Maven/docs/plans/2026-07-10-router-lfm-foundation.md

# Plan — Sub-project 1: Router-as-LFM + Foundation

> **Historical/completed foundation.** The shared llama-server client, router,
> fallback and replier described here were implemented. The resident-model
> decision changed on 2026-07-18 from LFM to locally trained Qwen3-1.7B. Do not
> use the embedded LFM model paths or old single-object examples as current ops
> guidance; see `2026-07-18-qwen3-resident-training-eval.md`.

> Scope from `REARCH.md`. Make Maven trustworthy: the LFM becomes the router
> (fixes "messes up queries" / "doesn't take notes"), the engine actually runs
> (fixes stub replies), dates stop being read as "number dot number dot number",
> and telegram becomes a reach channel. NOT in scope: on-demand 4B reasoner,
> digestion worker, voice-quality/custom-TTS.

## Findings that shape the plan

- **Notes + reminders already work in code** (`applyAction` → `WriteNote` /
  `CreateReminder`, `cmd/mavend/voice.go:613,548`). The user's "she doesn't take
  notes / remind" is (a) the embedder-classifier misrouting the utterance away
  from `IntentNote`/`IntentReminder`, and (b) `StubReplier` giving canned
  confirmations. The LFM router + LLM replier fix both.
- **Telegram sink already wired** (`cmd/mavend/main.go:306-321`,
  `delivery.Config.Telegram`). Only the deployed config lacks a `telegram` block.
- **Phraser runs llama-server** (`internal/phraser/llmphraser.go`) over
  `/v1/chat/completions` at base URL `p.port`. Deployed `mavend.json` has **no
  `phraser` block** → `phraser.NewStub()` (`main.go:261`). Engine is off.
- **Import direction:** `phraser → router`. So the LLM router lives *in*
  `router`, using a new dependency-free `internal/llm` client that both the
  phraser and router share (one llama-server, two callers). `router → llm` is
  acyclic; `phraser → llm` is acyclic.

## File structure

| File | Create/Modify | Responsibility |
|---|---|---|
| `internal/ttsnorm/ttsnorm.go` | create | RU date/number → speakable text, before TTS |
| `internal/ttsnorm/ttsnorm_test.go` | create | table tests for date/time/number forms |
| `cmd/mavend/voice.go` | modify | call `ttsnorm.Speakable` in `reply()`; wire LLM router + LLM replier |
| `internal/delivery/voicesink/voicesink.go` | modify | normalize nudge text before synth |
| `cmd/mavend/reactive_notes_test.go` | create | notes+reminders round-trip integration test |
| `internal/llm/client.go` | create | llama-server `/v1/chat/completions` client w/ GBNF grammar |
| `internal/llm/client_test.go` | create | httptest server round-trip + grammar passthrough |
| `internal/phraser/llmphraser.go` | modify | add `BaseURL()` accessor |
| `internal/router/llmrouter.go` | create | `Action`, GBNF grammar, prompt, parse, `LLMRouter` |
| `internal/router/llmrouter_test.go` | create | mock client: intent+slot mapping, fallback |
| `internal/router/router.go` | modify | `Config.LLM`; `Route` consults LLM after stage-0 miss |
| `cmd/mavend/replier_llm.go` | create | `llmReplier` (voice.Replier) — natural reactive replies |
| `cmd/mavend/replier_llm_test.go` | create | mock client: reply per intent, stub fallback |
| `deploy/mavend.json` | modify | add `phraser` + `telegram` blocks |
| `docker-compose.yml` | modify | mount LFM model + telegram env |
| `AGENTS.md` | modify | document LFM model download + router-LFM |

---

## Task 1 — RU date/number TTS normalizer

**Goal:** dates/times/bare number-dot-number strings are spoken as words, never
"number dot number dot number".

**Files:** create `internal/ttsnorm/ttsnorm.go`, `internal/ttsnorm/ttsnorm_test.go`;
modify `cmd/mavend/voice.go` (`reply`), `internal/delivery/voicesink/voicesink.go`.

**Acceptance criteria:**
- `Speakable("напомню 10.07.2026")` → `"напомню 10 июля 2026"`.
- `Speakable("встреча в 14:00")` → `"встреча в 14 часов 00 минут"`.
- `Speakable("это 3.2.1 версия")` → digits joined with " точка " (`"3 точка 2 точка 1"`) — the generic fallback so no `.` is voiced as a symbol.
- Idempotent: `Speakable(Speakable(x)) == Speakable(x)`.
- Pure, no deps beyond stdlib; RU month table.

**Verify:** `go test ./internal/ttsnorm/` → `ok`.

**Steps (TDD):**

1. Write `ttsnorm_test.go`:
```go
package ttsnorm

import "testing"

func TestSpeakable(t *testing.T) {
	cases := []struct{ in, want string }{
		{"напомню 10.07.2026", "напомню 10 июля 2026"},
		{"срок 01.01", "срок 1 января"},
		{"встреча в 14:00", "встреча в 14 часов 00 минут"},
		{"в 9:05 подъём", "в 9 часов 05 минут подъём"},
		{"это 3.2.1 версия", "это 3 точка 2 точка 1 версия"},
		{"без чисел", "без чисел"},
	}
	for _, c := range cases {
		if got := Speakable(c.in); got != c.want {
			t.Errorf("Speakable(%q) = %q, want %q", c.in, got, c.want)
		}
		if got := Speakable(Speakable(c.in)); got != Speakable(c.in) {
			t.Errorf("not idempotent for %q: %q", c.in, got)
		}
	}
}
```

2. Implement `ttsnorm.go`. Order matters: dates (with year, then without) →
times → generic numeric-dot fallback.
```go
// Package ttsnorm rewrites machine-formatted dates/times/numbers into RU text
// a TTS voice speaks naturally — so "10.07.2026" is not read as "number dot
// number dot number". Pure, deterministic; runs on reply/nudge text before synth.
package ttsnorm

import (
	"regexp"
	"strconv"
	"strings"
)

var months = [...]string{"", "января", "февраля", "марта", "апреля", "мая",
	"июня", "июля", "августа", "сентября", "октября", "ноября", "декабря"}

var (
	reDateY = regexp.MustCompile(`\b(\d{1,2})\.(\d{1,2})\.(\d{4})\b`)
	reDate  = regexp.MustCompile(`\b(\d{1,2})\.(\d{1,2})\b`)
	reTime  = regexp.MustCompile(`\b(\d{1,2}):(\d{2})\b`)
	reDots  = regexp.MustCompile(`\b\d+(?:\.\d+)+\b`)
)

// Speakable rewrites d.m.y, d.m, h:mm, and residual dotted-number runs.
func Speakable(s string) string {
	s = reDateY.ReplaceAllStringFunc(s, func(m string) string {
		p := reDateY.FindStringSubmatch(m)
		return spokenDate(p[1], p[2], p[3])
	})
	s = reDate.ReplaceAllStringFunc(s, func(m string) string {
		p := reDate.FindStringSubmatch(m)
		return spokenDate(p[1], p[2], "")
	})
	s = reTime.ReplaceAllStringFunc(s, func(m string) string {
		p := reTime.FindStringSubmatch(m)
		return p[1] + " часов " + p[2] + " минут"
	})
	s = reDots.ReplaceAllStringFunc(s, func(m string) string {
		return strings.Join(strings.Split(m, "."), " точка ")
	})
	return s
}

func spokenDate(dd, mm, yyyy string) string {
	mi, _ := strconv.Atoi(mm)
	if mi < 1 || mi > 12 {
		return dd + " " + mm + gap(yyyy) // out-of-range: leave numbers, drop dot
	}
	day := strconv.Itoa(mustInt(dd)) // strip leading zero: "01" → "1"
	out := day + " " + months[mi]
	if yyyy != "" {
		out += " " + yyyy
	}
	return out
}

func mustInt(s string) int { n, _ := strconv.Atoi(s); return n }
func gap(y string) string  { if y == "" { return "" }; return " " + y }
```

3. Wire into `reply()` (`cmd/mavend/voice.go:972`), before synth:
```go
func (h *reactiveHandler) reply(ctx context.Context, text string, _ []string) (voice.PushToTalkResp, error) {
	spoken := ttsnorm.Speakable(text)
	log.Printf("voice: reply → %q", text)
	audioOut, err := h.tts.Synthesize(ctx, spoken)
	if err != nil {
		log.Printf("voice: tts error: %v", err)
		return voice.PushToTalkResp{ReplyText: text, ReplyAudio: audio.Audio{}}, nil
	}
	return voice.PushToTalkResp{ReplyText: text, ReplyAudio: audioOut}, nil
}
```
(`ReplyText` stays raw for on-screen display; only audio is normalized.) Add the
`ttsnorm` import.

4. Wire into `voicesink.go` before its `Synthesize` call (same one-line
`ttsnorm.Speakable` on the nudge body). Add import.

5. `go test ./internal/ttsnorm/ ./cmd/mavend/ ./internal/delivery/...` green.

---

## Task 2 — Notes + reminders round-trip test (+ fix any real break)

**Goal:** prove — and lock with a test — that a note utterance persists a note
and a reminder utterance persists a reminder through the reactive handler.

**Files:** create `cmd/mavend/reactive_notes_test.go`. Modify `applyAction` /
seeds ONLY if the test exposes a real break.

**Acceptance criteria:**
- Driving `handleText` with a note utterance results in a `notes` row whose text
  matches; with a reminder utterance, a `reminders` row with a future `fire_ts`.
- Test uses the real `store.Store` (in-tmpfs or file), `ipc.NewStoreAPI(st)`, a
  `HashEmbedder`, and `StubReplier` — no LLM required.
- If routing (not the write) is what fails, the test forces intent by using a
  stage-0 grammar utterance (`"напомни завтра позвонить маме"`) and a directly
  constructed `IntentNote` decision path, isolating capture from routing.

**Verify:** `go test ./cmd/mavend/ -run TestReactiveNotesReminders -v` → PASS.

**Steps (TDD):**

1. Write the test: open a temp `store.Store`, build a minimal `reactiveHandler`
   (fields: `api: ipc.NewStoreAPI(st)`, `embedder: router.NewHashEmbedder(1024)`,
   `router: buildRouter(emb, matcher, thr)`, `replier: voice.NewStubReplier()`,
   `now: time.Now`, `memStore: memory.NewInMemoryStore()`, `dataStore: st`,
   `timeParser: router.NewPythonDateParser()`). Call
   `h.handleText(ctx, "напомни завтра в 9:00 позвонить маме")`; assert
   `st.DueReminders`/`ListReminders` returns one row with `fire_ts > now`.
   Then a note: construct `dec := router.Decision{Intent: router.IntentNote,
   Utterance: "запомни что кофе закончился"}`, call `h.applyAction(ctx, dec)`;
   assert `st.RecentNotes(ctx, 10)` contains the text.
2. Run it. If it fails at the **write**, fix the write in `applyAction`. If it
   fails only because routing sent the utterance elsewhere, that is Task 4/5's
   job — this test pins the *capture* contract, so force the intent as above and
   leave a `// routing covered by llmrouter_test` note.
3. Keep the test deterministic (no network: `NewPythonDateParser` shells out —
   if unavailable in CI, use a fixed `router.Slots{Time: now.Add(24h), HasTime:true}`
   decision for the reminder capture assertion instead of the parser path).

---

## Task 3 — `internal/llm` shared llama-server client

**Goal:** one client type both the phraser and the LFM router use to call the
already-running llama-server, with optional GBNF grammar for constrained output.

**Files:** create `internal/llm/client.go`, `internal/llm/client_test.go`;
modify `internal/phraser/llmphraser.go` (add `BaseURL()`).

**Acceptance criteria:**
- `Client.Complete(ctx, Req{System, User, Grammar, MaxTokens})` POSTs to
  `<base>/v1/chat/completions`, returns the assistant message content.
- When `Req.Grammar != ""`, the request body carries a top-level `"grammar"`
  field (llama.cpp extension) verbatim.
- `LLMPhraser.BaseURL()` returns the spawned server's base URL (`p.port`).

**Verify:** `go test ./internal/llm/` → `ok`.

**Steps (TDD):**

1. `client_test.go`: stand up `httptest.NewServer` that (a) asserts the decoded
   body has `messages[0].role=="system"`, `messages[1].content=="hi"`, and
   `grammar=="root ::= \"x\""`, and (b) replies
   `{"choices":[{"message":{"content":"ok"}}]}`. Assert `Complete` returns `"ok"`.
2. Implement `client.go`:
```go
// Package llm is the shared llama-server completion client — one seam both the
// phraser (talking back) and the router (routing) call. It does NOT spawn the
// server; the daemon owns one llama-server (spawned by the phraser) and hands
// its base URL here, so a single resident model serves both callers.
package llm

import (
	"bytes"
	"context"
	"encoding/json"
	"fmt"
	"net/http"
	"time"
)

type Client struct {
	base string
	http *http.Client
}

func New(baseURL string, timeout time.Duration) *Client {
	return &Client{base: baseURL, http: &http.Client{Timeout: timeout}}
}

type Req struct {
	System    string
	User      string
	Grammar   string // GBNF; empty ⇒ unconstrained
	MaxTokens int
}

type msg struct {
	Role    string `json:"role"`
	Content string `json:"content"`
}
type body struct {
	Messages  []msg  `json:"messages"`
	MaxTokens int    `json:"max_tokens,omitempty"`
	Grammar   string `json:"grammar,omitempty"`
	Temp      float64 `json:"temperature"`
}
type resp struct {
	Choices []struct {
		Message msg `json:"message"`
	} `json:"choices"`
}

func (c *Client) Complete(ctx context.Context, r Req) (string, error) {
	b, _ := json.Marshal(body{
		Messages:  []msg{{"system", r.System}, {"user", r.User}},
		MaxTokens: r.MaxTokens,
		Grammar:   r.Grammar,
		Temp:      0,
	})
	req, err := http.NewRequestWithContext(ctx, "POST", c.base+"/v1/chat/completions", bytes.NewReader(b))
	if err != nil {
		return "", err
	}
	req.Header.Set("Content-Type", "application/json")
	httpResp, err := c.http.Do(req)
	if err != nil {
		return "", err
	}
	defer httpResp.Body.Close()
	if httpResp.StatusCode != 200 {
		return "", fmt.Errorf("llm: status %d", httpResp.StatusCode)
	}
	var out resp
	if err := json.NewDecoder(httpResp.Body).Decode(&out); err != nil {
		return "", err
	}
	if len(out.Choices) == 0 {
		return "", fmt.Errorf("llm: no choices")
	}
	return out.Choices[0].Message.Content, nil
}
```
3. Add to `llmphraser.go`: `func (p *LLMPhraser) BaseURL() string { return p.port }`
   (the field already holds the full base URL incl. scheme+host+port).
4. `go test ./internal/llm/ ./internal/phraser/` green.

---

## Task 4 — LFM router core (in `internal/router`)

**Goal:** an agentic router that turns an utterance into a structured `Action`
(intent + raw slots + escalate flag) via one grammar-constrained LFM call, with
deterministic mapping to a `Decision`. Classifier stays as the fallback.

**Files:** create `internal/router/llmrouter.go`, `internal/router/llmrouter_test.go`;
modify `internal/router/router.go`.

**Acceptance criteria:**
- `LLMRouter.Route(ctx, utterance, now)` returns a `router.Decision` with
  `Intent` ∈ the existing enum, `Slots` populated (`Key/Value` for fact,
  `Text` for note/query/reminder, `Fn` left for the act matcher), and
  `Stage = 1`.
- Output is constrained by a GBNF grammar to a JSON object with an `intent`
  enum and optional string fields — malformed model output cannot escape the
  enum.
- On any error (LLM down, parse fail) `Route` returns `(Decision{}, false, err)`
  and the caller falls back to the classifier cascade.
- Time is NOT parsed here (kept deterministic): reminder actions carry the raw
  text in `Slots.Text`; the daemon's existing `timeParser` fallback
  (`voice.go:552`) resolves it.

**Verify:** `go test ./internal/router/ -run TestLLMRouter -v` → PASS.

**Steps (TDD):**

1. `llmrouter_test.go` with a mock completer (interface, not `*llm.Client`):
```go
type mockLLM struct{ out string; err error }
func (m mockLLM) Complete(ctx context.Context, r llm.Req) (string, error) { return m.out, m.err }

func TestLLMRouterFactMapping(t *testing.T) {
	lr := NewLLMRouter(mockLLM{out: `{"intent":"fact","key":"water","value":"выпил"}`})
	d, ok, err := lr.Route(context.Background(), "я выпил воду", time.Now())
	if err != nil || !ok { t.Fatalf("ok=%v err=%v", ok, err) }
	if d.Intent != IntentFact || d.Slots.Key != "water" || !d.Slots.HasKey {
		t.Fatalf("bad decision %+v", d)
	}
}
func TestLLMRouterNoteMapping(t *testing.T) { /* intent:"note" → Slots.Text set */ }
func TestLLMRouterBadJSONFallsBack(t *testing.T) {
	lr := NewLLMRouter(mockLLM{out: `garbage`})
	if _, ok, err := lr.Route(context.Background(), "x", time.Now()); ok || err == nil {
		t.Fatal("want ok=false, err!=nil on bad json")
	}
}
```

2. `llmrouter.go`:
```go
package router

import (
	"context"
	"encoding/json"
	"fmt"
	"strings"
	"time"

	"github.com/kami/maven/internal/llm"
)

// Completer — the LLM seam (mockable). *llm.Client satisfies it.
type Completer interface {
	Complete(ctx context.Context, r llm.Req) (string, error)
}

// LLMRouter — the agentic router. One grammar-constrained call classifies the
// utterance and pulls raw slots; deterministic parsers (time) refine downstream.
type LLMRouter struct{ c Completer }

func NewLLMRouter(c Completer) *LLMRouter { return &LLMRouter{c: c} }

// routeGrammar — GBNF constraining the model to a fixed-shape JSON object with
// an intent enum. Prevents free-form drift from a sub-1B model.
const routeGrammar = `
root   ::= "{" ws "\"intent\"" ws ":" ws intent ("," ws field)* ws "}"
intent ::= "\"fact\"" | "\"reminder\"" | "\"note\"" | "\"query\"" | "\"act\"" | "\"chat\"" | "\"system\""
field  ::= key ws ":" ws string
key    ::= "\"key\"" | "\"value\"" | "\"text\"" | "\"verb\""
string ::= "\"" ([^"\\] | "\\" .)* "\""
ws     ::= [ \t\n]*
`

const routeSystem = `Ты — маршрутизатор Maven. По реплике пользователя верни ОДИН JSON-объект:
{"intent": one of fact|reminder|note|query|act|chat|system, ...}
Правила: fact — состояние которое надо запомнить и отслеживать ("я выпил воду" → key=water, value=выпил); reminder — просьба напомнить в будущем (text = что напомнить); note — заметка/предпочтение без отслеживания (text = текст); query — вопрос (text = вопрос); act — команда выполнить действие на сервере (verb = глагол); chat — свободный разговор (text); system — время/дата/день. Только JSON, без пояснений.`

type routeAction struct {
	Intent string `json:"intent"`
	Key    string `json:"key"`
	Value  string `json:"value"`
	Text   string `json:"text"`
	Verb   string `json:"verb"`
}

func (lr *LLMRouter) Route(ctx context.Context, utterance string, now time.Time) (Decision, bool, error) {
	raw, err := lr.c.Complete(ctx, llm.Req{System: routeSystem, User: utterance, Grammar: routeGrammar, MaxTokens: 128})
	if err != nil {
		return Decision{}, false, err
	}
	var a routeAction
	if err := json.Unmarshal([]byte(strings.TrimSpace(raw)), &a); err != nil {
		return Decision{}, false, fmt.Errorf("llmrouter: parse %q: %w", raw, err)
	}
	d := Decision{Utterance: utterance, Stage: 1, Confidence: 1.0}
	switch Intent(a.Intent) {
	case IntentFact:
		d.Intent = IntentFact
		d.Slots.Key, d.Slots.Value = a.Key, a.Value
		d.Slots.HasKey = a.Key != ""
	case IntentReminder:
		d.Intent = IntentReminder
		d.Slots.Text = firstNonEmpty(a.Text, utterance) // time parsed downstream
	case IntentNote:
		d.Intent = IntentNote
		d.Slots.Text = firstNonEmpty(a.Text, utterance)
	case IntentQuery:
		d.Intent = IntentQuery
		d.Slots.Text = firstNonEmpty(a.Text, utterance)
	case IntentAct:
		d.Intent = IntentAct // Fn resolved by the act matcher in the daemon
		d.Slots.Text = firstNonEmpty(a.Verb, utterance)
	case IntentSystem:
		d.Intent = IntentSystem
	default:
		d.Intent = IntentChat
		d.Slots.Text = firstNonEmpty(a.Text, utterance)
	}
	return d, true, nil
}

func firstNonEmpty(a, b string) string { if strings.TrimSpace(a) != "" { return a }; return b }
```

3. `router.go`: add `LLM *LLMRouter` to `Config` and the `Router` struct; in
   `Route`, after the stage-0 loop and before `r.classifier.Classify`:
```go
	// stage 1a — LLM router (when wired). It reasons over the utterance instead
	// of nearest-centroid guessing. On any error/parse-fail, fall through to the
	// classifier cascade (never fail the turn on the model).
	if r.llm != nil {
		if d, ok, err := r.llm.Route(ctx, utterance, now); err == nil && ok {
			d.Utterance = utterance
			return d, nil
		} else if err != nil {
			log.Printf("router: llm route fell back to classifier: %v", err)
		}
	}
```
   Add `"log"` import; set `llm: cfg.LLM` in `New`.

4. `go test ./internal/router/` green (existing classifier tests still pass —
   `LLM` nil by default preserves the cascade).

---

## Task 5 — Wire the LFM router into the daemon

**Goal:** the daemon builds one `llm.Client` from the running phraser and injects
it into the router, so production routing goes through the LFM (with classifier
fallback). Act intents still resolve `Fn` via the existing matcher.

**Files:** modify `cmd/mavend/voice.go` (`wireVoice`, `buildRouter`), and pass
the phraser through (already a `wireVoice` param `phr phraser.Phraser`).

**Acceptance criteria:**
- When `phr` is an `*phraser.LLMPhraser`, `buildRouter` receives a non-nil
  `Completer` and the router's `Config.LLM` is set; when `phr` is the Stub,
  `LLM` is nil and routing uses the classifier (unchanged behavior / tests).
- Act `Fn` resolution: after an LLM `IntentAct` decision, the handler runs the
  utterance/verb through the existing `matcher.Match` (as stage-0 does) so
  `HasFn`/`Args` are filled before `applyAction`.

**Verify:** `go build ./cmd/mavend/ && go test ./cmd/mavend/` → ok; manual:
`handleText` on a paraphrased note/reminder (not matching any seed) routes
correctly (covered by an added `cmd/mavend` test using a mock completer).

**Steps:**

1. In `wireVoice`, after the phraser is known, build the optional client:
```go
	var llmRouter *router.LLMRouter
	if lp, ok := phr.(*phraser.LLMPhraser); ok {
		cli := llm.New(lp.BaseURL(), 20*time.Second)
		llmRouter = router.NewLLMRouter(cli)
	}
	rtr := buildRouter(emb, matcher, threshold, llmRouter)
```
   Add `internal/llm` import.

2. Extend `buildRouter` signature with `llmR *router.LLMRouter` and set
   `LLM: llmR` in `router.Config`.

3. Act `Fn` fill: in `applyAction`'s `IntentAct` branch, when `!dec.Slots.HasFn`
   and `dec.Slots.Text != ""`, try `h.tools`/matcher resolution before
   `proposeGap` (mirror stage-0's `actMatcher.Match`). Add a `matcher` field to
   `reactiveHandler` (set in `wireVoice`) and:
```go
	case router.IntentAct:
		if !dec.Slots.HasFn && dec.Slots.Text != "" && h.matcher != nil {
			if fn, args, ok := h.matcher.Match(dec.Slots.Text); ok {
				dec.Slots.Fn, dec.Slots.Args, dec.Slots.HasFn = fn, args, true
			}
		}
		if !dec.Slots.HasFn { return h.proposeGap(ctx, dec) }
		// ... unchanged executor path
```

4. Add a `cmd/mavend` test: build a handler with a mock-completer-backed
   `router.LLMRouter`, drive `handleText("надо не забыть купить молоко")`,
   assert it routes to `IntentNote`/`IntentReminder` (per mock) and persists.

5. `go test ./cmd/mavend/ ./internal/router/` green.

---

## Task 6 — LLM-backed reactive Replier (talking back)

**Goal:** reactive confirmations are phrased by the LFM ("записала: кофе
закончился" in Maven's voice), not `StubReplier` canned strings — the visible
payoff of the engine being on. Falls back to the Stub on any LLM error.

**Files:** create `cmd/mavend/replier_llm.go`, `cmd/mavend/replier_llm_test.go`;
modify `cmd/mavend/voice.go` (set `replier` to the LLM impl when the engine is on).

**Acceptance criteria:**
- `llmReplier.Reply(dec)` returns a short RU confirmation generated from the
  decision (intent + slots), feminine self-reference, ≤ ~120 chars.
- On LLM error/empty, returns `StubReplier.Reply(dec)` (never blank, never a
  wire error).
- Clarify decisions still yield a clarify prompt.

**Verify:** `go test ./cmd/mavend/ -run TestLLMReplier -v` → PASS.

**Steps (TDD):**

1. `replier_llm_test.go`: mock completer returning `"записала, кофе закончился"`
   → assert `Reply(noteDecision)` returns it; mock returning error → assert it
   equals `voice.NewStubReplier().Reply(dec)`.
2. `replier_llm.go`:
```go
package main

import (
	"context"
	"time"

	"github.com/kami/maven/internal/llm"
	"github.com/kami/maven/internal/router"
	"github.com/kami/maven/internal/voice"
)

type completer interface {
	Complete(ctx context.Context, r llm.Req) (string, error)
}

// llmReplier phrases reactive confirmations with the resident LFM. Stub is the
// floor on any error (offline-safe). Maven speaks as "she", feminine RU.
type llmReplier struct {
	c    completer
	stub *voice.StubReplier
}

func newLLMReplier(c completer) *llmReplier { return &llmReplier{c: c, stub: voice.NewStubReplier()} }

const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Одной короткой фразой (≤120 симв) подтверди действие пользователю тепло и по-русски. Без кавычек и пояснений.`

func (r *llmReplier) Reply(d router.Decision) string {
	if d.Clarify {
		return r.stub.Reply(d)
	}
	ctx, cancel := context.WithTimeout(context.Background(), 8*time.Second)
	defer cancel()
	out, err := r.c.Complete(ctx, llm.Req{System: replySystem, User: replyContext(d), MaxTokens: 64})
	if err != nil || out == "" {
		return r.stub.Reply(d)
	}
	return out
}

// replyContext renders the decision into a compact RU description for the model.
func replyContext(d router.Decision) string {
	switch d.Intent {
	case router.IntentFact:
		return "записала факт: " + d.Slots.Key + " " + d.Slots.Value
	case router.IntentNote:
		return "сохранила заметку: " + d.Slots.Text
	case router.IntentReminder:
		return "поставила напоминание: " + d.Slots.Text
	default:
		return string(d.Intent) + ": " + d.Slots.Text
	}
}
```
   Note: `Reply` satisfies `voice.Replier` (sync). The 8s timeout bounds a slow
   sub-1B turn; the Stub floor keeps the round-trip alive if it lapses.
3. In `wireVoice`, when `llmRouter != nil` (engine on), set
   `replier: newLLMReplier(cli)` instead of `voice.NewStubReplier()` (reuse the
   same `cli`). Keep Stub when the engine is off.
4. `go test ./cmd/mavend/` green.

---

## Task 7 — Turn the engine on + telegram reach (deploy config)

**Goal:** homesrv actually runs the LFM (router + phraser + replier live) and can
reach the user over telegram. This is the commit that flips Maven from stub to
real on the deploy box.

**Files:** modify `deploy/mavend.json`, `docker-compose.yml`, `AGENTS.md`.
Ops: place the LFM gguf on homesrv; `docker compose up -d`.

**Acceptance criteria:**
- `deploy/mavend.json` has a `phraser` block pointing at the LFM gguf
  (`LFM2.5-1.2B-Instruct-Q4_K_M.gguf`) with `bin_path: llama-server`,
  `n_gpu_layers` tuned for the Vega iGPU (start `0` = CPU; the 1.2B is real-time
  on CPU, avoids fragile ROCm), `n_ctx: 2048`.
- `deploy/mavend.json` has a `telegram` block (`bot_token`, `chat_id`) — token
  from env, not committed.
- On homesrv: `docker logs maven-mavend-1` shows a phraser/llama line and
  `voice listening`, and no fall-through to Stub.
- A push-to-talk (or `/api/chat`) "запомни что кофе закончился" returns an LFM
  phrased confirmation AND persists a note (check `/dash` recent notes).
- A sev4 nudge with no live voice session lands in telegram.

**Verify:**
```sh
ssh kami@192.168.1.104 'docker logs maven-mavend-1 2>&1 | grep -iE "llama|phraser|listening"'
# expect a llama-server spawn line + "voice listening on ..."
ssh kami@192.168.1.104 'curl -s localhost:9201/api/chat -d "{\"text\":\"запомни что кофе закончился\"}"'
# expect a non-stub RU confirmation; then /dash shows the note
```

**Steps:**

1. Add to `deploy/mavend.json`:
```json
"phraser": {
  "model_path": "/opt/maven/models/llm/LFM2.5-1.2B-Instruct-Q4_K_M.gguf",
  "bin_path": "llama-server",
  "n_gpu_layers": 0,
  "n_ctx": 2048,
  "timeout": "20s"
},
"telegram": { "bot_token": "${TELEGRAM_BOT_TOKEN}", "chat_id": "${TELEGRAM_CHAT_ID}" }
```
   (Match the existing config's env-substitution convention; if it doesn't
   support `${...}`, wire the two telegram values via compose env + a small
   config env-read, consistent with `db_key_env`.)
2. `docker-compose.yml`: mount `./models/llm:/opt/maven/models/llm:ro` on
   `mavend`; add `TELEGRAM_BOT_TOKEN` / `TELEGRAM_CHAT_ID` to its env (from a
   gitignored `deploy/telegram.env`, mirroring `deploy/db_key.env`). Ensure
   `llama-server` is on PATH in the image (it's already used by design; confirm
   the Dockerfile installs it — if not, add it).
3. Ops: copy `models/llm/LFM2.5-1.2B-Instruct-Q4_K_M.gguf` to
   homesrv `/opt/maven/models/llm/` (the box already has `/usr/local/bin/llama-server`).
   If the LFM gguf isn't on disk yet, download it (HF, sha-checked) — document in
   AGENTS.md next to the embedder section.
4. `docker compose up -d mavend && docker logs -f maven-mavend-1` — confirm the
   phraser spawns and no `phraser: NewStub` path. Run the two verify curls.
5. Update `AGENTS.md`: LFM model download + note that routing is now LFM-first
   with classifier fallback (`REARCH.md` is the design of record).

---

## Self-review

- **Placeholders:** none — every code step carries real code. Deploy steps (Task
  7) are concrete commands, not "configure somehow".
- **Type consistency:** `Completer` interface (`Complete(ctx, llm.Req) (string,
  error)`) is used identically in Task 4 (router), Task 5 (wiring), Task 6
  (replier); `*llm.Client` satisfies it. `router.Decision`/`Slots` field names
  match the real structs (`HasKey`, `Text`, `Fn`, `Args`). `LLMRouter.Route`
  returns `(Decision, bool, error)` consistently across Task 4 def and Task 4
  router.go call site.
- **Fallback discipline:** every LLM seam (router, replier) falls back to the
  existing deterministic floor on error — a slow/absent model degrades to today's
  behavior, never a broken turn. Engine-off path is byte-for-byte unchanged
  (LLM nil, Stub replier), so all existing tests hold.
- **Scope:** no 4B reasoner, no digestion worker, no custom TTS. `escalate` is
  intentionally omitted from the grammar this sub-project (the reasoner it would
  escalate to doesn't exist yet) — added in sub-project 2.

## Ordering & deps

1 (ttsnorm) and 2 (notes/reminders test) are independent — do first, any order.
3 (llm client) → 4 (llm router) → 5 (wire) is the critical chain. 6 (replier)
depends on 3. 7 (deploy) depends on 3–6 landed. Each task = one commit.