so provenance cannot tell a chat turn from a typed one. Adding a rung to the ladder in runTurn means adding its name to preRouteLadder in cmd/mavend/decisiontrace.go, or that rung is silently missing from the record.

A route now says where the answer lives, not only that the turn is a question (V-655, 07-08-2026). query was a shrug. The cascade sorted an utterance into one of seven intents, with stage 0, the resident model and the classifier behind it. Then IntentQuery handed the turn to querySources in the daemon. That is twenty-two branches deciding by seed similarity in a fixed order. It has no fixture and no accuracy number, no model arm and no floor. Decision.Source (internal/router/source.go) is the second half of the route. Twelve destinations, not twenty-two. The three recall passes plus fact-by-key are one destination from outside. So are search, Kiwix and the URL reader.

SourceUnknown is a real value and it is the floor. Nothing named a destination, so the daemon walks the whole chain. That is byte-for-byte what shipped before the field existed. The classifier arm names nothing, so a box whose model is down routes queries exactly as it did.

queryWalk in cmd/mavend/actions_query.go takes sources out and moves none. That is the safety argument and it is not negotiable. The table's order is load-bearing. Every comment on it argues a reason between two sources, and above all it carries "the owner's data first, then the world". Naming SourceWorld does not send the turn outside on its own. His notes and his facts still run first, because they look rather than guess.

The personal boundary is the one exception and it is deliberate. It guesses, so naming SourceWorld drops it. That is what stops it answering "кто такой Линус Торвальдс?" with "не нашла у тебя такой записи", which it did on 2026-08-07. The cost is that a destination a model wrote can now take the boundary off a turn. A question about him that the model calls world reaches SearXNG, where today the boundary stops it. Only the utterance leaves the box, never his notes or history, so this widens what is asked and not what is sent. TestNamingRecallKeepsTheBoundary pins the other half: naming SourceRecall keeps the boundary in front of the world. Whether a model may drop it at all is the owner's call and has not been made.

What comes out is only the sources that guess. Those decide a turn is theirs by cosine against frozen seeds, then answer whatever they claimed. They hold no table that could come back empty. Weather is the pure case and has no local data at all. It was measured on the box on 2026-08-07 (docs/evals/2026-08-07-week-of-usage.md section 4). It answered both "что такое TCP?" and "сколько будет 17 на 23?" with "для какого города?". The feed answered "какой у меня любимый язык?" with kernel headlines. The personal boundary answered "кто такой Линус Торвальдс?" with "не нашла у тебя такой записи". A source that guesses is marked guesses: true in the table. One that looks is not, and it is always asked.

Stage 0 fills the destination where a rule already knows it. WorldQueryGrammars() (internal/router/worldquery.go) claims "что такое X" and "сколько будет 17 на 23". It is wired after the agenda rules and before the feed and list rules. "что такое лента" is a definition question, and the feed rule would take it on the noun alone. calendar-query and event-time-query name the calendar. The possessive agenda rules deliberately do not. "что у меня в списке покупок" matches agenda-query, and naming the calendar there would take the list source off the turn.

Fixture unchanged at 69/91 classifier+ONNX, measured both sides. That is the expected result, because it scores intent and no case here changes intent.

The destination has its own fixture and its own number as of 08-08-2026 (V-659, docs/evals/2026-08-08-destination-fixture.md). This section used to say it had neither. want_source on eval.Case is a pointer, because the destination has three states and a bare string has two. Absent is every intent but query, which never reaches queryWalk. Present and empty is the SourceUnknown contract: name nothing and walk the chain. Present and named is a destination the route must produce. Thirty-three of ninety-six cases carry one.

A destination miss does not fail the case. It lands in Outcome.SourceReason and never in Reasons, so Accuracy and IntentAccuracy mean what they meant and SourceAccuracy is a second number over the labelled cases only. Intent and destination are two decisions, and one number hides which one moved. A route that lost its intent scores no destination hit, or a clarify would satisfy an empty label for free.

Measured classifier+ONNX: intent 73/96 (76.0%), destination 12/33 (36.4%). The split is the finding. World is 5/5, because a stage 0 rule names it. The SourceUnknown floor is 5/7. Calendar is 2/6, because the possessive agenda rules deliberately do not name it. And recall is 0/15, because nothing anywhere names it. Those turns are still answered, since the chain walks recall early. Recall is the number the fourth head has to move.

Seven cases assert the floor and five of them are homelab operations. They cluster because SourceRecall, SourceNetwork and SourceAttention overlap on every question about the box. The other two are ru-query-005 and ru-query-014. No query source reads the reminder store, and a deadline could sit in tasks, the calendar or Praxis. mavpoll writes its netdata and uptime-kuma observations into the fact store recall reads. That is a finding about the enum, not a gap in the labelling. The owner confirmed all seven floor labels on 08-08-2026, so they are a decision rather than an agent's guess.

baselineGrammars in eval_test.go mirrors buildRouter and had drifted: WorldQueryGrammars was wired into the daemon by V-655 and not into the mirror, so the fixture scored a grammar set nobody runs. Fixed by V-659, worth 3 points of destination and nothing else. Check that function when adding a grammar.

The model arm landed the same day (V-660, docs/evals/2026-08-08-destination-model-arm.md). routeGrammar carries a source rule closed over router.Sources plus the empty floor, so the model cannot emit a destination that does not exist. The prompt lists the twelve in Russian and says "" is a normal answer to give often. LLMRouter.Route reads it back through ValidSource and on IntentQuery alone. Against gemma-4-12b on the workstation the cascade scores destination 24/33 (72.7%) with intent unmoved at 84.4%, and recall goes 0/15 to 14/15. The resident Qwen3-1.7B is unmeasured, because it binds --port 0 inside the container.

Stage 0 now costs four destination points. It did not before. The four cases the cascade loses and the model alone wins are all calendar. The possessive agenda rules claim them first and name nothing on purpose. That caution was free while nothing downstream could name anything either. It is not free now, and the fix is the owner's call rather than a quiet edit.

The last arm is V-546. Intent, mood and BIO slot tags were already three heads on one forward pass of the resident e5-small. Destination is a fourth head on the same pass, and 72.7% from a 12B teacher is the label source for training it.

LLM output contract

All phrasing paths emit {"response":"...","mood":"..."}, with fallback to plain text when the model skips the JSON. One parser, parseResponseMood in internal/phraser/parse.go, and every path reaches it: the six LLMPhraser methods, PhraseWorld, and Replier.PhraseReply, which cmd/mavend/replier_llm.go wraps — that file holds the stub fallback and no parsing of its own. The legacy {"body","summary"} fallback was deleted on 2026-08-06 (V-397): it was the contract before {"response","mood"} replaced it, no prompt asks for that shape, the GBNF cannot emit it, and no test covered it. Mood is a fixed enum. Router prompt is a separate contract: [{"intent":<enum>, key?, value?, text?, verb?}, ...], 7 intents (fact, reminder, note, query, act, chat, system). llm/check_prompt_parity.py in the training workspace enforces that the Go and relabelling prompts remain identical.

Russian patterns — three mechanisms, no fourth

Hand-written Russian stem patterns were swept out on 2026-08-04 (owner's call: not a pattern, and the resident model cannot be asked per turn either). A regex whose output is a fact or a route is the defect; a regex over structured input — HTML, MIME, JSON, a URL, an argv list — is not. Before writing a Russian word list, pick one of these:

• internal/lexicon — closed classes, in lexicon_ru_v1.json. Interrogatives, capture verbs, reminder verbs, cardinals, day offsets, parts of day, weekdays, months, spoken hours. Editing a word is a data change, and there is exactly one copy: months used to live in three files. Cardinals carry the oblique forms, because a spoken time declines and в семь / к семи are one hour.
• internal/morph — grammar, from the vendored golem Russian dictionary. IsVerbForm and SameWord. Note that lemma matching is BROADER than stem-plus-one-ending, so a verb slot that means the imperative must be matched exactly — говори and говорил are one lemma and only one of them is a command (cmd/mavend/quiet_toggle.go).
• cmd/mavend/topics.go and the embedder — open sets, where the question is what a turn is ABOUT. Frozen seeds per subject plus a real other class, scored against the turn's own query vector. Same shape as the personal boundary in personalboundary.go, with one difference: a topic must clear the runner-up by topicMargin, because a false claim here spends a network scan rather than one honest "не знаю". The old keyword tests stay as the offline floor and may remain narrow, since they are no longer the only answer.
• The ecosystem trio — when the answer is not in the utterance at all. Identity is Nexus's, never a local pattern.

Seeds are scoring data. Editing one moves a recogniser and must be re-measured against the TestONNX* tests, not eyeballed.

Non-goals (hard constraints)

Not a nag, not autonomous. Maven's persona is feminine — Russian self-reference must use feminine forms — рада, not рад; поняла, not понял. The owner is male and is addressed informally: "ты", singular, never "вы"/"ваш" and never "он"/"его" (she talks TO the owner, not about the owner). Pet names ("милый", "дорогой") are forbidden; the name ("Ками") is not. The eval enforces this: CheckAddress, CheckFeminine and CheckCringe in internal/phraser/eval/checks.go, scored by make eval-phrasing.

"Never phones home" is DEPRECATED (owner's call, 2026-07-31). It used to be a hard constraint and it is not one any more: a 0.8B — and a 1.7B — does not know enough to answer world questions, so she needs to read external sources. What replaces it:

• No telemetry, no cloud model, no third-party account. That part never changes. Nothing about Maven is reported to anyone, and inference stays on the box.
• The owner's data first, then the world. Every source that reads the owner's facts, notes, calendar, tasks or house runs before anything outside, and the personal boundary sits between them. Reading beats recalling for a small model.
• In the world, live search leads and the ZIMs are the fallback (owner's call, 2026-08-02). A self-hosted SearXNG (search block) answers first; the Kiwix ZIMs on homesrv answer when the search is empty, unreachable, or the line is down. Verified with the line down on 2026-08-05 (V-508, docs/evals/2026-08-05-kiwix-offline-fallback.md): a stopped SearXNG costs nothing, the ZIM answers in the same turn budget. A blackholed host cost 8 seconds the owner waited through. So the connect phase alone is capped at dialTimeout (1.5s), while a slow instance that did connect keeps the full 8. A Russian question reads wikipedia_ru_all_maxi_2026-02 verbatim through kiwix.book_ru. The RU→EN rewriter is the workaround for an English book and is skipped there. Kiwix catalog names come from the filename, not the <name> field. Response.Empty() is the whole gate and there is no quality threshold in front of it: the three signals one could read were measured on 2026-08-05 and none of them separate a real question from an invented one. Token overlap would cost "столица Франции" its answer, because the answer is Париж and that word is not in the question. See docs/evals/2026-08-05-search-quality-signals.md (V-539). Which query source claimed a turn is readable on /chat as a badge beside the reply, carried on ipc.ChatReply.Source and noted by noteQuerySource in cmd/mavend/querysource.go. It rides the context, so handleText keeps the one string signature the mic, telegram and the web share.
• External search is allowed and off unless configured, like the weather and telegram capabilities. The code default is still off. deploy/mavend.json now ships a search block (owner's call, 2026-08-02), so it is on for this box and deleting the block turns it off again.
• The owner's notes and facts are never search input. Looking up why the sky is blue and sending the owner's stored personal notes to an upstream engine are different acts. Only the utterance goes out, never the persona block, history, or matched notes.

Web UI conventions

Server-rendered pages share cmd/mavweb/static/ui.css (served at /ui.css) and the shell partial in cmd/mavweb/shell.html: a page opens with {{template "shellTop" "<page-key>"}} and closes with {{template "shellBottom"}}, and the key marks the active sidebar link. Every page is its own embedded .html file next to main.go — no page markup lives in Go, and the sidebar is data (sidebarSections, pageIcon) the template renders. No per-page <style> beyond true one-offs. Wrap every table in <div class=scroll> so wide data pans on a phone. Local preview + headless screenshot recipe is in AGENTS.md.

Vikunja

This repo is project Maven (ID 2) in Vikunja. MCP: http://localhost:9100/mcp (or http://192.168.1.104:9100/mcp from workpc). Feature/bug/deploy tasks go there.

Vikunja is the durable task store. A task holds the goal, the constraints and the assumption ledger. Work without a task id is work nobody can resume, so a session that has no id asks for one before it starts.

The MCP tool schemas are deferred, so load the four you actually use in ONE call at the start of a session rather than one lookup per first use:

ToolSearch("select:mcp__vikunja__list_tasks,mcp__vikunja__get_task_details,mcp__vikunja__create_task,mcp__vikunja__update_task")

Close a finished task with done: true and nothing else (owner's call, 07-08-2026). Do not write a completion summary into the description on the way out. It is lost anyway, and the durable record is the commit messages and the merged PR. Note that update_task carrying a description resets done to false, which is why a write-up ever took two calls.

Session workflow

~/.local/bin/task owns the branch, the commit identity and the PR. One task, one session, one PR.

task start <vikunja-id>   # branch off origin/master, write TASK.md, fetch review comments
task pr                   # push, open or refresh the PR, label Vikunja, notify
task comments             # re-pull this branch's review comments into .task/

Around that, /pickup opens a session and /wrap closes it. Wrap at roughly half context rather than letting the session compact.

Five stores, and each one owns something the others must not hold:

Store	Holds	Lifetime
Vikunja task	goal, constraints, assumption ledger, status	durable
CLAUDE.md, AGENTS.md	what an agent must know before touching code	durable
docs/	design, measurements, decisions	durable
TASK.md	the brief for this branch, written by task start, immutable	one branch
HANDOFF.md	only what the next agent needs to resume	one session

TASK.md and .task/ are excluded through .git/info/exclude. HANDOFF.md is gitignored and injected at session start. If a line in the handoff would still matter next week, it is in the wrong file.

Docs are tiered by path, so staleness is visible from the filename. Files directly under docs/ are living and carry a Last verified: <date> @ <sha> line. Files under docs/evals/ are dated measurements and are never edited after the day, so a newer number is a new file. Files under docs/archive/ are dead and read by nobody by default.

Git guards

Two hooks in .githooks/, tracked, wired with core.hooksPath. Fresh clone:

git config core.hooksPath .githooks

• pre-commit refuses master, and refuses more than 300 changed lines in non-markdown files. Markdown is exempt and may land as one batch.
• commit-msg requires the subject to end with (V-<id>). V- and not #, because Gitea autolinks #123 to a Gitea issue, which is the wrong tracker.

Two more guards live outside the repo, in ~/.claude/hooks/. diff-budget.sh blocks further edits past 600 changed lines on a task/ branch. prose_lint_hook.py checks prose on every write. Both measure against origin/master, so a local master that is ahead of the remote makes the diff budget read high.

--no-verify exists. Using it means saying why in the commit body.